Optical_Fiber_Telecommunications_IV-B

OPTICAL FIBER TE LEC 0 M MU N ICAT I SYSTEMS AND IMPAIRMENTS OPTICAL FIBER TELECOMMUNICATIONS IV B SYSTEMS AND IMPAIRMENTS OPTICAL FIBER TELECOMMUNICATIONS IV B SYSTEMS AND IMPAIRMENTS Edited by IVAN P KAMINOW . Bell Laboratories (retired) Kaminow Lightwave Technology Holmdel, New Jersey TINGYE LI AT&T Labs (retired) Boulder, Colorado ACADEMIC PRESS An Elsevier Science Imprint San Diego San Francisco New York Boston London Sydney Tokyo This book is printed on acid-free paper. @ Copyright @ 2002, Elsevier Science (USA). All rights reserved. No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system,without permission in writing from the publisher. The appearance of the code at the bottom of the h s t page of a chapter in this book indicates the Publisher’s consent that copies of the chapter may be made for personal or internal use of speci6c clients. This consent is given on the condition, however, that the copier pay the stated per copy fee through the Copyright Clearance Center, Inc. (222 Rosewood Drive, Danvers, Massachusetts 01923), for copying beyond that permitted by Sections 107 or 108 of the U.S. Copyright Law. This consent does not extend to other kinds of copying, such as copying for general distribution, for advertising or promotional purposes, for creating new collective works, or for resale. Copy fees for pre-2001 chapters are as shown on the title pages. If no fee code appears on the title page, the copy fee is the same as for current chapters. $35.00. Academic Press An Elsevier Science Imprint 525 B Street, Suite 1900, San Diego, California 92101-4495, USA http://www.academicpress.com Academic Press Harcourt Place, 32 Jamestown Road,London NW17BY, UK http://www.academicpress.com Library of Congress Control Number: 2001098830 International Standard Book Number: 0-12-395173-9 PRINTED IN CHINA 02 03 04 05 06 07 RDC 9 8 7 6 5 4 3 2 1 For Florence and Edith, with love Contents Contributors xi Chapter 1 Overview Ivan 19 Kaminow 1 Chapter 2 Growth of the Internet Kerry G. Coflman and Andrew M. Odbzko 17 Chapter 3 Optical Network Architecture Evolution John Strand 57 Chapter 4 Undersea Communication Systems Neal S. Bergano 154 Chapter 5 High-Capacity, Ultra-Long-Haul Networks John Zyskind, Rick Bany, Graeme Pendock, Michael Cahill, and Jinendra Ranka 198 Chapter 6 Pseudo-Linear Transmission of High-speed TDM Signals: 40 and 160 Gb/s Red-Jean Essiambre, Gregory Raybon, and Benny Mikkelsen 232 vii viii Contents Chapter 7 Dispersion-Managed Solitons and Chirped Return to Zero: What Is the Difference? Curtis R. Menyuk, Gary M. Carter; WilliamL. Kath, and Ruo-Mei Mu 305 Chapter 8 Metropolitan Optical Networks Nasir Ghani, Jin-B Pan, and Xin Cheng 329 Chapter 9 The Evolution of Cable TV Networks Xiaolin Lu and OIeh Sneizka 404 Chapter 10 Optical Access Networks Edward Harstead and Pieter H. van Heyningen 438 Chapter 1 Beyond Gigabit: Application and Development of 1 High-speed Ethernet Technology Cedric E Lam 514 Chapter 12 Photonic Simulation Tools Arthur J. Lowery 564 Chapter 13 Nonlinear Optical Effects in WDM Transmission Polina Bayvel and Robert Killey 61 1 Chapter 14 Fixed and Tunable Management of Fiber Chromatic Dispersion Alan E. Willner and Bogdan Hoanca 642 Chapter 15 Polarization-ModeDispersion 725 Herwig Kogelnik, Robert M. Jopson, and Lynn E. Nelson Contents ix Chapter 16 Bandwidth-Efficient Modulation Formats for Digital Fiber Transmission Systems 862 Jan Conradi Chapter 17 Error-Control Coding Techniques and Applications 902 E! Vjay Kumar, Moe Z. Win, Hsiao-Feng Lu, and Costas N. Georghiades Chapter 18 Equalization Techniques for Mitigating Transmission Impairments 965 Moe Z. Win, Jack H. Winters, and Giorgio M. Etetta Index to Volumes IVA and IVB 999 Contributors D. A. Ackerman (A:587), Agere Systems, 600 Mountain Avenue, Murray Hill, New Jersey 07974 Daniel Y. Al-Salameh (A295), JDS Uniphase Corporation, 100 Willowbrook Road, Bldg. 1,Freehold, New Jersey 07728-2879 Rick Barry (B: 198), Sycamore Networks, 10 Elizabeth Drive, Chelmsford, Massachusetts 01824-4111 Polina Bayvel (B:61 l), Optical Networks Group, Department of Electronic and Electrical Engineering, University College London (UCL), Torrington Place, London WCl E 7JE, United Kingdom Neal S. Bergano (B: 154), Tyco Telecommunications,250 Industrial Way West, Eatontown, New Jersey 07724-2206 Lee L. Blyler (A:17), OFS Fitel, LLC, 600 Mountain Avenue, Murray Hill, New Jersey 07974 Raymond K. Boncek (A: 17), OFS Fitel, LLC, 600 Mountain Avenue, Murray Hill, New Jersey 07974 Michael Cahil (B: 198), SycamoreNetworks, 10 Elizabeth Drive, Chelmsford, Massachusetts 01824-4111 Gary M. Carter (B:305), Computer Science and Electrical Engineering Department, TRC-201A, University of Maryland Baltimore County, 1000 Hilltop Circle, Baltimore, Maryland 21250 and Laboratory for Physical Sciences, College Park, Maryland Connie J. Chang-Hasnain (A:666), Department of Electrical Engineering and Computer Science, University of California, Berkeley, California 94720 and Bandwidth 9 Inc., 46410 Fremont Boulevard, Fremont, California 94538 Young-Kai Chen (A:784), Lucent Technologies, High Speed Electronics Research, 600 Mountain Avenue, Murray Hill, New Jersey 07974 Xin Cheng (B:329), Sorrento Networks Inc., 9990 Mesa Rim Drive, San Diego, California 9212 1-2930 x i xu Contributors Dominique Chiaroni(A:732), Alcatel Research & Innovation, Route de Nozay, F-9 1461 Marcoussis cedex, France Kerry G. Coffman (B:17), AT&T Labs-Research, A5-1D03, 200 Laurel Avenue South, Middletown, New Jersey 07748 Jan Conradi (B:862), Director of Strategy,Corning Optical Communications, Corning Incorporated,MP-HQ-Wl-43, One River Front Plaza, Corning, New York 14831 Santanu J Das (A:17), OFS Fitel, LLC, 600 Mountain Avenue, Murray Hill, L New Jersey 07974 Emmanuel Desurvire (A:732), Alcatel Technical Academy, Villarceaux, F-9 1625 Nozay cedex, France David J. DiGiovanni (A:17), OFS Fitel, LLC, 600 Mountain Avenue, Murray Hill, New Jersey 07974 Christopher R Doerr (A:405), Bell Laboratories, Lucent Technologies, 791 Holmdel-Keyport Road, Holmdel, New Jersey 07733 Adam Ellison (A:80), Corning, Inc., SP-FR-05, Corning, New York 14831 L. E. Eng (A:587), Agere Systems, Room 2F-204, 9999 Hamilton Blvd., Breinigsville, Pennsylvania 18031-9304 Turan Erdogan (A:477), Semrock, Inc., 3625 Buffalo Road, Rochester, New York 14624 RenkJean Essiambre (B:232), Bell Laboratories, Lucent Technologies, 79 1 Holmdel-KeyportRoad, Holmdel, New Jersey 07733 Costas N. Georghiades (B:902), Texas A&M University, Electrical Engineering Department, 237 Wisenbaker, College Station, Texas 77843-3128 Nasir Ghani (B:329), Sorrento Networks Inc., 9990 Mesa Rim Drive, San Diego, California 92121-2930 Steven E. Golowich (A:17), Bell Laboratories, Lucent Technologies, Room 2C-357,600 Mountain Avenue, Murray Hill, New Jersey 07974 Christoph S. Harder (A:563), Nortel Networks Optical Components, Binzstrasse 17, CH-8045 Zurich, Switzerland Edward Harstead (B:438), Bell Laboratories, Lucent Technologies, 101 Crawford Corners Road, Holmdel, New Jersey 07733 Bogdan Hoanca (B:642), Phaethon Communications, Inc., Fremont, California 96538 Contributors xiii J. E. Johnson (A587), Agere Systems, 600 Mountain Avenue, Murray Hill, New Jersey 07974 Robert M. Jopson (B:725), Crawford Hill Laboratory, Bell Laboratories, Lucent Technologies, 79 1 Holmdel-Keyport Road, Holmdel, New Jersey 07733 Ivan P. Kaminow (A: 1,B: l), Bell Laboratories (retired), Kaminow Lightwave Technology, 12 Stonehenge Drive, Holmdel, New Jersey 07733 Bryon L. Kasper (A:784), Agere Systems, Advanced Development Group, 4920 Rivergrade Road, Irwindale, California 91706-1404 William L. Kath (B:305), Computer Science and Electrical Engineering Department, University of Maryland Baltimore County, 1000 Hilltop Circle, Baltimore, Maryland 21250 and Applied Mathematics Department, Northwestern University, 2145 Sheridan Road, Evanston, Illinois 60208-3125 L. J. P. Ketelsen (A:587), Agere Systems, 600 Mountain Avenue, Murray Hill, New Jersey 07974 P. A. Kiely (A:587), Agere Systems, 9999 Hamilton Blvd., Breinigsville, Pennsylvania 18031-9304 Robert Killey (B:61 l), Optical Networks Group, Department of Electronic and Electrical Engineering, University College London (UCL), Torrington Place, London WClE 7JE, United Kingdom Herwig Kogelnik (B:725), Crawford Hill Laboratory, Bell Laboratories, Lucent Technologies, 79 1 Holmdel-Keyport Road, Holmdel, New Jersey 07733 StevenK Korotky (A295), Bell Laboratories, Lucent Technologies,Room H O 3C-351,101 Crawfords Corner Road, Holmdel, New Jersey 07733-1900 P. Mjay Kumar (B:902), Communication Science Institute, Department of Electrical Engineering- Systems,University of Southern California, 3740 McClintock Avenue, EEBSOO, Los Angeles, California 90089-2565 and Scintera Networks, Inc., San Diego, California Cedric E Lam (B:514), AT&T Labs-Research, 200 Laurel Avenue South, Middletown, New Jersey 07748 Bruno Lavigne (A:732), Alcatel CIT/ Research & Innovation, Route de Nozay, F-91461 Marcoussis cedex, France Olivier Leclerc (A732), Alcatel Research & Innovation, Route de Nozay, F-91460 Marcoussis cedex, France xiv Contributors David S. Levy (A295), Bell Laboratories, Lucent Technologies, Room H O 3B-506, 101 Crawfords Corner Road, Holmdel, New Jersey 07733-3030 Arthur J. Lowery (B:564), VPIsystems Inc., Design Center Group, 17-27 Cotham Road, Kew, Melbourne 3101, Australia Xiaolin Lu (B:404), Morning Forest, LLC, 8804 S. Blue Mountain Place, Highlands Ranch, Colorado 80126 Hsiao-Feng Lu (B:902), Communication Science Institute, Department of ElectricalEngineering- Systems, University of Southern California, 3740 McClintock Avenue, EEBSOO, Los Angeles, California 90089-2565 Amaresh Mahapatra (A:258), Linden Corp., 10 Northbriar Road, Acton, Massachusetts 01720 T. G. B. Mason (A:587), Agere Systems, 9999 Hamilton Blvd., Breinigsville, Pennsylvania 18031-9304 Curtis R. Menyuk (B:305), Computer Science and Electrical Engineering Department, TRC-201A, University of Maryland Baltimore County 1000 Hilltop Circle, Baltimore, Maryland 21250 and PhotonEx Corporation, 200 MetroWest Technology Park, Maynard, Massachusetts 0 1754 Benny Mikkelsen (B:232), Mintera Corporation, 847 Rogers Street, One Lowell Research Center, Lowell, Massachusetts 01852 John Minelly (A:80), Corning, Inc., SP-AR-02-01, Corning, New York 14831 Osamu Muuhara (A:784), Agere Systems, Optical Systems Research, 9999 Hamilton Blvd., Breinigsville, Pennsylvania 18031 Stefan Mohrdiek (A:563), Nortel Networks Optical Components, Binzstrasse 17, CH-8045 Ziirich, Switzerland Ruo-Mei Mu (B:305), Tyco Telecommunications, 250 Industrial Way West, Eatontown, New Jersey 07724-2206 Edmond J. Murphy (A:258), JDS Uniphase, 1985 Blue Hills Avenue Ext., Windsor, Connecticut 06095 Timothy 0. Murphy (A:295), Bell Laboratories, Lucent Technologies, Room H 3D-516, 101 Crawfords Corner Road, Holmdel, New Jersey 07733O 3030 Lynn E. Nelson (B:725), OFS Fitel, Holmdel, New Jersey 07733 Andrew M. Odlyzko (B:17), University of Minnesota Digital Technology Center, 1200 Washington Avenue S., Minneapolis, Minnesota 55415 Contributors xv Jin-Yi Pan (B:329), SorrentoNetworks Inc., 9990 Mesa Rim Drive, San Diego, California 92121-2930 Sunita 1 . Pate1 (A:295), Bell Laboratories, Lucent Technologies, Room H 8 O 3D-502,101 Crawfords Comer Road, Holmdel, New Jersey 07733-3030 Graeme Pendock (B: 198), Sycamore Networks, 10 Elizabeth Drive, Chelmsford, Massachusetts 01824-4111 Jinendra Ranka (B: 198), Sycamore Networks, 10 Elizabeth Drive, Chelmsford, Massachusetts 01824-4111 Gregory Raybon (B:232), Bell Laboratories, Lucent Technologies, 79 1 Holmdel-Keyport Road, Holmdel, New Jersey 07733 Gaylord W. Richards (A:295), Bell Laboratories, Lucent Technologies,Room 6L-219,2000 Naperville Road, Naperville, Illinois 60566-7033 Karsten Rottwitt (A:213), Orsted Laboratory, Niels Bohr Institute, University of Copenhagen, Universitetsparken 5, Copenhagen dk 2 100, Denmark Bertold E. Schmidt (A:563), Nortel Networks Optical Components, Binzstrasse 17, Ch-8045 Zurich, Switzerland Oleh Sniezko (B:404), Oleh-Lightcom, Highlands Ranch, Colorado 80126 Leo H. Spiekman (A:699), Genoa Corporation, Lodewijkstraat 1A, 5652 AC Eindhoven, The Netherlands Atul K. Srivastava (A:174), Onetta Inc., 1195 Borregas Avenue, Sunnyvale, California 94089 Andrew J. Stentz (A:213), Photuris, Inc., 20 Corporate Place South, Piscataway, New Jersey 08809 John Strand (B:57), AT&T Laboratories, Lightwave Networks Research Department,RoomA5-106,200LaurelAvenue, Middletown, New Jersey 07748 Thomas A. Strassser (A:477), Photuris Inc., 20 Corporate Place South, Piscataway, New Jersey 08854 Yan Sun (A:174), Onetta Inc., 1195 Borregas Avenue, Sunnyvale, California 94089 Eric S. Tentarelli (A:295), Bell Laboratories, Lucent Technologies, Room H O 3B-530,101 Crawfords Corner Road, Holmdel, New Jersey 07733-3030 Pieter H. van Heyningen (B:438), Lucent Technologies NL, PO. Box 18, Huizen 1270AA,The Netherlands xvi Contributors Giorgio M. Vitetta (B:965), University of Modena and Reggio Emilia, Department of Information Engineering, Via Vignolese 905, Modena 41100, Italy W. White (A:17), OFS Fitel, LLC, 600 Mountain Avenue, Murray Hill, New Jersey 07974 Alan E. Willner (B:642), University of Southern California, Los Angeles, California 90089-2565 Moe Z. Win (B:902, B:965), AT&T Labs-Research, Room A5-1D01, 200 Laurel Avenue South, Middletown, New Jersey 07748-1914 Jack H. Winters (B:965), AT&T Labs-Research, Room 4-147, 100 Schulz Drive, Middletown, New Jersey 07748-1914 Martin Zirngibl (A:374), Bell Laboratories, Lucent Technologies, 79 1 Holmdel-KeyportRoad, Holmdel, New Jersey 07733-0400 John Zyskind (B: 198), Sycamore Networks, 10 Elizabeth Drive, Chelmsford, Massachusetts 0 1824-4111 Ivan P Kaminow . Bell Laboratories (retired),Kaminow Lightwave Technology, Holmdel, New Jerscy Introduction Modern lightwave communications had its origin in the first demonstrations of the laser in 1960. Most of the early lightwave R&D was pursued by established telecommunications company labs (AT&T, NTT, and the British Post Office among them). By 1979, enough progress had been made in lightwave technology to warrant a book, Optical Fiber Telecommunications(OFlJ, edited by S. E. Miller and A. G. Chynoweth, summarizing the state of the art. Two sequels have appeared: in 1988, OFT 11, edited by S. E. Miller and I. P. Kaminow, and in 1997, OFT 111 (A & B), edited by I. P. Kaminow and T. L. Koch. The rapid changes in the field now call for a fourth set of books, OFTW (A & B). This chapter briefly summarizes the previous books and chronicles the remarkably changing climates associated with each period of their publication. The main purpose, however, is to summarize the chapters in OFT IV in order to give the reader an overview. History While many excellent books on lightwave communications have been published, this series has developed a special character, with a reputation for comprehensiveness and authority, because of its unique history. Optical Fiber Telecommunications was published in 1979, at the dawn of the revolution in lightwave telecommunications. It was a stand-alone work that aimed to collect all available information on lightwave research. Miller was Director of the Lightwave Systems Research Laboratory and, together with Rudi Kompfner, the Associate Executive Director, guided the system research at the Crawford Hill Laboratory of AT&T Bell Laboratories; Chynoweth was an Executive Director in the Murray Hill Laboratory, leading the optical fiber research. Many groups were active at other laboratories in the United States, Europe, and Japan. OFT, however, was written exclusively by Bell Laboratories authors, who nevertheless aimed to incorporate global results. 1 OPTICAL FIBER TELECOMMUNICATIONS, VOLUME IVB Copyright 0 2002, Elsevier Science (USA). n All rights of reproduction i any form reserved. ISBN 0-12-395173-9 2 Ivan P Kaminow . Miller and Chynoweth had little trouble finding suitable chapter authors at Bell Labs to cover practically all the relevant aspects of the field at that time. Looking back at that volume, it is interesting that the topics selected are still quite basic. Most of the chapters cover the theory, materials, measurement techniques, and properties of fibers and cables (for the most part, multimode fibers). Only one chapter covers optical sources, mainly multimode AlGaAs lasers operating in the 800- to 900-nm band. The remaining chapterscover direct and externalmodulation techniques,photodetectors and receiver design, and system design and applications Still, the basic elements of the present day systems are discussed: low-loss vapor-phase silica fiber and double-heterostructurelasers. Although system trials were initiated around 1979, it required several more years before a commercially attractive lightwave telecommunications system was installed in the United States. The AT&T Northeast Corridor System, operating between New York and Washington, DC, began service in January 1983, operating at a wavelength of 820 nm and a bit rate of 45 Mb/s in multimode fiber. Lightwave systems were upgraded in 1984 to 1310nm and 417 or 560 Mb/s in single-mode fiber in the United States as well as in Europe and Japan. The year 1984 also saw the Bell System broken up by the court-imposed “Modified Final Judgment” that separated the Bell operating companies into seven regional companies and left AT&T as the long distance camer as well as a telephone equipment vendor. Bell Laboratories remained with AT&T, and Bellcore was formed to serve as the R&D lab for all seven regional Bell operating companies (RBOCs). The breakup spurred a rise in diversity and competition in the communications business. The combination of technical advances in computers and communications, growing government deregulation, and apparent new business opportunities all served to raise expectations. Tremendoustechnical progresswas made during the next few years, and the choice of lightwave over copper coaxial cable or microwave relay for most longhaul transmission systems was assured. The goal of research was to improve performance, such as bitrate and repeater spacing, and to find other applications beyond point-to-point long haul telephone transmission. A completely new book, Optical Fiber Telecommunications1 ,was published in 1988 to sum1 marize the lightwave R&D advancesat the time. To broaden the coverage, nonBell Laboratories authors from Bellcore (now Telcordia), Corning, Nippon Electric Corporation, and several universities were represented among the contributors. Although research results are described in OFT 1 , the emphasis 1 is much stronger on commercial applications than in the previous volume. The initial chapters of OFT 1 cover fibers, cables, and connectors, deal1 ing with both single- and multimode fiber. Topics include vapor-phase methods for fabricating low-loss fiber operating at 13 10 and 1550 nm, understanding chromatic dispersion and nonlinear effects, and designing polarization-maintaining fiber. Another large group of chapters deals with 1.Overview 3 a range of systems for loop, intercity, interoffice, and undersea applications. A research-oriented chapter deals with coherent systems and another with possible local area network designs, including a comparison of time-division multiplexing (TDM) and wavelength division multiplexing (WDM) to efficiently utilize the fiber bandwidth. Several chapters cover practical subsystem components, such as receivers and transmitters and their reliability. Other chapters cover photonic devices, such as lasers, photodiodes, modulators, and integrated electronic and integrated optic circuits that make up the subsystems. In particular, epitaxial growth methods for InGaAsP materials suitable for 1310 and 1550nm applications and the design of high-speed single-mode lasers are discussed in these chapters. By 1995, it was clear that the time had arrived to plan for a new volume to address recent research advances and the maturing of lightwave systems. The contrast with the research and business climates of 1979 was dramatic. Sophisticatedsystem experiments were being performed utilizing the commercial and research components developed for a proven multibillion-dollar global lightwave industry. For example, 10,000-kmlengths of high-performance fiber were assembled in several laboratories around the world for nonreturn-to-zero (NRZ), soliton, and WDM transmission demonstrations. Worldwide regulatory relief stimulated the competition in both the service and hardware ends of the telecommunications business. The success in the long-haul market and the availability of relatively inexpensive components led to a wider quest for other lightwave applications in cable television and local access network markets. The development of the diode-pumped, erbium-doped fiber amplifier (EDFA) played a crucial role in enhancing the feasibility and performance of long-distance and WDM applications. By the time of publication of OFT 1 1 1 in 1997, incumbent telephone companies no longer dominated the industry. New companies were offering components and systems and other startups were providing regional, exchange, and Internet services. In 1996, AT&Tvoluntarilyseparated its long distance serviceand telephone equipment businesses to better meet the competition. The former kept the AT&T name, and the latter took on the name Lucent Technologies. Bell Labs remained with Lucent, and AT&T Labs was formed. Bellcore was put up for sale, as the consolidating and competing RBOCs found they did not need a joint lab. Because of a wealth of new information, OFTIII was divided into two books, A and B, covering systems and components, respectively. Many topics of the previous volumes, such as fibers, cables, and laser sources, are updated. But a much larger list of topics covers fields not previously included. In A , for example, transceiver design, EDFAs, laser sources, optical fiber components,planar (silica on silicon) integrated circuits, lithium niobate devices, and photonic switching are reviewed. And in By SONET (synchronous optical network) standards, fiber and cable design, fiber nonlinearities, polarization effects, solitons, terrestrial and undersea systems, high bitrate transmission, analog 4 Ivan P Kaminow . cable systems,passive optical networks (PONS),and multiaccess networks are covered. Throughout the two books, erbium amplifiers and WDM are common themes. It is difficult to overstate the impact these two technologies have had in both generating and supporting the telecommunications revolution that coincided with their commercialintroduction. The EDFA was first reported in about 1987 by researchers at Southampton Universityin the UK and at AT&T Bell Labs. In 1990, driven by the prospect of vast savings offered by WDM transmission using EDFAs, Bell Labs began to develop long-haul WDM systems. By 1996, AT&T and Alcatel had installed the first transatlantic cable with an EDFA chain and a single 5 Gb/s optical channel. AT&T installed the first commercialterrestrial WDM system employing EDFAs in 1995. Massive deployment of WDM worldwide soon followed. WDM has made the exponential tratlic growth spurred by the coincident introduction of the Internet browser economically feasible. If increased TDM bitrates and multiple fibers were the only alternative, the enthusiastic users and investors in the Internet would have been priced out of the market. Optical Fiber TelecommunicationsIV BA CKGROWD There was considerableexcitementin the lightwaveresearchcommunityduring the 1970s and early 1980s as wonderful new ideas emerged at a rapid pace. The monopoly telephone system providers, however, were less enthusiastic. They were accustomed to moving at their own deliberate pace, designingequipment to install in their own systems, which were expected to have a long economic life. The long-range planners projected annual telephone voice traffic growth in the United States at about 5-lo%, based on population and business growth. Recent years, on the other hand, have seen mind-numbing changes in the communication business-especially for people brought up in the telephone environment. The Internet browser spawned a tremendous growth in data traillc, which in turn encouraged visions of tremendous revenue growth. Meanwhile, advances in WDM technology and its wide deploymentsynergistically supported the Internet traffic and enthusiasm. As a result, entrepreneurs invested billions of dollars in many companies vying for the same slice of pie. The frenzy reached a peak in the spring of 2000 and then rapidly melted down as investors realized that the increased network capacity had already outstripped demand. As of October 2001, the lightwave community is waiting for a recovery from the current industry collapse. Nevertheless, the technical advances achieved during these last five years will continue to impact telecommunications for years to come. Thus, we are proud to present a comprehensive and forward-lookingaccount of these accomplishments. 1.Overview 5 Survey of O F T W A and B Advances in optical network architectures have followed component innovations. For example, the low loss fiber and double heterostructure laser enabled the first lightwave system generation; and the EDFA has enabled the WDM generation. Novel components (such as tunable lasers, MEMS switches, and planar waveguide devices) are making possible more sophisticatedoptical networks. At the same time, practical network implementationsuncover the need for added device functionality and very low cost points. For example, 40 Gb/s systems need dynamic dispersion and PMD compensationto overcome system impairments. We have divided OFTIV into two books: book A comprises the component chapters and book B the system and system impairment chapters. BOOKA: COMPONENTS Design of Optical Fibers for Communications Systems (Chapter 2) Optical fiber has been a key element in each of the previous volumes of OFT. The present chapter by DiGiovanni, Boncek, Golowich, Das, Blyler, and White reflects a maturation of the field: fiber performance must now go beyond simple low attenuation and must exhibit critical characteristicsto support the high speeds and long routes on terrestrial and undersea systems. At the same time, fiber for the metropolitan and access markets must meet demandingprice points. The chapter reviews the design criteria for a variety of fibers of current commercial interest. For the traditional long-haul market, impairments such as dispersion slope and polarization mode dispersion (PMD) that were negligible in earlier systems are now limiting factors. If improved fiber design is unable to overcome these limits, new components will be required to solve the problem. These issues are addressed again from different points of view in later systems and components chapters in O F T N A and B. The present chapter also reviews a variety of new low-cost fiber designs for emerging metropolitan and access markets. Further down the network chain, the design of multimode glass and plastic fiber for the highly cost-sensitivelocal area network market are also explored. Finally, current research on hollow core and photonic bandgap fiber structures is summarized. New Materials for Optical Amplifiers (Chapter 3) In addition to transport, fiber plays an important role as an amplifying medium. Aluminum-doped silica has been the only important commercial host and erbium the major amplifying dopant. Happily, erbium is soluble in AI-silica and provides gain at the attenuation minimum for silica transmission fiber. Still, researchers are exploring other means for satisfying demands 6 Ivan P. Kaminow for wider bandwidth in the 1550nm region as well as in bands that might be supported by other rare-earth ions, which have low efficiency in silica hosts. Ellison and Minelly review research on new fiber materials, including fluorides, alumina-doped silica, antimony silicates, and tellurite. They also report on extended band erbium-doped fiber amplifiers (EDFAs), thulium-doped fiber amplifiers, and 980 nm ytterbium fiber lasers for pumping EDFAs. Advances in Erbium-Doped Fiber Amplifiers (Chapter 4) The development of practical EDFAs has ushered in a generation of dense WDM (DWDM) optical networks. These systems go beyond single frequency or even multifrequency point-to-point links to dynamic networks that can be reconfigured by adddrop multiplexers or optical cross-connects to meet varying system demands. Such networks place new requirements on the EDFAs: they must maintain flatness over many links, and they must recover from sudden drops or adds of channels. And economics drives designs that provide more channels and denser spacing of channels. Srivastava and Sun summarize recent advances in EDFA design and means for coping with the challenges mentioned above. In particular, they treat long wave L-band amplifiers, which have more than doubled the conventional C-band to 84nm. They also treat combinations of EDFA and Raman amplification, and dynamic control of gain flatness. Raman Amplification in Lightwave Communication Systems (Chapter 5) Raman amplification in fibers has been an intellectual curiosity for nearly 30 years; the large pump powers and long lengths required made Raman amplifiers seem impractical. The advent of the EDFA appeared to drive a stake into the heart of Raman amplXers.Now, however, Raman amplifiers are rising along with the needs of submarine and ultralong-haul systems. More powerful practical diode pumps have become available; and the ability to provide gain at any wavelength and with low effective noise figure is now recognized as essential for these systems. Rottwitt and Stentz review the advances in distributed and lumped Raman amplifierswith emphasis on noise performance and recent system experiments. Electrooptic Modulators (Chapter 6) Modulators put the payload on the optical carrier and have been a focus of attention from the beginning. Direct modulation of the laser current is often the cheapest solution where laser linewidth and chirp are not important. However, for high performance systems, external modulators are needed. Modulators based on the electrooptic effect have proven most versatile in meeting performance requirements, although cost may be a constraint. 1.Overview 7 Titanium-diffusedlithium niobate has been the natural choice of material, in that no commercial substitutes have emerged in nearly 30 years. However, integrated semiconductor electroabsorption modulators are now offering strong competition on the cost and performance fronts. Mahapatra and Murphy briefly compare electroabsorption-modulated lasers (EMLs) and electrooptic modulators. They then focus on titaniumdiffused lithium niobate modulators for lightwave systems. They cover fabrication methods, component design, system requirements, and modulator performance. Mach-Zehnder modulators are capable of speeds in excess of 40Gb/s and have the ability to control chirp from positive through zero to negative values for various system requirements. Finally, the authors survey research on polymer electroopticmodulators, which offer the prospect of lower cost and novel uses. Optical Switching in Transport Networks: Applications, Requirements, Architectures, Technologies, and Solutions (Chapter 7) Early DWDM optical line systems provided simple point-to-point links between electronic end terminals without allowing access to the intermediate wavelength channels. Today’s systems carry over 100 channels per fiber and new technologies allow intermediate routing of wavelengths at add/drop multiplexers and optical cross-connects. These new capabilities allow “optical layer networking,” an architecture with great flexibility and intelligence. AI-Salameh, Korotky, Levy, Murphy, Patel, Richards, and Tentarelli explore the use of optical switchingin modern networking architectures. After reviewing principles of networking, they consider in detail various aspects of the topic. The performance and requirements for an optical cross connect (OXC) for opaque (with an electronic interface and/or electronic switch fabric) and transparent (all-optical) technologies are compared. Also, the applications of the OXC in areas such as provisioning, protection, and restoration are reviewed. Note that an OXC has all-optical ports but may have internal electronics at the interfaces and switch fabric. Finally, several demonstration OXCs are studied, including small optical switch fabrics, wavelength-selective OXCs, and large strictly nonblocking cross connects employing microelectromechanical system (MEMS) technology. These switches are expected to be needed soon at core network nodes with 1000 x 1000 ports. Applications for Optical Switch Fabrics (Chapter 8) Whereas the previous chapter looked at OXCs from the point of view of the network designer, Zirngibl focuses on the physical design of OXCs with capacities greater than 1Tb/s. He considers various design options including MEMS switch fabrics, transparent and opaque variants, and nonwavelength-blocking 8 Ivan P Kaminow . configurations He finds that transport in the backplane for very large capacity (bitrate x port number) requires optics in the interconnectsand switch fabric. He goes beyond the cross-connect application, which is a slowly reconfigurable circuit switch, to consider the possibility of a high-capacity packet switch, which, although schematically similar to an OXC, must switch in times short relative to a packet length. Again the backplaneproblem dictatesan optical fabric and interconnects. He proposes tunable lasers in conjunction with a waveguide grating router as the fast optical switch fabric. Planar Lightwave Devices for WDM (Chapter 9) The notion of integrated optical circuits, in analogy with integrated electronic circuits, has been in the air for over 30 years, but the vision of large-scale integration has never materialized. Nevertheless, the concept of small-scale planar waveguide circuits has paid o f handsomely. Optical waveguiding prof vides efficientinteractions in lasers and modulators, and novel functionalityin waveguide grating routers and Bragg gratings. These elements are often linked together with waveguides. Doerr updates recent progress in the design of planar waveguides, starting with waveguide propagation analysis and the design of the star coupler and waveguide grating router (or arrayed waveguide grating). He goes on to describe a large number of innovativeplanar devices such as the dynamic gain equalizer, wavelength selective cross connect, wavelength adddrop, dynamic dispersion compensator, and the multifrequency laser. Finally, he compares various waveguide materials: silica, lithium niobate, semiconductor, and polymer. Fiber Grating Devices in High-Performance Optical Communication Systems (Chapter 10) The fiber Bragg grating is ideally suited to lightwave systems because of the ease of integrating it into the fiber structure. The technology for economically fabricating gratings has developed over a relatively short period, and these devices have found a number of applicationsto which they are uniquely suited. For example, they are used to stabilize lasers, to provide gain flattening in EDFAs, and to separate closely spaced WDM channels in adddrops. Strasser and Erdogan review the materials aspects of the major approaches to fiber grating fabrication. Then they treat the properties of fiber gratings analytically. Finally, they review the device properties and applications of fiber gratings. Pump Laser Diodes (Chapter 11) Although EDFAs were known as early as 1986, it was not until a high-power 1480nm semiconductorpump laser was demonstratedthat people took notice. 1.Overview 9 Earlier, expensive and bulky argon ion lasers provided the pump power. Later, 980nm pump lasers were shown to be effective. Recent interest in Raman amplifiers has also generated a new interest in 1400nm pumps. Ironically, the first 1480nm pump diode that gave life to EDFAs was developed for a Raman amplifier application. Schmidt, Mohrdiek, and Harder review the design and performance of 980 and 1480nm pump lasers. They go on to compare devices at the two wavelengths, and discuss pump reliability and diode packaging. TelecommunicationLasers (Chapter 12) Semiconductor diode lasers have undergone years of refinement to satisfy the demands of a wide range of telecommunication systems. Long-haul terrestrial and undersea systems demand reliability, speed, and low chirp; short-reach systems demand low cost; and analog cable TV systems demand high power and linearity. Ackerman, Eng, Johnson, Ketelsen, Kiely, and Mason survey the design and performance of these and other lasers. They also discuss electroabsorption modulated lasers (EMLs) at speeds up to 40 Gb/s and a wide variety of tunable lasers. VCSELs for Metro Communications (Chapter 13) Vertical cavity surface emitting lasers (VCSELs) are employed as low-cost sources in local area networks at 850nm. Their cost advantage stems from the ease of coupling to fiber and the ability to do wafer-scale testing to eliminate bad devices. Recent advances have permitted the design of efficient long wavelength diodes in the 1300-1600 nm range. Chang-Hasnain describes the design of VCSELs in the 1310 and 1550nm bands for application in the metropolitan market, where cost is key. She also describes tunable designs that promise to reduce the cost of sparing lasers. Semiconductor Optical Amplifers (Chapter 14) The semiconductor gain element has been known from the beginning, but it was fraught with difficulties as a practical transmission line amplifier: it was difficult to reduce reflections, and its short time constant led to unacceptable nonlinear effects. The advent of the EDFA practically wiped out interest in the semiconductor optical amplifier (SOA) as a gain element. However, new applications based on its fast response time have revived interest in SOAs. Spiekman reviews recent work to overcome the limitations on SOAs for amplification in single-frequency and WDM systems. The applications of main interest, however, are in optical signal processing, where SOAs are used in wavelength conversion, optical time division multiplexing, optical phase 10 Ivan P Kaminow . conjugation, and all-optical regeneration. The latter topic is covered in detail in the following chapter. All-Optical Regeneration: Principles and WDM Implementation (Chapter 15) A basic component in long-haul lightwave systems is the electronic regenerator. It has three functions: reamplifying, reshaping, and retiming the optical pulses. The EDFA is a 1R regenerator; regenerators without retiming are 2R; but a full-scale repeater is a 3R regenerator. A separate 3R electronic regenerator is required for each WDM channel after a fixed system span. As the bitrate increases, these regenerators become more expensive and physically more difficult to realize. The goal of ultralong-haul systems is to eliminate or minimize the need for electronic regenerators (see Chapter 5 in Volume B). Leclerc, Lavigne, Chiaroni, and Desurvire describe another approach, the all-optical3R regenerator. They describe avariety of techniques that have been demonstratedfor both single channel and WDM regenerators. They argue that at some bitrates, say 40 Gb/s, the optical and electronic alternatives may be equally difficult and expensive to realize, but at higher rates the all-optical version may dominate. High Bitrate Tkansmitters, Receivers, and Electronics (Chapter 16) In high-speed lightwave systems, the optical components usually steal the spotlight. However, the high bitrate electronics in the terminals are often the limiting components. Kasper, Mizuhara, and Chen review the design of practical high bitrate (10 and 40 Gb/s) receivers, transmitters, and electronic circuits in three separate sections. The first section reviews the performance of various detectors, analyzes receiver sensitivity, and considers system impairments. The second section covers directly and externally modulated transmitters and modulation formats like return-to-zero (RZ) and chirped RZ (CRZ). The final section covers the electronic circuit elements found in the transmitters and receivers, including broadband amplifiers, clock and data recovery circuits, and multiplexers. BOOK B: SYSTEMS AND IMPAIRMENTS Growth of the Internet (Chapter 2) The explosion in the telecommunicationsmarketplace is usually attributed to the exponentialgrowth of the Internet, which began its rise with the introduction of the Netscape browser in 1996. Voice traffic continues to grow steadily, but data traffic is said to have already matched or overtaken it. A lot of selfserving myth and hyperbole surround these fuzzy statistics. Certainly claims of doubling data t r a c every three months helped to sustain the market frenzy. 1.Overview 11 On the other hand, the fact that revenues from voice traffic still far exceed revenues from data was not widely circulated. Coffman and Odlyzko have been studying the actual growth of Internet traffic for several years by gathering quantitative data from service providers and other reliable sources. The availability of data has been shrinking as the Internet has become more commercial and fragmented. Still, they find that, while there may have been early bursts of three-month doubling, the overall sustained rate is an annual doubling. An annual doubling is a very powerful growth rate; and, if it continues, it will not be long before the demand catches up with the network capacity. Yet, with prices dropping at a comparable rate, faster traffic growth may be required for strong revenue growth. Optical Network Architecture Evolution (Chapter 3) The telephone network architecture has evolved over more than a century to provide highly reliable voice connections to a global network of hundreds of millions of telephones served by different providers. Data networks, on the other hand, have developed in a more ad hoc fashion with the goal of connecting a few terminals with a range of needs at the lowest price in the shortest time. Reliability, while important, is not the prime concern. Strand gives a tutorial review of the Optical Transport Network employed by telephone service providers for intercity applications. He discussesthe techniques used to satisfy the traditional requirements for reliability, restoration, and interoperability.He includes a refresher on SONET (SDH). He discusses architectural changes brought on by optical fiber in the physical layer and the use of optical layer cross connects. Topics include all-optical domains, protection switching, rings, the transport control plane, and business trends. Undersea Communication Systems (Chapter 4) The oceans provide a unique environment for long-haul communication systems. Unlike terrestrial systems, each design starts with a clean slate; there are no legacy cables, repeater huts, or rights-of-way in place and few international standards to limit the design. Moreover, there are extreme economic constraints and technological challenges For these reasons, submarine systems designers have been the first to risk adopting new and untried technologies, leading the way for the terrestrial ultralong-haul systemdesigners (see Chapter 5). Following a brief historical introduction, Bergano gives a tutorial review of some of the technologies that promise to enable capacities of 2Tbh on a single fiber over transoceanic spans. The technologies include the chirped RZ (CRZ) modulation format, which is compared briefly with NRZ, RZ, and dispersion-managed solitons (see Chapters 5,6, and 7 for more on this topic). He also discusses measures of system performance (the Q-factor), forward 12 Ivan P Kaminow . error correcting(FEC) codes (see Chapters 5 and 17), long-haulsystem design, and future trends. High Capacity, Ultralong-Haul Transmission (Chapter 5) The major hardware expense for long-haul terrestrial systems is in electronic terminals, repeaters, and line cards. Since WDM systems permit traffic with various destinationsto be bundled on individualwavelengths, great savingscan be realized if the unrepeatered reach can be extended to 2000-5000 km, allowing traffic to pass through nodes without optical-to-electrical (O/E) conversion. As noted in connection with Chapter 4, some of the technologypioneered in undersea systems can be adapted in terrestrial systems but with the added complexities of legacy systems and standards. On the other hand, the terrestrial systems can add the flexibilityof optical networkingby employingoptical routing in add/drops and OXCs (see Chapters 7 and 8) at intermediatepoints. Zyskind, Barry, Pendock, and Cahill review the technologies needed to design ultralong-haul (ULH) systems. The technologies include EDFAs and distributed Raman amplification, novel modulation formats, FEC, and gain flattening. They also treat transmission impairments (see later chapters in this book) such as the characteristics of fibers and compensators needed to deal with chromatic dispersion and PMD. Finally, they discuss the advantages of optical networking in the efficient distribution of data using IP (Internet Protocol) directly on wavelengths with meshes rather than SONET rings. Pseudo-Linear Transmission of High-speed TDM Signals: 40 and 160 Gb/s (Chapter 6) A reduction in the cost and complexity of electronic and optoelectroniccomponents can be realized by an increase in channel bitrate, as well as by the ULH techniquesmentioned in Chapter 5. The higher bitrates, 40 and 160Gb/s, present their own challenges, among them the fact that the required energy per bit leads to power levels that produce nonlinear pulse distortions. Newly discovered techniques of pseudo-linear transmission offer a means for dealing with the problem. They involve a complex optimization of modulation format, dispersion mapping, and nonlinearity. Pseudo-linear transmission occupies a space somewhere between dispersion-mapped linear transmission and nonlinear soliton transmission (see Chapter 7). Essiambre, Raybon, and Mikkelsen first present an extensive analysis of pseudo-linear transmission and then review TDM transmission experiments at 40 and 160Gb/s. Dispersion Managed Solitons and Chirped RZ: What Is the Difference? (Chapter 7) Menyuk, Carter, Kath, and Mu trace the evolution of soliton transmission to its present incarnation as Dispersion Managed Soliton (DMS) transmission 1.Overview 13 and the evolution of NRZ transmission to its present incarnation as CRZ transmission. Both approaches depend on an optimization of modulation format, dispersion mapping, and nonlinearity, defined as pseudo-linear transmission in Chapter 6 and here as “quasi-linear” transmission. The authors show how both DMS and CRZ exhibit aspects of linear transmission despite their dependence on the nonlinear Icerr effect. Remarkably, they argue that, despite widely disparate starting points and independent reasoning, the two approaches unwittingly converge in the same place. Still, on their way to convergence, DMS and CRZ pulses exhibit different characteristicsthat suit them to different applications: For example, CRZ produces pulses that merge in transit along a wide undersea span and reform only at the receiver ashore, while DMS produces pulses that reform periodically, thereby permitting access at intermediate adddrops. Metropolitan Optical Networks (Chapter 8) For many years the long-haul domain has been the happy hunting ground for lightwave systems, since the cost of expensive hardware can be shared among many users. Now that component costs are moderating, the focus is on the metropolitan domain where costs cannot be spread as widely. Metropolitan regions generally span ranges of 10 to 100km and provide the interface with access networks (see Chapters 9, 10, and 11). SONETBDH rings, installed to serve voice traffic, dominate metropolitan networks today. Ghani, Pan, and Chen trace the developing access users, such as Internet service providers, local area networks, and storage area networks. They discuss a number of WDM metropolitan applications to better serve them, based on optical networking via optical rings, optical adddrops, and OXCs. They also consider 1P over wavelengths to replace SONET. Finally, they discuss possible economical migration paths from the present architecture to the optical metropolitan networks. The Evolution of Cable TV Networks (Chapter 9) Coaxial analog cable TV networks were substantially upgraded in the 1990s by the introduction of linear lasers and single-modefiber. Hybrid Fiber Coax (HFC) systems were able to deliver in excess of 80 channels of analog video plus a wide band suitable for digital broadcast and interactive services over a distance of 60 km. Currently high-speed Internet access and voice-over-IP telephony have become available,making HFC part of the telecommunications access network. Lu and Sniezko outline past, present, and future HFC architectures. In particular, the mini fiber node (mFN) architecture provides added capacity for two-way digital as well as analog broadcast services. They consider a number of mFNvariants based on advancesin RF, lightwave,and DSP (digital signal processor) technologies that promise to provide better performance at lower cost. 14 Ivan P Kaminow . Optical Access Networks (Chapter 10) The access portion of the telephone network, connecting the central office to the residence, is called the “loop.” By 1990 half the new loops in the United Stateswere served by digital loop carrier (DLC), a fiber severalmiles long from the central office to aremote terminal in aneighborhoodthat connects to about 100 homes with analog signals over twisted pairs. Despite much anticipation, fiber hasn’t gotten much closer to residences since. The reason is that none of the approachesproposed so far is competitivewith existing technology for the applicationspeople will buy. Harstead and van Heyningen survey numerous proposals for Fiber-in-theLoop (FITL) and Fiber-to-the-X (FTTX), where X = Curb, Home, Desktop, etc. They consider the applications and costs of these systems. Considerable creativity and thought have been devoted to fiber in the access network, but the economics still do not work because the costs cannot be divided among a suflicient number of users. An access technology that is successful is Digital Subscriber Line (DSL) for providing high-speed Internet over twisted pairs in the loop. DSL is reviewed in an Appendix. Beyond Gigabit: Development and Application of High-speed Ethernet Technology (Chapter 11) Ethernet is a simple protocol for sharing a local area network (LAN). Most of the data on the Internet start as Ethernet packets generated by desktop computers and system servers. Because of their ubiquity, Ethernet line cards are cheap and easy to install. Many people now see Ethernet as the universal protocol for optical packet networks. Its speed has already increased to 1000Mb/s, and 10 Gb/s is on the way. Lam describes the Ethernet system in detail from protocols to hardware, including 10 Gb/s Ethernet. He shows applications in LANs, campus, metropolitan, and long distance networks. Photonic Simulation Tools (Chapter 12) In the old days, new devices or systems were sketched on a pad, a prototype was put together in the lab, and its performance tested. In the present climate, physical complexity and the expense and time required rule out this bruteforce approach, at least in the early design phase. Instead, individual groups have developed their own computer simulators to test numerous variations in a short time with little laboratory expense. Now, several commercial vendors offer general-purposesimulators for optical device and system development. Lowery relates the history of lightwave simulators and explains how they work and what they can do. The user operates from a graphic user interface (GUI) to select elements from a library and combine them. The simulated device or system can then be run and measured as in the lab to determine 1.Overview 15 attributes like the eye-diagram or bit-error-rate. In the end, a physical prototype is required because of limits on computation speed among other reasons. THE PRECEDING CRAPTERS RAVE DEALT WITH SYSTEM DESIGN; THE REMAIMNG CHAPTERS DEAL WlTH SYSTEM IMPAIRMENTS AND METHODS FOR MITIGATING THEM Nonlinear Optical Effects in WDM Systems (Chapter 13) Nonlinear effects have been mentioned in different contexts in several of the earlier system chapters. The Kerr effect is an intrinsic property of glass that causes a change in refractive index proportional to the optical power. Bayvel and Killey give a comprehensive review of intensity-dependent behavior based on the K n effect. They cover such topics as self-phase mode ulation, cross-phase modulation, four-wave mixing, and distortions in NRZ and RZ systems. F id and "unable Management of Fiber Chromatic Dispersion % e (Chapter 14) Chromatic dispersion is a linear effect and as such can be compensated by adding the complementary dispersion before any significant nonlinearities intervene. Nonlinearities do intervene in many of the systems previously discussed so that periodic dispersion mapping is required to manage them. Willner and Hoanca present a thorough taxonomy of techniques for compensating dispersion in transmission fiber. They cover fixed compensation by fibers and gratings, as well as tunable compensation by gratings and other novel devices. They also catalog the reasons for incorporating dynamic as well as k e d compensation in systems. Polarization Mode Dispersion (Chapter 15) Polarization mode dispersion (PMD), like chromatic dispersion, is a linear effect that can be compensated in principle. However, fluctuationsin the polarization mode and fiber birefringence produced by the environment lead to a dispersion that varies statistically with time and frequency. The statistical nature makes PMD difficult to measure and compensate for. Nevertheless, it is an impairment that can kill a system, particularly when the bitrate is large (> 10 Gb/s) or the fiber has poor PMD performance. Nelson, Jopson, and Kogelnik offer an exhaustive survey of PMD covering the basic concepts, measurement techniques, PMD measurement, PMD statistics for first- and higher orders, PMD simulation and emulation, system impairments, and mitigation methods. Both optical and electrical PMD compensation (see Chapter 18) are considered. 16 Ivan P Kamhow . Bandwidth Efficient Formats for Digital Fiber Transmission Systems (Chapter 16) Early lightwave systems employedNRZ modulation; newer long-haul systems are using RZ and chirped RZ to obtain better performance. One goal of system designersis to increase spectral efficiencyby reducing the RF spectrum required to transmit a given bitrate. Conradi examines a number of modulation formats well known to radio engineers to see if lightwave systems might benefit from their application. He reviews the theory and DWDM experiments for such formats as M-ary ASK, duo-binary, and optical single-sideband.He also examines RZ formats combined with various types of phase modulation, some of which are related to discussions of CRZ in the previous Chapters 4-7. Error-Control Coding Techniques and Applications (Chapter 17) Error-correctingcodes axe widelyused in electronics, e.g., in compactdisc players, to radically improve system performance at modest cost. Similar forward error correcting codes (FEC) are used in undersea systems (see Chapter 4) and are planned for ULH systems (Chapter 5). Win, Georghiades, Kumar, and Lu give a tutorial introduction to coding theory and discuss its application to lightwave systems. They conclude with a critical survey of recent literature on FEC applications in lightwave systems, where FEC provides substantial system gains. Equalization Techniques for Mitigating Transmission Impairments (Chapter 18) Chapters 14and 15describe optical means for compensatingthe linear impairments caused by chromatic dispersion and PMD. Chapters 16 and 17 describe two electronic means for reducing errors by novel modulation formats and by FEC. This chapter discusses a third electronic means for improvingperformance using equalizer circuits in the receiving terminal, which in principle can be added to upgrade an existing system. Equalization is widely used in telephony and other electronic applications. It is now on the verge of application in lightwave systems. Win, Vitetta, and Winters point out the challengesencounteredin lightwave applications and survey the mathematical techniques that can be employed to mitigate many of the impairments mentioned in previous chapters. They also describe some of the recent experimental implementations of equalizers. Additional discussion of PMD equalizers can be found in Chapter 15. Duty Cycle lOO%(Ng) 50% 33% 20% 10% 40 Gbls 2 Single channel Distance of 80 km TrueWave@ with D = 4 ps/(km nm) -250 -125 0 -250 -125 0 -250 -125 0 -250 -125 0 -250 -2-1 0 I 2 3 Eye Closure Penalty (dB) -12J I 0 Pceromp Ipdnmm) k c a m p (p4nml k r o m p (pshml Prccomp lpdnrni k c o m p lpdnrnl Plate 1 Eye closure penalties as a function of modulation formats and launch (average) power for 40-Gbh single-channel transmission over 80 km of TrueWaveTMfiber [TrueWaveTMparameters are given in Table 6.1 except for the value of D = 4ps/(km nm) here]. Full dispersion and dispersion slope compensation are assumed before the postcompensation. Each plot in the matrix of plots shows the color-coded eye closure penalties Ceyeas a function of pre- and postcompensation. Only a small improvement in eye opening (essentially the back-to-back difference in eye opening) can be seen when decreasing the duty cycle. Only when the duty cycle is reduced to a value as low as 10% is a significant improvement in transmission observed. Amplifier noise is not included. --- Duty Cycle 33% 20% 10% I 5 I I RFrornp lpdnm) F-omp lpdnml Rsomp lpdnml F-omp (~Jn,nm) R-mp Ipdoml Plate 2 Identical to Plate 1 except for STD unshifted fiber (parameters given in Table 6.1). For large duty cycles (NRZ and 50% duty cycle), transmission is limited to lower powers than TrueWaveTMfiber but rapidly increases as the duty cycle decreases. 40 Gb/s Single channel 12 dBm 8 spans of 80 !un TrueWaveTM/DSF with D = 2 ps/(km nm) - I O 1 2 3 4 Eye Closure Penalty (dB) Recomp. (pdnm) Reeomp. (pdnm) Recomp. (pdnm) Recomp. (pdnm) Plate 3 Eye closure penalties after 8 spans of 80 km as a function of residual dispersion per span and modulation format. The transmission fiber has the same parameters as TrueWaveTM (Table 6.1) except we use here D = 2 ps/(km nm). The dashed lines are the points of zero net residual dispersion. Two regimes of transmission are present. A first regime (solitonic regime) has optimum transmission with a net positive residual dispersion. In this regime the solitonic effect is at play and is responsiblefor the compensationof dispersionby nonlinearity (even for NRZ!). The second regime (pseudo-linearregime) has its optimum transmission at zero net residualdispersion. Residual DisDersion per Span 0 40 Gb/s Single channel 12 dBm 8 spans of 80 km TrueWavem I with D = 4 pd&m nm) -la, 1 0 1 2 3 4 Eye Closure Penalty (dB) -200 -200 -100 0 a -100 0 Reeomp. (pdnm) Precomp. (pdnm) -200 -100 0 Reeomp (plnm) Precomp. (pdnm) Plate 4 Same as Plate 3 except D = 4ps/(km nm). 0 Residual Dispersion per Span 8ps/m 16 ps/nm 24 pdnm _____________ -_____________ 40 Gb/s Single channel 12 dBm 8 spans of 80 km TrueWavem with D = 8 pd&m nm) M..., ,. . -1 0 1 2 3 1 Eye Closure Penalty (dB) m -100 o -200 -100 o -200 -100 o Recomp. (jdnm) PReomp (pdnm) Reeomp. (pslnm) Plate 5 Same as Plate 3 except D = 8 ps/(km nm). Residual Dispersion per Span 40 Gb/s Single channel 12 dBm 8 spans of 80 km STD Unshifted Fiber with D = 17 pd@m nm) - 1 0 1 2 3 4 Eye Closure Penalty (dB) Rsomp. (pS/nrnl Rsomp. (pslnml F’recomp. ( g n m l Pr~comp. (pdnm) Plate 6 Same as Plate 3 except for STD unshifted fiber D = 17ps/(km nm) and A,ff = 80 km2. Fiber 1 35 Fiber 2 I .- 2 5 ‘ 2 g 20 F 15 10 15,J 1520 1530 1540 1550 Wavelength [nm] 1560 1510 1520 1530 1540 1550 Wavelength [nm] 1560 Plate 7 Contour plots of the simultaneous DGD measurements of two fibers in the same embedded cable over a 36-day period. The mean DGDs averaged over time and wavelength were 2.75 and 2.89 ps for fibers 1 and 2, respectively. Data is courtesy of Magnus Karlsson. Chapter 2 Growth of the Internet Kerry G. Coffman AT&T Labs-Research, Middletown, New Jersey Andrew M. Odlyzko University of Minnesota, Minneapolis, Minnesota Abstract The Internet is the main cause of the recent explosion of activity in opticalfiber telecommunications.The high growth rates observed on the Internet, and the popular perception that growth rates were even higher, led to an upsurge in research, development, and investment in telecommunications.The telecom crash of 2000 occurred when investors realized that transmission capacity in place and under construction greatly exceeded actual traffic demand. This chapter discussesthe growth of the Internet and compares it with that of other communication services. It also presents speculations about future developments. Internet traffic is growing, approximatelydoublingeach year. There are reasonable arguments that it will continue to grow at this rate for the rest of this decade. If this happens, then in a few years we may have a rough balance between supply and demand. 1. Introduction Optical fiber communication was initially developed for the voice phone system. The feverish level of activity that we have experienced since the late 199Os, though, was caused primarily by the rapidly rising demand for Internet connectivity. The Internet has been growing at unprecedented rates. Moreover, because it is versatile and penetrates deeply into the economy, it is affecting all of society, and therefore has attracted inordinate amounts of public attention. The aim of this chapter is to summarize the current state of knowledge about the growth rates of the Internet, with special attention paid to the implications for fiber optic transmission. We also attempt to put the growth rates of the Internet into the proper context by providing comparisons with other communications services. The overwhelmingly predominant view has been that Internet traffic (as measured in bytes received by customers) doubles every 3 or 4 months. 17 OPTICAL FIBER TELECOMMUNICATIONS, VOLUME TVB Copyright 0 2002, Elsevier Scienm (USA). ALI rightsof reproduction in any form reserved. ISBN &12-395173-9 18 Kerry G. Coffman and Andrew M. Odlyzko Such unprecedented rates (corresponding to traffic increasing by factors of between 8 and 16 each year) did prevail within the United States during the crucial 2-year period of 1995 and 1996, when the Internet first burst onto the scene as a major new factor with the potential to transform the economy. However, as we pointed out in [CoffmanOl] (written in early 1998, based on data through the end of 1997), by 1997those growth rates subsided to approximate the doubling of traffic each year that had been experienced in the early 1990s. A more recent study [CoffmanO2] provided much more evidence, and in particular more recent evidence, that traffic has about doubled each year since 1997. (We use a doubling of traffic each year to refer to growth rates between 70 and 150% per year, with the wide range reflecting the uncertainties in the estimates.) Other recent observers also found that Internet traffic is about doubling each year. The evidence was always plentiful, and the only thing lacking was the interest in investigatingthe question. By 2000, though, the myth of Internet traffic doubling every 3 or 4 months was getting hard to accept. Very simple arithmetic shows that such growth rates, had they been sustained throughout the period from 1995 (when they did hold) to the end of 2000, would have produced absurdly high tr&c volumes. For example, at the end of 1994, traffic on the NSFNet backbone, which was well instrumented, came to about 15TB/month. Had just that traffic grown at 1300% per year (which is what a doubling every 3 months corresponds to), by the end of 2000, there would have been about 250,000,000 TB/month of backbone traffic in the United States. If we assume there are 150million Internet users in the United States, that would produce a data flow of about 5Mb/s for each user around the clock. The assumption of a doubling of traffic every 4 months produces traffic volumes that are only slightly less absurd. Table 2.1 shows our estimates for traffic on the Internet. The data for 1990 through 1994 is that for the NSFNet backbone, and therefore is very precise. It is incomplete only to the extent of neglecting what is thought to have been small fractions of traffic that went completely through other backbones. The data for 1996 through 2000 are our estimates, and the wide ranges reflect the uncertainties caused by the lack of comprehensive data. Table 2.2 presents our estimates of the tr&c on various long-distance networks at the end of 2000. The voice network still dominated, but it will likely be surpassed by the public Internet within a year or two. (For details of the measurements used to convert voice traffic to terabytes and related issues, see [CoffmanOl].) In terms of bandwidth, the Internet is already dominant. However, it is hard to obtain good figures, since, as we discuss later, the bandwidth of Internet backbones jumps erratically. In terms of dollars, though, voice still provides the lion’s share (well over 80%) of total revenues. We concentrate in this chapter (as in our previous papers [CoffmanOl, CoffmanO21) on the growth rates in Internet tr&c, as measured in bytes. For many purposes, it is the other measures, namely bandwidth and revenues, that are more important. 2. Growth of the Internet Table 2.1 Traffic on Internet Backbones in the United States. Data are estimated traffic in terabytes (TB) during December of that year 19 Year 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 TB/month 1 .o 2.0 4.4 8.3 16.3 ? 1,500 2,5004,000 5,00&8,000 10,00&16,000 20,00&35,000 Tuble 2.2 Traffic on U.S. Long-Distance Networks, Year-End 2000 U.S. voice Internet Other public data networks Private line 53,000 20,00&35,000 3,000 6,000-1 1,000 The reason we look at traffic is that we find more regularity there, and in the long run, we expect that there will be direct (although not linear) relations between traffic and the other measures. In particular, based on what we have observed so far, we expect capacity to grow somewhat faster than traffic. The studies of [CoffmanOl, Coffman021 led to the proposal of a new form of Moore’s Law, namely that a doubling of Internet tr&c each year is a natural growth rate. This hypothesis is supported by the estimates of Table 2.1, as well as by evidence presented in [CoffmanOl, CoffmanO2] of many institutions whose data traffic has been growing at about that rate for many years. This “law” is discussed further in Section 8. It is not a law of nature, but rather, like the Moore’s Law for semiconductors,a reflection of the complicated interactions of technology, economics, and sociology. Whether this “law” continues to hold or not will have important implications for the fiberoptic transmission industry. 20 Kerry G. Coffman and Andrew M. Odlyzko Much of this chapter, especially Sections6 8 , is based on our earlier studies [CoffmanOl,CoffmanO2]. In Section 2, we present yet more evidence of how often popular perception and subsequent technologyand investmentdecisions are colored by myths that are easy to disprove, but which nobody had bothered to disprove for an astonishingly long time. In Section 3, we look at historical growth rates of various communication services and how they compare to the much higher growth rate of the Internet. Section4 is a brief review of the history of the Internet. Section 5 discusses some of the various types of growth rates that are relevant in different contexts. Section 6 presents the evidence about Internet traffic growth rates we have been able to assemble. Section7 is devoted to new sources of traffic that might create sudden surges of demand, such as Napster. Section 8 discusses the conventional Moore’s Law and the analog we are proposing for data traffic. Section 9 suggests a way of thinking about data-traffic growth, based on an analogy with the computer industry. Finally, Section 10 presents our conclusions. 2. Growth Myths and Reality Internet growth is an unusual subject, in that it has been attracting enormous attention but very little serious study. In particular, the general consensus has been that Internet traffic is doubling every 3 or 4 months. Yet no real evidence of that astronomical rate of growth was ever presented. As we discuss later, Internet traffic did grow at such rates in 1995 and 1996, but before and since it has been about doubling each year. At this point, we would like to point out the need for careful quantitative data in evaluating any claims about growth rates. Some examples of public claims that do not match reality are presented in [Coffman02].Here we discuss another case, this one concerning the widely held belief that any capacity that is installed will be quickly saturated. The British JANET network, which provides connectivity to British academic and research institutions, will be discussed in more detail later. What is important is that it is large (with three OC3 links across the Atlantic at the end of 2000), and has traffic statisticsgoing back several years available at http://bill.ja.net/. A press release, available at http://~.ja.net/press_release/archive~nnounce/index.html “Increase in as Transatlantic Bandwidth-28 May 1998” (but actually dated 3 June 1998), describedwhat happened when JANET’Stransatlanticlink was increasedfrom a single T3 to two T3s: With effect from Thursday 28 May 1998, JANET has been running a second T3 (45 Mbit/s) link to the North American Internet, bringing the total transatlantic bandwidth available to JANET to 90 Mbit/s. . . . Usage of the new capacity has been brisk, with the afternoon usage levels reaching in excess of 80 Mbit/s This is of course evidence of the suppressed demand imposed by the single T3 link 2. Growth of the Internet 21 operating previously. The fact that usage has risen so quickly on this occasion is also indicative of the improved domestic infrastructures.. .that now exist. This quote certainly appears to support the claim that demand for bandwidth is inexhaustible. One could easily conclude that traffic essentially doubled as soon as capacity doubled. The quote is imprecise, though, since it does not say how often those “afternoon usage levels” are “in excess of 80 Mbit/s,” nor does it say how those usage levels are measured. The usage statistics for JANET, available at http://bill.ja.net/, enable us to obtain precise information. Table 2.3 shows the transfer volumes on the more heavily utilized United States to United Kingdom part of the link for several days before and after the doubling of capacity of the link. (No data for May 27 is available, and the figures for May 28, the day the second T3 was put into operation, are suspiciously low, probably reflecting incomplete measurements, so those are not included.) Table 2.3 Traffic from the United States to the JANET Network during Late Spring 1998, When the Capacity Was Doubled Day GB 272.7 275.5 265.1 202.7 189.8 211.2 267.2 UtiZizution (%) 58.8 59.4 57.1 43.7 40.9 45.5 57.6 Wed 5/20 Thu 5/21 Fri 5/22 Sat 5/23 Sun 5/24 Mon 5/25 Tue 5/26 Wed 5/27 Thu 5/28 Fri 5/29 Sat 5130 Sun 5/31 Mon 6/01 Tue 6102 Wed 6/03 Thu 6/04 Fri 6/05 Sat 6/06 Sun 6/07 Mon 6/08 Tue 6/09 286.6 209.7 199.9 318.1 319.2 295.9 343.2 322.4 208.3 202.7 338.0 307.2 30.9 22.6 21.5 34.3 34.4 31.9 37.0 34.7 22.4 21.8 36.4 33.1 22 Kerry G. Coffman and Andrew M. Odlyzko What we observe is that although there was substantial growth in traffic after the capacity increase, suggesting that the transatlantic link had been a bottleneck, this increase was far more moderate than the popular Internetgrowth mythology or the JANET press release would make one think. While capacity doubled, traffic increased by less than a third. 3. Growth Rates of Other Communication Services Telecommunicationshas been a growth industry for centuries, but growth rates have generally been modest, except for a few episodes, such as the beginnings of the electric telegraph (see [Odlyzko2]). For example, the number of pieces of mail delivered in the United States grew by a factor of over 50,000 between 1800 and 2000, but that was a growth rate of about 5.6% per year. (If we adjust for population increase, we find a growth rate of about 3.5% in the mail volume per capita.) The number of phone calls in the United Statesgrew by a factor of over 230 between 1900 and 2000, for a compound annual growth rate of 5.6%. (The per capita growth rate was 4.2% during this period.) Long-distance calls grew faster, about 12% per year between 1930 and 2000, and transatlantic calls faster yet. (There was just one voice circuit between the United States and Europe in 1927, when service was inaugurated. It used radio to span the ocean. This single low quality link grew to 23,000 voice circuits to Western Europe by 1995, for a compound annual growth rate of capacity of 16%.) One communications industry that has been growing very rapidly recently is wireless communication. Table 2.4 shows the growth of the U.S. cell phone industry, with the number of subscribers as of June of each year, and the revenue figures obtained by doubling those of the fmt 6 months of each year (and thus seriously understating the full-year figure). In many other countries, wireless communication has developed faster and plays a bigger role than it does in the United States. Still, even in the United States, at the end of 2000, there were close to 100 million cell phones in use, and the rate of growth was far higher than for traditional wired voice services. The cell phone example is worth keeping in mind, because it shows that volume of traffic or even the number of users has only a slight correlation to value. In the United States (unlike several other countries), there were more Internet users than cell phone subscribers at the end of 2000 (around 150 million vs. about 100 million). However, the revenues of the cell phone industry were far higher than those of the Internet. If we take a rough estimate of 60 million residential Internet users and assume they pay an average of $20 per month (both slight overestimates), we find that the total revenues from this segment come to about $15 billion. Business customers, with dedicated connections to the Internet, pay considerably less than that. For example, the 2000 revenues from business Internet connections of WorldCom (whose UUNet unit has the largest backbone in the world, often thought to carry over 2. Growth of the Internet Table 2.4 Growth of the U S Cell .. Phone Industry Year 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 23 Number of Subscribers (miZlions) 0.20 0.50 0.89 1.61 2.69 4.37 6.38 8.89 13.07 19.28 28.15 38.20 48.71 60.83 76.28 97.04 Revenues (millions) $352 721 959 1,772 2,813 4,253 5,307 7,267 9,639 13,038 17,499 22,388 26,270 30,573 38,737 49,291 30% of the total backbone traffic) were just $2.5 billion (up from $1.6 billion in 1999). The conclusion of the previous paragraph is that even in the United States, basic Internet transport revenues are less than half those of cell phones. Yet volumes of traffic are far higher on the Internet. The average daily time spent by a subscriber on a cell phone in the United States is about 8 minutes. If we count wireless communicationas taking 8 Kb/s (since compression is used), we find that the total volume of traffic generated by cell phone users in the United States at the end of 2000 was only about 1500TB/month, a tiny fraction of the 20,000 to 35,000 TB/month traffic on United States Internet backbones. (Moreover, this comparison overestimates wireless traffic, since most of the mobile calls are local, whereas backbone traffic is by definition long distance.) The comparison of revenues from Internet connectivity to those of the cell phone industry leads naturally to the next topic, namely a comparison with the entire phone industry. As we saw earlier, Internet revenues were under $25 billion in the United States in 2000. On the other hand, the revenues of the entire telephone industry (including wireless communicationand data services such as private lines leased by corporations) were around $300 billion that year. Thus, in terms of revenues, the Internet is still small. Furthermore, it is so 24 Kerry G. Coffman and Andrew M. Odlyzko intimately tied to the phone industry that it is difficult to see what its roleis. The basic technologies (fiber transmission, SONET, and so on) that are used for Internet transport were developed initially for voice telephony, but were easily adopted for data. (Some, such as SONET, will likely turn out to be redundant, but are still widely used.) At the transport level, voice has been carried as bits for a long time. What happened is that during the late 1990s, the long-distance telecommunications infrastructure changed. It used to be dominated by the demands of voice transport, and data was a small part of what it carried. Now, however, its developmentis driven by data, especially Internet data. For quite a long time, the volume of data was extremely small, so that even though the growth rate was higher than for voice, this did not affect the overallgrowth rate of the infrastructure. That was one reason the telecommunications industry was repeatedly surprised by the demand for bandwidthin the 1990s. Moreover, the transition from voice to data domination was complicated by the presence of several types of data, with substantially different growth rates. We discuss this in more detail below. Another reason that the recent upsurge in demand for bandwidth was a surprise is that there had been several previous false predictions that data tralTic was about to explode. The excitement of the early 1990sabout the “telecommunications superhighway”and “500 channels to the home,” to be accomplished through technologies such as hybrid fiber-coax,certainly led to large financial losses and serious disappointments (see Woll21). However, there were even earlier periods of extremely rapid growth followed by sudden deceleration. For example, the number of modems in the United States grew between 1965 and 1970 at about 60% per year, to over 150,000 at the end of that period [WalkerM]. Had that growth rate been maintained, we would have had about 200 billion modems in the United States by the end of 2000, clearly an absurd number. Instead, it appears that growth in the 1970s followed the projections made around 1970 (p. 297 of [WalkerM]),which predicted annual increases of 25 to 30%. It is interestingto read the speculations in [DunnL]about the supposedly rosy prospects for electronic cash, distance education, and other data services (as well as for Picturephone) that were supposed to power the growth of networks. In general, predicting what communications services society will accept and how it will use them has been difficult (see [Luckyl, Odlyzko2]). In particular, even recent history is littered with technologies that seemed extremely promising at one point, such as ISDN (see meinrock3, WuL]) or SMDS (Switched Multimegabit Data Services-a high speed packet switched WAN technology), but never attained more than a marginal role. There are two aspects of the inability to forecast the prospects of communications technologies that are worth discussing at greater length. One goes back to the earlier discussion of wireless telephony and how the mobility offered by cl phones appears to be more important for many people than broadband el Internet access. Sometimes, though, higher bandwidth did prevail. In the early days of telephony, there was widespread lack of appreciation of how attractive 2. Growth of the Internet 25 it would eventually prove to be. The telephone was used primarily for business purposes, and the telegraph appeared to be adequate for that to many. Yet it was the phone that won, even though it appeared to use bandwith very wastefully when compared to the telegraph, and even though it encouraged what was often dismissed as “idle chatter.” The attractions of instantaneous personal interactions turned out to be crucial in leading to an almost universal penetration of the telephone in industrialized countries. In the last four decades of the twentieth century though, the telecommunications industry attempted several times to extend its success with the voice telephone by introducing videotelephony. This service appeared to offer the attraction of an even deeper level of communication than voice. Yet prospective users have not only not embraced it, but have in many cases treated it with hostility. There is a growth of videoconferencing, but even that is far slower than its proponents had forecasted. For a variety of reasons that have not been completely explained, videotelephony does not appeal to people for person-to-person communication. On the other hand, mobile narrowband voice flourishes. The other aspect of the dismal record in forecasting the prospects of communications technologies that we now consider is that of the nature of traffic carried. Data networks, which in commercial settings go back about four decades, have spent essentially all this time in the shadow of the much larger voice telephone network. (They also benefited from being able to use the infrastructure of the phone network, and were also constrained by its limitations, but that is less relevant for us here.) It was therefore natural for networking experts to continuously think of voice traffic, and in particular of the possibility of eventually carrying it as data. Looking further out, to a stage where the progress of technology appeared to offer the possibility of data networks becoming much larger than the phone networks, it was also natural to think of enriching the communicationsmedium through the addition of video. (See the projections of Estill Green [Green, Lucky21 and Hough [Hough], for example.) Later, the huge volume of broadcast data (radio and especially television) offered further possibilities for traffic that could be carried on data networks. The key point is what was seen as eventually filling data network was streaming multimedia traffic. The Internet’s rise to dominance was a surprise for many reasons, but one of the main ones was that it did not fit this model. Although much current work on Internet technologies is devoted to streaming multimedia, there are good reasons, to be discussed later, why such traffic is not likely to dominate. Althoughit has proven difficult to forecast which technologieswill be widely adopted, once a service had been successfullyintroduced, it often showed regular growth rates for extended periods of time [Odlyzko2].The approximately 30% annual growth rate that had been projected in 1970 for data transmission (or, to be more precise, for the proxy for actual transmission that is offered by the number of modems) appears to have held not just in the 1970s, but in the 1980s and most of the 1990s as well. There are no comprehensive 26 Kerry G. Coffman and Andrew M. Odlyzko statistics (and there are measurement problems, in that private lines, whose bandwidth is often taken as a measure of the data traffic, can also be used for voice transmission). However, there are a few pieces of evidence supporting those growth rates around 1980 in [deSolaPITH]. Those same growth rates appeared to also hold for long-distance private line transmission in the mid 1990s [CoffmanOl] and for local data bandwidth in the late 1980s and most of the 1990s [Galbi]. The comprehensive data summarized in [Galbi] is especially interesting. During the late 1980s and most of the 199Os, installed computer power came close to doubling each year, and the new “Information Economy” was taking root, but this was not reflected in the volume of data traffic. This low rate of growth in data transmission may have come from the high cost and poor quality of data transmission or from other causes, such as lack of uniform standards that would enable easy data communication between companies. It may also have been caused to a large extent by the slow rate at which computation and communication technologies were adopted. Whatever the reasons, this low growth rate of approximately 30% a year (low by comparison to growth of computing power) in data transmission was higher than that of voice networks. Hence by the mid 199Os, the bandwidth of long-distance data networks (primarily private lines used for intracompany communication) was already comparable to that of the voice network [CoffmanOl]. The Internet has historically had a growth rate of close to 100% per year in the traffic it carried. As Table 2.1 shows, it was growing with striking regularity in the early 1990s at this rate. Then it experienced a period of astronomical growth in 1995 and 1996, and then reverted to an approximate doubling each year in 1997, and has continued growing at about that rate through the end of 2000. The big question is how fast it will grow in the future. While the l overwhelming preponderance of opinion a l through the end of 2000 was that Internet traffic was doubling every 3 or 4 months, by early 2001 the consensus started changing. Some analysts even began projecting declines in the growth rates to the 50% per year range by around 2005. And indeed, some sources of growth did dry up. With the crash of telecom stocks (caused largely by the realization that expected demand and revenues were not materializing), investments slowed, and many dot-coms that had been busily filling transmission pipes with their content disappeared. In a related development, corporate managements started asking for detailed justifications for new data networking expenditures instead of rushing to endorse any proposals that came along. At various enterprises, the growth rates of data traffic, which had been close to doubling every year in the late 199Os, began to slow down toward doubling every 18 or 24 months. It is not inconceivable that overall data traffic growth may be moving back to its historical rate of around 30% per year. We do not think this will occur, but before considering the reasons why (presented in detail in Sections 6 to 9), we look at the general history of the Internet and its growth rates. 2. Growth of the Internet 27 At this point we just remark that the dominant role of the Internet in communications, whether in terms of bandwidth of networks or popular consciousness, is a fairly recent phenomenon. There had been extensive discussions of the “Information Superhighway” and the “National Information Infrastructure” for a long time. Leading thinkers foresaw the possibilities for much improved communication offered by new technologies, and there was tremendous effort devoted to various systems. However, the general expectation was that the “Information Superhighway” would be composed of a very heterogeneous collection of (interconnected)networks. This was true even as late as the beginning of the Clinton presidency in 1993 and 1994 (see [NII]). It was only in the mid to late 1990s that the Internet was perceived as evolving toward an all-encompassing network, carrying all types of traffic. 4. Internet History Over the past 5 to 10 years, we have witnessed not only an explosion of activity, but the creation of entirely new sectors within the optical industry. As the concept of wavelength division multiplexing (WDM) began to emerge, many new companies developing WDM transport equipment came into existence. The newer enterprises pushed the older established equipment vendors to more aggressive deployment schedules, and a constant downward trend for the corresponding prices of WDM transport equipment followed. In what appeared to be an almost insatiable demand for more bandwidth, a situation arose that allowed the creation of the new companies and the accompanying innovation. Not only did new equipment vendors emerge, but also new national-scale carriers were created. This trend is continuing as the concept of optical layeringhetworking is gaining acceptance and new optical equipment companies are being formed on a regular basis. They deal not only with “traditional” WDM transport equipment, but also with terrestrial ultra long-haul systems, regional and metro optimized systems, and various incarnations of optical cross connects. There were hundreds of developments and contributions enabling this burst of activity. Many of the technical innovations are described in this book and its predecessors. However, perhaps the greatest single factor that fueled this phenomena was the belief and perception that traffic, and hence needed capacity, were growing at explosive rates. This is a remarkable fact, especially when one recalls that around 1990 both the traditional carriers and most of their equipment vendors still expected the traffic demands to not vary much from the voice demand growths (which historically was around 10% per year). In fact, both carriers and equipment vendors were arguing that WDM would not be needed and that going to individual channel rates of at most 10 Gb/s would be adequate. Also, around 1995, the conventional wisdom was that 8-channel WDM systems would suffice well into the foreseeable future. Now it almost 28 Kerry G. Coffman and Andrew M. Odlyzko appears as if the pendulum has swung the other way. Is too much capacity being deployed?Are many of the reported traffic growth rates correct? And if so, will they continue? As we explainedin the previous section, the early skepticism about the need for high-capacity optical transport was rooted in the reality of the telecommunications networks. Up until 1990, they were dominated by voice, which was growing slowly. Then, by the mid 1990s, they came to be dominated (in terms of capacity) by private lines, which were growing three or four times as fast. And then, in the late 1990s, they came to be dominated by the Internet, which was growing faster still. Before we go through the analyses for the tr&c growth on the Internet, we must first at least define the Internet and describe its history and structure. This is paramount in helping put much of the later described growth analyses into perspective. When one now speaks of the Internet, it is usually described as an evolution from ARPANET to NSFNet, and finally to the commercial Internet that now exists. Arguably, the phenomenal growth of the Internet started in 1986 (more than 17 years after its “birth”) with NSFNet. However, the path was very complicated and full of many twists and turns in its roughly 40-year history [Cerf, Hobbes, Leiner]. From the very early research in packet switching, academia, industry, and the U.S. government have been intertwined as partners. Ironically, the beginnings of the Internet can trace itself back to the Cold War and specihlly to the launch of Sputnik in 1957. The U.S. government formed the Advanced Research Project Agency (ARPA; the name was later changed to DARPA, Defense Advanced Research Project Agency, and later back to ARPA) the year after the launch with the stated goal of establishing a U.S. lead in technology and science (with emphasis on military applications). As ARPA was establishing itself, there were several pivotal works [Kleinrockl, Baran] in the early 1960s on packet switching and computer communications. These works and the efforts they spawned laid many of the foundations that enabled the deployment of distributed packet networks. J. C. R. Licklider (of MIT) [LickC] wrote a series of papers in 1962 in which he “envisioned a globally interconnected array of computers which would enable ‘everything’ to easily access data and programs from any of the sites.” Generically speaking, this idea is not much diflterent from what today’s Internet has become. Of importance is the fact the Licklider was the first head of the computer research program at DARPA (beginning in 1962), and in this role he was instrumental in pushing his concept of networks. Kleinrock published both the first paper on packet switching and the first book on the subject. In addition, Kleinrock convinced several key players of the theoretical feasibility of using packets instead of circuits for communications. One such person was Larry Roberts, one of the initial architects for the ARPANET. In the 1965 to 1966time frame, ARPA sponsoredstudies on a “cooperativenetwork of [users] 2. Growth of the Internet 29 sharing computers” [Leiner], and the first ARPANET plans were begun, with the first design papers on ARPANET being published in 1967. Concurrently, the National Physical Laboratory (NPL) in England deployed an experimental network called the NPL Network that made use of packet switching. It utilized 768 Kb/s lines. A year before the moon landing, in 1968, the first ARPANET requests for proposals were sent out, and the first ARPANET contracts were awarded.Two of the earliest contractswent to UCLA to develop the Network Measurement Center, and to Bolt, Beranek, and Newman (BBN) for the Packet Switch contract (to construct the Interface Message Processors or IMPs-effectively the routers). Kleinrock headed the Network Measurement Center at UCLA and it was selected as the first node on the ARPANET. The first IMP was installed at UCLA and the first host computer was connected in September of 1969. The second node was at Stanford Research Institution (SRI). Two other nodes were added at UC Santa Barbara and in Utah, so that by the second half of 1969,just months past the lkst moon landing, the initial four-node ARPANET became functional. This was truly the initial ARPANET, and thus a case can be made that this was when the Internet was born. The first message carried over the network went from Kleinrock’s lab to SRI. Supposedly the first packet sent over ARPANET was sent by Charley Kline, and as he was trying to log in the system crashed as the letter “ G of “LOGIN” was entered. One of the next major innovations for the fledgling Internet (i.e., ARPANET) was the introduction of the first host-to-host protocol, called Network Control Protocol, or NCP, which was first used in ARPANET in 1970. By 1972, a l of the ARPANET sites had finished implementing NCP. l Hence the users of ARPANET could finally begin to focus on the development of applications-another paramount driver for the phenomenal growth and sustained growth of the Internet. It was also in 1970 that the first crosscountry link was established for ARPANET by AT&T between UCLA and BBN (at the blinding rate of 56 Kbh). By 1971, the ARPANET had grown to 15 nodes and had 23 hosts. However, perhaps the most influential work that year was the creation of an e-mail program that could send messages across a distributed network. (E-mailwas not among the original design criteria for the ARPANET, and its success caught the creators of this network by surprise.) Ray Tomlinson of BBN developed this application, and his original program was based on two previous ones [Hobbes]. Tomlinson modified his program for ARPANET in 1972, and at that point its popularity quickly soared. In fact, it was at this time that the symbol “@,, was chosen. Arguably, Internet e-mail as we know it today can trace its origins directly to this work. Internet e-mail was clearly one of the key drivers for the popularity (and hence the phenomenal traffic-growthdemands) of the Internet and was the first “killer app” for the Net. It was every bit as critical to the Internet’s “success” as spreadsheet applications were to the popularization of the PC. Internet e-mail provided 30 Kerry G. Coffman and Andrew M. Odlyzko a new model of how people could communicate with each other and alter the very nature of collaborations. Although there was already considerable work being done on packet networks outside the United States, the first international connections to the ARPANET (to England via Norway) took place in 1973. To put the time frame in perspective, this was the same year that Robert Metcalfe did his PhD that described his idea for Ethernet. Also during this year, the number of ARPANET c‘users”was estimated to be 2,000 and that 75% of all the ARPANET t r a c (in terms of bytes) was e-mail. One needs to note that in only 1 to 2 years from its introduction onto the Internet, e-mail became the predominant type of traffic. The same behavior took place several years later for HTML (Hypertext Markup Language) (i.e., Web traffic), and to a somewhat lesser degree, this was seen for Napster-like traffic within many networks a few years later. Several other key developments began to take place in the mid 1970s. The initial design specification for TCP (Transfer Control Protocol) was published by Vint Cerf and Bob Kahn in 1974 [CerfK]. The NCP protocol, which was being utilized at the time, tended to act like a device driver, whereas the future TCP (later TCP/IP) would be much more like a communications protocol. As is discussed later, the evolution from ARPANET’s NCP protocol to TCP (which in 1978 was split into TCP and IP (Internet Protocol)) was critical in allowing the future growth and scalability of today’s Internet. DARPA had three contracts to implement TCPLIP (at the time still called TCP), at Stanford (led by Cerf), BBN (led by Ray Tomlinson), and UCLA (led by Kirsten). Stanford produced the detailed specification and within a year there were three independent implementations of TCP that could interoperate. It is noted that the basic reasons that led to the separation of TCP (which guaranteed reliable delivery) from IP actually came out of work that was done trying to encode and transport voice through a packet switch. It was found that a tremendous amount of bdering was needed in order to allow for the appropriate reassembly after transmission was completed. This in turn led to trying to find a way to deliver the packets without requiring a guaranteed level of reliability. In essence, the UDP (User Datagram Protocol) was created to allow users to make use of IP. In addition, it was also in 1978 that the first commercial version of ARPANET came into existence when BBN opened Telenet. In 1981-1982, the first plans were made to “migrate”from NCP to TCP. It is claimed by some that it was this event (TCP was establishedas the protocol suite for ARPANET) was truly the birth of the InternetAefined as a connected set of networks, specifically those with TCP/IP. A few years later (in 1983) another major development occurred, which later enabled the Internet to scale with the “explosive” growth and popularity of the future Internet. This was the development of the name server, which evolved into the DNS [Cerf, Leiner]. The name server was developed at the University of Wisconsin [Hobbes]. This made it easy for people to use the network because hosts were assigned names 2. Growth of the Internet 31 and it was not necessaryto remember numeric addresses. Much of the credit for the invention of the DNS (Domain Name Server) is given to Paul Mockapetris of USC/ISI [Cerfl. The year 1983 was also the date for two other key developments on ARPANET. The first one was the cutover from NCP to TCP on the ARPANET. Secondly, ARPANET was split into ARPANET and MILNET. Although the road was convoluted, this split was one of the key bifurcation points that later allowed NSFNet to come into existence. Soon thereafter (in 1984), the number of hosts on ARPANET had grown to 1,000, and the next year in 1985 the first registered domain was assigned in March. In 1985, NSFNet was created with a backbone speed of 56 Kb/s. Initially, there were five supercomputing centers that were interconnected. One of the paramount benefits of this was that it allowed an explosion of connections (most importantly from universities) to take place. Two years later in 1987, NSF agreed to work with MERIT Network to manage the NSFNet backbone. The next year (1988), the process of upgrading the NSFNet backbone to one based on T1 (Le., 1.5 Mb/s links) was begun. In 1987, the number of hosts on the Internet broke 10,000. Two years later in 1989, this had grown to around 100,000, and 3 years after that, in 1992, it reached the 1,000,000 value. It is noted that if you look at how the number of hosts had been growing from 1984 to 1992, that it was still pretty much tracking a growth curve that was less than tripling each year (i.e., doubling every 9 months). In the 1985-1986 time frame, a key decision was made that had very long-term impact: that TCP/IP would be mandatory for the NSFNet program. In the 1988-1990 time frame, a conscious decision was made to connect the Internet to electronicmail carriers, and by 1992,most of the commerciale-mail carriers in the United States were “like the Internet.” This was still another development that cemented e-mail as the single most important application to take advantage of the Internet. In 1990, the ARPANET ceased to exist, and arguably NSFNet was the essence of the Internet. The following year, commercial Internet Service Providers (ISPs) began to emerge (PSI, ANS, Sprint Link, to name a few), and the Commercial Internet Xchange (CIX) was organized in 1991 by commercial ISPs to provide transfer points for traffic. NSF’s lifting the restriction on the commercial use of the Net was again one of the pivotal decisions. This was again a key bifurcation point, in that this helped set the stage for the complete commercialization of the Net that would follow only a few years later. In 1991, the upgrading of the NSFNet backbone continued as the work to upgrade to a T3 (Le., 45 Mb/s links) began. It is also interesting to note that it was the next year, 1992, that the term “surfing the Internet” was first coined by Jean Armour Polly [Polly], only 2 years before the ARPANEThternet celebrated its 25th anniversary. It was in the 1993-1995 time period that several major events seemed to emerge that fueled an almost explosivegrowth in the popularity of the Internet. 32 Kerry G. Coffman and Andrew M. Odlyzko One of the key ones was the introduction of “browsers,” most notably Mosaic. This led to the creation of Netscape, which went public in 1995. Even as early as 1994, WWW (i.e., predominantly HTML) tr&c was increasing in volume on the Net. By then it was the second most popular type of traffic, surpassed only by FTP (File Transfer Protocol) tr&c. However, in 1995 WWW tra& surpassed FTP as the greatest amount of tr&c. In addition, the traditional online dial-up systems such as AOL (America Online), Prodigy, and CompuServe began to provide Internet access. In 1996, the Net truly became public with the NSFNet being phased out. Soon thereafter, major infrastructure improvements were made within the transport part of the Internet. The Internet began to upgrade much of its backbone to OC3-0C12 (up to 622Mb/s) links, and in 1999, upgrades began for much of the Net to OC48 (2.5 Gb/s) links. 5. The Many Internet Growth Rates The Internet is very hard to describe. By comparison, even the voice phone system, which is a huge enterprise, far larger in terms of revenues than the Internet, is much simpler. In the phone system, the basic service is well defined and simple to describe. The users have only limited ability to interact with the system. The Internet is completely different. Users interact with the system in a multiplicity of ways, on widely different time scales, and there are many complicated feedback loops. The paper [FloydP] is an excellent overview of the problems that arise in attempting to simulate the Internet. The problems of measuring the Internet are also formidable. There are many different measures that are relevant. In this chapter,just as in the papers [CoffmanOl,CoffmanO2], we will concentrate on traffic as measured in bytes. For the optical fiber telecommunications industry, it is capacity that is most relevant. Unfortunately, there are numerous problems in measuring capacity. Much of the fiber is not lit, and even when it is lit, often only a few wavelengths are lit. Finally, much of the potential capacity is used for restoration, through SONET or other methods. In addition, even at the levels of links used for providing IP traffic, it is hard to obtain accurate capacity measurements, because few carriers provide detailed data. Further, this type of capacity has a tendency to jump suddenly, as bandwidth is usually increased in large steps (such as going from OC3 to OC12, and then 0048, a phenomenon that contributes to the low utilization of data links [Odlyzkol]). Thus, there is little regularity in capacity growth figures. On the other hand, we do find astonishing regularity in traffic growth, which leads us to propose that a form of Moore’s Law applies. In the long run, we expect that capacity will grow slightly faster than traffic, as we explain later. For many purposes other measures are important, such as the number of users, how they spend their time, how many and what types of commercial 2. Growth of the Internet 33 transactions they engage in, and so on. There are many sources of such data, and useful references can be found at [Cyberspace, MeekerMJ, Nua]. 6. Internet Traffic and Bandwidth Growth Whether Internet traffic doubles every 3 months or just once a year has huge consequences for network design as well as the telecommunicationsequipment industry. Much of the excitement about and funding for novel technologies appears to be based on expectations of unrealistically high growth rates ([Bruno]). In this section we briefly examine avariety of examples in an attempt to understand the traffic-growth rates that the Internet has experiencedover its lifetime. There are places where the traffic is growing at rates that exceed 100% per year. One such example is LINX (London Internet Exchange). Its online data, available at http://ochre.linx.net/, clearly shows a growth rate of about 300% from early 1999 to early 2001. There are also examples of even higher growth rates, although those tend to be for much smaller links or exchange points. However, there are also numerous examples of much more slowly growing links. In this section we briefly present growth rates from avariety of sources and attempt to put them into context. In an earlier study [CoffmanOl] in 1997, we found that the evidence supported a traffic growth rate of about 100% per year (doubling annually). Four years later, the general conclusion is that Internet traffic still appears to be growing at about 100% per year. In other words, we have not found any substantial slowdown in the growth rate. Some recent reports and projections conclude that Internet traffic is only about doubling each year, but claim that it was growing much faster until recently, and that its growth rate will continue to slow down. In that view, the telecom crash of 2000 was associated with a sudden decline in the growth rate of traffic. As far as we can tell, that is not accurate. The general rate of growth of traffic appears to have been remarkably stable throughout the period 19972000. As one of the most convincing pieces confirming this claim, we cite the news story ([Cochrane]) based on official figures from Telstra, the dominant Australian telecommunications carrier. This story reports that Telstra’s IP traffic was almost exactly doubling each year between November 1997 and November 2000. (The printed version of this news story, but not the one available online at the URL listed in [Cochrane], shows a very regular growth, about 100% per year, from the beginning of 1997 to November 2000.) Hence our conclusion is that the problems the photonics industry is experiencingare not caused by any sudden slowdown in traffic, but rather by a realization that the astronomical growth rates that people had been assuming were fantasies. Most of this section is drawn from the more detailed account in [Coffman02]. There are only a few new pieces of information. For example, the China Internet Network Information Center has statistics (at www.cnnic.net.cn/develst./e-index.shtm1) the Internet bandwidth between of 34 Kerry G. Coffman and Andrew M. Odlyzko China and the rest of the world. It grew from 84.64Mb/s in June 1998 to 2,799 Mbls in December 2000, for a compound growth rate of 305% per year. Thus, even in a rapidly growing economy like that of China, where Internet penetration is low and is trying to catch up with the industrialized world, traffic is only doubling about every 6 months. The comparison of the international bandwidth for Australia and China is instructive. In December 2000, Telstra had about 1,000Mb/s to the rest of the world, about a third of Chinese bandwidth. Thus, making allowances for other Australian carriers, we can speculate that Australia may be exchanging half as much traffic with international destinations as China does, even though the latter has over 60 times the population. This shows the degree to which countries can differ in their intensity of Internet usage. The data in [Cochrane], showing that Telstra’s IP traffic in November 2000 reached about 270 TBlmonth, also shows that our general estimates for U.S. backbone traffic are reasonable, because the United Statesis not only larger than Australia, but also richer on aper capita basis and has a better developed telecommunications infrastructure. In the remainder of this section we examine some of the data and trends from ISPs, exchange points, and residential traffic patterns, along with traffic from “stable sources,” such as corporate, research, and academic networks. It is noted that the data for the first two sources (ISPs and exchange points) are not nearly as complete nor reliable as only a few years ago. However, much better data are available for the “stable sources,” and several are examined in much more detail later. As a brief note on conversion factors, traffic that averages 100Mb/s is equivalent to about 30 TB/month. (It is 32.4 TB for a 30-day month, but such precision is excessive given the uncertainties in the data we have.) Unfortunately, the largest ISPs do not release reliable statistics. This situation was better even a couple of years ago. Much of the older data was used in previous studies ([CoffmanOl]).For example, MCI used to publish precise data about the traffic volumes on their Internet backbone. Even though they were among the first ISPs to stop providing official network maps, one could obtain good estimates of the MCI Internet backbone capacity from public presentations. These sources dried up when MCI was acquired by WorldCom, and the backbone was sold to Cable &Wireless. As was noted in [CoffmanOl], the traffic-growthrate for that backbone had been in the range of 100% a year before the change. Today, one can obtain some idea of the sizes (but not trafEc) of various ISP networks through the backbone maps available from Boardwatch. However, even those are not too reliable. The only large ISP in the United States to provide detailed network statistics is AboveNet, at http:lluww.above.netltrafficl. Therefore, we looked at this ISP in moderate detail. We have recorded the MRTG (Multi-RouterTraffic Grapher)[MRTG]data for AboveNet for March 1999, June 1999, February 2000, June 2000, November 2000, and April 2001. 2. Growth of the Internet 35 The average utilizations of the links in the AboveNet long-haul backbone during those 4 months were 18, 16, 29, 12, 11, and lo%, respectively. (The large drop between February and June 2000 was caused by deployment of massive new capacity, including four OC48s. One of the reasons we concentrate on traffic and not network sizes in this chapter is that extensive new capacity is being deployed at an irregular schedule and is often lightly utilized. Thus, it is hard to obtain an accurate picture of the evolution of network capacity.) If one just adds up the volumes for each link separately, one finds that between March 1999 and April 2001, the total volumes of traffic increased at an annual growth rate of about 200%. However, this figure has to be treated with caution, as actual traffic almost surely increased less than 200%. During this period, AboveNet expanded geographically, with links to Japan and Europe, so that at the end it probably carried packets over more hops than before. Because we are interested in end-to-end traffic as seen by customers (which can be thought of as the ingress and/or egress traffic into and/or out of “the network”), we have to deflate the sum of traffic volumes seen on separate backbone links by the average number of hops that a packet makes over the backbones (perhaps around three). Even when there is reliable data for a single carrier, such as AboveNet, some of the growth seen may be coming from gains in market share, both from gains within a geographical region and from greater geographical reach, and not from general growth in the market. We next look at Internet exchangepoints. When the NSF Internet backbone was phased out in early 1995, it was widely claimed that most of the Internet backbone traffic was going through the Network Access Points (NAPs) (which are effectivelyinterconnectionvehicles), which tended to provide decent statistics on their traf3ic. Currently it is thought that only a small fraction of backbone traffic goes through the NAPs, while most goes through private peering connections. Furthermore, NAP statistics are either no longer available or not as reliable. This is in sharp contrast to the situation in 1998 [CoffmanOl]. As documented elsewhere [CoffmanO2], there is very little that can be reliably concluded about current growth rates of Internet traffic by examining the statistics of the public NAPs in the United States. However, the situation was slightly better when we examined a large number of international exchange points. These included LINX, AMS-IX (the Amsterdam Internet exchange), the Slovak Internet exchange, HKIX (a commercial exchange created by the Chinese University of Hong Kong, BNIX (located in Belgium), the INEX (an Irish exchange), and FICX (the Finish exchange). Some of these show growth rates of only about doubling per year while others show much faster growth rates. Trafiic interchange statistics are hard to interpret, unless one has data for most exchanges, which is virtually impossible to obtain. Much of the growth one sees can come from ISPs moving from one exchange to another, moving their traffic from one exchange to another, or coming to an exchange in preference to buying transit from another ISP. Consider the specific case of LINX. A large part of its growth 36 Kerry G. Coffman and Andrew M. Odlyzko is almost surely caused by more ISPs exchanging their traffic there. Between March 1999 and March 2000, the ranks of ISPs that are members of LINX have grown by about two-thirds, based on the data on the LINX home page. Hence the averageper-member traffic through LINX may have increased only around 120% during that year. The traffic from residential U.S. customers will probably begin to increase at a faster rate in the near future. The growth in the number of users is likely to diminish as we reach saturation. (You cannot doublethe ranks of subscribersif more than half the people are already signed up!) However, broadband access, i the shape of cable modems and DSL (and to a lesser extent fixed wireless n links), will stimulate usage. The evidence so far is that users who switch to cable modem or DSL access increase their time online by 50 to loo%, and the total volume of data they download per month by factors of five to ten. A five- or tenfold growth in data traEic would correspond to a doubling of traffic every 4 months if everyone were to switch to such broadband access in a year. However, that is not going to happen. At the end of 1999, there were about 3 million households in the United States with broadband access. The most ambitious projections for cable modem and DSL access call for about 13 million households to have such links in 2003, and between 5 M O million in the year 2007. That is approximatelya doubling each year. (There was apparently almost a tripling in the ranks of households with broadband access in 2000, but the telecom crash that wiped out many of the ADSL providers has led to a slowdown in the pace of deployment in 2001.) The traffic from a typical residential broadband customer is likely to grow beyond the level we see today as more content becomes available and especially as more content that requires high bandwidth is produced. Still, it is hard to see average traffic per customer among those with broadband connections growing at more than 50% a year. Together with a doubling in the ranks of such customers, this might produce a tripling of traffic from this source. Because the ranks of customers with regular modems are unlikely to decrease much, if any, and because their traffic dominates, it appears that the most likely scenario will be for the total residential customer traffic to grow no faster than 200% per year, and probably closer to 100% per year. (Access from information appliances,which are forecast to proliferate, is unlikely to have a major impact on total traffic, since the mobile radio link will continue to have small bandwidth compared to wired connections.) We next consider traf€ic at various stableinstitutions-corporate, academic, and governmental. Growth in traffic can be broken down into growth in the number of traffic sources and growth in traffic per source. For LINX, much of the increase in traffic may be coming from an increase in member ISPs. For individual ISPs, much of the increase in traffic may also be coming from new customers. Yet in the end, that kind of growth is limited, as the market becomes saturated. The rest of this section focuses on rates of growth in traffic from stable sources. Now nothing is completely stable, as 2. Growth of the Internet 37 the number of devices per person is likely to continue growing, especially with the advent of information appliances and wireless data transmission. Hence we will consider growth in traffic from large institutions that are already well wired, such as corporations and universities. Most corporations do not publicize information about their network traffic, and many do not even collect it. However, there are some exceptions. For example, Lew Platt, the former CEO of Hewlett-Packard, used to regularly cite the HP intranet in his presentations.. The last such report, dated September 7, 1998, and available at http://www.hp.com/financials/textonly/personnel/ceo/~es.ht~, stated that this network carried 20 TB/month, and a comparison with previous reports shows that this volume of traffic had been doubling each year for at least the previous 2 years. (As an interesting point of comparison, the entire NSFNet Internet backbone carried 15TB/month at its peak at the end of 1994.) Several other corporations have provided data showing similar rates of growth for their Intranet trafllc, although some indicated their growth has slowed, and a few have had practically no recent growth. Internal corporate tr&c appears to be growing much more slowly than public Internet traffic. Data for retail private lines as well as for Frame Relay and ATM (Asynchronous Transfer Mode) services show aggregate growth in bandwidth (and therefore most likely also traffic) in a range of 3040% per year. The growth is slow for retail private lines and fast for Frame Relay and ATM. These rates are remarkably close to the growth rate observed in the late 1970s in the United States, which was around 30% per year [deSolaPITH]. Thus, it is the corporate traffic to the public Internet that is growing at 100% per year. It is also important to note that in the year 2000, over two-thirds of the volume on the public Internet appeared to be business to business. Thus, the accelerationof the overallgrowth rate of data trafficto about 100%per year from the old 30% or so a year appears to be a consequence of the advantages of the Internet, with its open standards and any-to-anyconnectivity. For the remainder of this section we concentrate on publicly available information, primarily about academic, research, and government networks. These might be thought of as unrepresentative of the corporate or private residential users. Our view is just the opposite, in that these are the institutions that are worth studyingthe most, since they normally alreadyhave broadband accessto the Internet, tend to be populated by technically sophisticated users, and tend to try out new technologies first. The spread of Napster through universities is a good example of the last point. We believe that Napster and related tools, such as Gnutella and Wrapster, are just the forerunners of other programs for sharing of general information, and not just for disseminating pirated MP3 files. As we explained elsewhere, there is already much more digital data on hard disks alone than shows up on today’s Internet. Further, this situation is likely to continue. The prevalent opinion appears to be that in data networks, “If you build it, they will fill it.” Our evidence supports this, but with the important 38 Kerry G. Coffian and Andrew M. Odlyzko qualiiication that “they” will not fill it immediately. That certainly has been the experience in local area networks, LANs. The prevalence of lightly utilized long-distance corporate links was noted in [Odlyzkol]. That paper also discussed the vBNS (very High Speed Backbone Network) research network, which was extremely lightly loaded. Here we cite another example of a large network with low utilizations and moderate growth rates. Abilene is the network created by the Internet2 consortium of U.S. universities. Its backbone consists of 13 OC48 (2.4 Gb/s) links. Moreover, most of the consortium members had OC3 links to it. The average utilization in June 2000 was about 1.5%, and by April 2001 it had grown to about 4.1%. Thus, in spite of the uncongested access and backbone links, tr&c did not explode. Even on more congested links, it often happens that an increase in capacity does not lead to a dramatic increase in traffic. This is supported by several examples. Such examples include the University of Waterloo, the SWITCH network, the NORDUNet network, the European TEN-155 network, the Merit network, the University of Toronto, Princeton University, and the University of California at Santa Cruz [CoffmanO2]. Later we go into moderate detail for these networks. Figure 2.1 shows statistics for the traffic from the public Internet to the University of Waterloo over the last 7 years. Detailed statistics for the Waterloo network are available at http://www.ist.uwaterloo.ca/cn/#Stats,but Fig. 2.1 is based on additional historical data provided to us by this institution. Just as for the JANET network discussed previously and the SWITCH network to be discussed later, as well as most access links, there is much more traffic from the public Internet to Traffic from the Internet to the University of Waterloo 03 I I I I 1993 1994 1995 1996 1997 1998 1999 2000 year Fig. 2.1 T r a c on the link from the public Internet to the University of Waterloo. The line with circles shows average traffic during the month of heaviest traffic in each school term. The step function is the full capacity of the link. 2. Growth of the Internet 39 the institution than in the other direction. Hence we concentrate on this more congested link, because it offers more of a barrier. We see that even substantial jumps in link capacity did not affect the growth rate much. Traffic has been about doubling each year for the entire 7-year period. (Overall, the growth rate at the University of Waterloo has slowed, about 55% from early 1999 to early 2000. This was at least partially the result of official limits on individual users that were imposed, limits we will discuss later.) The same phenomenon of traffic doubling each year, no matter what happens to capacity, can be observed in the statistics for the SWITCH network, which provides connectivity for Swiss academic and research institutions. The history and operations of this network are described in [Harms, ReichlLS], and extensive current and historical data are available at http:// www.switch.ch/lan/stat/. data used to prepare Fig. 2.2 was provided to us The by SWITCH. As is noted in [ReichlLS], the transatlantic link has historically been the most expensive part of the SWITCH infrastructure, and at times was more expensive than the entire network within Switzerland. It is therefore not surprising that this link tends to be the most congested in the SWITCH network. Even so, increasing its capacity did not lead to a dramatic change in the growth rate of traffic. If we compare increases in volume of data received between November of one year and January of the following year, there was an unusually high jump (420/0)from November 1998to January 1999.This was in response to extreme congestion experienced at the end of 1998, congestion that produced extremely poor service, with packet loss rates during peak periods exceeding 20%. However, over longer periods of time, the growth rate has been rather steady at close to 100% per year and independent of the capacity Capacity and traffic on SWITCH transatlantic link H I 1996 1997 1998 1999 2000 2001 3 2 year Fig. 2.2 Capacity of link between the Swiss SWITCH network and the United States and traffic on it toward Switzerland. 40 Kerry G. Coffman and Andrew M. Odlyzko of the link. More detailed data about other types of SWITCH traffic can be found at http://www.switch.cMan/stat/ through the “Public access” link. The listings available there as of mid 2000, as well as those from previous years, show that various transmissions tended to grow at 100 to 150% per year. It is worth noting that capacity grew faster than traffic, but not too much faster. Merit Network is a nonprofit ISP that serves primarily Michigan educational institutions.It has data availableonline at http://www.merit.net/michnet/ statistics/direct.htmlthat goes back to January 1993. This data was used to construct the graph in Fig. 2.3. The data for January 1993 through June 1998 shows only the number of inbound IP packets. The data for months since July 1998 is more complete, but it is so complete, with details of so many interfaces, that we have not yet determined the best way to use it. Hence we have used only the earlier information for January 1993 through June 1998. The resulting time series is a reasonable, although imperfect, representation of a straight line, modulated by the periodic variations introduced by the academic calendar. The growth rate is almost exactly 100% per year. The research networks that were examined have low utilizations. It should be emphasized that this is not a sign of inefficiency. Many novel applications required high bandwidth to be effective. That, along with some additional factors, such as the high growth rate, lumpy capacity, and pricing structure, contributes to the much lower utilization of data networks than of the longdistance voice network [Odlyzkol]. The general conclusion that can be drawn from the examples listed in this section (along with numerous other examples) is that data traffic has a remarkable tendency to double each year. There are of course slower and faster growth rates. Overall though, they tend to cluster in the Vicinity of 100% per year. MichNet traffic 1993 1994 1995 1996 1997 1998 1999 year Fig. 2.3 Traffic from Merit Network to customers 2. Growth of the Internet 41 To date, the authors have not seen any large institutions with traffic doubling anywhere close to every 3 or even 4 months. The growth rates that are cited here are often affected strongly by restrictions imposed at various levels. As described elsewhere [CoffmanOl, CoffmanO2], some of the explicit limits are imposed by network administrators. The arrival of Napster (discussed in Section 7) led many institutions to either ban its use or else limit traffic rates to some parts of the campus (typically student dormitories). Push technologies were stifled at least partially because enterprise network administrators blocked them at their firewalls. E-mail often has size restrictions that block large attachments (and in some cases all attachments are still banned). Teleconferencing is only slowly being experimented with on corporate intranets, and even packetized voice sees very limited (although growing) use. Similar constraints apply to most of the content seen on the Web. As long as a large fraction of potential users have limited bandwidth, such as through dial modems, managers of Web servers will have an incentive to keep individual pages moderate in size. Thus, one can see that Internet traffic is subject to a variety of constraints at different levels. Some are applied by network managers, others by individual users, and the interaction of these constraints with the rising demands is fundamental in understanding what produces the growth rates observed. The ability to sustain the high growth rate of Internet traffic will require the creation of new applications that will generate huge volumes of traffic. At current growth rates, by 2005 there will be eight times as much Internet as voice traffic (on the U.S. long-haul networks). If voice were packetized, in all likelihood the voice traffic would only account for about 3% of the Internet traffic. Thus, voice traffic will not fiU the pipes that are likely to exist, and neither will traditional Web surfing. This will create a dilemma for service providers, network administrators, and equipment suppliers: To sustain the growth rates that the industry has come to depend on, and to accommodate the progress in technology, new technologies are needed. Such applications will appear disruptive to network operations today, and as such, they often have to be controlled. However, in the long run, they must be encouraged. 7 Disruptive Innovation . It is often said that everything changes so rapidly on the Internet that it is impossible to forecast far into the future. The next “killer app” could disrupt any plans that one makes. Yet there have been just two “killer apps” in the history of the Internet: e-mail and the Web (or, more precisely, Web browsers, which made the Web usable by the masses). Many other technologies that had been widely touted as the next “killer app,” such as push technology have fizzled. (Push technology allows the sending of information directly to one’s 42 Kerry G. Coffman and Andrew M. Odlyzko computer instead of the computer needing to actively go out and obtain it.) Furthermore, only the Web can be said to have been truly disruptive. From the first release of the Mosaic visual browser around the middle of 1993, it apparently took under 18 months before Web trafKc became dominant on Internet backbones. It appears overwhelmingly likely that it was the appearance of browsers that then led, in combination with other developments, to that abnormal spurt of a doubling of Internet trafKc every 3 or 4 months in 1995 and 1996. What were the causes of the 100-fold explosion in Internet backbone traffic over the 2-year period of 1995 and 1996? We do not have precise data, but it appears that there were four main factors, all interrelated. Browsers passed some magic threshold of usability, so many more people were willing to use computers and online information services. Users of the established online services, primarily AOL, CompuServe, and Prodigy, started using the Internet. The text-based transmissions of those services, which probably averagedonly a few hundred bits per second per connected user, were replaced by the graphicsrich content of the Web, so transmission rates increased to a few thousand bits per second. Finally, flat rate access plans led to a tripling of the time that individual users spent online [Odlyzko3], as well as faster growth in number of users. The Internetwas able to support this explosionin use because it was utilizing the existinginfrastructure of the telephone network. At that time, the Internet was tiny compared to the voice network.It is likely that the data network that handles control and billing for the AT&T long-distance voice services by itself was carrying more traffic than the NSF Internet backbone did at its peak at the end of 1994. Today, by contrast, the public Internet is rapidly moving toward being the main network, so quantum jumps in traffic cannot be tolerated so easily. In late 1999, a new application appeared that attracted extensive attention and led to many predictions that network traffic would see a major impact. It was Napster. At the time, numerous articles in the press cited Napster’s ability to “overwhelm Internet lines,” and have claimed that it has forced numerous universities to ban or limit its use. The impression one got from those press reports was that Napster was causing a quantumjump in Internet traffic, and was driving the traffic growth rates well beyond the normal range. However, upon close examination this does not appear to be completely accurate, and the use of Napster has not increased growth rates much beyond the annual doubling or tripling rates, even within university environments,where Napster is most popular. That is not to say that it has not resulted in huge amounts of traffic, nor that it has not had serious impact on several major networks. Napster provides software that enables users connected to the Internet to exchange andor download MP3 music files. The Napster Web site matches users seeking certain music files with other users who have those files on their computer. The Napster system preferentially uses machines that have high 2. Growth of the Internet 43 bandwidth connections as sources of files. This means that universities are the primary sources, since other organizations with fast dedicated links, mainly corporations, do not allow such traffic. The result is that although college students are often cited as the greatest users of MP3 files, it is the traffic from universities that gets boosted the most. (Because that direction of traffic is typically much less heavily used than the reverse one, the impact of Napster is much less severe than if the dominant direction of traffic were reversed.) Regular modem users are usually not affected, because their connections are too slow. However, the proliferation of cable modems and DSL connections that have “always-onyy high-bandwidth connectivity is leading to problems for some residential users, especially since the uplink is the one that invariably has the more limited bandwidth. A key reason that Napster is of great interest to us is that similar types of sharing applications effectively turn consumers of information into providers of information. u h e World Wide Web was designed for such information sharing, but for some types of files Napster and its kin are preferable.) These applicationswill effectively turn traditional consumerPCs into Internet servers that will output large amounts of traffic to other users. In Napster’s case this has been predominantly MP3 musicfiles, but other programs, such as Gnutella, work with more general data. It is highly probable that such applicationscould be one of the key applications that fuel the continued annual doubling or tripling of data traffic. Napster first became noticeable in the summer of 1999. Its share of the total Internet traffic on many of the university networks has grown from essentially nothing to around 25% of the total traffic by mid to late 2000. In [CoffmanO2] the traffic generated by Napster and its impact on various networks was examined. The amount of Napster traffic that is reported by several university networks (such as University of California at Santa Cruz, University of Michigan, University of Indiana, University of California at Berkeley, Northwestern University, and Oregon State University to name a few) ranges from around 20 to 50%. However, the reported numbers are often very preliminary, and in some cases they compare Napster traffic to total traffic, whereas in others it appears that the high values may represent a comparison only to the out traffic. In any event, this is a phenomenal growth rate for any single application. Since it started from zero and our data only goes out to about a year from that time, it is risky to extrapolate this initial explosion out indefinitely. In most cases [CoffmanO2], Napster has had a noticeable effect on the growth rate of traffic on this campus, but not an outlandish one. Several networks, such as that of the University of Wisconsin-Madison that report Napster traflk making up as much as 30% of the total, are not doing anything to limit Napster because they claim that they still have plenty of bandwidth. Others have imposed limits on the total bandwidth available to the dormitories. 44 Kerry G. Coffman and Andrew M. Odlyzko Aside from Napster, occasionally even a large institution will experience a local perturbation in its data traffic patterns caused by one particular application. For example, the SETI@home distributed computing project (http://setiathome.ssl.berkeley.edu) uses idle time on about three million PCs (as of mid 2001) to search for signs of extraterrestrial intelligence in signals collected by radio telescopes. This project is run out of the Space SciencesInstitute at the University of California at Berkeley, and within a year of inception it accounted for about a third of the outgoing campus traffic [McCredie]. (Moreover, this was extremely asymmetrical traffic, with large sets of data to be analyzed going out to the participating PCs and small final results coming back. That most of the data went away from campus made this application less disruptive than it would have been otherwise.) Its disruptive effect is moderated by limiting its transmission rate to about 20 Mb/s. At the University of California at Santa Cruz, a complete copy of the available genome sequence was made available for public download in early July 2000. This, combined with coverage in the popular press and on Slashdot, led to an immediate surge in traffic, far exceeding the effects of Napster. If the interest in this database continues, it will require reengineering of the campus network. The SETI@home project is interesting for several reasons. It is cited in [McCredie] as a major new disruptive influence. Yet it contributes only about 20 Mb/s to the outgoing traffic. An increasing number of PCs and workstations are connected at 100Mb/s, and even Gigabit Ethernet (1,000 Mb/s) is coming to the desktop. This means that for the foreseeable future, a handful of workstations will, in principle, be capable of saturating any Internet link. Given the projections for bandwidth, a few thousand machines will continue to be capable of saturating all the links in the entire Internet. Thus control on user traffic will have to be exercised to prevent accidental as well as malicious disruptions of service. However, it seems likely that such control could be limited to the edges of the network. In fact, such control will pretty much have to be exercised at the edges of the network. QoS (Quality of Service) will not help by itself, since a malicious attacker who takes control of a machine will be able to subvert any automatic controls. Finally, after considering current disruptions from Napster and SETI@home, we go back and consider browsers and the Web again. They were cited as disruptive back in 1994 and 1995. (Mosaic was first released unofficially around the middle of 1993, officially in the fall of 1993, and took off in 1994.) However, when we consider the growth rates for the University of Waterloo, for MichNet [CoffmanOl], or for SWITCH (which apparently had regular growth throughout the 1990s according to [Harms]), we do not see anything anomalous, just the steady doubling of traffic each year or so. If we consider the composition of the traffic, there were major changes. For example, Fig. 2.4 shows the evolution of traffic between the University of Waterloo and the Internet. (It is based on analysis of traffic during the third week in each March, and more complete results 2. Growth of the Internet 45 are available at http://www.ist.uwaterloo.ca/cn/Stats/ext-prot.html.) Web The did take over, but much more slowly than on Internet backbones. There are no good data sets, but it has been claimed that by the end of 1994, Web traffic was more than half of the volume of the commercial backbones. On the other hand, the data for the NSFNet backbone, available at http://www.merit.edu/merit/archive/nsfnet/statistics/index.html, that show Web traffic was only approaching 20% there by the end of 1994, a level similar to that for the University of Waterloo. Thus, at well-wired academic institutions such as the University of Waterloo and others that dominated NSFNet traffic, the impact of the Web was muted. Perhaps the main lesson to be drawn from the discussion in this section is that the most disruptive factor is simply rapid growth by itself. A doubling of traffic each year is very rapid, much more rapid than in other communication services. Figure 2.4 shows e-mail and netnews shrinking as fractions of the traffic at the University of Waterloo, from a quarter to about 5%. Yet the byte volume of these two applications grew by a factor of 12 during the 6 years covered by the graph, for a growth rate of over 50% per year, which is very rapid by most standards. If we are to continue the doubling of traffic each year, new applicationswill have to keep appearing and assuming dominant roles. An interesting data point is that even at the University of Wisconsin in Madison, which analyzes its data traffic very carefully, about 40% of the transmissions escape classification.That is consistent with information from a few corporate networks, where the managers report that upwards of half of their traffic is of unknown types. (A vast majority of network managers do not even attempt to perform such analyses.) This shows how difficult coping with rapid growth is. internet traffic at the University of Waterloo 0 E a u - 0 0 , m c 9 0) 0 m o m p n 0 cu I I I I I I 2000 1994 1995 1996 1997 year 1998 1999 Fig. 2.4 Composition of traffic between the University of Waterloo and the Internet based on data collected in March of each year. 46 Kerry G. Coffman and Andrew M. Odlyzko 8. Moore’s Law for Data Traffic The approximate doubling of transmission capacity of each fiber that is described in [CoffmanO2] is analogous to the famous Moore’s Law in the semiconductorindustry. In 1965, Gordon E. Moore, then in charge of R&D at Fairchild Semiconductor,made a simple extrapolation from three data points in his company’s product history. He predicted that the number of transistors per chip would about double each year for the next 10 years. This prediction was fulfilled, but when Moore revisited the subject in 1975, he modified his projection for further progress by predicting that the doubling period would be closer to 18 months. (For the history and fuller discussion of Moore’s Law, see [Schaller].) Remarkably enough, this growth rate has been sustained over the past 25 years. There have been many predictions that progress was about to come to a screeching halt (including some recent ones), but the most that can be said is that there may have been some slight slowdown recently. (For example, according to the calculations shown in FlderingSE], the number of transistors in leading-edge microprocessors doubles every 2.2 years. On the other hand, the doubling period is lower for commodity memories.) Experts in the semiconductor area are confident that Moore’s 1975 prediction for rate of improvement can be fulfilled for at least most of the next decade. Predictions similar to Moore’s had been made before in other areas, and in [Licklider]they were made for the entire spectrum of computing and communications. However, it is Moore’s Law that has entered the vernacular as a description of the steady and predictable progress of technology that improves at an exponential rate (in the precise mathematical sense). Moore’s Law results from a complex interaction of technology, sociology, and economics. No new laws of nature had to be discovered, and there have been no dramatic breakthroughs. On the other hand, an enormous amount of research had to be carried out to overcome the numerous obstacles that were encountered.It may have been incremental research, but it required increasing ranks of very clever people to undertake it. Furthermore, huge investments in manufacturing capacity had to be made to produce the hardware. Perhaps even more important, the resulting products had to be integrated into work and lifestyles of the institutions and individuals using them. For further discussions of the genesis, operations, and prospects of Moore’s Law, see [ElderingSE, Schaller]. The key point is that Moore’s Law is not a natural law, but depends on a variety of factors. Still, it has held with remarkable regularity over many decades. Although Moore’s Law does apply to a wide variety of technologies, the actual rates of progressvary tremendouslyamong Merent areas For example, battery storage is progressing at a snail‘s pace’ compared to microprocessor improvements. This has signscant implications for mobile Internet access, limiting processor power and display quality. Display advancesare more rapid than those in power storage, but nowhere near fast enough to replace paper 2. Growth of the Internet 47 as the preferred technology for general reading, at least not at any time in the next decade. (This implies, in particular, that the bandwidth required for a single video transmission will be growing slowly.) Dynamic Random Access Memories (DRAMS) is growing in size in accordance with Moore’s Law, but their speeds are improving slowly. Microprocessors are rapidly increasing their speed and size (which allows for faster execution through parallelism and other clever techniques), but memory buses are improving slowly. For some quantitative figures on recent progress, see [Grays]. From the standpoint of a decade ago, we have had tidal waves of just about everything: processing power, main memory, disk storage, and so on. For a typical user, the details of the PC on the desktop (MHz rating of the processor, disk capacity) do not matter too much. It is generally assumed that in a couple of years a nav and much more powerful machine will be required to run the new applications, and that it will be bought for about the same price as the current one. In the meantime, the average utilization of the processor is low (since it is provided for peak performance only), compression is not used, and wasteful encodings of information (such as 200 KB Word documents conveying a simple message of a few lines) are used. The stress is not on optimizing the utilization of the PC’s resources, but on making life easy for the user. To make life easy for the end user, though, clever engineering is employed. Because the tidal waves of different technologies are advancing at different rates, optimizing user experience requires careful architectural decisions [Grays, HennessyP]. In particular, since processing power and storage capacity are growing the fastest, while communication within a PC is improving much more slowly, elaborate memory hierarchies are built. They start with magnetic hard disks and proceed through several levels of caches, invisibly to the user. The resulting architecture has several interesting implications, which are explored in [Grays]. For example, mirroring disks is becoming preferable to RAID (Redundant Arrays of Inexpensive Disks) fault-tolerant schemes that are far more efficient but slower. The density of magnetic disk storage increased at about 30% per year from 1956 to 1991, doubling every 2; years [Economist]. (Total deployed storage capacity increased faster, as the number of disks shipped grew.) In the 1990s, the growth rate accelerated, and in the late 1990sincreased yet again. By some accounts, t h densities in disk drives are about doubling each year. For our ~ purposes, the most relevant figure will be total storage of disk drives. Table 2.5 shows data from an IDC study, which shows storagecapacity shippedeach year just about doubling through the year 2000, and then slowing down. However, that study was prepared in 1998, and since then IDC has revised upwards its estimatesfor disk storage systems toward a continuation of the doubling trend. Similar projections from Disk/Trend (http://www.disktrend.com/)also suggest that the total capacity of disk drives shipped will continue doubling through at least the year 2002. Given the advances in research on magnetic storage, it seems that a doubling each year until the year 2010 might be achievable (with 48 Kerry G. Coffman and Andrew M. Odlyzko Table 2.5 Worldwide Hard Disk Drive Market (based on September 1998 and August 2000 IDC reports) Year 1995 1996 1997 1998 1999 2000 200 1 2002 2003 2004 Revenues (bizfions) $21.593 24.655 27.339 26.969 29.143 32.519 36.219 40.683 Storage Capacity (terabytes) 76,243 147,200 334,791 695,140 1,463,109 3,222,153 7,239,972 15,424,824 30,239,756 56,558,700 some contributionfrom higher revenues, as shown in Table 2.5, but most coming from better technology).After about 2010, it appears that magnetic storage progress will face serious limits, but by then more exotic storage technologies may become competitive. It seems safest to assume that total magnetic disk storage capacity will be doubling each year for the next decade. However, even if there is a slowdown, say to a 70% annual growth rate, this will not affect our arguments too much. The key point is that storage capacity is likely to grow at rates not much slower than those of network capacity. Furthermore, total installed storage is already immense. Table 2.5 shows that at the beginning of the year 2000, there were about 3,000,000TB of magnetic disk storage. If we compare that with the estimates of Table 2.1 for network traffic, we see that it would take between 250 and 400 months to transmit all the bits on existing disks over the Internet backbones. This comparison is meant as just a thought exercise. The backbones considered in Table 2.1 arejust those in the United States, whereas disks counted in Table 2.5 are spread around the world. A large fraction of the disk space is spare, and much of the content is duplicated (such as those hundreds of millions of copies of Windows 98), so nobody would want to send them over the Internet. Still, this thought exercise is useful in showing that there is a huge amount of digital data that could potentially be sent over the Internet. Further, this pool of digital data is about doubling each year. An interesting estimate of the volume of information in the world is presented in [Lesk]. It shows that already in the year 1997we were on the threshold of being able to store all data that has ever been generated (meaning books, movies, music, and so on) in digital format on hard disks. By now we are well 2. Growth of the Internet 49 past that threshold, so future growth in disk capacities will have to be devoted to other types of data that we have not dealt with before. Some of that capacity will surely be devoted to duplicate storage (such as a separate copy of an increasingly bloated operating system on each machine). Most of the storage, though, will have to be filled by new types of data. The same process that is yielding faster processors and larger memories is also leading to improved cameras and sensors These will yield huge amounts of new data that have not been available before. It appears impossible to predict precisely what type of data this will be. Much is likely to be video storage from cameras set up as security measures or ones that record our every movement. There could also be huge amounts of data from medical sensors on our bodies. What is clear, though, is that “[tlhe typical piece of information will never be looked at by a human being” [Lesk]. There will simply not be enough of the traditional “content” (books, movies, music) nor even enough of the less formal type of “content” that individuals will be generating on their own. Huge amounts of data that is machine generated for machine use suggests that data networks will also be dominated by transfers of such data. This was already predicted in [deSolaPITH],and more recently in [Odlyzko2,StArnaud, StArnaudCFM].Given an exponential growthrate in volume of data transfers, it was clear that at some point in the future most of the data flying through the networks would be neither seen nor heard by any human being. Thus, we can expect that streaming media with real-time quality requirements will be a decreasing fraction of total traffic at some point within the next decade. There will surely be an increase in the raw volume of streaming real-time traffic, as applications such as videoconferencing move onto the Internet. However, as a fraction of total trafiic, such transmissionswill not only decrease eventually, but may not grow much at all even in the intermediate future. (Recall that at the University of Waterloo over the last 6 years, the volume of e-mail grew about 50% a year, but as a fraction of total trafiic it is almost negligible now.) The huge imbalance in volume of storage and capacities of long-distance data networks means that even the majority of traditional “content” will be transmitted as files, and not in streaming form. For more detailed arguments supporting this prediction, see [Odlyzko2]. This development, in which “content”is sent around as files for local storage and playback, is already making its appearance with MP3, Napster, and related programs. The huge hard disk storagevolumes also mean that most data will have to be generated locally. There will surely also be much duplication (such as operating systems, movies, and so on that would be stored on millions of computers). Aside from that, there will surely be huge volumes of locally generated data (e.g., from security cameras and medical sensors) that will be used (if at all) only in highly digested form. The examples in [Coffman02] support the notion that there is a “Moore’s Law” for data traffic, with transmission volumes doubling each year. Even at large institutions that already have access to state-of-the art technology, 50 Kerry G. Cofitnan and Andrew M. Odlyzko data traffic to the public Internet tends to follow this rule of doubling each year. This is not a natural law, but, like all other versions of Moore’s Law, reflects a complicated process, the interaction of technology and the speed with which new technologies are absorbed. A Moore’s Law for data traffic is different from those in other areas, since it depends in a much more direct way on user behavior. In semiconductors, consumer willingness to pay drives the research, development, and investment decisions of the industry, but the effects are indirect. In data traffic, though, changes can potentially be much faster. A residential customer with dial-up modem access to the Internet could increase the volume of data transfer by a factor of about five very quickly. All it would take would be the installation of one of the software packages that prefetch Web sites that are of potentialinterest and that fill in the slack between transmissions initiated by the user. Similarly, a university’s T3 connection to the Internet could potentially be filled by a single workstation sending data to another institution. Thus any Moore’s Law for data traflic is by nature much more fragile than the standard Moore’s Law for semiconductors,for example. Thus it is remarkable that we see so much regularity in growth rates of data transfers. Links to the public Internet are usually the most expensive parts of a network, and are regarded as key choke points They are where congestion is seen most frequently at institutional networks. Yet the “mere” annual doubling of data traffic even at institutions that have plenty of spare capacity on their Internet links means that there are other barriers that matter. The obvious one is the public Internet itself. It is often (some would say usually) congested. A terabit pipe does not help if it is hooked up to a megabit link, and so providing a lightly utilized link to the Internet does not guarantee good end-to-end performance. Yet that is not the entire explanation either, since corporate Intranets, which tend to have adequate bandwidth and seldom run into congestion, tend to grow no faster than a doubling of traflic each year. There are other obstructions, such as servers, middleware, and, perhaps most important, services and user interfaces People do not care about getting many bits. What they care about are the applications. However, applications take time to be developed, deployed, and adopted. To quote J. Licklider (who probably deserves to be called “the grandfather of the Internet” for his role in setting up the research program that led to the Internet’s creation): A modern maxim says: “People tend to overestimate what can be done i one n year and to underestimate what can be done in five or ten years.” Picklider] “Internet time,” where everything changes in 18 months, has a grain of truth, but is largely a myth. Except for the ascendancy of browsers, most substantial changes take 5 to 10 years. As an example, it has been at least 4 years since voice over IP was first acclaimed as the ‘‘next big thing.” Yet its impact so far has been surprisinglymodest. It is coming, but it is not here today, and it won’t be here tomorrow. People take time to absorb new technologies. 2. Growth o f the Internet 51 What is perhaps most remarkable is that even at institutions with congested links to the Internet, traffic doubles or almost doubles each year. Users appear to find the Internet attractive enough that they exert pressure on their administration to increase the capacity of the connection. Existing constraints, such as those on e-mail attachments, or on packetized voice or video, as well as the basic constraint of limited bandwidth, are gradually loosened. Note that this is similar to the process that produces the standard Moore’s Law for PCs. Intel, Micron, Toshiba, and the rest of the computer industry would surely produce faster advances if users bought new PCs every year. Instead, a typical PC is used for 3 to 4 years. On one hand there is pressure to keep expenditures on new equipment and software under control, and also to minimize the complexity of the computing and communicationssupportjob. On the other hand, there is pressure to upgrade, either to better support existing applications or to introduce new ones. Over the last three decades, the conflict between these two pressures has produced a steady progress in computers. Similar pressures appear to be in operation in data networking. In conclusion, we cannot be certain that Internet t r a c will continue doubling each year. All we can say is that historically it has tended to double each year. Still, trends in both transmission and in other information technologies appear to provide both the demand and the supply that will allow a continuing doubling each year. Since betting against such Moore’s laws in other areas has been a loser’s game for the last few decades, it appears safest to assume that data traffic will indeed follow the same pattern, and grow at close to 100% per year. 9. Further Economic and Technical Considerations A frequently asked question concerns the elasticity of demand for data transmission capacity. However, for long-range projections it might be more useful to think of analogies with the computer industry. In that industry, product managers clearly do think about elasticities in the short or intermediate terms. From a long-range perspective, though, what dominates are the effects of Moore’s Law. Table 2.6 (drawn from [FishburnO]) shows a dozen years from the history of Intel. The leading microprocessor sold for roughly a constant price all during this period. However, its power was increasing at the exponential rate given by Moore’s Law. Intel’s total revenues (and profits) grew, as more processors were being sold, but this growth rate was considerably more modest than that of the computing power. Users found the increasing computational power of new PCs sufficiently attractive that they not only bought new PCs, but increased their total spending. They did this even though most of that power was sitting idle, and it was only the occasional bursts of recomputing a spreadsheet or bringing up a presentation package that mattered. A similar 52 Kerry G. Coffman and Andrew M. Odlyzko Table 2.6 Intel and Its Microprocessors(each year lists the most powerful General Purpose Microprocessors Sold by Intel, Its Computing Power, Price at the End of the Year (in Dollars), and Intel’s Revenues and Profits for That Year (in Millions of Dollars)) Price (dollam) 300 Year 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 Processor 386 DX (16MHz) 386 DX (20 MHz) 386 DX (25 MHz) 486 DX (25 MHz) 486 DX (33 MHz) 486 DX (50 MHz) DX2 (66 MHz) Pentium (66 MHz) Pentium (100MHz) Pentium Pro (200 MHz) Mips 5 6 8 20 27 41 54 112 166 400 Revenue (millions of dollars) 1,265 1,907 2,875 3,127 3,922 4,779 5,844 8,782 11,521 16,202 20,847 25,070 Net Profit (millions of dollars) - 173 248 453 391 650 819 1,067 2,295 2,266 3,566 5,157 8,945 950 950 644 600 898 935 1,325 735 Pentium I1 (300 MHz) 600 evolution might take place in networking. Total spendingmay (subjectto business cycles) increase at a moderate pace, while the bandwidth and traffic grow at rates determined by technological progress. If that happens, we are likely to see traffic and capacity about doubling each year, with capacity growth faster than that of traffic. 10. Conclusions Much of the almost hyperactivitywithin the optical fiber telecommunications industry over the past few years can be traced to the perceived and real growth of the traffic on the Internet. We maintain that the overall growth rate of the Internet for most of its existence (despite some excursions) was remarkably close to “doubling every year,” and we anticipate that this rate will continue into the foreseeable future. In effect, we see a type of Moore’s Law associated with the growth of data traffic. This type of growth rate is in sharp contrast to the historical growth rates of various methods of communications (including conventional mail, telegraph service, and traditional voice phone service) that tended to be no greater (and typically much less) than about 10Y0per year. Still, even though a doubling each year represents very fast growth, it is only comparable to the rate of progress in transmission capacity. Hence we are 2. Growth of the Internet 53 unlikely to see the huge increases in spending on optical communication that many business plans had been based on. Throughout the history of the Internet there have only been two “killer applications”: e-mail and the Web (including Web browsers). Several events conspired that allowed an unprecedentedexplosion (roughly 100-fold increase) in Internet traffic in the 1995-1996 time frame, and the Internet was able to handle this because it made use of the existing telephone industry infrastructure. Because the Internet is quickly approaching the point at which it is the predominant network, it is very unlikely that such huge growth rates could be so easily supported in the future. It also appears that, aside from short-range perturbations, there will be neither a “bandwidth glut” nor a “bandwidth shortage” in the foreseeable future, in that supply and demand will be growing at comparable rates. As such, it is very likely that pricing will begin to play an even more important role in the evolution of traffic. Throughout most of the 1990s, data transmission prices were increasing. However, there are recent signs that they are beginning to decrease, and in some cases, especially across the Atlantic and on major transcontinental routes in the United States, they have decreased dramatically. If they begin to decrease rapidly in general, then many of the constraints on usage that exist today may very likely start to ease. We are likely to see capacity growing somewhat faster than traffic, a continuation of the trend we have already seen in the last few years. We also believe that “file” transfers, and not real-time streaming,will remain dominant on the network. Streaming real-time transmissionswill undoubtedly grow in absolute terms, and as a fraction of the total traffic it may increase for a while. However, in all likelihood it will eventually begin to decline as the demand for this type of traffic will not grow as fast as network capacity. We foresee sharing applications as a likely candidate to fuel traffic growth. One of the first major examples of this was Napster, because it effectively turned consumers of information into providers of information. It is extremely likely that such file sharing applications will be some of the key applications that continue to fuel the annual doubling of data traffic. References [Abbate] [Baran] [Boardwatch] [Bruno] J. Abbate, Inventing the Internet, MIT Press, 1999. I? Baran, On distributed communications network, IEEE Trans. Comm. Systems, vol. 12, 1964, pp. 1-9. Boardwatch. Available at http://www.boardwatch.com. L. Bruno, Fiber optimism: Nortel, Lucent, and Cisco are battling to win the high-stakes fiber-optics game, Red Herring, June 2000. Available at http://www.herring.com/mag/issue79/mag-fiber79.html. 54 Kerry G. Coffman and Andrew M. Odlyzko V. G. Cerf, A brief history of the Internet and related networks. Available at http://www.isoc.org/internethistory/cerf html. V. G. Cerf and R. E. Kahn, A protocol for packet network interconnection, IEEE Trans. Comm. Tech., vol. COM-22, May 1974, pp. 627-641. N. Cochrane, We’re insatiable: Now it’s 20 million million bytes a day, Melbourne Age, January 15, 2001. Available at http:llwww .it.fairfax.com.au/networking/200 115/A13694-2001Jan15.html. 10 K. G. Coffman and A. M. Odlyzko, The size and growth rate of the Internet, First Monday, October 1998. Available at http://fkstmonday.org/. Also available at http://www.dtc.umn.edu/ -0dlyzko. K. G. Coffman and A. M. Odlyzko, Internet growth: Is there a “Moore’s Law” for data traffic?, Handbook o Massive Data Sets, f J. Abello, P M. Pardalos, and M. G. C. Resende, eds, Kluwer, . 200 1, in press. Available at http://www.dtc.umn.edu/-odlyzko. CTIA (Cellular Telecommunications Industry Association), SemiAnnual FEreleSs Industry Survey, June 1985 to June 2000. Available at http:l/www.wow-com.codwirelesssurvey/. Geography of Cyberspace Directoiy: Internet Trafic and Demographic Statistics. Available at http://www.cybergeography.org/ statistics.html. I. de Sola Pool, H. Inose, N. Takasaki, and R. Hurwitz, Communications Flows: A Census in the Unitedstates andJapan, North-Holland, 1984. D. A. Dunn and A. J. Lipinski, Economic considerations in computer-communication systems, pp. 371- 422 in ComputerCommunication Networh, N. Abramson and F. F. Kuo, eda, Prentice-Hall, 1973. Not Moore’s Law, The Economist, July 12,1997. C . A. Eldering, M. L. Sylla, and J. A. Eisenach, Is there a Moore’s Law for bandwidth? IEEE Communications Magazine, October 1999, pp. 2-7. I? C . Fishburn and A. M. Odlyzko, Dynamic behavior of differential pricing and Quality of Service options for the Internet, pp. 128-139 in Proc. First Intern. Con$ on Information and Computation Economies (ICE-98), ACM Press, 1998. Available at http://www.dtc.umn.edu/-odlyzko. S. Floyd and V Paxson, Dimculties in simulating the Inter. net, IEEE/ACM Damactions on Networking, in press. Available at ht tp://www.aciri.orglfloydpapers.html. D. Galbi, Bandwidth use and pricing trends in the U.S., Telecommunications Policy, vol. 24, no. 11, December 2000. Available at http://www.galbithink.org. 1Cerf-I [CerfKI [Cochrane] [CoffmanOl] [Coffman02] [CTIA] [Cyberspace] [deSolaPITH] PWLI [Economist] [ElderingSE] [Fishburno] FlOYdPI [Galbi] 2. Growth of the Internet 55 J. Gray and P. Shenoy, Rules of thumb in data engineering, Proc. 2000 ZEEE Intern. Con$ Data Engineering. Available at http://research.microsoft.com/-gray. E. E. Green, Communicationsspectra by the wholesage2012 A.D., Proc. IRE, vol. 50,1962, pp. 585-587. Reprinted i Proc. ZEEE, vol. n [Green] [Harms] [HennessyP] [Hobbes] [Houghl [Kleinrockl] 87,1999,1293-1295. J. Harms, From SWITCH to SWITCH*-extrapolating from a case study, Proc. INET '94, pp. 341-1 to 341-6. Available at http:l/ info.isoc.org/iso~whatis/conferences/ine~94lpapers/index. html. J. L. Hennessy and D. A. Patterson, Computer Architecture: A Quantitative Approach, Morgan Kaufmann, 1990. Hobbes Internet Emeline. Available at http:/lwww.zakon.org/ robert/internet/timeline/. R. W. Hough, Future data traffic volume, ZEEE Computer, September-October 1970, pp. 6-12. L. Kleinrock, Information flow in large communicationsnetworks, RLE Quarterly Progress Report, July 1961, pp. 1-35. Available at http://www.lk.cs.ucla.edu/LK/Bib/REPORT/PhD. [Kleinrock2] [Kleinrock3] [Leiner] L. Kleinrock, Communication Nets; Stochastic Message Flow and Delay, McGraw-Hill, 1964. L. Kleinrock, ISDN-The path to broadband networks, Proc. ZEEE, 1991, pp. 112-117. B. M. Leiner, V. G. Cerf, D. D. Clark, R. E. Kahn, L. Kleinrock, D. C. Lynch, J. Postel, L. G. Roberts, and S. Wolf€, A Brief History of the Internet, Version 3.31, August 4, 2000. Available at http://www.isoc.org/internetlhistory/brief.html. M. Lesk, How much information is there in the world? 1997, unpublished paper. Available at http://www.lesk.com/mlesk/diglib.html. J. C. R. Licklider, Libraries of the Future, MIT Press, 1965. J. C . R. Licklider and W. Clark, On-Line Man Computer Communications, August 1962. R. W. Lucky, New communications services-What does society want? Proc. ZEEE, vol. 85, 1997, pp. 15361543. R. W. Lucky, Through a glass darkly-Viewing communicationsin 2012 from 1961, Proc. IEEE, vol. 87, 1999, pp. 1296-1300. J. McCredie, UC Berkeley must manage campus network growth, The Daily Californian, March 14, 2000. Available at http://www.dailycal.org/article.asp?id=1912&ref=news M. Meeker, M. Mahaney, and D. Joseph, The Internet userlusage ecosystem framework, Morgan Stanley Dean Witter report, January 24, 200 1. Available at http://www.msdw.com/techresearch/ index.htm1. The Multi-Router Traflc Grapher of Tobias Oetiker and Dave Rand, V O ~ .79, [Lesk] [Licklider] [LickC] [Lucky1 1 [Lucky21 [McCredie] [MeekerMJ] information and links to sites using it at http://ee-staff.ethz.ch/ -oetiker/webtools/mrtg/mrtg.html. 56 Kerry G. Coffman and Andrew M. Odlyzko U. S. Information Infrastructure Task Force, The NationalZnformation Znffastructure.Available at http://www.ibiblio.orglniiltoc.html. [Nolll] A. M. Noll, Znhoduction to Telephones and Telephone Traflc, 2nd ed., Artech House, 1991. A. M. Noll, Highway of Dreams: A Critical Appraisal of the Commu[No1121 nications Superhighway, Lawrence Erlbaum Associates, 1997. pol131 A. M. Noll, Does data traffic exceed voice traffic? Comm. ACM, June 1999, pp. 121-124. Nua Internet Surveys. Available at http://www.nua.com. [Nual A. M. Odlyzko, Data networks are lightly utilized, and wl stay il [OdlYZko11 that way. Available at http://m.dtc.umn.edu/-odlyzko. [Odlyzko2] A. M. Odlyzko, The history of communications and its implications for the Intenet. Available at http://www.dtc.umn.edu/-odlyzko. [Odlyzko3] A. M. Odlyzko, Internet pricing and the history of communications, Computer Networh, vol. 36, 2001, pp. 493-517. Also available at http://www.dtc.umn.edu/-odlyzko. J. A. Polly, Surfing the Internet: An introduction, Wilson Library Bulletin, June 1992, pp. 38-42. Available at http://www.netmom .com/about/surfing.shtml. [ReichlLS] P Reichl, S. Leinen, and B. Stiller, A practical review of pricing and . cost recovery for Internet services, Proc. 2nd Internet Economics Worhhop Berlin (IEW '99), Berlin, Germany, May 28-29, 1999. Available at http://www.tik.ee.ethz.ch/-cati/. R. R. Schaller, Moore's law: Past, present, and future, IEEE Spec[Schaller] trum, vol. 34: no. 6, June 1997, pp. 52-59. Available through Spectrum online search at http://www.spectrum.ieee.org. [StAmaud] B. St. Arnaud, The future of the Internet is NOT multimedia, Network World, November 1997. Available at http://www.canarie.ca/ -bstarn/publications. html . [StAmaudCFM] B. St. Arnaud, J. Coulter, J. Fitchett, and S. Mokbel, Architectural and engineering issues for building an optical Internet. Short version in Proc. SOC.Optical Engineering, 1998. Full version available at http://www.canet3.net. [Standage] T. Standage, n e Kctorian Internet: n e Remarkable Story of the Telegraph and the Nineteenth Century5. On-line Pioneers, Walker, 1998. S. Taggart, Telstra: The prices fight, Wired News, http://www.wired .com/news/politics/O,1283,32961,OO.html. €? M. Walker and S. L. Mathison, Regulatory policy and future data transmission services, pp. 295-370 in ComputerCommunication Networh, N. Abramson and F. E Kuo, eds, Prentice-Hall, 1973. W W. Wu and A. Livne, ISDN A snapshot, Proc. ZEEE, vol. 79, . 1 9 9 1 , ~103-111. ~. PI11 Chapter 3 Optical Network Architecture Evolution John Strand AT&T Laboratories, Middletown, New Jersey 1. Introduction An Optical Transport Network (OTN) is composed of interconnectednetwork elements (NEs), plus software and operational processes that must function together to provide services. Ongoing advances in technology will cause significant changes in OTN architecture; however, equally important will be the growth and evolution of the services it transports, particularly the Internet. In addition, business changesin the telecommunicationsindustry will be very critical. This chapter tries to weave the technology, services, and business stories together to indicate how they are shaping the architecture of the emerging optical network. Because most of the readers of these volumes are primarily technologists, tutorial material has been included. 1.1. WHAT IS AN OPTICAL TRANSPORTNETWORK? The very definition of “Optical Transport Network” illustrates why these stones are interrelated. All would agree that optical fiber will be the physical layer of an OTN. There is less agreement on whether a network making extensive use of electronics for regeneration and switching should be called an OTN. Most of the audience for this volume are presumably most interested in all-optical OTNs; however, for a variety of business and service-related reasons that will be discussed later, many of the first generation “optical crossconnects” perform electronics-based functions like multiplexing DS-3s and slower speed SONET OC-ns into OC-48s and OC-192s. From this perspective, SONET is an “opaque one-wavelength”optical network, as Green [47]pointed out, even though much of its characteristic functionality is implemented in electronics. We will take a “broad church” approach to defining an OTN. Given the nature of this volume, we will concentrate primarily, when possible, on networks built from optical components; however, when necessary we will use the term to include networks transporting STS-1 (52 Mb/s) and larger connections that have an optical physical layer. When necessary to be precise, we will use the term “photonic” instead of “optical” to indicate that an all-optical The views expressed are those of the author and not necessarily those of AT&T. 57 OPTICAL FIBER TELECOMMUNICATIONS, VOLUME IVB Copyright 0 2002, Elsevier Science (USA). AU rights of reproduction in any form reserved. ISBN: 0-12-395173-9 58 John Strand situation is being discussed. Thus, a “PhotonicTransport Network” (PTN) will be an all-optical OTN, and “Photonic Cross-Connect” (PXC) and “Photonic Add/Drop Multiplexer” (PADM) will be used when all-optical equipment is under discussion. We will call a communications channel through an OTN a “circuit.” It is frequentlycalled a “wavelength”or a “lightpath,”but because of the possibility that a connection might be partially electronic,we will stick with “connection.” If it is all-optical,we will also use the term “optical channel” (OCh). 1.2. SCOPE OF THIS CHAPTER Because other chapters in the current volume deal with submarine systems, metropolitan OTNs, and access OTNs, we will focus primarily on intercity networks. Because of the background of the author, most of the discussion will have a strong U.S.flavor: For example, I w luse SONET rather than SDH i l terminology, and I will frequently discuss issues of most interest to networks with diameters of thousands of kilometers, even though such large networks are only relevant to a handful of countries. To keep the material to be covered within bounds, only technology likely to be commercially available within the next few years is considered. Many interesting areas, such as optical packet switching, are therefore omitted. The reader should keep in mind the large error bars surrounding this whole enterprise. As an example, the architecture of actual OTNs even a few years in the future depends profoundly on the Internet: If its growth were to slow, the rate of introduction of new architectures would certainly slow, and the functionality desired (and hence the underlying technologies)might well change. 2. Technology Advances and Trends The important underlying optical technologiesare for the most part discussed elsewhere in this volume; the reader interested in the technologies per se should turn to the relevant sections. Our purpose here is twofold: (1) identify and define at a system level the building blocks on which the OTN architecture rests, and (2) point out major system-leveltrends we need to deal with in our later architecture development. 2.1. SONETBDH REFRESHER SONET/SDHis important for optical networking for several reasons: (1) The overwhelming proportion of connections in long-haul OTNs are SONET or SDH formatted, and (2) many aspects of OTN architectures have been modeled on SONET/SDH concepts. This section gives a very brief overview of a few key aspects of these protocols that we will need later. For more complete overviews, see [18,72,73]. 3. Optical Network Architecture Evolution 59 Table 3.1 Selected SONET Signal Rates SONET Name STS-1 STS-3 STS-12 STS-48 STS-192 STS-768 Name When Transported Optically oc-1 OC-3 oc-12 OC-48 OC- 192 OC-768 SDH Name STM-1 STM-4 STM-16 STM-64 STM-256 Signal Rate (Mbhec) 51.84 155.52 622.08 2,488.32 9,953.80 39,813.12 User Rate’ (IMb/sec) 49.54 148.61 594.43 2,377.73 9,510.91 38,043.65 The “User Rate” is the bandwidth actually available for user data. The difference between it and the “Signal Rate” is due to overhead information as discussed in the text. ’ Synchronous Optical NETwork (SONET) is a North American standard for networking developed in the mid-1980s primarily by Bellcore and standardized by ANSI [79,99, 1001. It defines the interface between two SONET network elements (NEs). More specifically, it defines a digital hierarchy of synchronous signals, including their formats and mappings of asynchronous signals (e.g., DS-1, DS-3) into these formats, and defines the electrical and optical characteristics of the interface. The Synchronous Digital Hierarchy (SDH) is a closely related standard developed by the ITU [78]. The basic SONET entity is the Synchronous Transport Signal-1 (STS-1). It operates at 51.84Mb/secondY which 49.5 Mb/sec is usable payload and of the rest overhead. The STS-1 frame structure is byte-oriented and has 9 rows and 90 columns, 4 of which are used for overhead purposes. The frame rate is 8000/second (125 ps/frame). Normally all SONET signals are bidirectional and run at the same rates in each direction. An STS-N signal (n > 1) is formed by byte-interleavingn STS-1s together. When an STS-N is transported electrically, it is called an “EC-n”; when transported optically, it is an c‘OC-n.y’ (EC stands for “Electrical Carrier,” OC for “Optical Carrier.”) SDH is virtually identical to SONET, except that its base frame is three times larger, corresponding to a SONET STS-3. It is called “STM-1,” where STM stands for “Synchronous Transfer Module.” The key SONET signal rates are summarized in Table 3.1. In their simplest form, the three basic SONET network elements are shown in Fig. 3.1. The Digital Cross-Connect System (DCS) is the most general of these NEs. In its simplest form, it has a fabric capable of cross-connecting STS-M and STS-N line-side2ports. Normally N is greater than the M , but it can be the “Line-side” refers to the ports that are closest to the interoffice fibers; “drop-side” refers to those closest to the customer. 60 Johnstrand STS-N I I k STS-N STS-M Fabric (M<=N) STS-N -t E S sE P A I S T llloopll STS-M Ports N:M Digital Cross-Connect System (DW STS-M Ports N:M AddlDrop Multiplexer (ADMI STS-M Ports N:M Multiplexer (“End Terminal”) Fig. 3.1 Generic SONET network elements. same. If N > M , an incoming STS-N is first demultiplexed into a number (N/M at most) of STS-Ms, which are switched through the STS-M fabric and reassembled into STS-N if they are continuing on or are dropped in the office as STS-Ms through the ports shown at the bottom. There could be many (thousands) of STS-N and STS-M ports on a single DCS. The Add-Drop Multiplexer (ADM), shown in the middle of Fig. 3.1, is a special case of a DCS. It has only two pair of bidirectional line-side ports, which are often designated “east” and “west.” Each pair has a service STS-N (“S” in the figure), and also a protection STS-N (“P”)that is normally in standby mode until needed to recover from a failure. The protection capacity may also be used for “extra traffic”--connections that will be preempted in the event of a failure. A basic multiplexer (see the right side of Fig. 3.1; often called an “end terminal”) is a further specialization. There is either a single line-side STS-N port (as shown) or a service/protection pair. It is used to multiplex a number of lower-speed signals into a single STS-N for transport through the network. All three of these types of NEs are very widely deployed today and continue to be deployed in large volumes (billions of dollars per year in the United States alone). In intercity networks, a typical ADM installed today would likely have STS-48 or STS-192 line-side ports and a mix of STS-3 and STS-12 drop-side ports. A number of specialized deployment configurations are also specified by the SONETEDH standards. The most important configuration is the Self-Healing Ring (SHR). One such SONET SHR configuration is called a “line-switched ring” for reasons that will become apparent shortly. In this configuration, up to 16 SONET ADMs are configured in a ring topology, as shown at the left in Fig. 3.2. 3. Optical Network Architecture Evolution 61 P=‘ !? : P O ................ P @I--I 0 S S ADM At Office F Fig. 3.2 SONET line-switched ring example. The ADMs are labeled “A” through “F,” with the service and protection STS-Ns interconnecting them labeled “S” and “P,” respectively. A connection is shown entering at A, then passing through F and E before exiting the ring at D. (Solid line is also labeled “1.”) The right side shows this connection passing through the fabric from one service port to the other. If the service connection between F and E fails, ADMs F and E would cooperate to reroute the connection over their protection connection. (Dashed line in the figure, also labeled “2.”) If there is a route failure that affects both S and P between F and E, the connection is rerouted the opposite way around the ring. (Dotted line, labeled “3.”) In both cases, the initial routing of the connection is reestablished at E. These recovery mechanisms are triggered by standardized signaling messages between F and E that are carried in the SONET overhead bytes mentioned earlier. The rerouting is done by the ADM switch fabrics as illustrated at the right in Fig. 3.2. The specialized structure of the configuration allows very rapid reaction (50 ms on rings under 1200km in circumference and with no extra traffic [79]. Sometimes 150-200 ms is used as a bound for an arbitrary ring.) In some cases the rigid restoration discipline generates convoluted restoration paths. For example, if for some reason an A-B connection was routed the long way around the ring (A-F-E-D-C-B) and E-F failed, then the restored connection would be routed A-F-A-B-C-D-E-D-C-B, even though A-B would have restored the failed connection! This is done to keep the protocol and implementations more manageable. The protection capacity is shared in this type of ring. This means that if a different link failed, the same protection capacity would be used for the connections affected. For this reason they are also called “Shared Protection Rings” (SPRING). Another important type of SONET ring is called a “path-switched ring,” for reasons discussed below. In this type of ring each connection has dedicated protection capacity. In Fig. 3.2, if the A-F-E-D connection were protected in this manner, there would have been dedicated protection capacity for the 62 John Strand connection that would be routed A-B-C-D. The ADMs at A and D (the entry and exit points) would monitor the health of the connection and switch the connection to the A-B-C-D path if necessary. For this type of ring there is no need to know exactly where the failure occurred, because the switch is between entry and exit points. A path-switched ring is closely related to what is called “1 1 protection.” As in the path-switched ring, the connection has two dedicated paths, but in the 1 1 case, copies of the signal are continuously sent along both paths and the better quality signal is selected (at D in our example for the A to D signal). This is extremely fast and has the added advantage that no signaling between nodes is required. The drawback of 1 + 1 protection is that no extra preemptible traffic is possible on the protection capacity. SONET functionality is divided into three layers, not all of which need to be implemented by every SONET NE. The layers and their principal functions are: + + 0 Path. Maps specific services into a SONET payload; end-to-end error 0 0 0 and status monitoring; path protection switching. Adds path overhead. Line. Multiplexing multiple paths into a STS-N; synchronization;error and status monitoring; line protection switching. Adds line overhead. Section. Framing, scrambling, other functions associated with the preparation for physical transport. Adds section overhead. Physical. Electrical-to-Opticalconversion; actual optical transmission. Normally an ADM would be a line and section terminating NE, whereas a regenerator would terminate only the section. A path can ride on multiple lines in series, a line on multiple sections. See Fig. 3.3. Here, PTE, LTE, and STE stand for path, line, and section terminating equipment, respectively. SONET is a byte-structured protocol. An STS-1 frame is transmitted as a string of 810 bytes that are divided into 9 “rows” of 90 bytes each. The frame is Fig. 3.3 SONET layering. 3. Optical Network Architecture Evolution 63 divided into payload and overhead sections. The resulting frame is illustrated in Fig. 3.4. POH stands for “payload overhead.” Each of the overhead areas contain parity information to allow error checking and data communication channels to allow peer NEs to communicate. In addition, the section overhead contains framing bytes to allow frame alignment to be verified, the line overhead contains fields to control protection switching, and the path overhead identifies the type of payload. Four columns are dedicated to overhead and 86 to payload. We mentioned earlier that STS-Ns are formed by byte interleaving N STS-1s. This fragments the payload into N pieces, which is undesirable for data communications (as discussed later). To deal with this, a “concatenated” frame is also deked. Denoted STS-Nc (e.g., STS-48c or 0 C - 4 8 ~ it~ ) has one large payload rather than N smaller ones. However, it still has 3N columns of overhead, which are mostly unused. SDH is virtually identical functionally to SONET,but unfortunately the two standards use different terminology. Table 3.2 gives their correspondences. Overhead p Section 0 Line Overhead ,, Payload 1SDH Term A 9 Rows i Fig. 3.4 STS-1 frame structure. Table 3.2 SONET and SDH TerminologyRelationship SONET Term STS-N Path Line Section Line-switchedring Path-switched ring STM-N/3 Transmission path Multiplex section Regenerator section Shared-protectionring (MSEPRING) Dedicated-protectionring (PaWDPRING) 64 John Strand Receivers Frequency registered transmitters Fig. 3.5 Basic Optical Transport System (OTrS). 2.1.1. Multivendor Interworking One of the goals of SONET and SDH was to define standards that would let equipment from multiple vendors interwork. This has been achieved in simple point-to-point configurations, but in more complex configurations, such as shared protection rings, progress has been frustratingly slow. Hardware interoperability has by and large not been a serious problem; instead, software interworking has been limiting, particularly the exchange of state information between the ADMs and the handling of “operations, administration, and maintenance” functions. As a result, transport people tend to be suspicious of proposals for multivendor “mid-span meets.” 2.2. OPTICAL TRANSPORT SYSTEMS A basic Optical Transport System (OTrS) is shown in Fig. 3.5. In its most basic form, it consists of an optical multiplexer and demultiplexer and a number of optical amplifiers (OAs) between them.3 OAs in terrestrial systems usually have a nominal spacing (span length) of about 25dB (roughly 80km after allowance for splices, etc., and assuming a nominal span loss of 0.25 dB/km). Wider spacings can significantly reduce the first costs of a route, but cause difficulties for 10 Gb/sec and faster connections and also impose limits on the number of wavelengths that can be supported, therefore most operators appear to be sticking with 80 km or shorter spacings. Impairments force regeneration after about 5-7 spans (400-560 km). This is a wavelength rather than an OTrS constraint, so that if a wavelength traversed two such systems in series without regeneration between the systems, the span limit would apply to the sum of the spans on the systems. 2.2.1. Capacity Trends OTrS capacity has been increasing very rapidly. In fact, by some estimates the rate of increase is considerably faster than Moore’s Law for electronics The system shown (and the underlying technology, usually) is unidirectional; however, they are normally deployed in pairs so as to support bidirectional SONET/SDH circuits. 3. Optical Network Architecture Evolution Gbls 65 1996 1998 2000 2002 2004 Fig. 3.6 Capacity of a single fiber (derived from data in [SI). (see [54]). The underlying trends are analyzed in [55], from which Fig. 3.6 is derived. The years given are for commercial deployment (actual and projected). Over the 10-year period shown, capacity doubles approximately eight times due to increasesin the bits per channel, reductions in wavelength spacing, and increases in the total usable fiber bandwidth. In a recent talk [104], Rick Barry of Sycamore used bandwidth times distance as a relevant metric. He traced the evolution of capacity from 120-160 Terabits/sec*km with conventional systems (e.g., 80 OC-48s over 600km or 40 OC-192s over 400km) to today's 160&2400Tb/sec*km (e.g., 80 OC-192 over 3000km) and projected a next generation providing 3200-4800 Tb/sec*km. Important architectural implications from these trends: 0 0 0 The increase in wavelength counts make the introduction of some form of mechanized cross-connect a necessity if operations costs and complexity are to be kept under control. Total OTrS costs are rising significantly more slowly than capacity. Hence unit cost ($/Gigabit) is declining. The combination of these two trends-rapidly rising capacity and declining unit costs-have formed a synergistic relationship with the rapid growth of the Internet. Internet growth has allowed the new technologies to be economicallyjustified, while the rapid capacity increases and cost decreases have been essential enablers for the growth of the Internet. 2.2.2. Ultra-Long-Haul Systems The OTrS just described requires each wavelength to be regenerated roughly every 500 km. The optical-electrical-optical(OEO) functionality required for regeneration is quite expensive; if a transponder is put on each wavelength of the OTS shown in Fig. 3.5, their cost could exceed the total costs of all the 66 John Strand MdDemux and OAs combined. Hence extending the regeneration distance is an appealing possibility. Raman amplification, strong forward error correction (FEC), and advances in dynamic power management are making this possible. The resulting “ultra-long-haul” technology can support OTrS with dozens of 80-km spans, leading to regeneration distances of 3000 km or more. However, this comes at a cost: To get longer regeneration distances, the perwavelengthmultiplexing, demultiplexing,and amplificationscosts are likely to significantlyexceed those of traditional systems. We can expect ultra-long-haul systems to be competitive, therefore, only for long systems where these perwavelength costs can be counterbalancedby regeneration savings. A strawman ultra-long system based on 2000-2001 products is shown in Fig. 3.7. This system shown in Fig. 3.7 illustrates a number of architecturally interesting features we shall return to in the architecture discussion. a Adaptation. The boxes labeled “A” are adaptation functions. Using an OEO transponder function, they map one or more inputs (the a’s, typically standard short-reach signals) into a long-reach OCh or group of OCh‘s that will pass transparently to a distant adaptation function. Adaptation options include: a. Multiplexing. Either electrical or optical TDM may be used to combine the inputs into a single wavelength. This is done to increase effective capacity. After multiplexing, the combined signal must be routed as a group to the distant adaptation function. b. Adaptation grouping. In this technique, groups of k (e.g., 4) inputs are managed as a group (an “adaptation grouping,” increasingly called a “wave group”) within the system and normally must be addeddropped as a group. Tight spacing is used for wavelengths within a group, with larger guard bands between groups. Grouping is done to simplify power management. It may also be possible to largely contain nonlinear effects such as four-wave mixing within the groups. Wavelength spacing may vary between groups, as may the X PADM Y PADM 6 11 Olk ........................ =a1 T..............F IA D AN =Nl =Nk ok a Fig. 3.7 Strawman ultra-long-haul optical transport system (OAs not shown). 3. Optical Network Architecture Evolution 67 number of wavelengths in each group. Note that either option places limits on connectivity (which 0’s are delivered to which output port), because the inputs involved must be routed as a group to a compatible output adaptation function. Photonic add-drop multiplexers (PADM). X and Y are “photonic add-drop multiplexers”-the all-optical analog to the SONET ADM discussed in Section 3.1. They allow economical dropping and insertion of limited numbers of adaptation groupings without requiring demultiplexing and remultiplexing all the other wavelengths. Depending on the filtering architecture, it may be possible to reuse frequencies so that, for instance, the same frequency could be used for a D to X grouping, an X to Y grouping, and a Y to E grouping. “Domain oftransparency.” The dotted line encloses an all-optical “domain of transparency,” an all-optical subnetwork. The adaptation functions just discussed optically isolate the domain. 0 2.2.3. More Complex Domains of Transparency Since the PADMs in Fig. 3.7 are all-optical, it is possible to build more complex all-optical domains, as shown in Fig. 3.8. In Fig. 3.8, the basic ultra-long system D-X-Y-E from Fig. 3.7 has had branches added at the PADM’s X and Y, with further branching at PADM U. In this configuration, there is an all-optical path, A-Y-X-U-Z, connecting A to Z . Transponders to optically isolate the domain would need to be present on the boundary of the domain and would serve to define the boundary of the domain. There are no “loops” in Fig. 3.8. If a U-Y link were added, the domain would turn from a topological “tree” (only one path between any two points) into a more general “mesh.” The alternate paths that result might be useful for restoration, but they also might complicate considerably the management of impairments. OPADM Fig. 3.8 Larger domain of transparency (OAs not shown). 68 John Strand 2.3. RECONFIGURATION CAPABILITIES 2 3 1 Why Reconfigurability Is Important ... Returning to the example given in Fig. 3.7, for fixed wavelengths, hk, and a fixed configuration of the PADMs, each input port is in effect hard-wired to some distant output port. The connectivity between ports is fixed. In fact, the set of transmitters and receivers that are tuned to a specific frequency can be thought of as a plane that cannot be interconnected within the domain to planes d e h e d by other frequencies. We will see later that this can be a serious problem in some situations. In addition, the lack of reconfigurability can make it difficult to effectively use the capacity of a complex domain of transparency such as that in Fig. 3.8. In this figure, for example, if a specific frequency is in use between D and E, then this frequency cannot be used for the path A-Y-X-U-Z connecting A to Z-it is blocked on the X-Y link. If it were possible to reconfigure this connectivity, the OTrS would in effect be turned into a distributed switch or cross-connect, which could be used for software controlled provisioning or for restoration after some types of failures. These types of functional capabilities are at the heart of the business rationale for deploying the optical network. There are a number of ways in which reconfigurability may be achieved: 0 0 0 0 Laser/receiver tunability. The lasers producing the LR wavelengths (the Ai in Fig. 3.8) may have a fixed frequency, may be tunable over a limited range, or be tunable over the entire range of wavelengths supported by the DWDM. Tunability speeds may also vary. Tunability may give additional connectivity options, and allow ports that could not otherwise be connected to do so. Wavelengthconversion. Internal to a domain of transparency, it might be possible to change the frequency of a connection, thereby in effect interconnecting the planes described earlier. This could be done by converting to the electrical domain and then modulating a laser (fixed frequency or tunable) with the signal. However, this is apt to be quite expensive and may degrade the signal. Conversion in the optical domain has been the subject of considerable research but products with functionality, reliability, performance, and cost adequate to make them attractive are not yet available (see [5] and [13] for surveys). Switchfabrics. A switching fabric could be placed in the PADM in Fig. 3.7 or a cross-connect could be inside a domain of transparency or be placed between two domains. This technology is discussed next. Adaptation grouping adjustments. If the boundaries between groupings are dynamically adjustable, bandwidth could be moved between groupings. 3. Optical Network Architecture Evolution 69 2.3.2. Optical Layer Cross-Connects (OLXC) An OLXC is the optical layer’s equivalent of the Digital Cross-Connect (DCS) discussed in Section 2.1. OLXCs can be placed in a number of places, as the three options in Fig. 3.9 illustrate. The transponders are the demarcation point between the short-reach optical signals, all at the same frequency, that are often used for intraoffice connectivity and the proprietary frequency registered longreach OChs used interoffice. For simplicity, connectivity to routers and other services and TDM equipment is not shown; consider all signals as coming in from the right in the figure and then looping back to the right through one of the cross-connect options. Options B and C really only make sense if they are all-optical, whereas Option A could have either an electrical or optical fabric. Option C (often called a “fiber switch”) is switching the very wide-band proprietary multiwavelength signals produced by WDM multiplexers. This has the advantage that it handles many fewer signals and so needs many fewer ports and a much smaller fabric; however, transmission impairments and technology and vendor incompatibilities impose complex and potentially very restrictive limits on the connections that it can establish; at present, only single-vendor DWDM-DWDM connections can be made, and even with a single vendor, technology differences (e.g., different frequency grids) can prevent other connections. Consequently, Option C does not seem to be getting much attention at present, at least for long-haul applications. Option B (often called a “wavelength selective cross-connect”) also deals with wavelengths that must conform to the proprietary frequency grid and transmission constraints imposed by the DWDM equipment. Consequently, it is not really an option at present, except in the interior of a domain of SR (1 frequency) LR (N frequencies) Multiplexed I U Transponders Fig. 3.9 Optical Layer cross-connect (OLXC) options. 70 John Strand transparency. Where feasible, however, it does offer the promise of substantial cost savings because it does eliminate transponders. If the domain implements adaptation groupings, this option could switch these groupings as an entity, thus reducing the number of ports and fabric cross-points required. Note also that in the absence of all-optical wavelength conversion only channels of the same color may be interconnected. Option A is optically isolated from the DWDM equipment and crossconnects wavelengths with a standard frequency. It thereforeprovidesthe most connectivity.It is insensitive to the optical architectures used by the DWDM vendors, and hence can sit between proprietary “domains of transparency.” It has two disadvantages, however: (1) It requires transponders on each port, these are expensive and also constrain the formats and bit rates that can be cross-connected; and (2) it may not scale as well as the other options because it requires a port per wavelength. Fault detection and localization can be done in Option A by the transponders, which have electrical access to the SONET/SDH overhead bytes. These functions are trickier in the all-optical environments of the other two options. In Option A, an important design choice is whether the transponders are functionally integrated into the DWDM, the OTS, or are stand-alone. Fast reaction to faults is really only possible if there is some level of integration; otherwise it is necessary for an alarm to be sent from the transponder through a time-consumingsequence of softwarelayersbefore restoration can be initiated. With the exception of all-optical network vendors, Option A has received the most attention to date. 2 3 3 Optical Add-Drop Multiplexers (OADM) ... The OADM’srole was mentioned in our discussion of ultra-long-haul OTrSs. The specifics of its rol+how many wavelengths need to be dropped, what constraints (if any) should be placed on which wavelengths can be dropped, what sort of rapid reconfiguration capabilities are needed-are not yet clear, and there are many technological choices to be made. Some of these choices are illustrated in Fig. 3.10.4 2.4. INTELLIGENT OPTICAL NETWORKS Optical Transport Systems have been software intensive since their inception, but by and large this software has not been externally visible. This appears to be changing. The basic component technologies from which optical systems are built are usually availableto all systems integrators, so to differentiatetheir products they are turning to intelligent networking and management software. This software has the potential to allow network operators to reduce their operations costs and also to better customize their servicesfor their end users. Developed by Cedric Lam of A B T Laboratories. TE 3. Optical Network ArchitectureEvolution OADM completely flexible (any collection of wavelengths can be dropped) I 71 I wavelength selective (constraints on wavelengths dropped) I No banding each node can add/dro~ I only one adddrop band allowed per site I multiple adddrop bands allowed per site WavelengthsThatCan Be Dropped PerBand (5number of wavelengths included in a band) I ports not configurable I Ports remotely configurable (requiretunable laser) m rn”*n..r.hlm r d . ! I I Wavelength Reusability (tiwedlaser) I Fig. 3.10 OADM architecture choices. Intelligence is appearing in several areas: 0 0 0 “Soft optics” internal to individual systems to allow things like dynamic power balancing and automatic discovery and reconfiguration of optical subsystems. “Network is the database” functionality that relieves the network operator of the expense of determining network state; instead, the network performs this function and makes the information available to queries. Automatic “point and click” provisioning of new connections using vendor provided software and Graphical User Interface (GUI). These developments are potentially very attractive to network operators and are receiving a lot of attention. We will return to this topic later (Section 5). 2.5. OPTICAL FAULT MANAGEMENT SONET/SDH has very mature and time-tested methods for detecting faults and isolating the source of the problem, primarily based on electronic detection of bit errors using CRCs carried at each level of the frame overhead. A significant concern of network operators has been the possible existence 12 JohnStrand of failure modes that are difficult or perhaps impossible to isolate by optical means alone. One can envision monitoring optical power or signal-to-noise ratio, but these analog measurements will not detect pulse distortion arising from nonlinearity and dispersion, for example. Maeda [121gives as an example an OLXC failure that resulted in the delivery to the proper port of a signal with the correct optical characteristics but incorrect digital content. This is a serious matter, perhaps even a show-stopper, for network operators, particularly as network growth and cost pressures continue to stretch operations resources ever thinner. A number of approaches are being tried to deal with the issue: 0 0 0 Persuasion. Green [47l points out that this issue was encountered when all-optical OAs replaced OEO regenerators that did electronic fault monitoring approximately every 40 miles. Anxious to realize the enormous economic and capacity benefits offered by OAs, operators were persuaded that pump power level and other optical parameters were adequate. Green hopes that history will repeat itself and make looking at the bits unnecessary. SignaZ splitting. A small amount of the optical signal could be diverted and examined electronically. Containment. All-optical subnetworks could be kept small and optically contained, with electronic monitoring of all connections entering/ leaving such a subnetwork. A related step would be to keep the topology of all-optical subnetworks simple-for example, require them to be topological “trees” without loops. (Fig. 3.8 is such a tree.) 2.6. FUNCTIONAL CONSOLIDATION Advances in silicon technology have created the possibility of integrating transport functions that traditionally had been in discrete boxes. For example, in SONET/SDH networks cross-connects and add-drop multiplexers have traditionally been physically separate network elements (NEs), as have DWDM terminals. However, it is now possible to implement a full ADM on a DCS line card. In the intercity network, so-called “optical layer cross-connects” have appeared: They have a large number (thousands) of OC-48/192 ports but an STS-1 fabric and the ability to do all the traditional ADM functions as well as provide an OC-48/192 cross-connect capability. In the metropolitan market, products (often called “multi-service provisioning platforms”) are appearing that carry this trend further, with optical ring functionality and the ability to groom individual DS-1 (1.5 Mb/sec) signals and even ATM switch and IP router capabilities added to the functionality mix in a single NE. The drivers for this trend are compelling: Significantcost, power, and floorspace savings result from the eliminationof the line cards and cabling necessary to interconnect network elements. In addition, maintenance costs tend to be 3. Optical Network Architecture Evolution 73 proportional to the number of NEs, and would therefore benefit from this trend. When functions are integrated in the same NE, layer boundaries get blurred. In the case of the OTN, it is likely that it will be difficult to cleanly separate the management of the OTN from the other transport networks residing in the same NE. If OTN software and operations cannot be cleanly separated from the massive and complex multilayer legacy transport network, the introduction of new OTN technology and functionality will be significantly impeded. 3. Service and Business Trends Changes and trends in the telecommunications services mix and also in the structure of the telecom industry will be critically important determinants of tomorrow's OTN architecture. In this section we will look at a few of the most important trends and their architectural implications. 3.1. SERVICE BASICS The breakdown of bytes of U.S. long-distance traffic by major service grouping is summarized in Fig. 3.1 1. The two bars on the left show the service breakdown at the end of 1997 and 1999, respectively. The bar on the right shows the breakdown of the growth in this period. The predominance of voice traffic even at the end of 1999 may seem surprising to some; much of this is due to differences in utilization, which are discussed later. The Internet segment is clearly growing more rapidly; in [55] it is estimated that average annual growth rates for voice are about 109'0, for the Internet about loo%, and for private line about 30%. Because of this, Internet traffic is expected to exceed voice traffic by early 2002. Note that even in the 1997-1999 period, network growth was driven by the Internet segment. After 1999, Internet growth is expected to become increasingly dominant. From an architectural perspective, this means DataNetworks EOY 1997 EOY 1999 1997-99 Growth Fig. 3.11 Traffic on U.S. long-distance networks, 1997-1999 (from [SI). 74 John Strand that new investmentwill need to be pointed increasingly to the needs of Internet trfic. The revenue perspective is quite different. A leading industry consulting group (RHK) estimates that voice revenue per megabit is seven times that for a T1 Internet connection; for the telecommunications industry as a whole, total voice revenues are projected to be much larger than data revenues for many years. Much of the voice revenue comes from a small number of very large corporate customers who are dependent on very sophisticated and complex software-based functionality in the legacy voice network, and therefore very hard to migrate to a new IP-based infrastructure without the same level of f~nctionality.~ introduces conflicts into the strategic planThis ning process, particularly for carriers with a large embedded base of voice customers, because there is a constant tension between investing in the future and satisfying the current customer base. 3.1.1. Ethernet We are focused on intercity transport, and so Local Area Networks (LANs) are out of scope. However, we should not lose sight of the fact that Ethernet is the dominant protocol in buildings and campuses, and that a large portion of data traffic is carried at least partly on Ethernet. Ethernet is a very rapidly evolving technology.6 It has benefited from enormous volumes and achieved unsurpassedfiber-porteconomies and is steadilyextendingits reach into metro and even long-haul networks, and therefore it bears careful watching as a potential service driver in the long-haul network in its own right, and even a technological competitor for some long-haul applications. 3.2. INTERNET TRAFFIC CHARACTERISTICS Since Internet traffic will be the driver for network growth and architectural change, it is important to understand the nature of this t r f i c . In this section we will look at some of its important characteristics. 3.2.1. Connection Bandwidth The size of the connections between routers will impact a number of aspects of the OTN architecture. For example, the necessity for leaving the optical IBM, for example, has thousands of call center agents averaging 95 sales calls and revenues of $63,000 a minute [107]. Such an operation depends heavily on sophisticated network call-management software to balance loads between call centers, among other functions. This softwm works in coordination with IBM’s internal computer telephony integration product, which provides a customer history screen-popwhen an incoming customer number is transferred to an mailable agent’s phone. See Chapter 11 on Ethernet in this volume. 3. Optical Network Architecture Evolution 75 layer at intermediate points to do TDM multiplexing at sub OC-48 rates will be eliminated if these connections are all at OC-48 or higher rates. The interfaces to voice-service switches normally are (1.5 Mb/s) DS-1 or channelized (45 Mb/s) DS-3. If there is a large community of interest between two such switches, there will be a large number of independent links. This is not the case for large routers, where a singlehigh-speed port normally is much more efficient than a number of lower-speed ports with the same aggregate capacity as the high-speed port. There are a number of reasons for this: (1) There is a significantgain in statisticalmultiplexingefficiencyin the one-port case; (2) the hardware design of routers has normally put a hard upper limit on the total number of interfaces that it can support. The effect of this on the service-providingfacilities7offered to today’s transport network has been dramatic. As recently as the late 1990s, the great bulk of demand growth was for DS-1s. By 2001, the bandwidth-weighted growth in most intercity networks is overwhelminglyfor unchannelized DS-3s and larger facilities; indeed, it is expected that the dominant size of service-providing facilities will become OC-48 (2.5 Gb/sec) and OC-192 as IP traffic starts to predominate. 322 Connection Length ... Connection-lengthdistributions have a subtle but profound effect on transport network architectures. For example, if lengths are short, the market opportunities for ultra-long DWDM systems (see Sections 2.2 and 4.2) and their enabling technologies such as Raman amplification are limited. The volume of voice traffic between two cities is determined by their “community of interest,” the propensity of people in them to want to communicate. This can be roughly modeled using a “gravity model,” which predicts that call volume between two cities is proportional to the product of their sizes (measured in people or total income, for example) divided by the square of the distance between them. This results in relatively short connections-the median length of an intercity DS3 carrying voice traffic, for example, is just a few hundred miles. Internet traffic, and particularly Web traffic, is quite different. Normally one has no idea or interest in the physical location of the server involved in a Web session. Furthermore, the determinants of server location are different, for example, Silicon Valley has an enormous concentration of servers because so many Web-based services were developed there. The net effect of this has been to lengthen average intercity Internet-related facility lengths by an order of magnitude compared to those deployed for voice services. A “service-providingfacility” is a circuit that terminates on a service-providing NE, such as a voice switch or a router. 76 John Strand R Z (a) Toll Switching Hierarchy (b) Internet ISP Hierarchy Fig. 3.12 Hierarchical and nonhierarchical routing. There are a number of other more speculative trends that may affect connection length. One is based on an analogy to the growth of the long-distance voice network. Originally, “toll” voice switches were hierarchically structured into a four-layer “tree.” This was done on a geographic basis, with parent and sibling switches connected together by relatively short trunks (called “final” trunks). This is illustrated in Fig. 3.12a. Calls were passed up the tree until a switch above both caller and called party was reached. In Fig. 3.12a, an A-Z call was routed A-B-C-R-X-Y-Z. However, if there was sufficient calling volume between two lower-level switches, some “high-usage trunks” (B-Y in the figure)would be built between them to avoid the cost and delay of followingthe rigid hierarchy. These trunks tended to be much longer than the final trunks. As this network scaled, the proportion of calls handled on these trunks rose, and the higher-level switches in the hierarchy lost their importance. Eventually they were not needed, and the hierarchical structure was replaced by a flat nonhierarchical arrangement. In this process, connection (trunk) lengths increased substantially. The Internet Service Providers (ISPs) that compose the Internet are today also structured hierarchically into local, regional, and national (nondefault) ISPs (Fig. 3.12b). As the Internet rapidly grows, it would be natural if lowerlevel ISPs would discover opportunities to build the optical equivalent of high-usage trunks that would avoid the expense and delay associated with the current structure. If this occurs, one would expect the net effect to be the further lengthening of Internet facilities.8 323 ... Tkaffic Symmetry A fundamental element of current TDM-based architectures is the symmetry of the connections supported: There is always the same unidirectional The drivers for “direct peering,” as this process is called, are discussed in [59], and its sigmficance is discussed in [58]. Currently a Web fetch requires two to seven round-trips to the server, each of which goes through a large number of routers (17 is the number given in [60]). By reducing the number of ISPs and routers in series, direct peering also has performance and reliability benefits that are encouraging this trend [60]. * 3. Optical Network Architecture Evolution 77 bandwidth dedicated to A to Z traffic as there is no traffic in the reverse direction. This originates in the design of traditional voice circuits, which are 64-kb/sec full-duplex connections. If this were to change, there would be opportunities to build more economical networks if asymmetry were better supported. There is no need for data traffic to be symmetric. A file transfer is unidirectional; on the Internet, the prevalent client-server architecture leads to large amounts of data sent from servers in response to small requests. Various studies have found twice as much traffic flowing from the United States to other countries, and much larger imbalances in flows from ISPs that specialize in supporting servers to those that are residentially oriented (see [48] and [29]). 3.2.4. Utilization By utilization, we mean the proportion of bandwidth that is actually used. Because there is hourly, daily, and seasonal variation in traffic intensity, it is appropriate to look at a time-averaged utilization. Odlyzko [40, 571 estimates that Internet backbones are about one-third as utilized as U.S. long-distance switched voice networks, and that private line data networks are only about 10% as utilized. There are many reasons for this that are discussed in the references; two of these provide opportunities for the OTN architecture to add value: 0 0 If it is possible to add capacity very rapidly, there is little reason for an ISP to keep a spare capacity buffer. However, if additions take a long time, it is prudent to order capacity so it is available well before it is likely to be needed to guard against unexpectedly rapid demand growth or unexpected delays in getting the capacity online. The higher the demand volatility or the more uncertain the capacity delivery process, the earlier a prudent manager will order new capacity. The Internet is noted for wild demand surges, and the time it takes to get an additional large (multi-megabidsecond) connection installed today is frequently measured in months, and there can be significant uncertainties, particularly when multiple operators are involved (e.g., a local telco and a long-distance telco). Hence early capacity ordering, which leads to low utilization on average, is standard operations procedure for many data network managers. Intranets and other business-oriented private data networks have usage concentrated during business hours. Because these networks use dedicated private lines, there is little ability to use the idle bandwidth for other purposes during off hours. Conversely, ISPs catering to residential customers are likely to see their usage higher in the evenings and on weekends. 78 John Strand 3.3. INDUSTRY STRUCTURE The US. telecommunicationsindustry is changing in ways that are profoundly affecting OTN architecture. This section identifies a few of the changes that affect network architecture. 3.3.1. Additional Intercity Optical Networks In the mid-1990s, there were basically three national scale OTNs in the United States, each vertically integrated into one of the major intercity service providers (AT&T, MCI, and Sprint). They accounted for about three-quarters of total intercity fiber deployment. By the end of the decade, however, they accounted for less than a third, with 39 new national carriers accounting for an equal amount [61]. This same source estimates that in 1999 there was a total of 400K route miles in long-haul networks, with an average of 46 fibers per cable. This trend has a number of architectural implications: 0 0 The new OTNs can provide a facility underpinning for nonfacility based ISPs and other service providers. The resulting service competition puts pressure on vertically integrated service providers to keep the unit cost of their OTNs competitive,for example, by introducing new technologies faster than they would otherwise have done. There are many opportunities for buying, selling, or swapping fiber, leading to a situation where competing OTN operators may share fiber in the same right of way or even the same fiber cable. 3.3.2. Additional Local Networks Hundreds of so-called “Competitive Local Exchange Companies” (CLECs) provide competition for the “Baby Bells.” Particularly in areas with a high density of telecom-intensivebusinesses such as lower Manhattan, these companies provide optical connectivity to many key customerlocations. They have more incentive to introducenew architecturesthan the incumbentcarriers. The long-termeconomicviability of many CLECs is hostage to politicdregulatory developments affecting their complex relationships with the “Baby Bells” 3.3.3. Web Hosting and Carrier Hotels Facilitiesto meet the need of carriers, ISPs, and Internet content providers are changing the geographic structure of the industry: 0 “Carrier hotels” are buildings run by third parties that meet the needs of new carriers for a physical presence in some city. They provide a 3. Optical Network Architecture Evolution 79 e telco-grade infrastructure, so all conduit, power, environmental, and security requirements are satisfied; provide an opportunity to share costs; and facilitate interconnection between CLECs, ISPs, and intercity carriers. Web hosting sites provide similar facilities for the enormous number of servers and routers deployed by Internet content providers. These sites provide enormous concentrations of potential business and potentially make it easier for start-up OTNs to find the service volumes they need. The “dot-com” downturn in 2001 has hit a number of them very severely, however. 3.3.4. Bandwidth Trading We mentioned earlier that carriers frequently lease dark fiber from each other. These deals are each separately negotiated by the parties to specify quality of service, any penalties for contract nonfulfillment, and other details, and the physical implementation typically requires engineering and construction to establish the physical connection. This process is time consuming and expensive. This is in marked contrast with the situation in energy markets like gas, oil, and electricity where there are enormous markets to facilitate trading. These markets are based on standardizing the product’s physical characteristics and quality, specifyinga clearing and settlement process for payments, and defining penalties for nonperformance. To facilitate trading, a “benchmark” product delivered at a specified location is used as a basis for establishing a price, and conversion factors are established to establish prices for other grades and physical delivery points. Once this is done, it is possible to establish forward markets that allow buyers and sellers to do financial risk management and also allow speculation. Markets that have been through this ‘ccommodification’y process have been profoundly changed, as anyone following the deregulation of the U.S. electricity industry is aware. Many sophisticatedand very well financed players are trying to commodify the bandwidth market: A Web search for “bandwidth trading” in January 2001 got 95,800 hits. A number of intermediaries have sprung up to help buyers and sellers of bandwidth to trade with each other efficiently. One approach is to act as a “matchmaker.” Many of these companies have Web sites where owners of underutilized capacity can post their offering^.^ A smaller number of players are actually deploying equipment to facilitatethe process. A possible architecture is shown in Fig. 3.13. An offering typically identifies the two end points, the type of circuit (OC-3, STM-1, etc.), and the price and availability date. Offerings can be either “IrrevocableRight to Use” (long-term or permanent) or on some sort of leasing arrangement, e.g., on a monthly basis. 80 John Strand Pooling Point Carrier F DCS CarrierK CLEC or Fig. 3.13 Pooling points for OC-n connections. A “pooling point” with a DCS and/or an OLXC provides the necessary connectivity. It might be located at a carrier hotel or a Web hosting site where colocated routers require a high volume of OC-n connections. It also could be connected by CLEC or ILEC facilities to remote customer locations. Carriers might be invited to connect together pooling points in different cities. The pooling point operator could then establish a trading operation to match buyers with sellers. In Fig. 3.13, a customer wanting an OC-n from A to Z could then use one carrier from A to B and another from B to Z . The bandwidth buyers would hope to gain lower prices through vendor competition. Colocated buyers especially could then hope for more rapid provisioning of their capacity. Among the carriers, start-ups with lots of unused capacity could be expected to gain customers. New ways to use the OTNmight also arise: A large private line network that had very low utilization at night and weekends could try to sell this bandwidth at off-peak periods to someone needing bandwidth to do computer back-ups, for example. Carriers might be able to reduce their sales and marketing expenses significantly. There are a number of issues regarding this: Competition would be increasingly price driven, and there would be pressure to provide a basic standardized product. This could make it difficult to introduce technologies and products differentiated by reliability, security, or customer service. The vertical integration of network with services currently prevalent in the telecom industry would come under pressure and new business models might be needed. Many of the criteria associated with successful commodity markets are not present. A study by the Boston Consulting Group [loll identified six critical criteria: (1) Vertical deconstruction. Vertically integrated industries are poor candidates. (2) Fragmented supplier base. Suppliers with market power can refuse to participate. 3. Optical Network Architecture Evolution 81 (3) Fragmented customer base. If there are few buyers, they can be targeted by suppliers, eliminating the need for a market. (4) Price volatility. Predictable prices reduce the need to hedge risk and reduce the incentives of market makers. (5) Common unit ofexchange. Efficient trading environments require common units of exchange and settlement contracts to fuel liquidity. (6) Delivery mechanism. A physical delivery mechanism is required to efficiently and quickly move commodities between buyers and sellers. This study concluded that in most of these areas, the bandwidth market was not yet ripe for commodification but might be in a few years. They did see immediate opportunities in a few specific areas, especially for private lines on high-volume, capacity constrained routes. In summary, bandwidth trading is not yet a significant factor for 0 7 3 s . However, it does appear that there are short-term opportunities and incentives for companies running carrier hotels and server farms, and also for some start-up carriers, to move in this direction. In the longer term, the prospects appear rosier.lo There are also significant architectural implications: Bandwidth trading would make rapid provisioning an essential network capability, would make it harder for both equipment vendors and carriers to establish proprietary product and service improvements not incorporated in the definition of the standard traded product, and would make network interworking much more important. 3.4. OPTICAL NETWORK SERVICES 3.4.1. Services Overview It should be clear from the discussion in Section 3.1, that OTN services in the future will be targeted primarily at meeting the needs of public and private data services, and particularly IP-based services. We will generically refer to the providers of these services as “ISPs.” From an architectural perspective, an OTN architect must make a choice: (1) Build a network with premium features such as very fast restoration in the hope that it will command a premium price and profit margin; (2) build a network offering only the functionality required for “commodity” bandwidth and hope to get higher volumes and efficiencies; or (3) build a network that lo The Web is the best place to do further research on bandwidth trading. Some market participants whose Web sites might be of interest are Band-X, Interxion, AIG, and Arbinet (all facilitiesbased) and Ratexchange, Bandwidth Market, and Bandwidth.com(nonfacilitiesbased). 82 Johnstrand can offer both commodity and premium services. We expect that there will be network operators opting for each of these options. A fruitful way to start thinking about the service/architectureinterrelationship is to identify ISP needs, identify the functionalitiesthat might possibly be provided by the network, and then match functionalitiesto needs. 3.4.2. ISP Needs ISP needs axe as follows: 0 0 0 0 Price. The most important need is undoubtedly low price. Availability. ISPs are struggling to keep up with exponential growth. Therefore, the availability of bandwidth is a key need-whether it can be provided at all between the desired locations, and if so, how quickly. ISP cost displacement. Displacement of internal ISP costs is also desirable. By this we mean functionality provided by the OTN that will allow the ISP to reduce their internal costs. Some potential areas for this are: a. Reduce the cost of physical interfaces on routers. b. Provide bandwidth at the speeds optimal for the router. c. Assume some of the costs of reliability and failure recovery. d. Assume the responsibility for providing exactly the capacity required when it is required. e. Allow flexible peering with other ISPs. f. Provide network management capabilities that allow the ISP to be aware of the state of their connections at all time and to reconfigure them as appropriate. Additional revenue opportunities. Provision of capabilities that enable new services. For example, a highly reliable Virtual Private Network (VPN) offering conceivably might be based on an optical layer restoration capability. Coverage of special events might be facilitated by rapid OTN provisioning of extra bandwidth. 3.4.3. Possible Functionality Areas Possible areas of functionality include the following: 0 0 Additional connection oflerings. Higher bit rates (e.g., OC-768); different formats (Ethernet, Digital Wrapper); more flexible bandwidth configurations, such as asymmetric or unidirectional connections; flexible concatenation of standard SDHBONET connectionsto provide additional bandwidth options; and inverse multiplexing. More rupidprovisioning. Software control of optical cross-connects and other reconfigurableONES.User-Network Interface (UNI) allowing 3. Optical Network Architecture Evolution 83 direct signaling by a router or customer controller requesting immediate additional connections. Larger networkfootprint. More ubiquitous connectivity can be provided by interworking between networks and by providing OTN access from more locations. Additional restoration options. A range of restoration speeds and threat coverage; assistance in dealing with service layer failures such as router or service outages. 3.4.4. Relating Needs and Possible Functionality Table 3.3 attempts to identify possible relationships between needs and functionality. The justification for many of the entries can be found in the subsections of Section 3.4. 3.4.5. A Carrier Perspective on Service Functionality The Carrier Working Group within the Optical Interworking Forum (OIF) recently produced an “Optical Services Framework and Associated Requirements” document [43] that provides a good snapshot of current services thinking in the carrier community as it relates to optical networking and software control. What follows is excerpted from this document.’* 3.4.6. Value Statement Optical networkingpermits camers to provide new types of network services not available with other technologies, enabling sophisticated transport applications of (D)WDM based networks (featuring a variety of topologies such as pointto-point, ring and mesh). These new generation networks provide means for the improved use of network resources and the support of high-bandwidth services. Dynamic bandwidth allocation, fast restoration techniques and flow-through provisioning give birth to an assortment of services. Intelligent OTNs contain distributedmanagement capabilityand subsume many provisioning and data basing functions currently performed by carrier Operations Systems (OS). This allows the rapid establishment and reconfiguration of connections, potentially reducing provisioning times from months to seconds, thus lowering operating costs and providing the means to set and guarantee SLAs12 and QoS configured on a per-connection basis to better meet customer’s specificneeds. I I The current author was the chair of the OIF group producing this report. It is a working text and not an official OIF Technical Report, and is not binding on the OIF or its members. l2 SLA Service Level Agreement. Defines the details of the service to be provided, particularly its availability and reliability. Table 3.3 Relations Between Needs and Potential OTN Functionality Additional Connection Offerings Price 0 Rapid Provisioning Reduced network operations expense Larger Footprint Lower internetwork coordination costs Additional Restoration Options Asymmetric connections Concatenation (e.g., OC-15) 0 Availability Shortened provisioning interval More optically reachable locations 0 Faster multinetwork provisioning 0 cost displacement Revenue opportunities Less expensive router line cards 0 Higher utilization 0 Higher utilization Higher utilization Higher availability 0 Lower MTBF 0 Lower delay Reconfigure for special events 0 Add temporary busy hour bandwidth 0 3. Optical Network Architecture Evolution The large capacity and great flexibility of such networks enables the support of several degrees of transparency to user traffic at lower cost to the end customer. The new services expectedto be enabled as aminimumare bandwidth on demand, point and click provisioning of optical connections, and optical virtual private networks. The standardized interface between the optical layer and the higher layer data service layers such as IP, ATM, SONET/SDH enables the end-to-end internetworking of the optical channels for conveying user information of varying formats. The use of standardized protocols will make the benefits of the intelligent OTNs available end-to-end, even if several networks are involved. 85 This document also defined business models for three types of service offerings they hoped to see supported. These were: Provisioned Bandwidth Service: Enhanced leaseaprivate line services. Provisioning is done at the customer request by the network operator. . . . This is basically the “point and click” type of service currently proposed by many vendors.. . . Billing will be based on the bandwidth, restoration and diversity provided, service duration, quality of service, and other characteristics of the connection.. . . No customer visibility into the interior of the OTN is required; however, information on the health of provisioned connection and other technical aspects of the connection may in some circumstances be provided to the user network as a part of the service agreement. . .may involve multiple networks, e.g., both access networks and an intercity network. In this case provisioning may be initiated by whichever network has primary service responsibility. . . Bandwidth-&-Demand Service: OC-n/STM-n and other facility connections are established and reconfigured in real time. Signalingbetween the user NE and the optical layer control plane initiates all necessary network activities. A real-time commitment for a future connection may also be established.A standard set of “branded” service options is available. . . . Optical Ertual Private Network The customer contracts for specific network resources (capacity between OLXCs, OLXC ports, OLXC switching resources) and is able to control these resowces to establish, disconnect, and reconfigure opticalconnection connections.In effect they would have a dedicatedoptical subnetwork under their control.. . . Billing will be based on the network resources contracted. Network connection acceptancewould involve only a check to ensure that the request is in conformance with capacities and constraints specified in the OWN service agreement.. .. Real-time information about the state of all resources contracted for would be made availableto the customer. Depending on the service agreement,this may includeinformationon both in-effectconnections and spare resources accessible to the customer. 86 Johnstrand 4. Optical Network Architectures 4.1. INTRODUCTION 4.1.1. Unit Costs Architecture is basically about achieving a proper balance between costs and capabilities.A common mistake is to equate “cost” with equipment cost. This is far from the truth. For a typical traditional long-distance voice service, for example, a typical cost breakdown is as follows: Access (paid to the local exchange carrier) Network-related costs Customer care, billing, miscellaneous 35% 15% 50% The network-related costs typically divide roughly equally between the actual carrying costs associated with the equipment and the expenses associated with running the network. It is not unusual for the hardware cost of the NEs to be 10% or less of the total costs that need to be recovered. Thus it is crucial to always consider the non-hardware-cost implications of a l architecture l decisions. The revenue implications are also crucial. 4.2. TRANSPARENCY In an optical network, “transparency” refers to whether, or to what degree, an optical signal passes through the network optically. In today’s (2001) socalled “optical” networks, there are actuallymany OEO conversionsand a high degree of reliance on electronicprocessing. However,the vision of many optical networking researchersincludes a much larger role for all-opticalfunctionality. In this section we will explore this issue, the outcome of which will have a large role in shaping the use of optical technologyin future OTN architectures. 4.2.1. m e s of Transparency There are many shades of transparency. A categorizationused in the MONET project13was: Digital transparency. Transparency to intensity-modulated digital signals of arbitrary bit rate, frame format, and protocol. Amplitude transparency. Transparency to intensity-modulated digital or analog signals. Strict transparency. Transparency to any optical signal. l3 MONET w a s an ARPA-sponsored project established to define and demonstrate how best achieve multiwavelength optical networking of national scale. See www.bell-Iabs.com/ project/MONET or www.darpa.mil. to 3. Optical Network Architecture Evolution 87 Even if there is optoelectronic regeneration, there are a number of levels of transparency possible [11: 0 a 0 Regeneration with retiming and reshaping (3R). This involves acquiring the clock of the regenerated signal, and thus makes it quite difficult to handle multiple frame formats. Regeneration with reshaping but without retiming (2R). This offers bit rate and format transparency, but allows jitter to accumulate, thus limiting the number of regenerations that can be done. Regeneration without retiming or reshaping (IR). This has the worst performance but can handle a wide variety of signals, both analog and digital. We will use the word transparency to mean no optoelectronic conversions of any kind. 4.2.2. Potential Advantages of Transparency The potential advantages of transparency fall into two major categories: a 0 Format independence. A transparent network is largely indifferent to the details of the signal being transported so long as the power levels and other optical characteristics are within bounds. This has the major advantage of allowing new protocols (and also legacy protocols such as PDH) to be easily handled.I4 Without this independence, potentially each new format or bit rate requires standards changes and hardware and software to be developed and deployed. Less expense at intermediate nodes. A transparent network does not require expensive OEO functionality at intermediate nodes; electronics is bypassed by through wavelengths. 4.2.3. Limitations on Transparency [30,45,46] Unfortunately, transparency also has some serious problems: Impairments accumulate. Various forms of dispersion, nonlinearities, polarization-dependent loss, multipath interference, misalignment of lasers and WDM filters all degrade optical signals, and many of them pose increasingly serious performance limitations as bit rates increases channel spacing gets tighter, and the number of channels increases. Individually, these limitations constrain the distance a wavelength can l4 A thought-provoking example of this is quantum cryptography, an unbreakable form of cryptography that exploits the uncertainty principle of quantum theory, but requires the polarization of individual photons to be preserved. It has been demonstrated over 23 km of fiber [62]. 88 John Strand e e travel without regeneration and/or lead to a relentless tightening of component and fiber requirements. Furthermore they may combine to provide unexpectedly severe impairments. (See also [64].) All-optical interoperability is problematic. This problem has several dimensions. First, performance monitoring and fault location is problematic. As discussed in Section 2.5, our optical monitoring ability is limited. In a large network, particularly one involving multiple vendors or network operators, it is essential that faults be quickly detectable and the source of the problem be quickly identifiable. Second, it is unclear how to introduce new technology, such as an upgrade from 100 to 50 GHz wavelength spacing, incrementally. It appears that the technical specifications of an all-optical network must be fixed at the time it is first deployed if costly and difficult in-service upgrades of equipment are to be avoided. Wavelength interconnection is restricted. As discussed in Section 2.3, our ability to do wavelength translation in the optical domain is inadequate at present. As we shall see later in this section, this can make it difficult to use some of the capacity of the system. 4.2.4. Opaque Optical Networks An opaque OTN is one where each cross-connect and each OTrS is optically isolated by transponders. In its simplest form, an opaque network is formed by adding transponders at the interfaces to the OTrS shown in Fig. 3.5 (see Fig. 3.14). Figure 3.15 shows a cross-connect in an opaque network. Typically the optical signalsbetween the transponders and the OLXC would be short-reach or intermediate-reach, depending on the loss characteristics of the cross-connect, and would all be operating at the same frequency. Note that the OLXC could have either an electrical or optical fabric. Frequency Registered LRh Transponders Frequency Registered LRh Standard SRh 4---;--+ : 4--+-- + Standard SRh El El Receivers Span Mux Trai Fig. 3.14 Optical Transport System (OTrS) with transponders. 3. Optical Network Architecture Evolution 89 0 Transponders Fig. 3.15 Opaque node showing transponders and cross-connect. The strengths and weaknesses of opaque and transparent OTNs are mirror images of each other. An opaque OTN is very limited in the formats and bit rates of the signals it can carry, and it incurs significant costs for OEO functionality for each wavelength at each node. On the other hand, in an opaque network, impairments do not accumulate, interoperability is guaranteed, and wavelength translation is obtained as a by-product. All the large OTNs known to the author have found the case for opacity compelling to date. 4.2.5. Domains of Transparency The choice between opacity and transparency is not really black and white. The concept of a “domain of transparency”-a transparent subnetwork, optically isolated from the rest of the network by transponders-provides a means to control the drawbacks of transparency discussed above. This technique offers us the possibility of limiting the size of each domain, thereby keeping impairment-related problems in check. New technologies might be put in separate domains, thereby avoiding technology interworking problems. Organizational boundaries, such as those between network operators, can be aligned with domain boundaries. A DWDM system such as that shown in Fig. 3.14 is an example of a domain of transparency. If transponders are put on all the q, in the ultra-long OTrS shown in Figs. 3.7 or 3.8, a more interesting domain would be defined. In effect, this is what vendors of such systems are proposing. The technological trends enabling longer wavelengths and all-optical reconfigurability should make ever larger and more complex domains of transparency feasible. However, there are costs associated with introducing multiple domains of transparency. On each boundary, costly transponders must be installed. In addition, as we shall see later when we discuss control planes, additional complexity can be added to the processes that route and manage wavelengths. 90 John Strand 4.2.6. Economics of Transparency The economic attractiveness of a domain of transparency arises largely from the opportunity to reduce transponder (OEO) costs. These per-wavelength costs can be the largest single cost in an OTN, as is shown in Fig. 3.16. The cost breakdown is for an OTrS such as that in Fig. 3.14 that is bounded by transponders. The costs are based on typical vendor prices in 2000; they could vary considerably based on the vendor’s pricing strategy and the specific size and capabilities of the OTrS. The transponder costs are linear in the number of working wavelengths (utilization), but independent of the number of spans, whereas the other costs are independent of the utilization and linear in the number of spans (except for the MudDemux), hence the relationships shown in Fig. 3.16. In an opaque OTN, transponders need to be placed on the ports of each OTrS. OTrS lengths are limited by (1) technology (the maximum number of spans before regeneration is needed), and also (2) by the need for wavelengths to be added or dropped from the OTrS. To get some insight into the economic trade-offs involved in transparency, we will give an example related to (1). Consider the example given in Fig. 3.17. Figure 3.17a shows a sequence of standard OTrS and an ultra-long-haul OTrS. Both are assumed to have the same OA spacing and the same wavelength capacity. The standard OTrS we assume to be limited to a five-span configuration before transponders are required for regeneration; in (a) this happens at offices B and C. The ULH system merely needs an OA at these locations. To go further, the ULH has presumably been designed using additional costly technology (see Section 2.2). We model this parametrically by use of n I I Transponder Optical Amplifier h Utilization (%) # Spans (80 km) 50 100 3 3 50 7 100 7 Fig. 3.16 Cost breakdown, OTrS bounded by transponders. 3. Optical Network Architecture Evolution 0 91 Transponder 100 Y ULHlStandard Cost Ratio OTS Type Standard b (a) Reference Systems ._ C 0 ULH 1 3 5 7 9 No. Of Standard OTS Systems (5 span) In Series (b) Domains Of Application Fig. 3.17 Ultra-long-haul economics. a cost ratio (assumed the same for DWDMs and for OAs). The additional ULH costs are system costs, which are independent of the number of wavelengths that are actually equipped. The penalty for opaqueness incurred by the five-span configuration is partially per-system (the additional back-toback DWDMs required) but primarily per-wavelength (the transponders at intermediate nodes like B and C in Fig. 3.17). Therefore, the ULH system would be expected to become more competitive as the number of systems increases (more DWDMs for the competing solution) and also as the utilization increases (more transponders). This is quantified in Fig. 3.17b, which shows for various cost ratios15 the frontier between the regions where each alternative has an economic advantage. The ULH system is economically preferred above and to the right of the appropriate curve. For example, if the ULH system is 75% more expensive, the arrows indicate that ULH is preferred for fills above about 45% if 5 systems (25 spans) are needed. If the topology of the subnetwork is more complex, as in Fig. 3.8, an alloptical solution has the additional advantage that transponders are not needed at the branch points. Evaluation of this benefit is more complex and will not be discussed further here. Unfortunately, we are not aware of any efforts to systematically look at this effect. In metro areas, the economic issues are quite different. Normally OEO is not needed for transmission reasons, because the distances involved are relatively short. Instead, the ability to carry a wide variety of bit rates and formats becomes more important. 4.2.7. Incorporating Domains of Transparency in a Network Domains of transparency will be limited in size by transmission constraints, the inability until standards evolve significantly to have optical interworking l5 Representative year 2000 list costs for the "standard system" were used in this exercise. 92 John Strand Fig. 3.18 Express domain of transparency example. between vendors, and by economics. In a long-haul network, a likely initial use of a domain of transparency would be to provide an express backbone on which longer connections would be routed. Shorter connections would be routed on an opaque technology that is more economic over shorter distances. This situation is illustrated in Fig. 3.18. In this mesh network, two technologies are used, one all-optical (shown by dashed lines with squares showing offices with DWDMs or OADMs), the other one an opaque technology with electrical fabric OLXCs (circles) and with solid lines showing connectivity. On the left, the two technologies are shown with the topologies separate, and on the right, they are laid onto the conduit network used for both. As just discussed, the all-optical technology is most cost-effective for longer distances. In this case, if we wished to route a connection from A to Z, the best route might be to go from A to J using the opaque technology (top plane in Fig. 3.18a), then go J-N-P-Q-L using the all-optical technology, and complete the route from L to Z using the opaque technology. This example suggests that introducing such a domain of transparency will raise new issues for routing and probably for other operational areas. We consider routing next. 4.2.8. Routing and Wavelength Assignment in a Domain of Transparency Within an all-optical domain, “wavelength conversion” (changing the wavelength of a connection) is still expensive and not yet practical without an OEO conversion. Therefore, it is important to understand the architectural implications of limited (or no) wavelength conversion. This requires us to look at what is called the “Routing and Wavelength Assignment (RWA) Problem” [l]: Given one or more connections that need to be established in an all-optical domain, determine the routes over which each connection should be routed and also assign each connection a color. If the routes are already known, the problem is called the “Wavelength Assignment (WA) Problem.” The RWA problem has received extensive attention in the literature, mostly from a mathematical perspective. This literature is best approached through 3. Optical Network Architecture Evolution 93 a survey article (e.g., [l, 2, 14, 741). The underlying mathematical problem is very hard in general. The WA problem is easily seen to be equivalent to the problem of coloring the nodes of a graph so that no two nodes connected by an arc of the graph have the same color: Simply represent each connection by a node, and connect every pair of nodes whose corresponding connections ride on the same link (and so need to be assigned different wavelengths). This coloring problem is known to be NP-complete [75], which means that, in general, it is computationally intractable, but there are many fast but approximate (heuristic) algorithms for solving it. Before discussing some of the architectural implications of this problem, it should be emphasized that the results in the literature frequently depend crucially on subtle points of the problem definition. The followingdefinitional choices are particularly important: 0 0 0 Is there an accurate forecast of future connection requirements? If so the RWA process can take a more global approach to decision making. What is the expected holding time of a connection? Some analyses assume that connections are permanent, while others assume that holding times are quite short. Is it possible to rearrange connections?In situations where unexpected demands are likely, this makes a major difference. We feel that the most appropriate assumption at the present time is to assume that (1) we have minimal knowledge about future demands, therefore each demand must be routed and assigned a wavelength when it appears without knowledge of future demands, and (2) that holding times are very long. With these assumptions, we simulated a number of RWA algorithms on a realistic model of the U.S. intercity network [76] without rearrangement or wavelength conversion and reached the following conclusions: 0 0 0 It is important to choose the route and do the wavelength assignment simultaneously.If routing is done first and then a wavelength is assigned, significantly suboptimal results are obtained (e.g., low utilization). Even with simultaneous route and wavelength selection, a significant capacity penalty can be incurred if wavelength conversion is not available. This penalty can be significantly mitigated if wavelength conversion is available at a small subset of nodes. In particular, half of the lack-of-wavelength-conversionpenalty is removed if 21% of the nodes have conversion capabilities; 75% was removed if 35% of the nodes have conversion capabilities. 94 e John Strand Dual-OTrS scenarios, which provide two instances of each wavelength on each link, reduced this penalty by only 6%. A recent survey of the value of wavelength conversion [74] concluded that the performance improvementsit offered depended on a number of factors, including the network topology and size. It further concluded that in networks with tunable transmitters and receivers, limited wavelength conversion provides improvements close to that achieved with ideal wavelength conversion. 4.2.9. Effect of Impairments on Routing in a Domain of Transparency [19,95] As domains of transparency get larger, optical impairments such as amplifier spontaneous emission (ASE) and various types of dispersion may become an issue. Specifically: e Polarization Mode Dispersion (PMD). PMD imposes a limit on the maximum wavelength length that is inversely proportional to the square of the bit rate of the signal. For typical installed fibers, the limits are 400 km and 25 km for bit rates of 10 Gb/s and 40 Gb/s, respectively. With newer fibers assuming PMD of 0.1 ps/,/km, the limits are 10,000km and 625 km, respectively. e AmpIiJierSpontaneous Emission (ASE). ASE imposes a constraint on the number of spans that is inversely proportional to its optical bandwidth. e other polarization-dependent impairments. For example, many components have polarization-dependentloss (PDL) [11 that accumulates in a system with many components on the transmission path. The state of polarization fluctuates with time, and it is generally required to maintain the total PDL on the path to be within some acceptable limit. a Nonlinear impairments. As wavelengths get longer, nonlinear impairments such as four-wave mixing cause more problems at a given launch power. These constraints are summarized in Fig. 3.19. The specific constraints required in a given situation will depend on the design and engineering of the domain of transparency. For example: e e The effect of nonlinear impairments depends on complex factors, e.g., on the order in which specific fiber types are traversed, the specific types of fiber involved, and the characteristics of the other active wavelengths (see, e.g., [46] or [64]). The impact of chromatic dispersion may depend on whether it has been dealt with on a per-link basis, and whether the domain is operating in a linear or nonlinear regime. 3. Optical Network Architecture Evolution \ 95 PMD Constraint Bit Rate PMD Parameter Optical Bandwidth Minimum acceptable SNR at receiver Launch Power (PL) ** L6ther System Parameters Length Of All-Optical Path Fig. 3.19 Effect of impairments on wavelength muting. 4.2.10. Connectivity Limitations Associated with a Domain of Transparency The strawman ultra-long-haul system shown in Fig. 3.7 may be thought of as a large distributed switch. Depending on the configuration of the OADMs and the tunable lasersh-eceivers, the port-to-port connectivity of the a's will change. However, the connectivity is limited: 0 0 0 The adaptation function forces groups of input channels to be delivered together to the same distant adaptation function. Only adaptation functions whose laserdreceivers are tunable to compatible frequencies can be connected. The switching capability of the OADMs may also be constrained. For example: a. There may be some wavelengths that cannot be dropped at all. b. There may be a fixed relationship between the wavelength dropped and the physical port on the OADM to which it is dropped. c. OADM physical design may put an upper bound on the number of adaptation groupings dropped at any single OADM. For a fixed configuration of the OADMs and adaptation functions, connectivity will be fixed: Each input port will essentially be hard-wired to some specific distant port. However, this connectivity can be changed by changing the configurations of the OADMs and adaptation functions. For example, an additional adaptation grouping might be dropped at an OADM or a tunable laser retuned. In each case, the port-to-port connectivityis changed. This capability can be expected to be under software control. Today the control would rest in the vendor-supplied Element Management System (EMS), which in turn would be controlled by the operator's 0%. 96 John Strand 4.3. LAYERING AND THE OPTICAL LAYER The term “optical layer” is frequently used in architecture discussions. The term “layer” is taken from a modeling methodology that must be understood if one wants to read the technical literature on this subject or work on network architecture problems. This section gives a very brief overview of the methodology and then applies it to optical networks. The methodology described in this section was developed by the International Telecommunication Union (ITU), the international governmentsanctioned international standards organization. Reference [65] defines the basic approach, and [66] and [67] apply the approach to SDH and OTNs, respectively. 4.3.1. Layering Concept A transport network may be vertically decomposed into a number of layers that are related by a client-server relationship as shown in Fig. 3.20. Only two layers are shown. The upper layer is the client: It requests services from the lower layer, where a “service” is defined by its protocol, bandwidth, service quality, and perhaps other functionality. The lower server layer, in turn, provides capacity and the other aspects of the desired service. A couple of caveats about layering: 0 0 There is no single way to define layers. For example, layered views of the Internet normally collapse all the transport layers into one, whereas transport-oriented layerings normally collapse the multiple protocol levels describing the various Internet protocols into a single layer but show many transport layers. There is no necessary relationship between layers and hardware. A single box may perform the functions of several layers or a single layer may be implemented in a set of boxes. Well-Defined Interface Protocol Service Quality Functionality - Requested Layer N Capacity Fig. 3.20 Layering concept. 3. Optical Network Architecture Evolution 97 4.3.2. Transport Layering The layers in today’s transport network are shown in Fig. 3.21 and Table 3.4. The relationship between these layers is illustrated in Fig. 3.21. 4.3.3. Layers and Planes To function, a transport network needs a signaling and control infrastructure. These infrastructures are also layered, as illustrated in Fig. 3.22. The shaded stack labeled “User Plane” corresponds to the layers we have been discussing. Two additional “planes” are shown in the figure: Control plane. Responsible for activities such as connection set-up or tear down and restoration. Operates in a dynamic “real-time” environment that directly responds to user activities or network state changes such as failures or maintenance activities. Increasingly its functionality is distributed and implemented in software resident on each ONE or directly connected to it. It has its own infrastructure for signaling to other systems within the network or to users or other networks. It is important to note that there may be separate control planes for each layer, as shown in Fig. 3.22. Management plane. Gives the network operator visibility into and control of the other two planes. Normally implemented in centralized OSs with information exchange carried over operations communications channels. This results in a static control environment characterized by scheduled activities and delayed response to network state changes. DSl(1SMbls) DS3 (45 Mbk) STS-3 (155 Mbk) STS-12 (622 Mb/S) STS-48~ Gb/S)(2.5 PTP 1 n?- fin P . .L I\ *,*-,7LL (‘VU”,>, igital Transmissio Proprietary (20 Gbk -400+ GbA) Fig. 3.21 Transport layering. Table 3.4 Transport Layers Layer Services Sublayer Voice services Data services Service Requests Originatingat Layer calls Data transfer Private line DSO Private line DS 1 Private line DS3/STS3/STS12 Switch access lines, private lines Switch access lines, private lines Private line STS-48/STS-192 Dark fiber Typical Demand DSO Packets DSOs DSls DS3s Typical Nodes Circuit switch (4E) Packet, frame, cell switches DCS-1/0 W-DCS B-DCS Typical Links DSO “trunks” (Modulo DS1) DS0/1/3, STS-N(c) DS 1 DS3, STS-I, STS-3 DS3, STS-N OC-12,OC-48, OC- 192 OC-3,OG12,OC-48, OC-192 Optical transport systems (OTrS) Fiber pairs Digital Cross-Connect DCS 110 WS) Wideband DCS (W-DCS) Broadband DCS (B-DCS) Self-healing rings Linear add/drop or point-to point systems Digital Transmission DS3, STS-N Add-drop multiplexer DS3, STS-N Terminal multiplexer, add-drop multiplexer STS-48c, Optical ADM, optical STS-1 9 2 ~ cross-connect DWDM OTS Optical Media See text Fiber 3. Optical Network Architecture Evolution N Layers 3 . *‘< : 5 99 < < 2 1 Physical Medium (fiber) Fig. 3 2 Abstract layered network showing control and management planes. .2 4.3.4. Sublayers of the Optical Layer The sublayers of SONET and SDH were described earlier (Section 3.1). The ITU is in the process of defining a very similar structure for what they call the “Optical Transport Network” (OTN). Starting from the top, the ITU’s layers are: e Optical Channel (OCh) Layer. Provides end-to-end networking of optical channels for transparently conveying information of varying format such as SONET or G-Ethernet. In non-ITU circles, an OCh may be called by other names, of which “lightpath” seems to be the most popular. Today OCh’s typically carry STS-48 or STS-192 signals, with Gigabit- or 10-Gigabit Ethernet also receiving considerable attention along with STS-768. OLXCs provide cross-connect functionality for this sublayer. e Optical Multiplex Section (OMS) Layer. Multiplexes OChs into the multiwavelength optical signal that is passed between DWDMs. It also is responsible for wavelength shifting and management. A fiber switch (see Section 2.3) might provide cross-connect functionality for this sublayer. e Optical TransmissionSection (OTS) Layer. Provides basic transport functions for an OMS such as amplification and gain equalization. Overlaying this structure on Fig. 3.1 one gets the system shown in Fig. 3.23. In SONET and SDH, at the terminations of the equivalents of each OCH, OMS, and OTS, the signal is electricallyaccessible and so it is straightforward to add overhead bytes for management and signaling. This is not the case for the OMS and OTS, therefore the ITU is proposing an out-of-band optical “OTM Overhead Signal” (00s).In [69] the logical elements to be included are specified but the format is not. Since today OTSs are invariably singlevendor, the handling of OTS and OMS overhead is vendor proprietary. There is, however, some interest from carriers, particularly in Europe, for all-optical 100 John Strand interworking between vendors. This may ultimately make standardizing of the 00s of more interest. 4.3.5. Optical Channel Format Issues Optical channels (OCh) are bounded by 3R regeneration, and therefore are electrically accessible. This makes in-band overheads possible. ITU’s draft G.709 standard [69] has defined a frame structure to do this based on earlier proposals for a “digital wrapper” [3]. It encapsulates the payload (e.g., a SONET frame) within a larger frame containing fields reserved for framing, inter-ONE communications, and mechanisms for performance monitoring and propagating error information both forward and backward along the channel. Protection switching has been left for further study. A specific forward-error-correction(FEC) algorithm, a Reed-Solomon RS(255,239) code, is also specified. The proposed standard has a number of attractive features: 0 0 0 0 It is explicitly targeted at optical transport networking. It starts to provide the basis on which wavelengths or OChs can be effectively managed. This makes more realistic the elimination of service-specificequipment and processing inside the OTN, and thus brings all-optical networking closer to practicality. It opens the door to eventually phasing out SONETISDH functionality and eliminating a layer from the protocol stack. Unlike SONETISDH, it is not optimized for traditional voice services and arguably is less inhibiting for packet-over-fiber applications such as Gigabit Ethernet and 10-Gigabit Ethernet over fiber. However, whether this standard will eventually be widely adopted is unclear at this point: 0 A major driver for the approach taken is the assumption that there will be a wide variety of client data networks with varying protocols and 3. Optical Network Architecture Evolution 101 0 0 0 service needs [3]. The recommended approach is designed to allow a wide variety of client protocols (SONETEDH, ATM, IP, PDH, etc.) to be transported independent of their format. There are, however, many who argue that IP will become the common traffic convergence layer for all services. If this does indeed happen, the importance of flexibly handling a wide variety of protocols will decrease and a less general, IP-centric solution may emerge. FEC needs have not yet stabilized. For example, ultra-long-haul vendors are using strong FEC algorithms that require more FEC information than allowed in the standard. The timeliness of trying to standardize FEC now is therefore an issue.I6 At present, optical interworking between vendors is limited to intraoffice short-reach connections covered by SDH/SONET standards. In long-haul networks, vendors would be optically isolated from each other by transponders. In this architecture, FEC originates and is terminated by a single vendor, and therefore proprietary solutions are feasible and allow vendors to seek competitive advantage with innovative FEC solutions. SONET/SDH can be used to encapsulate a wide range of protocols including IP, ATM, and Ethernet, thereby preserving the large investment in SONET/SDH transponders and 0%. This is not to say that the proposed G.709 standards will not be widely adopted eventually; however, their future seems most promising if multiple data protocols flourish, FEC technology stabilizes, practical multivendor all-optical standards emerge, and rapid growth (which would allow network operators to justify writing off their investment in SONET/SDH transponders and line cards) continues. In the meantime, a vendor wanting some of the promised functionality might wish to use the proposed format within a single domain of transparency. This would avoid most of the objections listed previously, because at the boundary of the domain, an encapsulation of SONET/SDH into the new format could be done without necessarily affecting the larger network or service level interfaces. 4.4. EVOLVING ROLE OF THE OPTICAL LAYER 4.4.1. SONET/SDH and WDM One of the most important functions performed by SONET/SDH has traditionally been TDM-based intermediate multiplexing. This function is possible l6 G.709 is sensitive to this issue. There is an appendix that discusses alternatives to the recommended approach. 102 John Strand only if the SONET line rate is significantly larger than the data rate of the underlying service. For traditional 64-kb/sec voice services riding on 1.5-Mb/sec DS-ls, this is certainly t r u e - a STS-48/STM-16 can carry 1344 DS-1s. However, this is not true for large IP backbone networks, where routers are increasingly connected with STS-48 or STS-192connections. The situation is illustrated in Fig. 3.24, which shows the maximum availableTDM rate (solid line), single-fiber WDM rate (dotted line), and IP router fabric capacities. The gap between the line speed of IP routers and the maximum TDM-line rate represents the opportunity for TDM intermediate multiplexing. It has basically vanished and the prospects for it catching up again are uncertain at best. Because data trailicis driving network architectureevolution, this implies that SONET/SDH intermediate multiplexing capabilities will be of declining importance. Because of this change, the SONET networking layers (the cross-connect and digital transmission layers in Table 3.4) cannot perform their traditional role for much of the bandwidth growth expected to be generated by IP-based services. Instead, their functions need to be distributed between the IP services layer and the optical layer, as illustrated in Fig. 3.25. This redistribution will very likely have major impacts on equipment vendors as well as network operators. If the functionality largely goes into the IP layer, then vendors of IP equipment will be well positioned to control transport evolution and also to charge premium prices; if it largely goes into the optical layer, then the OTN vendors will be in the driver’s seat. The rest of this chapter largely is devoted to spinning out the implications of this change. There are three major functionality groupings that must be studied: 100 L 8 lo ..TDM Multiplexing 0pp-v 1 01 . Fig. 3.24 The declining opportunity for TDM multiplexing. 3. Optical Network Architecture Evolution 103 DS 1 (1 5 Mb/s) STS-12(622 Mbis) F D i g i t a lTransmissiox I I A -. Restoration Netwnrkinn STS-192c (10 Gb/s) I rl ‘ x Propnetary (2OGb/s--400+ G Fig. 3.25 Redistribution of SONET functionality for IP services. 0 0 0 Physical Layer Functions. These include framing and fault detection and isolation. Restoration. Recovering from a failure. Networking. Provisioning and network management. 4.4.2. Future of SONET Physical Layer Functions Even if the traditional SONET networking functions are no longer needed, it does not follow that SONET technology will not continue to be used for framing, fault detection, and localization, and all the other operations functions currently based on SONET. IP routers produce SONET-formatted signals, for example, and there is no reason why they could not continue to do so. The key arguments against this are: 0 0 0 SONET overhead imposes a “bandwidth tax” of approximately 4%, which is no longer justifiable. SONET chip sets are much more expensive than competitors for other technologies like Ethernet. The elaborate operational processes built over the years around SONET are no longer needed if its networking function disappears. The force of these arguments will vary by network and application. In metro area networks, for example, Ethernet is already a strong competitor for 104 John Strand SONET for the transport of data between LANs. At the other extreme, an established carrier with a large investment in SONET equipment and SONET-based operational processes is unlikely to be tempted to migrate legacy TDM-based services to a non-SONET infrastructure. The “bandwidth tax” argument is not very compelling. The 4% overhead is significantly less than the overhead associated with higher-level protocols, for example, TCP/IP alone has a 15% overhead on an average-sized packet.17 A number of alternatives to SONET framing have been proposed: 0 0 0 There have been a number of proposals to create a “SONET Lite” by stripping out all SONET networking functionality. However, several vendors reported to the Optical Interworking Forum that doing so would only save a few dollars per OC-48 chip set. Ethernet components benefit from enormous economies of scale because of their ubiquitous use in LANs. As mentioned earlier, this makes them a formidable competitor, particularly for LAN interconnects over relatively short distances. At present, Ethernet does not have the operational robustness that established intercity network operators require; however, it would be surprising if one of the start-up carriers did not try to differentiate themselves with a low-cost Ethernet-based network. In the optical layer, the overhead proposed for the OCh (discussed earlier) could be a solid foundation for a SONET replacement, particularly within a domain of transparency (as discussed earlier). Discussions of the alternative protocol stacks that could be used for IP in an OTN are discussed in [3] and [4]. The protocol stack alternatives are summarized in Fig. 3.26. In this figure, “Optical Layer” refers to OLXC-resident functionality for managing OCh provisioning and for restoration.18Additionally, it may include the OCh framing functionality discussed previously. (It is discussed in more detail later in this chapter.) “ATM” refers to Asynchronous Transfer Mode, a cell (fixed-sized packet) technology used frequently today in large IP networks to allow better traffic engineering and network management than is currently available in IP. Multi-Protocol Label Switching (MPLS) provides ATM-like capabilities within the IP framework. Each of the layers normally has its own management layer. In Fig. 3.26 “IP/MPLS” indicates that either IP alone or IP with MPLS might be used. l7 Mean packet size measured in [48], for example, is roughly 300 bytes including headers; IPv4 and TCP per-packet overheads are at least 20 bytes each. [4] puts the OXC in the “Physical Layer” and limits the “Optical Layer” here to what we will describe later as the “optical control plane.”The definitionused here is more compatible with the terminology used elsewhere in this chapter. ’* 3. Optical Network Architecture Evolution 105 IP ATM (b) IPMPLS (C) IPMPLS Optical Layer (d) SONET Optical Layer SONET Optical Layer IPMPLS 0 0 0 0 Stuck (a) is common today. It has the most framing overhead and the most management layers. Stuck (a) is often called “Packet over SONET.” It eliminates one level and the associated overheads and management systems. It also is commonly used today. (See [3,4] and [70].) Stuck (c) is what is often called “IP over WDM.” It further slims down the protocol stack and associated management systems. To be successful in large networks, the fault management and maintenance capabilities of one or both of the remaining layers need to be beefed up. Stack (d) is the simplest. It might be implemented by integrating a WDM terminal directly into an IP router. All the residual SONET/SDH fault management and operations capabilities will need to be assumed by IP/MPLS. 4.4.3. Architectural Role of Optical Layer Switching An office configured with and without an optical layer switch is shown in Fig. 3.27. Note that from an equipment cost perspective, the hard-wired office will be less expensive because it has one less network element. However, in a hard-wired office, each time a wavelength is added to the office, manual cabling must be installed between the two appropriate OTrS (through wavelength) or between an OTrS and the appropriate router or other service layer equipment (terminating wavelength). This is expensive, error-prone, and slow.l 9 It also is difficult to do this before a specific service request defining which ONESneed to be connected is received. If there is a switch in the office, as long as care is l 9 It is especially problematic for the increasing number of “dark offices”-locations that are normally unstaffed. For these offices a special visit must be scheduled if hard-wiring is required. 106 John Strand n Service Layers (a) Hard-Wired Office (b) Office With Switch Fig. 3.27 Office configurations with and without a switch. IP Router (a) OLXC Office Architecture (b) “BFR Office Architecture Fig. 3.28 Optical Layer switching alternatives-OLXC and “big fat router.” taken to ensure that a spare connection is always available from each ONE to the switch, none of these difficulties arise-the software-controlled switch can establish the desired connectivity on demand. Arguments along these lines seem to have led most carriers to conclude that the option in Fig. 3.27b is desirable in larger offices at least. In Section 2 we introduced optical layer cross-connects (OLXCs). These are certainly a candidate for use here. The principal competition for an OLXC-based architecture is likely to come from “big fat routers” (BFR; [80, 8 11). In this architecture, IP routers with line-side interfaces operating at the per-wavelength bit rate are directly connected to one another via pointto-point OTrS. The OTrS terminals might even be integrated into the IP router. Legacy services requiring constant bit-rate connections are supported with virtual leased-line services [84]. The two competing architectures are shown in Fig. 3.28. The BFR approach is most appealing if IP-based traffic continues its dramatic growth and dwarfs other types of traffic. In this scenario, it is argued that there is no need to interpose another layer between the OTrS and the IP routers. It is further argued that eliminating a layer of switching will simplify network management and will reduce capital costs by eliminating some “boxes.” 3. Optical Network Architecture Evolution Ports 8,Assumed Costs OLXC $x d IP Router $y 107 E 4 Terminating Through (a) OLXC Office Architecture (b) "BFR Office Architecture Fig. 3.29 OLXC and BFR configurations. The counterarguments in favor of the OLXC-based architecture are, briefly, (1) non-IP-based services are expected to be very important to carriers for some time (see Section 3.1); (2) it is not clear whether router technology will be able to scale to port counts consistent with multi-Tb/s capacities without unduly compromisingperformance, reliability,restoration speed, and software stability [80, 821; and (3) the OLXC architecture will have lower capital costs than the BFR architecture, even for IP traffic. We will return to the management issues when we discuss the control plane (Section 5). Here we look only at the capital cost comparison. The prices of fully loaded large routers and OLXCs are both dominated by line-card costs. We will therefore ignore other costs. For simplicity we will also assume all traffic is IP. Figure 3.29 identifies the cost elements in each architecture for terminating and for through connections. If each OLXC port is priced at $x and each IP router port at $y, and the proportion of connections that are through is a, then the average costs per connection are: OLXC: a(2x)+ (1 - a)(2x +y ) BFR: 4 2 ~ ) (1 - a)b) + Solving, the OLXC architecture is cheaper if x < ay.'O A typical value of a in a long-haul network would be 0.8, so OLXC line cards must be 20% cheaper than IP line cards to be competitive. Is this likely? Based on representative year 2000 prices from a number of vendors, the answer is definitely yes. The x / y ratio for electrical-fabricOLXCs was comfortably less than 0.5; [ll] quotes a value (1999) of 0.22. The ratio is expected to become more favorable to OLXCs when transparent OLXCs appear. A technology comparison makes this relationship quite intuitive: In a transparent OLXC there is virtually no line card functionality needed, while IP router architects have achieved scalability by in effect making each line card *O A more comprehensive theoretical analysis looking at various network topologies may be found in [83]. The conclusions reached are consistent with ours. 108 John Strand a mini-router. We conclude that an OLXC is the most economic way of doing optical layer (OC-48/OC-192) cross-connections. 4.5. SUR WABILITI: PROTECTION,AND RESTORATION 4.5.1. Background Survivability-the ability to deal with and recover from failures-is a critical architectural consideration. In Fig. 3.25 each of the layers shown, except the Media Layer, have survivability capabilities Unless carefully designed, this functional overlap introduces additional costs and coordination issues. The protection and restoration functions provided by SONET/SDH are critical to the survivabilityof today’s TDM-based transport networks. Before discussing how these functions relate to the IP and the Optical Layers, we will take a digression to introduce the general topic of survivability. There are many types of failures that can occur in an optical network, but two are of particular importance: equipment failures and route failures. Equipment failures, primarily failures of individual components, such as a laser, are far more numerous than route failures. In “carrier class” equipment, normally there are no single points of failure; if a component such as a laser fails, automatic protection switching (APS,discussed later) is used to restore service. Although much less frequent, route failures (also called fiber cuts) can have a much more devastating effect. The prototypical cause of a route failure is an errant backhoe that physically severs the fibers; other causes might be a natural disaster (flood, earthquake) or a train wreck. Normally all fibers in a cable will be affected. By electronic standards, these failures occur in slow motion. Initially fiber stretching might cause a period of transient errors; the actual severing of fibers can be spread out over a period of seconds. This can cause confusion and churning if a recovery process able to react in milliseconds locks in on a restoration plan before the full magnitude of the failure is known. The incidence of route failures varies widely. Rural areas have fewer problems, as do cables run through concrete conduits. As a rough rule of thumb, one fiber cut per thousand route kilometers per year might be used; thus a national-scale network in the United Statesmight expect, on average, a couple of incidents per month.21 Physical repair after a route failure requires dispatching a repair crew to the site, which may be quite distant, and then physically splicing the ruptured In 1997, when the total intercity fiber in the United States was 59,000 route miles, 136 fiber cuts were reported by various US.carriers to the Federal CommunicationsCommission [6, 711. Only cuts with serious service impacts are required to be reported, therefore the total number of cuts is no doubt somewhat higher. 3. Optical Network Architecture Evolution 109 fibers. The nominal time for doing this is on the order of 4 hours, but it may be much longer if there is a massive failure, for example, a building fire or an extensive wash-out. For many services, however, the overall availability requirements are on the order of 99.999% (6 minutedyear) or better. Therefore, additional survivability mechanisms must be provided. There are many possible mechanisms, some of which are appropriate for service restoration and some of which need to be resident in a transport layer. Because of their importance in transport networks, we will focus on recovery from route failures in this section. Optical Layer mechanisms will be discussed first, then mechanisms appropriate for other layers will be introduced and related to the optical layer. 4.5.2. Protection and Restoration Protection and restoration are both used in the process of recovering from a failure. A quick review of a few published articles found a number of dehitions, of which the most popular were as follows: 0 0 0 0 Protection refers to the primary recovery, whereas restoration refers to secondary methods that are invoked if the primary method fails. Protection refers to methods whose operation is specified prior to a failure; restoration refers to methods that dynamically determine how to proceed. Protection reroutes onto preassigned spares; restoration makes use of a pool of spare resources that are dynamically assigned after a failure. Protection refers to recovery from a component or subsystem failure by switching to a hot standby that is usually performed in a distributed way by the ONES.Recovery that requires rerouting facilities over a physically different route is called “restoration.” Protection is triggered by “defects”-the physical layer’s detection of a problem, whereas restoration is triggered by ‘Lalarmsyy-messages generated by ONESto be sent to external managementlcontrol functions. In general, “protectionyy historically referred to deterministic methods has that are very fast, whereas “restoration” referred to more complex methods whose precise behavior is determined at the time of failure, and which may be somewhat slower than protection. However, IP-inspired changes in the way transport networks are controlled are rapidly eroding this difference. In this section we will use “protection” to refer to preplanned methods using preassigned capacity and “restoration” for other methods. We will use the term “recovery” when we want to be generic. 110 John Strand 4.5.3. Recovery Basics Any method for recovering from a fiber failure requires four fundamental capabilities: (1) Fault detection mechanism to know when there is a problem and what connections have failed; (2) Spare capacity to replace the failed capacity; (3) Switchfabric to give the failed connections access to the spare capacity; and (4) Control logic to trigger the recovery and reroute the failed connections on the spare capacity. In addition, many recovery methods also require afault localization mechanism to determine what links should be avoided in the rerouting process. Consider Fig. 3.30. In this figure, a connection routed X-A-B-Y fails because of the failure of link A-B. Fault detection could either be done at the link level at A or B or it could be done at the ends of the connection (X or Y). To restore the connection, spare capacity on some alternative path bypassing the failure (A-C-D-B in the figure) is needed. To reestablish the connection, A and B need switch fabrics so the new X-A-C-D-B-Y path can be established. Where this switching capability needs to be depends on the location of the spare capacity. Finally, there needs to be a control capability. This could be located in the NEs at A and B, at the ends of the connection (X and Y), or in some network management system connected to A and B by reliable data links. A key criterion for evaluating recovery methods is the amount of spare capacity required. Network topology puts a lower bound on this, as shown in Fig. 3.31. In the figure, the “degree” of a node is the number of links incident on it. In each case shown, we assume that the link from A to Z fails, resulting in a need to reroute the A to Z traffic. In the degree 2 case there is only one possible restoration path, which is to go clockwise from A to Z. In this case, to restore all the A to Z traffic there must be a 100% overbuild, that is, as much spare capacity on each link as there is A to Z traffic to restore. In the Fig. 3.30 Basic recovery capabilities. 3. Optical Network Architecture Evolution 111 Z Degree 2 Nodes 100% Overbuild Degree 3 Nodes 50% Overbuild Degree N Nodes l/(N-1) Overbuild Fig. 3.31 Effect of topology on restoration capacity. second (degree 3) case, there are two potential restoration paths from A to Z. Therefore, if there is a 50% overbuild on each restoration path, all the A-Z traffic can be restored. In general, in a network with all degree N nodes a 1/(N - 1) overbuild is required (last figure). 4.5.4. Taxonomy of Recovery Methods Recovery can be done in many different ways, many of them proprietary to either a vendor or a network operator. It also can be done at different levelsservice, SONET, or optical. When confronted with a new method, it is useful to determine how it provides the four basic capabilities listed previously: (1) Fault detection. Is the fault localized before the recovery process starts or not? Localizing the failure can take time and be complex, but it allows the process to bypass only the failed elements, and thus may require less spare capacity. Without fault detection, the connection must be rerouted onto a diverse path end-to-end. (2) Spare capacity. 0 Is the capacity used for restoration dedicated to individual connections or is it shared? If each connection to be restored has dedicated restoration capacity, the restoration process can be very fast and the control logic can be quite simple; however, much more restoration capacity is likely to be needed. 0 Is the spare capacity diverse from the working capacity? If so, protection against a wider range of failures is provided; however, this comes at the cost of having to plan and coordinate across a wider geographical area involving more network elements. ( 3 ) Switchfabric. Are connections restored individually (this is normally called “path” restoration) or are connections restored in bulk (“link” restoration). Also of interest is where the switching is done. 112 John Strand (4) Control logic. 0 0 0 Is the network partitioned in multiple recovery domains with recovery controlled independently in each domain or is control global? Recovery time and control process complexity usually both increase at least linearly with network size, which suggests partitioning. Balancing this is the problem of restoring “seam” failures (failures of the links connecting two domains). In addition, if the capacity available for recovery is partitioned so spare in domain A is not available to domain B, additional spare may be needed. Within each domain, is the control process distributed down to the network element level or is it centralized? A centralized process is a potential bottleneck that may make “scaling”22dficult, and it also is a potential single point of failure. On the other hand, it is extremely difficult to test and debug a distributed process. Also, in-service software bugs in distributed systems may be harder to deal with. There have been several spectacular, long-lasting network-wide failures in such systems in recent years. Is the recovery process preplanned for each specific possible failure or is it determined dynamically at the time of failure? Preplanning may speed up the recovery process significantly. However, each time the network state is changed by the addition or removal of a connection, or the failure or placing in a maintenance state of some of the recovery capacity pool, it may be necessary to replan. Also, it is difficult, if not impossible, to have a contingency plan for each possible combination of failure events. 4.5.5. SONETISDH Recovery Methods Optical Layer recovery methods at this point are mostly proprietary and largely derived fairly directly from the standard SONET/SDH methods. We will therefore first discuss the most important SONET layer methods, and then discuss Optical Layer differences and issues. SONET techniques that aren’t relevant to the Optical Layer are not covered. More detail on these methods may be found in [18], [79], and [99]. Table 3.5 summarizes the key SONET recovery mechanisms in terms of the capabilities described in the previous section. 1 : 1, M :N , and 1 + 1 automatic protection switching configurations are normally used to handle line card, link, and laser failures. They are especially “Scaling” refers to the efficiency of a control process as the network size increases. Table 3.5 SONET Recovery Methods Fault Localization Needed M:Nand l+lAPS Path-switched rings Line-switched rings Mesh Dedicated Recovery Capacity Yes Possible Typical Location of Switching Function Ubiquitous ADM line card ADM DCS Diverse Recovery Path or Capacity Link No Multipre Recovery Domains Yes Yes Yes Possible Control within Preplanned Each Domain or Dynamic Distributed Distributed Distributed Centralized or distributed Preplanned Prcplanncd Preplanned Either No No Path Path Link Path Yes Yes Yes Yes No No Possible 114 John Strand Switch I Switch (a) 1 : 1 APS SWI (b) M: N APS (c) 1+1 APS Fig. 3.32 Automatic protection switching options. useful for protecting intraoffice connections. When used in interoffice communications, the service and protection capacity is normally transported together in the same cable. These methods are illustrated in Fig. 3.32. In each case, the first number indicates the number of protection channels, and the second number indicates the number of service (working) channels. Figure 3.32b is a generalization of Fig. 3.32a where there are M protection channels for N service channels ( M < N ) . It allows any M of the service channels to be restored simultaneously. 1 :N has been used much more extensively than configurations with M > 1. In Fig. 3.32c, two copies of the signal are sent and a selector at the receiver chooses the better. SONET 1 : 1 and 1 : N APS use in-band signaling (the K1/K2 overhead bytes in the SONET overhead) for coordination. 1 1 APS does not require any signaling, which is a simplification. The price paid for this is the constant use of the protection channel; in the other architectures it is possible to put preemptible connection^^^ on this channel. These could be used for low-priority services or they could be part of the restoration capacity pool for a back-up restoration scheme to be used if APS alone was inadequate. The 1 : 1, 1 :N , and 1 1 configurations have been standardized for both SDH and SONET. To the best of the author’s knowledge, M : N with M > 1 has not been standardized. The path-switched ring configuration is called “SDH subnetwork connection protection” by the ITU [79]. It differs from the 1 : 1 and 1 1configurations just described in two respects: (1) The protection capacity is normally routed diversely from the service connection, and (2) it operates at the path (endto-end) layer. Figure 3.33 gives an example of the most common form, a “Unidirectional Path-Switched Ring” (UPSR). Figure 3.33a shows the underlying network structure: four nodes connected by OC-48s. Each OC-48 is represented by a pair of shaded lines. These represent the two unidirectional transmission paths of the OC-48, with the outside one going clockwise and the inner one counterclockwise. Figure 3.33b shows a lower speed (OC-3 or OC-12) UPSR between nodes 1 and 3 that is routed on these OC-48s. Both the + + + 23 Called “extra traffic” in SONET. 3. Optical Network Architecture Evolution 115 (a) 4 Node OC-48 Ring (b) UPSR - Normal State (c) UPSR - Failure State Fig. 3.33 Path-switched ring (OC-48-based UPSR example). Fig. 3.34 Generic path-switched ring. 1-3 direction and the 3-1 direction are routed clockwise, on the outer transmission path. Thus the two directions are routed separately. At nodes 1 and 3 there is path termination equipment for the UPSR. Figure 3 . 3 3 ~ shows that if the service path from 4 to 1 fails, the 3-1 direction of the UPSR is rerouted onto the counterclockwise loop, thus restoring the connection. The switching is done at nodes 1 and 3, with no involvement by the other nodes. Only the failed direction (3-1) is rerouted. There could be multiple UPSRs between different nodes riding on the same set of OC-48s. More general connecting networks are possible, as indicated in Fig. 3.34. Here the SONET network “cloud” could be composed of a mixture of SONET rings and linear structures. The protection switching is done at the terminations of the path. A line-switched ring configuration was discussed in conjunction with Fig. 3.2 in Section 2.1. All of the recovery mechanisms described so far require that the network be divided into recovery domains with either ring or linear topologies4egree 2 topologies in Fig. 3.31. Planning and operating a network so divided is complex and frequently expensive, as we shall see shortly. Mesh recovery works in an arbitrary topology (degreeN in Fig. 3.31) without having to deal with the planning and operational complexities of a multi-recovery-domain environment. 116 John Strand Unfortunately, this flexibility makes it virtually impossible to standardize a mesh restoration method to the level that has been done with SONET rings. As a result, all mesh implementations are proprietary. One mesh implementation that has been described in the public literature is AT&T’s FAST Automated Restoration (FASTAR@)[20]. Originallydesigned for a pre-SONET network, it has been extended for SONET. We will briefly describe how it works to give a flavor of mesh restoration. FASTAR’s physical infrastructure consists of Restoration Network Controllers (RNCs)in each office; a centralized restoration controller [theRestoration and Provisioning Integrated Design (RAPID) system]; and a redundant data network to connect them together, and also to connect RAPID to the DCSs, which do the actual rerouting. When a failure occurs, each RNC collects alarms from the affected NEs in its office. After waiting for a brief period to allow all alarms to come in, and also to see if the problem is a transient, it then sends the alarms to RAPID. The RAPID database contains information on all spare capacity available for restoration.24When the first alarm arrives, RAPID opens up an “event window.” While the window is open,25alarms are collected. After the window is closed, RAPID: 0 0 0 Determines what spare capacity has survived the failure. Orders all the connections reported failed according to their restoration priority. Processes each connection in turn. For each, an alternate path with spare capacity is computed, the spare capacity is committed, and commands are sent to the appropriate DCSs to implement the reroute. FASTAR’s restoration speed is gated by the rate at which the DCS can do reroutes.26 As indicated in Table 3.5, there are many dimensionsin which mesh methods may vary: 0 Recovery actions may be determined dynamically after the failure (like FASTAR) or can be preplanned. A number of levels of preplanning are 2 FASTAR depends critically on the timeliness and accuracy of this database. Keeping a 4 centralized view of a network composed of many types of equipment and distributed around hundreds of nodes and links sufficiently accurate is probably the most difficult challenge a centralized-mesh-restoration developer faces. 25 It is closed when either a specified period elapses with no additional alarms or when a time specifyingthe maximum window period expires. 26 For a U.S. government description of FASTAR, including some restoration times, see h t t p : / l ~ . n c s g o v / n 5 _ h p / I n f o r m a t i o n A s s u r a S e chtm. 3. 3. Optical Network Architecture Evolution 117 0 0 possible. For example, the reroute paths may be preplanned but the specific capacity to be used along the path for a specific connection may be determined dynamically. Control may be centralized (like FASTAR) or may be distributed to the NEs. Rerouting may be done end-to-end (which is allowed by FASTAR) or constrained to stay near the site of the failure. Many of these alternatives are being vigorously pursued for Optical Layer restoration. We will discuss distributed mesh algorithms in Section 5. 4.5.6. Ring/Mesh Comparison ITU and SONET standards specify that restoration should be completed within 50 ms in most cases,27and 200 ms is an achievable upper bound for ring restoration. There are new mesh restoration schemes proposed for the optical layer, and also for IP-based networks, that may be competitive, but these are still mostly theoretical results. Centralized mesh algorithms are likely to take a minute or more to recover from a large failure. Rings are required to have uniform capacity on each of their links. This can cause difficulties, as is illustrated in Fig. 3.35. Figure 3.35a shows a topological structure common in intercity networks-a set of paralleling north-south and east-west fiber routes. If there is a roughly equal amount of demand on each link, then the same amount of capacity is needed in the upper (e.g., A-B-C-D) and the lower (C-D-E-F) “window panes.” But this leads to double capacity on the middle route. Figure 3.35b shows another example of the complexities introduced by the equal capacity constraint. Here, each dashed line shows the routing of some connections, each of which requires more than 50% of the capacity of a ring. Demands (25 DS3 Each): A-E F-L H-G H-L I I I I (a) Window Pane Network Shortest Distance Routing Optimized Routing (b) Effect Of Routing On Ring Efficiency Fig. 3.35 Ring examples. If a ring is more than 1200km in circumference, or if it is carrying extra traffic, an extra 100ms or so may be required. *’ 118 John Strand The two networks differ only in how the demands are routed. If each demand is routed on the shortest path (left), the lower loop (A-F-G-H-J-K) requires two rings because of the demand between G and H. The top two loops require a total of three rings because of the demands between E and F. Thus a total of five rings are needed. If the demands are routed as a group with an eye to minimizing the number of rings required (right), only three rings are required. To accomplish this, the demands A-E, H-L, and G-H have been rerouted onto circuitous paths. Routing algorithms sophisticated enough to achieve the optimized routing need either to have an accurate view of future pointto-point demands or the ability to continually do massive rearrangements to preserve optimality as new demands appear, neither of which is practical at this time. A number of studies of the relative economics of ring versus mesh have been done. In [15] it is stated that mesh is “20% to 60% more efficient” than rings. Studies at AT&T of representative national topologies and demand sets have yielded larger differentials-at least 50% more capacity required even if routing is optimized (as in the right routing in Fig. 3.35b), over 100% if it is not. RHK has estimated that ring restoration is 80% more expensive than mesh for the optical layer. A comparison of optical ring versus optical mesh for a very large future IP network also showed significant backbone savings for mesh over rings [105]. 4.5.7. Virtual Rings Each SONET line-switched ring has a dedicated NE (an ADM) at each of its nodes. This is acceptable as long as the number of rings in an office is small. The unprecedented growth rates experienced in recent years, however, has raised the specter of offices with hundreds or even thousands of ADMs. Consider the situation shown in Fig. 3.36. Figure 3.36b illustrates the situation at office A using ADMs. All connections wishing to switch rings or terminate in office A must traverse cabling from the ADMs [the small squares in (b)] to the DCS (diamond in the center), which becomes a bottleneck. Furthermore, each (a) Grid Network (b) ADM’s & DCS At Office A (c) DCS With Virtual Rings Fig. 3.36 Virtual rings. 3. Optical Network Architecture Evolution 119 ADM is physically cabled on the network side to specific physical facilities; recabling would be needed to change this. Figure 3 . 3 6 ~ illustrates what is called a “virtual ring.” Here, a single DCS switch fabric replaces all the ADM fabrics. Any two DCS ports may be associated together through the fabric to form the functional equivalent of an ADM. Ring interconnects also are made through this fabric and the line-switched ring-protection algorithm is implemented in the DCS control software. This has major advantages: (1) Intraoffice connections between rings no longer require intraoffice cabling to and from the DCS; (2) any two high-speed ports on the DCS can be made into a logical ADM via a software command; and (3) it is now physically possible for two virtual rings to share protection capacity. Proprietary virtual ring implementations have been made by a number of DCS and OLXC vendors. 4.5.8. Diversity For a ring to work properly, it is necessary that the links forming it are physically diverse from each other, otherwise a single failure might cut it in two places and cause a total failure. To determine whether two lightpath routings are diverse, it is necessary to identify single points of failure in the interoffice plant. To do so, we will use the following terms: AJiber cabZeis a uniform group of fibers contained in a sheath. An OTrS will occupy fibers in a sequence of fiber cables. Each fiber cable will be placed in a sequence of conduits-buried honeycomb structures through which fiber cables may be pulled-or buried in a right ofway (ROW). It is worth noting that for economic reasons, ROWs are frequently obtained from railroads, pipeline companies, or thruways. Often several carriers may lease ROWs from the same source; this makes it common to have a number of carriers’ fiber cables in close proximity to each other. Similarly, in a metropolitan network, several carriers might lease duct space in the same RBOC conduit. As discussed in Section 3.3, carriers also trade capacity. In each of these situations, there is the opportunity for a single event to disrupt more than one carrier’s facilities simultaneously.Thus, relying on carriers to be diverse from each other without detailed evaluation of the physical routings is chancy. In a typical U.S. intercity facility network there might be on the order of lo2 major offices. To accurately capture diversity, however, a network with an order of magnitude more nodes is required. In addition to Optical Amplifier (OA) sites, these additional nodes include: 0 0 0 Places where fiber cables enterneave a conduit or right of way; Locations where fiber cables cross; Locations where fiber splices are used to interchange fibers between fiber cables. 120 B-A John Strand (a) Fiber Cable Topology (b) Right-Of-WayKonduit Topology Fig. 3.37 Fiber cable vs ROW topologies. An example of the first is shown in Fig. 3.37. Here, the A-B fiber cable would be physically routed A-X-Y-B, and the C-D cable would be physically routed C-X-Y-D. This topology might arise because of some physical bottleneck: X-Y might be the Lincoln Tunnel connecting Manhattan and New Jersey, for example, or the Bay Bridge. The imminent deployment of ultra-long Optical Transport Systems introduces a further complexity: Two OTrSs could interact a number of times. For example a New York-Atlanta OTrS and a Philadelphia-Orlando OTrS might ride on the same ROW for x miles in Maryland and then again for y miles in Georgia. They might also cross at Raleigh or some other intermediate node without sharing right of way. 4.5.9. Unique Features of Optical Recovery In most respects, Optical Layer recovery options and considerations are the same as those encountered in SONETBDH. This should not be surprising because both are multiplexed, connection-oriented networks. There are, however, a few sources of difference, which we now discuss. 4.5.9.1. Opaque Optical Networks In an opaque OTN, and also when protection is provided for the connectivity between domains of transparency, there is no real difference between the optical situation and that encountered in traditional SONET/SDH networks. 4.5.9.2. Domains of Transparency Recovery options within a domain of transparency are limited in several ways: 0 Transmission constraints may eliminate some recovery options. For example, in Fig. 3.30 the recovery path (X-A-C-D-B-Y) could be significantly longer than the initial path (X-A-B-Y), so it might have unacceptable noise or other impairments. 3. Optical Network Architecture Evolution 0 121 If wavelength translation options are limited, the spare capacity that can be used to restore a particular connection will be limited. In the extreme, if there is no wavelength conversion capability at all, then the pool of recovery capacity fragments into separate pools for each frequency. These limitations are much less of a problem in metro networks or other situations where path lengths are limited. 4.5.9.3. Flexible Groupings SONET line- and path-switched rings normally switch the equivalent of optical channels (e.g., Figs. 3.2 and 3.33). Optical switching has the property that it is much less sensitive to the composition of the optical signal being switched; switching technologies such as MEMS could switch an aggregate multi-OCh signal as easily as they can switch a single OCh. This provides additional implementation options for recovery mechanisms. For example, PADMs can do the equivalent of SONET line switching in two ways, as shown in Fig. 3.38. Figure 3.38 shows a two-fiber PADM. If there is a failure just to the right of it, switching at the OMS level is represented by the loopback labeled “1”; at the OCh level, by “2.” Both options have the effect of looping the entire signal back from the service (S) to the protection (P) fiber. If the ring is implemented in a cross-connect rather than a PADM, Fig. 3.9 shows the equivalent options: OMS switching would be done by a fiber crossconnect (Option C in the figure), whereas OCh switching would be done by the switch labeled “Option B.” If the wavelength bundles called “adaptation groupings” in Section 2.2 (see Fig. 3.7) are being used, the flexibility of optical switching would allow protection switching to be done at this level also. 4.5.9.4. OMS Level Protection Switching This option would have a number of potentially attractive features: (1) Many all-optical switching technologies, such as MEMS, could switch an aggregate multi-OCh signal as easily as they can switch a single OCh. This should lead to P S 4 * * . P .. S Fig. 3.38 Photonic ADM protection switching options. 122 John Strand economic savings. (2) Since there are fewer OMS than OCh, OMS restoration may scale better than OCh restoration. (3) Wavelength conversion is not an issue. However, OMS restoration can only be done within a domain of transparency. Thus all the issues associated with transparency discussed in earlier sections,28particularly transmission constraints, multivendor interoperability, and problems associated with the introduction of new technologies arise here. Therefore, most attention is currently focused on OCh-level restoration. However, OMS restoration can be effectivein situations where these issues are not compelling, particularly in metro and access networks and when integrated into a single-vendorall-opticalnetworking product such as the ultra-long-haul systems discussed in Section 2.2. OMS protection switching does not protect against the failure of an individual signal before they are wavelength multiplexed. To gain this capability, it must be combined with an OCh-level mechanism, such as 1 :N protection switching, which is discussed next. 4.5.9.5. 1 : N A P S Protection Switching Providing equipment protection for wavelength-specificcomponents such as a transponder that is transmitting into a domain of transparency can be done in a number of ways. The setting is shown in Fig. 3.39. As shown in Fig. 3.39, an incoming signal (from the left) is fed to a working transponder (with output wavelength hl in the example) and also to a protection transponder (at bottom). Not shown are up to N - 1other incoming signals that are sharing the same protection transponder. Two of the implementation approaches described in [151 are: Transponders Fig. 3.39 1 :N APS protection setting. 28 Sections 2.3 and 4.2 in particular. 3. Optical Network Architecture Evolution 123 Fixed spare. In this case, Aspare is some other fixed wavelength AN+^. At the Demux point there would need to be an appropriate receiver. Upon failure of a working transponder, the incoming signal is routed by the 1 x N switch to the protection transponder. In addition, to coordinate the ends a fast signaling protocol like that used for SONET, APS would be needed. If the domain of transparency in the center is a large mesh, this protocol could get very complex. In addition, a wavelength is dedicated for protection purposes only. Tunable spare. In this case, the spare transponder is built around a tunable laser. Upon failure this laser would be tuned to the frequency of the failed transponder. In this case, neither the dedicated protection wavelength nor the extra receiver is needed, and no coordination is needed, as nothing in the network has to change when the spare laser is tuned to the failed frequency. The trade-off is the requirement for a tunable laser, which at present would be expensive and present other difficulties. 0 4.5.10. Multilayer Considerations Recovery mechanisms are available in the service layers, especially for services built on IP. Whenever there are multiple recovery mechanisms that might be invoked by the same failure, coordination of some sort is needed. If this is not done, problems of the sort illustrated by the following time-line are likely (Fig. 3.40). The Optical Layer is at the bottom, the Service Layer above, and time runs from left to right. A link failure is shown at the left, which is remedied through optical protection in a few milliseconds; however, before this was accomplished an alarm was detected by the Service Layer. This triggers many service layer reroutes, a process taking 10s of seconds in a large IP network. These reroutes could also cause user-visible service glitches. Rediscovery of the link is accomplished in IP by another protocol, and could take some time. IP-Based Service Layer Optical Layer . . . . . . . . 10s seconds, 10s seconds , Fig. 3.40 Counterproductive inter-layer recovery behavior. 124 John Strand When this finally happens, another period of service instability will ensue as the Service Layer reroutes are undone. In addition to these effects, uncoordinated multilayer restoration is likely to lead to excessivecapacity being provided for restoration. The problem here is that it is very difficult for layers to share this capacity; hence if both layers are provided with capacity to handle the same failure scenario, much of this expenditure is wasted. It is thus important both for performance and economic reasons to apply the restoration capabilities of each layer to only the types of failures for which it is most suited. The remainder of this section tries to apply this principal to the Optical Layer and the IP Services layers. The Optical Layer strengths in this respect are: 0 0 0 Line-curd costs. As discussed in Section 4.4, OLXC costs per unit capacity are expected to be significantlylower than those of an IP router. Multiservice support. As long as there are multiple client services, a single restoration mechanism at the Optical Layer reduces the need for separate mechanisms with their associated capacity requirements and operational support costs. Bundle size. The Optical Layer can restore in much larger bundles than can higher layers. Larger bundles implies fewer recovery reroutes need to be done, which should translate into faster overall recovery times. Relying on the IP layer for recovery has advantages: 0 0 Much finer granularity restoration is possible. This means that “best effort” traffic (which comprises a large proportion of total Internet traffic) can be left unrestored, with corresponding restoration capacity savings. IP layer restoration capacity may be provided to deal with router failures. If this capability is in place, the incremental capacity and complexity required to deal with all failures may be modest. Doverspike et ul. [l 1, 851 have modeled a number of OLXC and IP-based architectures to study the economic trade-offs more closely. Their conclusion [l 1 : “. .. generally, for single fiber failure, OLXC restoration tends to be less 1 expensive than IP Layer restoration except when low fractions of IP traffic need restoration and the OLXC layer is omitted or, possibly, when we need to provide extra capacity to protect against IP layer node (router) failures.” Rather than viewing the choice of restoration layer as an either/or choice, it is likely to be more profitable to look for hybrid strategies that make use of the 3. Optical Network Architecture Evolution 125 strengths of each layer. Gerstel and Ramaswami [15] have identified several cooperative strategies: Optical layer routing to enhance IP layer restoration. a. At its simplest level, this requires the OL to ensure that IP service and protection capacity is kept physically diverse; more generally it requires that the IP layer understand its redundancy constraints and communicate these constraints to the OL when requesting new capacity. (See [19], [37], and [lo81 for more detail on diverse routing.) b. Additional capacity efficiency can be obtained by treating some of the IP layer protection capacity as preemptible “extra traffic” connections in the OL. Careful design is required to ensure that these protection resources will be available when needed. a Multilayerprotection. Recovery responsibility can be divided between layers in various ways. This requires an escalation strategy defining how layers interwork to provide recovery. Some of the building blocks for cooperation are: a. Hold-oftimes. The IP layer could allow a short interval for the OL to execute its procedures before starting its own procedures. b. Controlplane coordination. As will be discussed in Section 5, a new OL control mechanism based on IP protocols is being defined. Common protocols should facilitate interlayer communication. c. Assign responsibilityfor each type offailure to a specijk layer. As a strawman, the OL might deal with single route failures while the IP layer deals with node failures of any type. Intraoffice (tie cable) failures can be efficiently dealt with in the OL using 1 1 APS. 0 Segregate protected and unprotected trafic. The IP layer could separate traffic requiring restoration onto specific OL connections and identify these connections to the OL. This would allow the OL to provide recovery capacity for these connections only. a + We have mentioned strategies that use IP layer functionality to reroute around optical layer failures. It also may be possible to use the Optical Layer to recover from IP layer failures in some cases. We give one example here.29 In a large ISP central office, there are typically at least two backbone routers for redundancy in case of router hardware or software failures and to simplify software upgrades. These routers aggregate all the traffic to or from all the provider edge routers that connect to the customers. Figure 3.41 gives an example of such an IP network, with detailed office architecture shown for office B only. 29 See [lo21 for more details. To the author’s knowledge, this scheme has not yet been implemented by any vendor. 126 John Strand Edge Fig. 3.41 Joint IPlOptical Layer router recovery scenario. The dotted lines show Optical Layer connections between one of these the routers, R B I ,and a router at office A and also between RBI and R B ~ , other backbone router at B. With current IP rerouting, when router RBIfails, traffic from office A to B needs to go around D, E, F, and C to reach office B via R B ~Similarly, traffic from office A to office C, which originally went through . office B, now needs to go around D, E, and F to reach C . Additional capacity may therefore be needed on all these links. If router RB1 fails, it brings down both interoffice link RA-RBI and intraoffice The link R B I - R B ~ . bandwidth these links occupy is unproductive for the duration of the router failure. This capacity could be reassembled by OLXCBinto a connection from RA directly to Rsz, thus avoiding the need for additional IP layer bandwidth to deal with this failure. Implementing this hybrid approach to router failure requires control plane changes in both the IP and the Optical Layers. Control planes are the topic of the next section. 5. Transport Control Plane In Fig. 3.22 a layered model of the transport signaling and control infrastructure was introduced. This infrastructure was divided into two parts: a “control plane” responsible for real-time control and fault management and a “management plane” providing the network operator with network visibility and control in a static control environment characterized by scheduled activities and delayed response to network state changes. 3. Optical Network Architecture Evolution 127 5.1. TRADITIONAL TRANSPORT NETWORK MANAGEMENT AND CONTROL Legacy transport telecommunications services have evolved to be paragons of reliability, providing guaranteed quality of service regardless of network load. Much effort was directed at optimizing the use of what was then a very expensive resource, bandwidth. All this was accomplished in an era when intelligence could only be provided by centralized Operations Systems (0%). Communications between them was normally by faxed “work orders.” In this environment, microprocessors first appeared in discrete, intelligent NEs communicating with the external world through ASCII-based “craft terminals.” Because intelligence was largely isolated, there was little value in conforming to standards; each network operator developed proprietary OSs, and the Element Management Systems (EMSs) embedded in NEs were largely vendor proprietary. The architecture is illustrated in Fig. 3.42.30 Two carriers, X and Y, are shown in Fig. 3.42. Within each are separate vendor “clouds,” each with one or more vendor-supplied EMSs that communicate with the individual NEs. The EMSs also communicate with a carrier-proprietary NMS (which actually is likely to be a large number of cooperating Operations Systems (OSs), each dedicated to a specific task like restoration, provisioning, or maintenance, and sharing network state information to some extent). Protocols like Common Management Information Protocol (CMIP), Simple Network Management Protocol (SNMP), and Common Object Request Broker Architecture (CORBA) have brought some level of standardization to these interfaces. Human operators track network status and apply controls through both the NMS and the EMS. Intercarrier ______ /--------. .--_--_________---NE: Network Element EMS: Element Management System NMS: Network Management System Fig. 3.42 Traditional transport control structure. 30 See [86]for the next level of detail. 128 John Strand management communications is likely to be between the human operators rather than between system^.^' There is a nontrivial control plane in this structure. Each node in a SONET ring, for example, has state knowledge about the ring, and the nodes cooperate to recover from failures. However, the general architecture puts most multinode functionality in the management plane (the EMS and NMS). This structure has some serious deficiencies from a network operator’s perspective: 0 0 0 0 0 0 NMS software typically is large, complex, costly, and requires a major ongoing software development and maintenance effort. The introduction of a new technology, vendor, or service is frequently slowed down significantly by the need to wait for the necessary NMS software changes. Creating a network-wide “database of record” containing accurate and up-to-date state information is greatly complicated by the proliferation of proprietary software at the EMS and NE level. NEs have minimal ability to automatically determine their connectivity to other NEs. This forces reliance on expensive, error-prone manual inputs. State change information may take some time to percolate up to the NMS level. This can cause problems when rapid action is required, for example, when there is a failure. The ad hoc intervendor interfaces make rapid provisioning or recovery very difficult when multiple vendors are involved. Frequently months are required to establish a complex high-capacity connection. Even when only a single vendor is involved, communications problems and data inconsistencies among all the OSs and EMSs involved in a provisioning introduces significant delays and errors. Together, these deficiencies are a significant ongoing problem for transport network providers. Ever-increasing competition, together with the evermore volatile needs of customers, make costs, long provisioning intervals, and delays in introducing new services and technologies increasingly intolerable. 5.2. EMERGENCE OF ALTERNATIVE CONTROL PLANE ARCHITECTURES The management and control planes for the Internet and for ATM data networks were developed more recently and have avoided many of these problems. 31 For voice services there is of course a rigorously standardized service-controlplane that allows calls to be established and managed across carriers. We are talking only of the transport layers here. 3. Optical Network Architecture Evolution 129 Network elements are assumed to be intelligent and able to support a much “thicker” control plane, and thus require less NMS and EMS functionality. For someone aware of the problems discussed in the previous section, these new control planes have some very attractive features: 0 0 0 0 NEs could automatically discover their neighbors. This information could be disseminated throughout the network, and each node could build its own view of the entire network. As a byproduct, this process would generate much of the accurate, up-to-date state information needed by the NMS. This has come to be called “The Network as the Database.” This information could be used to distribute provisioning, fault localization, and restoration functions to the NEs. In this approach, the necessary software would be provided by the equipment vendor; this spreads software development costs over all the vendor’s customers and reduces the possibility of network operator software development bottlenecks slowing down the introduction of new technologies and services. This software would implement standardized, proven protocols derived from the Internet and ATM worlds. The standardization might allow multivendor interworking and even a degree of multicarrier interworking. Furthermore, public domain or commercially available implementations of the basic protocols are readily available, and programmers familiar with this software are also readily available. If this software proves to be adaptable to the needs of the Optical Layer, this all should greatly speed up the realization of the desired capabilities. From a carrier’s perspective, then, a vision of an alternative management and control structure has begun to emerge. In this vision, vendor-supplied OLXC control-plane software will automatically do some key NMS functions such as database creation, provisioning, and restoration, thereby allowing a lower-cost, streamlined NMS. Furthermore, the new control plane had the promise of supporting much more rapid provisioning and less painful intervendor and even intercarrier interworking. In the longer term, there is a hope that if the Optical Layer control plane is compatible with that used by the Internet, then additional ways to be responsive to the Internet’s needs will emerge. OLXC vendors, particularly start-ups, were quick to see that they could inexpensively differentiate their hardware products by modifying existing implementations of the basic protocols. Today virtually every OLXC vendor claims to have a control plane based on this approach. Additional support for the use of Internet protocols came from the Internet community. Anticipating the necessity of incorporating optical switching 130 John Strand if they were to continue to scale to meet exploding demand, they greeted the possibility of common Internet-Optical Transport control protocols with enthusiasm. If the transport paradigm could be changed from one of fixed “pipes” that could only be reconfigured on a time scale of weeks to months to one of the transport network as a sort of giant distributed circuit switch with a dynamically reconfigurable backplane, new dimensions in traffic engineering might emerge. Vendors with substantial Internet equipment expertise were particularly enthusiastic, foreseeing the possibility of gaining a foothold in the massive telecom equipment market. Both established and start-up network operators have shown considerable interest in these developments. The Carrier Working Group within the Optical Interworking Forum (OIF) recently produced an “Optical Services Framework and Associated Requirements” document [43] that provides a good snapshot of current services thinking in the carrier community as it relates to optical networking and software control. They articulated the following objectives for a new control plane.32 a Promote a standardized optical network control plane, with its associated inter$aces andprotocols. Ensure that intelligent optical networking (OTN) products from different vendors or employing different a technologies can interwork at the control plane level. This is not meant to preclude vendor specific extensions so long as the extensions do not degrade total network performance. However, it is unacceptable to expect carriers to write software to conceal OL vendor or technology incompatibilities. Provide rapid automatic end-to-end provisioning within the optical network, including routing and signaling by the controlplane. “Endto-end” is an important part of this objective; the value of rapid provisioning in the core network is greatly reduced if traditional provisioning intervals are encountered in metro and access networks. Hence interdomain interworking is viewed as being key to realizing many of the hoped-for benefits. It is recognized that rapid interdomain provisioning must be built upon rapid automatic provisioning within a single domain. In this document “provisioning”refers to network provisioning only and does not necessarily include other customer-facing aspects of provisioning like the establishment of a new account. Provide restoration, diverse routing, and other Quality of Servicefeatures within the control plane on a per-service-path basis. Per-connection a 32 The current author was the chair of the OIF group producing this report. It is working text and not an official OIF Technical Report and is not binding on the OIF or its members. 3. Optical Network Architecture Evolution 131 e e e e e e e control of these capabilities is important because of the anticipated diversity of needs of the OL users. Provide policy-based call acceptance and peering policies and support usage-based accounting based on start and termination time, bandwidth, and other resources requested. These capabilities are necessary to support a pay-for-service business model and otherwise safeguard carrier resources, which carriers expect to be basic to their OL plans. Oger carrier-spec@c “branded” services. A fundamental carrier business need is the ability to put together packages of features, quality of service, support, and pricing, which they can then market. Rapid deployment of new technologies and capabilities with no network service disruption. In particular, deployment of a new vendor X capability should not depend on vendor Y’s willingness to upgrade their control plane software. It is recognized that the part of the network served by vendor Y equipment may not get the benefits of the new capability in this case. Protect the security and reliability o the optical layer, andparticularly f the controlplane. The damage, which could be done if this is not accomplished, is beyond measure. Provide the ability for the carrier to control usage of its network resources. The carrier will control what network resources are wailable to individual services or users. Therefore, in the event of capacity shortages this ability will allow the carrier to ensure that critical and priority services get capacity. Reduce the need for carrier-written OS software through heavy use o open f protocols and vendor and third-party software, particularly for network provisioning and restoration. Carrier OS software development bottlenecks frequently are on the critical path to technology and service innovation. Improvements in this area are at least as important to many carriers as is the prospect of very rapid provisioning. Note, however, that entire functions need to be off-loaded in most cases; if this is not done the carrier OS function remains, with the added requirement of interfacing with the new control plane software. Ensure the scalability o the Optical Network. Several aspects of f scalability are highly important in a carrier’s network: e Node scalability-the size of a single node is expected to be sized anywhere between several hundred and several thousands optical terminations. e Network scalability-the size of an interconnected network is expected to be anywhere from several to hundred or thousands of nodes. 132 John Strand 5.3. INTERNET PROTOCOL BACKGROUND To understand the proposed control plane, some exposure to a few Internet concepts are necessary. We provide only the barest possible peek into this vast area. Readers interested in learningmore could start with [87,88,89,90, or 911. 53.1. Internal Protocol (IP) IP transmits blocks of data called packets from sources to destinations, where sources and destinations are hosts identified by fixed-length globally unique addresses. Each packet contains a destination address and is passed through a sequence of switches called “IP routers.” In a router, each packet is processed independently; IP has no concept of a “connection.” Its address is matched against entries in aforwarding table. The matched entry in this table specifies the output port through which the packet should be sent. 5.3.2. Domains There are millions of IP addresses. Indeed, the 32-bit address space supported by the current version of IP (IPv4)is in danger of being exhausted. Furthermore, the Internet is in a constant state of flux. Therefore, it is impossible and undesirable for any node to have an up-to-date entry in its forwarding table for each of these addresses-the signaling and processing overhead would overwhelm the Internet. Consequently,the Internet is divided into many thousands of domains (formally, “Autonomous Systems (ASS)”; also called “Routing Domains”). Each router keeps detailed forwarding table entries for addresses within its domain. Packets for other addresses are sent to a “border router” that communicateswith its peers in other domains and handles the processing of packets entering or leaving the AS for other ASS.Each AS is under the control of a single administrative entity and comprises a “domain of trust” within which state information is freely shared and routing decisionsnot questioned. 5.33. Open Shortest Path First (OSPF) OSPF is a “routing protocol.” An instance of OSPF33runs in each router, where it constantly communicates with its peers to keep an up-to-date view of the network topology. Whenever there is a change, OSPF modifies the forwarding table, which immediately modifies the way subsequent packets travel through the network. OSPF is an Interior Gateway Protocol (IGP), which means that it handles the routing within a single domain only. 33 Or an equivalent protocol lie IS-IS. Within an AS a single IGP needs to be used. 3. Optical Network Architecture Evolution 133 5.3.4. Link State Advertisements (LSAs) OSPF routers send LSAs to all their peers in their domain. These define the “linksYy-onnections between adjoining routers. With this information, each router can maintain a complete and up-to-date picture (Link-State Database) of the global topology of the entire domain. It then runs an algorithm (a “Shortest Path First” algorithm, whence the name OSPF) that determines how to construct the forwarding table. Each time an LSA update is received, the algorithm is rerun. 5.3.5. Neighbor Discovery and Maintenance Each OSPF router periodically (by default, every 10 seconds) sends a “Hello” message out of each of its ports. It learns of the existence of a neighboring router when it receives the neighbor’s OSPF “Hello” in turn. If too much time elapses between receipt of these messages (nominally 40 seconds), the router will stop advertising the unresponsive link and also will start routing packets around the failure. This time dominates “convergence time”-the time it takes all the forwarding tables in a domain to stabilize after a failure. 5.3.6. Border Gateway Protocol (BGP) BGP4 is the interdomain routing protocol (commonly called an “Exterior Gateway Protocol”) currently used over the Internet. BGP nodes (called “speakers”) exchange summarized information about the addresses they can reach and the sequence of ASSthrough which the address would be reached. Trust between ASS is not assumed; each BGP speaker applies “policy-based controls” to determine which ASSit wishes to peer with and what traffic it is willing to accept from each peer. BGP does not do neighbor discovery; instead, peering relationships are manually configured. Unlike OSPF, where the onginating router did its routing based on a global-link-state database, each BGP speaker keeps a summary list of distant addresses and the distance to each through each adjacent peer (the “distance vector”). Distance is normally measured as a hop count. Then a speaker will normally route packets to a given destination through the peer that had advertised the fewest number of hops. 5.3.7. Multi Protocol Label Switching (MPLS) IP is a connectionlessprotocol, meaning that each packet is routed independently; thus if a block of data is broken up into multiple packets they might be routed independently. An analogy has been drawn to copying a novel onto a set of postcards. MPLS creates the equivalent of connections (called “Label Switched Paths,” or LSPs) by attaching an additional label to each packet in a flow. MPLS label processing is illustrated in Fig. 3.43. In Fig. 3.43, a packet with IP address X is shown entering from the left. A table look-up in the first MPLS node results in the label “6” being appended I 1: 110; 6 11 ;II Label Address Label Interface I X In I Y IP 7 O 0 n 0 MPLS Ingress Router MPLS Node MPLS Egress Router Fig. 3.43 MPLS label processing example. 3. Optical Network Architecture Evolution 135 at the front of the packet and the out interface 0 being used for the packet. In the second MPLS node the label 5 replaces the initial label and out interface 3 is chosen. In the final node, the MPLS label is “popped” off to reveal the initial packet, which is then routed over out interface 0. As long as the MPLS tables are unchanged, all packets with the same destination address that enter this ingress router will be routed in exactly the same fashion. An LSP is a “virtual” connection. Unlike a TDM connection, creating an LSP does not commit bandwidth or use bandwidth when idle. MPLS can make use of this feature to set up numerous LSPs for restoration purposes without consuming bandwidth. To OSPF, an MPLS LSP can be considered as a link in its link state database. MPLS has many other features of importance. One that should be mentioned is the ability to have nested LSPs. This provides a sort of multiplexing capability that is of importance for our application because it will allow an optical connection to be treated as an LSP on which other LSPs are routed. 5.3.8. Label Distribution Protocol (LDP) LDP allows MPLS routers to request that an LSP be set up to a specified address. This is done by sending a label request to the desired node. If no problems are encountered, a second message is passed back along the same path, assigning the labels and establishing the LSP. 5.3.9. Constraint-basedRouting Label Distribution Protocol (CR-LDP) This extension of LDP allows the LSP establishment process to be constrained to a specific physical path. It also allows the reservation of resources (bandwidth) along the path. There is an alternative protocol called RSVP (Reservation Protocol) that performs essentially the same function. 5.4. MPkS AND GENERALIZED MPLS (GMPLS) GMPLS (closely related to MPhS, or ‘‘MP Lambda S”) is an extension of MPLS and related protocols to form the basis for a control plane for Optical Networks [24, 49, 921. In fact, it is more ambitious than this: The goal is to support devices that do packet switching and also switching in the time, wavelength, and space dimensions by enhancing the connection-orientedprotocols used by MPLS and its traffic engineering extensions. For Optical Networking, GMPLS: 0 Replaces the MPLS label discussed earlier with the wavelength frequency, which acts as an implicit label, thus treating an OTN 136 John Strand 0 0 connection as an LSP. An OXC is then analogous to an MPLS node and wavelength translation to a label swap. MPLS’s label-nesting capability allows the multiplexing of multiple SONET connections onto a single wavelength to be handled naturally. Uses an extended IGP (e.g., OSPF) to advertise information about the OTN topology and resource availability (e.g., spare capacity). Its routing algorithm then also needs extension or replacement so that it can compute appropriate paths in the OTN. Uses an enhanced CR-LDP or RSVP to set up connections. GMPLS requires enhancementsto the protocols used in MPLS to deal with peculiarities of OTNs: 0 0 0 0 0 0 0 0 0 Optical bandwidth comes in a few discrete sizes (OC-58, OC-192, etc.), whereas MPLS LSPs can be any size. Optical bandwidth is real and physical, whereas MPLS LSPs are virtual and need not consume any bandwidth if idle. It is not possible to set up standby optical connections for restoration or other purposes without consuming bandwidth. Overbooking optical capacity is not possible; the number of wavelengths on an OTrS is a hard constraint. Optical links have a number of other attributes that affect routing that will need to be advertised, for example, the acceptable formats (SONET, Ethernet) and bit rates (is OC-768 supported on the link). (This is discussed in detail below in Section 5.6.) Optical connections are normally bidirectional while LSPs are unidirectional. Labels are no longer abstract identifiers. They need to be mapped to specific wavelengths or time slots. Fast fault detection and isolation is needed if restoration is to be supported. Unlike normal data applications, control traffic cannot be directly inserted into the user data stream without an OEO conversion and other changes. Thus there needs to be a separate data communications infrastructure for the control plane. All the routing complexities discussed in Section 4,particularly impairments and connectivity constraints, need to be dealt with. One area where the challenges for GMPLS are likely to be particularly daunting is restoration. This area is discussed in [lo31 in more detail. 3. Optical Network Architecture Evolution 137 5.5. IP-CENTRIC CONTROL PLANE GMPLS is an important element in a new IP-centric control plane, but there are other elements required as well. The overall network framework is shown in Fig. 3.44. GMPLS deals with routing within a single carrier cloud, and thus defines part of the IaDI-the interface between NEs within a carrier domain. It does not define the intercarrier interface (IrDI) or the User-Network Interface (UNI). We turn next to the UNI and alternative ways in which a user's provisioning needs can be communicated to the OTN. 5.5.1. Automated Provisioning Interface Alternatives Some of the principal alternative modes for automated provisioning that are being discussed are illustrated in Fig. 3.45. Figure 3.45a illustrates the two principal models being discussed: 0 The UNI model is client-server. The client requests a connection with certain attributes by sending signaling messages defined by the UNI UNI. User-Network Interface IaDI. Intra Domain Interface (aka Network-Node Interface) lrDl Inter-Domain Interface (aka Network-NetworkInterface) n Fig. 3.44 Network interfaces. UNI-C Client Connection UNI-N (b) UNI Configuration (Client to ONE) UNI-C Network Network Client (a) UNI 8 Peer Provisioning Models ([24]) UNI-N (c) UNI Configuration (3rd Patty Signaling) Fig. 3.45 Alternative provisioning models. 138 John Strand 0 standard, and the OTN responds with success and status information. Internal network topology and state information is not disseminated to the client. In the peer model, routers are “peersyy the ONES inside of the of network. Topology information (LSAs) and other state information necessary to do routing is shared. If a router wanted to establish a connection, it would compute the path itself and then send the appropriate CR-LDP/RSVP messages through the OTN to set up the connection. The peer model was proposed first. It seems to be quite appealing to many Internet people, but has run into major resistance from the carriers because it seemingly does not allow carriers to control their resources34The Optical VPN service concept proposed by the OIF Carriers Group (see Section 3.4 and [43]) is an attempt to meet the need without compromising network integrity. Figures 3.45b and 3.4% give several variants on UNI. Here “UNI-CYy and “ U N I - N represent the UNI client and network processes, respectively. In Fig. 3.45b the client device and the ONE directly interact across the UNI. In Fig. 3.4% the devices are not UNI-capable. Instead, separate client and network management entities negotiate the connection. This may be appropriate when one or both entities are SONET or other legacy equipment. 5.5.2. Connection Attributes The proposed control plane would allow fixed-bandwidth connections to be requested, torn down, and modified. To establish a connection, the desired attributes of the connection must be specified. In [93] the major types of attributes to be supported were identified as: 0 0 0 0 IdentiJcation attributes. Source and destination nodes, when the connection is desired. Basic connection type. Framing type (SONET, Ethernet, etc.), bandwidth, and directionality (unidirectional, bidirectional). Priority, preemptibility, protection, and restoration requirements. Routing constraints. This could include explicit routing or diversity constraints. 34 An interested reader might want to compare the peer model with the OIF Carrier Group objectives outlined in Section 5.2. There are a number of discords: It is unclear how policy-based call acceptance or carrier-specific “branded” services would be supported, for example. Security issues and the ability of the carrier to control usage of its own resources also may be troublesome. Thesemay not be show-stoppers,but at the least: refinement of the peer model and/or modification of carrier objectives will be needed. 3. Optical Network Architecture Evolution 139 5.5.3. Neighbor Discovery This refers to the process that determines the port-to-port connectivity of the ONES. Both ONE-ONE connectivity within the OTN and client-ONE connectivityneeds to be determined. Automation of this process is particularly important because of the hundreds or even thousands of fiber connections expected in a single office. There are many types of interfaces currently in use in OTNs that differ in many respects, therefore a variety of methods are required to accomplish this [24]. The basic functions, however, are similar to those discussed in Section 5.3. 5.5.4. Service Discovery Those adjoining ONESand client devices connected to the OTN need to determine what the capabilities of each of their links are, for example, whether both OC-48 and OC-192 are supported on the link. 5.5.5. Topology and Resource Discovery OSPF extensions are needed to determine and globally disseminate information on spare bandwidth, multiplexing capabilities, and protection needed for route computations. 5.5.6. Standards Activities There are four groups involved in standardslimplementationagreement work in the optical arena: 0 0 0 The International TelecommunicationsUnion (ITU),35and particularly its Telecommunication Standardization Sector (ITU-T), which is an international government-sanctioned group that has traditionally controlled telecommunications standards. The ITU-T has a reputation for being thorough, but slow; they are trying to speed up their process. The Internet Engineering Task Force (IETF), and particularly its IP-over-Optical(IPO) working The IETF is the driver for all Internet-related standards and the nexus of IP-related activities. They tend to move rapidly. Before standardizing anything they require two working implementations. Their standards are called “Requests for Comments” (RFCs). Working papers are called “Internet Drafts.” The Optical Domain Services Interconnect (ODSI)37is an informal association of more than 100 companies, primarily equipment vendors. 35 See itu.int. 36 See ietEorghtm.charters/ipo-charter.htm1. 37 See odsi-coalition.com. 140 John Strand They have developed a functional definition of a mechanism that will allow electrical layer devices to automatically signal the optical network to establish high-speed bandwidth on demand. They anticipate turning their results over to formal standards activities. The Optical Interworking Forum (OIF)38is an informal group with over 300 members, both equipment vendors and carriers. Its goal is to foster the development and deployment of interoperable products and services for data switching and routing using optical networking technologies. The OIF is working on both physical layer and control plane (OIF) documents. The OIF and ODSI are not sanctioned standards groups. Instead, they are trylng to develop “implementation agreements”-specific focused specs that will allow implementation to move onward. The relationships between all these bodies is confusing and there clearly are overlaps. Many people seem to be following the strategy of submitting their standards contributions to multiple bodies in the hope that somehow they will have the desired impact. The proper relationships among all these groups are unclear and controversial. There are some who would like to see all standards work carried out in the ITU. Others would like to see more specialization. This might lead the IETF to be the arena for dealing with protocols, for example, and the ITU the arena for dealing with physical layer standards and for detailing and formalizingthe major standards. The author would like to see the carriers take a more active role in ensuring that the requirements driving the various bodies are consistent. 5.6. ROUTING COMPLEXITIESARISING FROM DOMAINS OF TRQNSPARENCY Plans for an IP-centric control plane have incorporatedsome of the idiosyncracies of optical technology. However, the work so far is principally relevant to opaque networks with OLXC nodes. This greatly simplifies the control-plane design. In this section we look at some of the additional considerations that arise if we wish to apply this control plane within a domain of transparency. 5.6.1. Routing Complications Section 4.2 summarized the state of routing and wavelength assignment in large domains of transparency. Several issues for the control plane arise from this discussion: 0 Impairments cannot be ignored in large domains. Additional constraints will need to be imposed on routing. These constraints see oiforum.com. 38 3. Optical Network Architecture Evolution 141 0 0 0 0 depend on additional physical parameters that will need to be determined and advertised, and the specific constraints required in a given situation may depend on the design and engineering of the domain. Port-to-port connectivity across a domain of transparency will be limited by the level of wavelength translation supported within the domain. It will also be constrained by implementation specifics such as adaptation grouping capabilities. At a minimum, significantly more connectivity information will need to be advertised. A number of additional software-controllable components (beyond the OLXC) that can change the internal connectivity of the domain can be expected. These include PADMs, tunable lasers and filters, and dynamic devices that allow changes in the bandwidth assigned to an adaptation grouping and the wavelength spacing within the grouping. At a minimum, information about the connectivity changes resulting whenever these components are reconfigured will need to be advertised. In addition, these components potentially could be configured by the control plane. Because the route selected must have the chosen wavelength available on all links, this information needs to be considered in the routing process. This is discussed in [94], where it is concluded that advertising detailed wavelength availabilities on each link is not likely to scale. Instead they propose an alternative method that probes along a chosen path to determine which wavelengths (if any) are available. This would require a significant addition to the routing logic normally used in OSPF. Choosing a path first and then a wavelength along the path is known to give adequate results in simple topologies such as rings and trees [74]. This does not appear to be true in large mesh networks under realistic provisioning scenarios, however. Instead, significantly better results are achieved if wavelength and route are chosen simultaneously [76]. This approach would, however, also have a significant affect on OSPF. 5.7. TRADING OFF THE CONTROL PLANE, SYSTEM DESIGN, AND CARRTER PLANNINGYENGINEERING As domains of transparency become both larger and more software reconfigurable, these complexities become increasingly of concern. It is important to note that at present this evolution is largely technology driven. Vendors pushing the technology envelope are competing fiercely to provide solutions that have higher capacity, can go further all-optically,are more reconfigurable, and are more cost-effective. As vendors pursue their diverse visions, it is quite plausible that the Optical Layer of the future will be made up of heterogeneous technologies that differ significantly in their control-plane needs. 142 John Strand Carriers will wish to take advantage of the new technologies; however, they are also going to want to gain the service and operational benefits offered by the new control plane concepts, which are made more difficult to achieve by this evolution. What are the choices? Alternative approaches that deserve consideration are: Control-planesolutions. The control plane could be enhanced in a number of ways: a. All-encompassing intradomainprotocols. GMPLS could attempt to include the full range of technological options and constraints and evolve as these evolve. This will be, at best, a very challenging approach, in this author’s opinion, which will be made more difficult by the need to do periodic in-service software upgrades in a multivendor network. b. Per-domain routing. In this approach, each domain could have its own tuned approach to routing. Interdomain routing would be handled by a multidomain or hierarchical protocol that allowed the hiding of local complexity. Single vendor domains might have proprietary intradomain routing strategies. This option is discussed further in [19], [36], and [37]. System architecture solutions. Vendor system designers could be less aggressivewith respect to nonlinearities, and therefore somewhat suboptimal, in exchange for improved scalability, simplicity, and flexibility in routing and control-plane design. As a hypothetical example, if control-plane protocols did not deal with chromatic dispersion, carriers would require their vendors to provide transport systems engineered to keep this impairment from ever being binding. Transportplanning/transmission engineering solutions. a. Transmission engineers could be required to only deploy domains where every possible route met all constraints not handled explicitly by the control plane, even if the cost penalties were severe. b. At (selected) OLXCs within a domain of transparency, the control plane could insert O/E/Oregeneration into routes with transmission problems. This might make all routes feasible again, but at the cost of additional cost and complexity, and with some loss of rate and format transparency. The author is not aware of any studies that evaluate these alternatives generically. 5.8. HETEROGENEOUS TECHNOLOGIESAND MULTIPLE DOMAINS One of the approaches for dealing with the routing and control problems associated with complex domains of transparency that were identified in the 3. Optical Network Architecture Evolution 143 preceding section was to divide a network into multiple routing domains, each with relatively homogeneous technology (e.g., single vendor). Then a multidomain protocol (the optical equivalent to BGP or multiarea OSPF for the Internet) would be used for multidomain interworking. There are a number of other scenarios where multiple routing domains are likely to be needed. Two of these are: e e Metro-core interworking. Deployment of intelligent optical networking technology is planned or under way in both intercity (core) and metro portions of a number of carrier networks. It is essential that the metro and core subnetworks be able to interwork as soon as possible if we are to realize the fast provisioning potential of these deployments, because a large proportion of the anticipated connections will need to traverse both metro and core subnets. Multi-carrier interconnection.As optical networking becomes established, opportunities and needs for routing optical connections over multiple competing carrier networks are presented. This is done today in the United States with private lines, many of which are routed through an intercity carrier network and one or more RBOC networks. Internationally also, multiple carriers are often involved. As the number of optical networks supporting rapid provisioning increases, the needs for this sort of interconnection can only increase. More details on these applicationsand some initial requirementsfor amultidomain optical protocol can be found in [97]. At the time of writing, work to define the necessary protocols is just starting (see, e.g., [96], [98]). 5.9. ALTERiVATIVE CONTROL-PLANE APPROACHES The distributed, IP-inspired control-plane work described in the last few sections is not the only possible approach to controlling and managing an optical network. In this section some dissenting opinions and alternative approaches will be mentioned. Gerstel [22] argues that it would be a mistake to adopt Internet-style distributed network control. He argues that a telecom-stylenetwork management interface augmented with a minimal control plane and a service-layer interface between management systems would be a better choice. The control plane would be distributed but would be limited to fault management. Connection setup would only be provided for automatic protection purposes. The key functions addressed by the control plane discussed in the previous sections would be performed by network management systems. There would be multiple single-vendor domains, each with its own EMS; they would interoperate 144 John Strand at the NMS level using a protocol based on the CORBA interface standards developed in the telecom industry. d s l i Hjdmtjrsson and his collaborators [go], on the other hand, argue that a thorough-going integration of the IP and Optical layersis a better architecture. In this architecture, each node integrates a router and a PXC. Intelligence remains entirely in the router, which is responsible for all networkingfunctions, including optical resource management. They argue that collapsing these two layers together would remove a major source of complexity. Their proposal is consistent with a strong form of the GMPLS peer model. Mukherjee [lo51 summarizes an alternative control plane approach for domains of transparency that is not based on each node keeping a global Link-State Database. In this distributed-routingapproach, routes are selected in a distributed fashion without knowledge of the overall network topology. Each node maintains a routing table that specifies the next hop and the cost associated with the shortest path to each destination on a given wavelength. The connection request is routed one hop at a time, with each node along the route independently selecting the next hop. If a node on the path is unable to reserve the desired wavelength on a link, it sends a negative acknowledgment back along the reverse path. Once the destination node is reached, an acknowledgment is sent back along the reverse path; this triggers OLXC configurations as it goes. This approach has the advantage of not requiring distant nodes to be cognizant of impairments in distant parts of the network, but has the disadvantage that it makes constraint-based routing, such as is required for diversity, more complex. 6 Summary . We have tried to emphasize the interconnections between technology, network structure, services, and economics. Low cost is critically important, but this does not imply that a vendor can focus strictly on hardware costs. Because most telecommunications costs are in operations rather than in hardware, the maintenance and provisioning costs associated with a design must always be a major consideration. Control-planearchitects ignore the many unique aspects of optical technology at their peril. The unique aspects of IP-based services must be foremost in the minds of everyone. Both network operators and equipment vendors will be struggling to differentiate their offerings. It appears that much of this struggle will focus on software-based fmctionality--"soft optics" and new control and management planes. Of critical importance will be the sorting out of functionality between the Optical and IP-based layers. Where should restoration be done? Who should be responsible for fault detection and management? Can a rapidly reconfigurable optical network compete with Tier 1 ISPs to provide IP connectivity? 3. Optical Network Architecture Evolution 145 In this context, some of the most interestingtechnological developments are those that make Optical Transport Systems more flexible and reconfigurablesoftware-reconfigurable OADMs, tunable lasers and receivers, and the like. They have the potential to turn an OTS into a sort of giant distributed optical cross-connect. Equally important are the developmentsin all-opticalswitching fabrics. As optical technology matures, all-optical “domains of transparency” are likely to become more important. They promise considerable cost savings by reducing the need for expensive transponders, and they may offer additional advantages in the form of better scalability and service flexibility. Complex architectural and economic trade-offs will need to be made to get the right balance of domain diameter, unit cost, and operational complexity. Fault management in all-optical networks is a continuing concern. There is also a tension between optimizing the use of optical technology and unduly complicating the control plane that must determine in real-time how to configure and manage the network. It is likely that this tension will lead to domain-specific control with some sort of distributed or centralized interdomain network management. The future of all-optical domains is complicated by another trend-the economic and operational improvements that can be attained by integrating layers. In one direction, IP vendors may see reasons to tightly couple optical functionality into their routers while keeping the discrete optical layer as simple and “dumb” as possible. In another direction, particularly in metro networks multiservice provisioning, platforms integrating TDM ,optical, and data fabrics into a single box are getting a lot of attention. Both of these trends couple optical network design with other worlds, and therefore may complicate the efforts to optimize the use of optical networking capabilities. Changes in the structure of the industry are also of great importance for the evolution of optical network architecture. Of particular note is the fragmentation of the long-haul transport infrastructure as many new entrants appear. Also important is the structure of the Internet world-its preference for very large bandwidth connections,as well as the emergence of server farms and carrier hotels, which may change the relationship between customers and their network providers and offer new business models, such as bandwidth trading, to get established. Acronyms ADM ANSI AS ATM BGP Adddrop multiplexer American National Standards Institute Autonomous system Asynchronous transfer mode (protocol) Border gateway protocol (Internet routing protocol) 146 John Strand CLEC CRC DCS DPRING EMS EO EOY FEC GUI IaDI IGP ILEC IP IrDI ISP ITU ITU-T LAN LEC LR LSA LSP LSR MAN MPLS MTBF MTTR NE OA OADM OCh OE OEO OH OL OLXC OMS ONE OSPF OTN OTrS OTS Competitive local exchange carrier Cyclic redundancy code Digital cross-connect system Dedicated protection ring (SDH terminology) Element management system Electrical-to-Optical(conversion) End of year Forward error correction Graphical user interface Intra-domain interface (ITU terminology) Interior gateway protocol Incumbent local exchange carrier Internet protocol (Internet layer 2 protocol) Inter-domain interface (ITU terminology) Internet service provider International telecommunicationunion Telecommunication standardization sector of the ITU Local area network Local exchange carrier Long reach (referringto lasers) Link state advertisement Label switched path (an MPLS connection) Label switched router (MPLS node) Metropolitan area network Multiprotocol label switching Mean time between failures Mean time to repair Network element Optical amplifier Optical add-drop multiplexer (either optical or electrical fabric) Optical channel (ITU terminology) Optical-to-Electrical(conversion) Optical-electrical-optical (as in a regenerator) Overhead (control information added to a packet or frame) Optical layer Optical layer cross-connect (may have optical or electrical fabric) Optical multiplex section (ITU terminology) Optical network element Open shortest path first (an Internet routing protocol) Optical transport network Optical transport system Optical transmission section (ITU terminology) 3. Optical Network Architecture Evolution 147 PADM PDH PN PXC QOS RWA SDH SHR SLA SONET SPRING SR TCP VPN WAN Photonic (all-optical) adddrop multiplexer Plesiochronous digital hierarchy (DSO, DS1, DS3) Photonic (all-optical) network Photonic (all-optical) cross-connect Quality of service Routing and wavelength assignment Synchronous digital hierarchy (International ITU analogue of SONET) Self-healingring Service level agreement Synchronous Optical NETwork Shared protection ring Short reach (referring to lasers) Transmission control protocol (principal Internet layer 3 protocol) Virtual private network Wide area network (refers to intercity networks) References Note: References are primarily to survey and overview material. References to primary research may be found in these articles. [l] R. Ramaswami and K. N. Sivarajan, Optical Networks: A Practical Perspective, San Francisco: Morgan Kaufmann, 1998. [2] B. Mukherjee, Optical CommunicationsNetworks, New York McGraw Hill, 1997. [3] P. Bonenfant and A. Rodriguez-Moral, “Optical Data Networking,” ZEEE CommunicationsMagazine, vol. 38, no. 3, March 2000, pp. 63-70. [4] N. Ghani, S. Dixit, and T-S. Wang, “On IP-over-WDM Integration,” IEEE CommunicationsMagazine, vol. 38, no. 3, March 2000, pp. 72-84. [5] J. M. H. Elmirghani and H. T. Mouftah, “All-OpticalWavelength Conversion: Technologies and Applications in DWDM Networks,” IEEE Communications Magazine, vol. 38, no. 3, March 2000, pp. 86-92. [6] 0. Gerstel and R. Ramaswami, “Optical Layer Survivability: A Services Perspective,” ZEEE Communications Magazine, vol. 38, no. 3, March 2000, pp. 104-113. [7] J. M. H. Elmirghani and H. T. Mouftah, “Technologies and Architectures for Scalable Dense WDM Networks,” ZEEE Communications Magazine, vol. 38, no. 2, Feb. 2000, pp. 58-66. [8] S. Yao, B. Mukherjee, and S. Dixit, “Advances in Photonic Packet Switching: An Overview,” IEEE CommunicationsMagazine, vol. 38, no. 2, Feb. 2000, pp. 84-94. [9] D. Cavendish, “Evolution of Optical Transport Technologies: From SONETI SDH to WDM,” IEEE CommunicationsMagazine, June 2000, pp. 164-172. 148 John Strand [lo] Y Ye and Sudhir Dixit, “On Joint ProtectionRestoration in IP-Centric DWDM-Based Optical Transport Networks,” ZEEE CommunicationsMagazine, vol. 38, no. 6, June 2000, pp. 174-183. [l 1 R. D. Doverspike, S. Phillips, and J. R. Westbrook, “Future Transport Network 1 Architectures,” IEEE CommunicationsMagazine, vol. 37, no. 8, August 1999, pp. 96-101. [12] M. W Maeda, “Management and Control of Transparent Optical Networks,” . ZEEEJ. on SelectedAreas in Communications,vol. 16, no. 7, Sept. 1998, pp. 10081023. [13] B. Ramamurthy and B. Mukherjee, “Wavelength Conversion in WDM Networking,” IEEE J. on Selected Areas in Communications, vol. 16, no. 7, Sept. 1998, pp. 1061-1073. [14] H. Zang, J. P. Jue, and B. Mukherjee, “A review of routing and wavelength assignment approaches for wavelength-routed optical WDM networks,” Optical , Networks Magazine ZEEE J. on Selected Amas in Communications, vol. 1 no. 1, Jan. 2000. [15] 0. Gerstel and R. Ramaswami, “Optical Layer Survivability-An Implementation Perspective,”IEEE J. on Selected Areas in Communications, vol. 18, no. 10, Oct. 2000, pp. 1885-1899. [ l a S. Baroni, J. 0. Eaves, M. Kumar, M. A. Qureshi, A. Rodriguez-Moral, and D. Sugerman, “Analysis and Design of Backbone Architecture Alternatives for IP Optical Networking,” IEEE J. on Selected Areas in Communications, vol. 18, no. 10, Oct. 2000, pp. 1980-1994. [17] D. Zhou and S. Subramaniam, “Survivability in Optical Networks,” ZEEE Network, vol. 14, no. 6, Nov. 2000, pp. 16-23. [18] T. H. Wu, Fiber NetworkSurvivability, Nonvood, MA: Artech House, 1992. [19] J. L. Strand, A. L. Chiu, and R. Tkach, “Issues For Routing In The Optical Layer,” ZEEE CommunicationsMagazine, vol. 39, no. 2, Feb. 2001, pp. 81-87. [20] C. W. Chao, P M. Dollard, J. E. Weythman, L. T. Nguyen, and H. Eslambolchi, . “FASTAR-A Robust System for Fast DS3 Restoration,” Proc. ZEEE GLOBECOM1991, Phoenix, Arizona, Dec. 1991, pp. 1396-1400. [21] A. Fumagalli and L. Valcarenghi, “IP Restorationvs. WDM Protection: Is There an Optimal Choice?’ ZEEE Network, vol. 14, no. 6, Nov. 2000, pp. 34-41. [22] 0. Gerstel, “Optical Layer Signaling: How Much Is Really Needed?’ ZEEE CommunicationsMagazine, vol. 38, no. 10, Oct. 2000, pp. 154-160. [23] S. Verma, H. Chaskar, and R. Ravikanth, “Optical Burst Switching: A Viable Solution for Terabit IP Backbone,” IEEE Network, vol. 14, no. 6, Nov. 2000, pp. 48-53. [24] G. M. Bernstein, J. Yates, and D. Saha, “IP-Centric Control and Management of Optical Transport Networks,” IEEE CommunicationsMagazine, vol. 38, no. 10, Oct. 2000, pp. 161-167. [25] A. Copley, “Optical Domain Service Interconnect: Defining Mechanisms for Enabling On-Demand High-speed Capacity from the Optical Domain,” ZEEE CommunicationsMagazine, vol. 38, no. 10, Oct. 2000, pp. 168-174. 3. Optical Network Architecture Evolution 149 [26] B. Rajagopalan, D. Pendarakis, D. Saha, R. S. h a m o o r t h y , and K. Bala, “IP Over Optical Networks: Architectural Aspects,” IEEE Communications Magazine, vol. 38, no. 9, Sept. 2000, pp. 94-103. [27] C. Qiao, “Labeled Optical Burst Switching For IP-Over-WDM Integration,” IEEE Communications Magazine, vol. 38, no. 9, Sept. 2000, pp. 104115. [28] D. K. Hunter and I. Andonovic, “Approaches to Optical Packet Switching,” IEEE Communications Magazine, vol. 38, no. 9, Sept. 2000, pp. 116-122. [29] M. Listanti and R. Sabella, “Architectural and Technological Issues for Future Optical Internet Networks,” IEEE Communications Magazine, vol. 38, no. 9, Sept. 2000, pp. 82-92. [30] R. Tkach, E. Goldstein, J. Nagel, and J. Strand, “Fundamental Limits of Optical Transparency,” Optical Fiber Communication ConJ, Feb. 1998, pp. 161-162. [31] E. Modiano and A. Narula-Tam, “Mechanisms for Providing Optical Bypass in WDM-Based Networks,” OpticalNetworh Magazine, vol. 1, no. 1, Jan. 2000, pp. 11-18. [32] H. Zang, J. P. Jue, and B. Mukherjee, “A Review of Routing and Wavelength Assignment Approaches for Wavelength-Routed Optical WDM Networks,” Optical Networks Magazine, vol. 1, no. 1, Jan. 2000, pp. 4740. [33] C. P. Laresen and F 0. Anderson, “Signal Quality Monitoring in Optical ! Networks,” Optical Networks Magazine, vol. 1, no. 4, Oct. 2000, pp. 17-23. [34] G. Hjilmtjrsson, P Sebos, G. Smith, and J. Yates, “Simple IP Restoration for . IP/GbE/lO GbE Optical Networks,” Optical Fiber Communication ConJ, 2000, vol. 4,2000, pp. 275-277. [35] S. Verma, H. Chaskar, and R. Ravikanth, “Optical Burst Switching: A Viable Solution for Terabit IP Backbone,” IEEE Network, vol. 14, no. 6, Nov. 2000, pp. 48-53. [36] J. Strand, “Optical Layer Services Framework,” Optical Interworking Forum oif2000.108, May, 2000; available from the author. [37] J. Strand and M. Lazer, “Some Routing Constraints,” Optical Interworking Forum oif 2000.109, May, 2000; available from the author. [38] M. Eiselt, “The Quest For Optimum Transmission Capacity: Is There An Optimum Fiber Choice,” OFC 2000 tutorial. [39] A. Odlyzko, “Internet Growth Myth and Reality, Use and Abuse,” iMP: Information Impacts Magazine, November 2000. [40] A. Odlyzko, “The Internet and Other Networks: Utilization Rates and Their Implications,” Feb. 26, 2000. Available at http://www.research.att.com/-amo/ doc/complete. html. [41] See http://www.canet2.net/library/papers.html. [42] B. Ramamurthy, H. Feng, D. Datta, J. P. Heritage, and B. Mukherjee, “Transparent vs. Opaque vs. Translucent Wavelength-routed Optical Networks,” Optical Fiber Communication Conference, 1999, and the International Conference on Integrated Optics and Optical Fiber Communication. OFC/IOOC ‘99.TechnicalDigest, vol. 1, 1999, pp. 5 M 1 . 150 John Strand [43] OIF Carrier Group, Yong Xue, and Monica Lazer, eds, “Carrier Optical Services Framework and Associated Requirements for UNI,” Optical Interworking Forum oif2000.155, Sept. 13, 2000. Also available as working documents in TlX1.5 and the IETF. [44] G. Mohan and C. S. R. Murthy, “Lightpath Restoration in WDM Optical Networks,” IEEE Network, vol. 14, no. 6, Nov./Dec. 2000, pp. 24-33. [45] E. L. Goldstein, L. Y Lin, and R. W. Tkach, “MultiwavelengthOpaque OpticalCrossconnect Networks,” IEICE Duns. Commun., vol. E82-B, no. 8, Aug. 1999, pp. 1095-1104. [46] F. Bentivoglio and E. Iannone, “The Opaque Optical Network,” OpticaZ Networks, vol. 1, no. 4, Oct. 2000, pp. 24-31. [47] P. Green, “Progress in Optical Networking,” IEEE Communications Magazine, vol. 39, no. 1, Jan. 2001, pp. 54-61. [48] G. Miller, K. Thompson, and R. Wilder, “Wide-Area Internet Traffic Patterns and Characteristics,” IEEE Network, vol. 11, no. 6, Nov. 1997. [49] A. Banerjee, J. Drake, J. F! Lang, B. Turner, K. Kompela, and Y Rekhter, “Generalized Multiprotocol Label Switching: An Overview of Routing and Management Enhancement,” IEEE Communications Magazine, vol. 39, no. 1, Jan. 2001, pp. 144-150. [50] P. Arijs, B. van Caenegem, P. Demeester, F LaGasse, W. Van Parys, and ! P. Achten, “Design of Ring- and Mesh-Based WDM Transport Networks,” Optical Networks, vol. 1, no. 3, July 2000, pp. 25-38. [51] R. H. Cardwell, 0. J. Wasem, and H. Kobrinski, “WDM Architectures and Economics in Metropolitan Areas,” Optical Networks, vol. 1, no. 3, pp. 41-50. [52] B. St. Amaud, “Overview of the Latest Developments in Optical Internets,” Optical Networks, vol. 1, no. 3, pp. 51-54. [53] C. Blaizot, I. Cerutti, A. Fumagalli, M. Tacca, L. Valcarenghi, R. Jagannathan, A. Lardies, F Masetti, and M. Garnot, “Design and Planning of Transport Net. works Based on SDH/SONET and WDM Optical Networks: Methods, Tools and Applications,” Optical Networks, vol. 1, no. 3, pp. 55-65. [54] Vinod Khosla, “Terabit Tsunami,” Keynote Address at OFC 2000, Baltimore, MD, March 2000. [55] K. G. Coffman and A. M. Odlyzko, “Internet Growth Is there a “Moore’s Law” for Data Traffic?” Preliminary version, 7/11/00. Available at: research.att.com/-amo. [56] L. Xu, H. Perros, and G. Rouska, “Techniques for Optical Packet Switching And Optical Burst Switching,”IEEE CommunicationsMagazine, vol. 39, no. 1, pp. 136-145. [57] A. Odlyzko, “Data Networks Are Mostly Emptly And For Good Reason,” ITPro, MarcWAprill999, pp. 6749. [58] B. St. Amaud, “Scaling Issues on Internet Networks” (draft), Canarie, Inc., Jan. 13,2001, available at www.canet2.nefflibraqdpapers. 3. Optical Network Architecture Evolution 151 [59] W Norton, “Internet Service Providers and Peering:’ North American Net. work Operators Group (NANOG), 2000, available at www.nanog.org/mtg001Oltree.doc. [60] G. Gilder, “The Post-Diluvian Paradigm,” Gilder Technology Report, vol. V, no. 2, Feb. 2000. [61] “Fiberoptic Networks of Long-Distance Carriers in North America: Market Developments And Forecast,” KMI Corp., Newport, R.I., Nov. 1999. [62] S Singh, The Code Book, New York: Anchor Books, 2000. . [63] A. Saleh, L. Benmohamed, and J. Simmons, “Proposed Extensions to the UNI for Interfacing to a Configurable All-Optical Network,” Optical Interworking Forum contribution oif 2000.278, Nov. 7,2000. [64] L. Garrett, R. Vodhanel, S. Patel, R. Tkach, and A. Chraplyvy, “Dispersion Management in Optical Networks,” Proc. Eur. Cod. on Optical Comxnun., Edinburgh, Scotland, 1997. [65] ITU-T G.805, “GenericFunctional Architectureof Transport Networks,” 1995. [66] ITU-T G.803, “Architecture of Transport Networks Based on the Synchronous Digital Hierarchy,” 2000. [67] ITU-T G.872, “Architecture of the Optical Transport Network,” 1999. [68] J. Heiles, “Alignment of the UNI with ITU terminology:’ OIF Contribution oif2000.197, Sept. 1,2000. [69] ITU-T G.709, “Network Node Interface For The Optical Transport Network,” initial version, Feb. 2001. [70] A. Malis and W. Simpson, “PPP over SONETlSDH,” IETF RFC 2615, June 1999. [71] J. Kraushaar, “Fiber Deployment Update: End of Year 1998,” U.S. Federal Communications Commission, 1999. [72] R. Ballart and Y-C. Ching, “SONET: Now It’s the Standard Optical Network,” IEEE Communications Magazine, vol. 29, no. 3, 1989, pp. 8-15. [73] J. Berthold, “SONET and ATM,” in I. Kaminow and T. Koch, eds, OpticaI Fiber CommunicationsIIIA, Boston: Academic Press, 1997. [74] J. Yates, J. Lacey, and M. Rumsewicz, “Wavelength Converters in DynamicallyReconfigurableWDM Networks,” IEEE Communications Surveys, vol. 2, no. 2, 1999, www.comsoc.org/pubs/surveys. [75] M. Garey and D. Johnson, Computers And Intractibili-A Guide to the Theory of Np-Completeness, New York W. H. Freeman, 1979. [76] J. Strand, R. Doverspike, and G. Li, “Importance Of Wavelength Conversion In An Optical Network,” Optical Networks Magazine, vol. 23, MaylJune 2001, () pp. 3344. [77l T.-H. Wu, “Emerging Technologiesfor Fiber Net Survivability,” IEEE Communications Magazine, vol. 35, no. 2, 1995, pp. 58-74. [78] ITU-T, G.841, “Types And Characteristics Of SDH Network Protection Architectures,” 1998. [79] ANSI, TI. 105.01, “SynchronousOptical Network (SONET)-Automatic Protection Switching,” 1999. 152 John Strand [80] G. HjilmtJisson, J. Yates, S. Chaudhuri, and A. Greenberg, “Smart RoutersSimple Optics: an Architecture for the Optical Internet,” IEEE/OSA Journal o f Lightwave Technology, Dec. 2000, pp. 37-50. [81] B. Doshi, S. Dravida, P. Harshavardhana, and M. A. Qureshi, “A Comparison of Next-Generation IP-Centric Transport Architectures,”Bell Labs Tech. J., vol. 3, no. 4,Oct.-Dec. 1998, pp. 137-143. [82] M. Reardon and S. Saunders, “Terabit Trouble,” Data Communications, Aug. 1999, pp. 11-16. [83] S. Chaudhuri and E. Goldstein, “On The Value of Optical-Layer Reconfigurability in IP-Over-WDM Lightwave Networks,” ZEEE Photonics Technology Letters, vol. 12, no. 8, Aug. 2000, pp. 1097-1099. [84] B. Gleeson, A. Lin, J. Heinanen, G. Armitage, and A. Malis, “A Framework for IP-Based Virtual Private Networks,” RFC 2764, IETF, Feb. 2000. [85] R. Doverspike, S. Phillips, and J. Westbrook, “Transport Network Architectures in an IP World,” Proc. INFOCOM-2000, March 2000, Tel-Aviv, Israel. [86] 0. Gerstel, “Optical Layer Signaling: How Much Is Really Needed,” ZEEE CommunicationsMagazine, vol. 38, no. 10, Oct. 2000, pp. 154-160. [87l D. Comer, Internetworking with TCP/IR El. I : Principles, Protocols, and Architecture, Upper Saddle River, NJ: Prentice Hall, 1995. [88] U. Black, Z Routing Protocols: Ne OSPF:BGE P W I & Cisco Routing Protocols, P Upper Saddle River, NJ: Prentice Hall, 2000. [89] J. Moy, OSPF:Anatomy OfAn Internet Routing Protocol, Boston: Addison Wesley Longman, 1998. [90] J. Stewart 1 1 BGP4: Inter-Domain Routing in the Internet, Boston: Addison1, Wesley Longman, 1998. [91] B. Davis and Y Rekhter, MPLS: Technology And Applications, San Francisco: Morgan Kaufmann, 2000. [92] D. Awduche et al., “Multiprotocol Lambda Switching:Combining MPLS Traffic Engineering with Optical Crossconnects,” Internet draft, draft-awduche-mplste-optical-02.txt, March 2000, work in progress. [93] L. McAdams and J. Yates, eds., “User to Network Interface Service Definition and Connection Attributes,” Optical Interworking Fonun Contribution oif2000.061, Dec. 15,2000. [94] S. Chaudhuri, G. HjPlmtJisson,and J. Yates, “Control of Lightpaths in an Optical Network,” work in progress, Internet draft, draft-chaudhuri-ip-olxc-controlOO.txt, 2000. [95] A. Chiu, J. Strand, R. Tkach, and J. Luciani, “Features and Requirements for the Optical Layer Control Plane,” work in progress, Internet draft, draft-chiustrand-unique-olcp.tt, Feb. 2001, available from the author also. [96] M. Blanchet, E Parent-Viagenie, and B. St-knaud, “Optical BGP (OBGP): InterAS lightpath provisioning,” work in progress, Internet draft, draft-parentobgp-OO.txt, Jan. 2001, available at www.canet2.netAibraxy/papers. [97] J. Strand and Y Xue, “Routing for Optical Networks With Multiple Routing Domains,” Optical Interworking Forum Contribution oif 2001.046, Jan. 2001; available from the authors. 3. Optical Network Architecture Evolution 153 [98] G. Bernstein and B. Rajagopalan, “Optical Inter-Domain Routing Requirements,” Internet draft, draft-bernstein-optical-bgp.txt,Feb. 2001, work in progress. [99] Bellcore, “SONET Bidirectional Line-Switched Ring Equipment: Generic Criteria,” GR-1230-CORE,Issue 2, Nov. 1995. [1001 Bellcore, “SynchronousOptical Network (SONET) Generic Criteria,” GR-253CORE, Issue 1, Dec. 1994. [loll Quoted by Russell McGuire, VP, Strategic Development Williams Corp., in a 12/3/1999 talk to the Risk Conference on Bandwidth Trading, New York. [lo21 A. Chiu and J. Strand, “Joint IP/Optical Layer Restoration After a Router Failure,” OFC2001, Anaheim, vol. 1, March 2001, pp. MN5-1-MN5-2. R. Doverspike and J. Yates, “Challenges for MPLS in Optical Network [lo31 Restoration,” IEEE Communications Magazine, vol. 39, no. 2, Feb. 2001, pp. 8%96. [lo41 R. Barry, “Optical Networking Technologies: Driving the Evolution of the Public Network,” NFOEC, 1999. [lo51 B. Mukherjee, “WDM Optical Communication Networks: Progress and Challenges,” IEEE Journal on Selected Areas In Communications, vol. 18, no. 10, Oct. 2000, pp. 181CL1824. 11061 S. Baroni, J. Eaves, M. Kumar, M. Qureshi, A. Rodriguez-Moral, and D. Sugrarman, “Analysis and Design of Backbone Architecture Alternatives for IP Optical Networking,” IEEE Journal on Selected Areas In Communications, vol. 18, no. 10, Oct. 2000, pp. 198CL1994. [lo71 E. Messmer, “Holding the Line on Call-Center Sprawl,” Network World, Apr. 2, 2001. [1081 R. Bhandari, Survivable Networks--Algorithm for Diverse Routing, New York: Kluwer Academic Publishers, 1999. Chapter 4 Undersea Communication Systems Neal S. Bergano ljco Telecommunications.Eatontown, New Jersey 4.1 Introduction In today’s Internet age information flows across continents as easily as it flows across the office. With so many “point and click” virtual connections, it is easy to forget that the world’s communicationsneeds are made possible by real systems based on fiber-optic cables. This comes as no surprise to those of us in the optics community; however, many others underestimate the importance of undersea fiber-optic cables for intercontinental telecommunications. The earth‘s continents are connected with a web of undersea fiber-optic cables that join the world’s major population centers. Anyone who makes international phone calls, sends international faxes, or simply surfs the Web at sites in other continents uses undersea fiber-optic cables. Although the debate regarding the merits of cable systems versus satellite systems for international communications ended many years ago, with cable systemsthe clear economicand technological winner, many people still assume that overseas communicationsoccur via satellite. But consider this: since 1990 over 600,000 km of undersea fiber-optic cable has been installed across the world’s oceans. Today, the vast majority of international telecommunications is carried on undersea cable^.^ This chapter reviews concepts for the design of long-haul transmission systems based on optical amplifierrepeaters and wavelength division multiplexing (WDM) techniques. Important strides have been made in areas of dispersion management, gain equalization, and modulation formats, which have made possible the demonstration of capacities exceeding 2TB/s on a single fiber. This chapter includes sections on the history of undersea cable, the amplified transmission line, dispersion and nonlinearity management, transmission formats, measures of system performance, error correcting codes, polarization effects, long-haul system design, experimental techniques, and future trends in long-haul optical transmission systems. ‘ 3 ’ 4.2 A Rich History of Undersea Cables Undersea cable has been in use for over 130 years. The installation of the first transatlantic telegraph cable was completed in 1858 (Fig. 4.1). Unfortunately, 154 OPTICAL FIBER TELECOMMUNICATIONS, VOLUME IVB Copyright Q 2002, Elsevier Science (USA). All rights of reproduction in any form reserved. ISBX 0-12-395173-9 4. Undersea Communication Systems 155 Fig. 4.1 The crew of the 3200-ton HMS Agameemnon, which laid the first, briefly successful cable in 1858, watches anxiously, knowing that the mere wake of a passing whale might be forceful enough to break the cable. (The Atlantic Cable, Burndy Library, Norwalk, CT, 1959.) this cable only worked for a few weeks before a problem with the cable rendered it unusable. The first successful transatlantic telegraph cable connected North America to Europe and went into service in 1866,34 years after Samuel Morse invented the telegraph. At that time an experienced telegraph operator could send about 17 words per minute, at a cost of about $5 per word.4 It took nearly 90 years from the time of the first telegraph cable to install the first transatlantic telephone cable. In 1956, the first TAT (Trans Atlantic Telephone) cable went into service, providing 48 telephone circuits between Newfoundland and Scotland.’ These analog systems were based on coaxial cables with electronic amplifiers. They eventually grew in capacity to over 4200 voice circuits for systems installed as recently as 1983. During this time, the capacity of transatlantic cable’s circuits increased at an annual rate of -20% (Fig. 4.2). These circuits were transmitted by frequency division multiplexing many circuits over an electrical bandwidth of a few tens of megahertz. The signals were boosted in “wide-band’’ electrical amplifiers that were placed in repeaters and spaced every 9.5 km. In an interesting twist of fate, the cable in the coaxial systems was linear, while great design efforts were expended to cope with the nonlinearity of the electronic amplifiers, which is opposite from the present optically amplifier fiber-optic systems. The first undersea fiber-optic systems were installed across the Atlantic and Pacific Oceans in 1988-1989 and had a capacity of 280 Megabithec on each 156 Neal S. Bergano TWO-WAY CABLE CIRCUITS 64Kbk Digital Circuits 3KHz Analoa Circuits 1M 100K- > I 00% Annual Growth 10K1K- 100 - 1960 1955 1965 1970 1975 1980 1985 1990 1995 2000 2005 YEAR Fig. 4.2 The cumulative capacity that has been installed crossing the Atlantic Ocean over the past 4 decades. The capacity is given in 3-kHz circuits for the analog systems, and 64-kB/s circuits for the digital systems. of three fiber pairs.6 These systems were actually hybrid optical systems in the sense that the repeaters converted the incoming signals from optical to electrical, regenerated the data with high-speed integrated circuits, and retransmitted the data with a local semiconductor laser. The transmission capacity of the regenerated fiber cables eventually increased to 2.5 Gigabidsec, and repeater spacing increased with the switch from 1.3 pm multifrequency lasers to 1.55 pm single-frequencylaser diodes. These first undersea fiber systems revolutionized international telephony, bringing costs down and greatly expanding capacity and quality. However, the ability of the regenerator systems to exploit the large fiber bandwidth was limited by the capacity bottleneck in the high~~ speed electronics of undersea repeaters. Beginning in the mid- 1 9 9 0 undersea fiber-optic systems using erbium-doped fiber-amplifiersin the repeaters were deployed. These systems removed the electronic bottleneck and provided the first clear optical channel connectivity between the world’s continent^.^ Today’s undersea systems use erbium-doped fiber amplifiers (EDFAs) to compensate the attenuation in the optical fiber cable. These optical amplifiers are in repeaters that are typically spaced every 50 km along the cable and have an optical bandwidth that is wide enough to support many optical channels using WDM techniques. Data signals coming from the land-based systems (typically referred to as “terrestrialyy systems) are converted to an optical format that is more robust for transmission over transoceanic distances. Systems being designed today for deployment in the next few years will support many 158 Neal S. Bergano ,*\ OPTICAL FIBER STRENGTH WIRES Fig. 4.4 Cut away diagram of a cable section. proliferation of lightwave cable systems. The EDFA is a nearly ideal building block for providing optical gain in a lightwave communications systems.8 EDFAs can be made with a variety of gains in both the “Conventional” wavelength band (C-band) from 1526-1566 nm and the “Long” wavelength band (L-band) from 1566-1606 nm.* Because the EDFA is a fiber device, it can be easily connected to telecommunication fiber with low loss and low polarization dependence. Most importantly, EDFAs can be manufactured with the 25-year reliability that is required for use in undersea systems. From an optical standpoint, the undersea portion of the system (Fig. 4.5) is sometimes referred to as an “amplifier chain”; it is a concatenation of optical amplifiers and cable spans. The attenuation in optical fiber is 0.2 dB/km (at 1550nm); thus, in a 6000-km-long transatlantic system, the optical amplifiers will compensate a total of 1200dB of cable attenuation! For proper system operation where there is a tight tolerance on the power launched into the transmission fiber, the gain in the amplifiers must exactly match the attenuation in the fiber. At first this might seem like a difficult task; however, the solution comes almost for free from the gain characteristics of the optical amplifier. One of the many advantages of the EDFAs is the ability to be operated deep into gain compression without any significant distortion of the high-speed data signals being transmitted. A natural “automatic gain control” occurs by * These are approximate values, since the wavelength range ofthe C- and L-bands are somewhat arbitrary. 4. Undersea Communication Systems 159 Fig. 4.5 A typical undersea cable link. The undersea equipment consists of cable sectionsjoined by repeaters, which house the EDFAs. The terminal equipment grooms the “terrestrial-grade” signals for transmission across the ocean. t Gain (Span Loss)-’ Operating Point Output Power Fig. 4.6 The amplifier’s output power is controlled by operating each EDFA with gain compression. Steady-state operation occurs where gain equals inverse attenuation in the spans. designing the small signal gain of the amplifier to be larger than the attenuation in the cable section (Fig. 4.6). Thus, if for some reason the power should drop in any particular section of cable (Fig. 4.7), the following amplifier will see a smaller input power, resulting in an increase in gain so as to establish proper operating power in the rest of the optical path. (Note: Strictly speaking, this simple automatic gain control (AGC) argument holds only for narrowband operation. For wide-band amplifier chains, any mismatch between the amplifier’s gain and the span’s attenuation results in gain tilt.) 160 Neal S Bergano . Repeater output Power r I, 7 ti f r r , 1 - _ _ _ _ ____-------- --- --- ---- r Normal Output Power _ ____ _ _ _ _ _ _ _ ___ ---- --- --- - Position of Repeater in System Fig. 4.7 Repeater output powers versus transmission distance. The repeater’s output power is controlled by gain compression in the amplifiers. The real challenge is to equalize the power over a wide optical bandwidth to allow for many WDM channels. Here we are not as fortunate as with the total power control given by the EDFA gain compression. Most of the gain equalization is accomplished using passive gain-flattening filters placed along the length of the amplifier chain.9These filters can be packaged with the individual amplifiers to provide the majority of the gain equalization. Depending on the required level of equalization, additional “clean-up” filters can be placed at regular intervals (i.e., once every 10-20 amplifiers) along the amplifier chain. Figure 4.8 shows the gain shape of a single-stage EDFA before and after gain equalization. This amplifier was designed to have usable gain throughout the entire C-band. The amplified spontaneous emission (ASE) noise accumulation can be controlled by properly selecting the distance between amplifiers. In a long transmission line, ASE noise generated in the EDFAs can accumulateto power levels similar to the data-carrying signal. The accumulated noise can influence the system’s performance by reducing the level of the signal and signal-tonoise ratio. The noise power out of an optical amplifier is proportional to the amplifier’s gain and is given by:’O where hv is photon energy, g is gain in linear units, nsp is the excess noise factor related to the amplifier’s noise figure, and Bo is optical bandwidth. For a chain of identical amplifier spans, the total accumulated ASE noise out of the N* amplifier is simply NPnoise (assuming no signal decay caused by noise accumulation). As a consequence, the spectral density of the accumulated noise at the end of the system depends on the repeater gain and fiber loss. Consider a linear system where we treat only noise accumulation, with a fiber loss of 0.2 dB/km. 4. Undersea Communication Systems Erbium Doped Fiber V 161 xmxwx Isolator Tap Coupler Gain Flattening Filter Signal out * Wavelength Selective *O 1 Coup’er 1525 1530 1535 1540 1545 1550 1555 1560 1565 1570 Wavelength (nm) Fig. 4.8 Measured gain vs wavelength for a full C-band EDFA. The gain is measured w t and without the gain-flatteningfilter for a “flat” input power. ih A 150-km repeater spacing would require 30-dB amplifiers, whereas a 50-km spacing needs three times as many 10-dB amplifiers. A 30-dB amplifier (1000x gain) generates about 100 times more noise per unit bandwidth than a 10-dB gain amplifier (lox gain). Thus, the 30-dB gain system would have 33 times more noise than the 10-dBgain system. The relationshipbetween accumulated noise and amplifier gain imposes an interesting engineering tradeoff; longer systems require shorter repeater spacing to keep the same output signal-tonoise ratio (SNR). Gordon and Mollenauer described this excess noise generated as a consequence of the amplifier’s gain.“ The excess noise is the factor by which the amplifier’s output power must increase to maintain a constant received SNR. Lichtman embellished this to include excess loss in the amplifiers.I2The excess noise is given by: Excess Noise = g s - 1 (4.2) Bhg ’ where is post-amplifier loss. The /3 term is added because in a real amplifier there is always some loss that follows the erbium-doped fiber such as a gainflattening filter or an isolator. The most important result of adding the postamplifier loss is that the optimum repeater spacingis not zero (pure distributed gain), rather it occurs between 10-20 km.In addition, other effects such as the ih difficulty in making low noise figure optical amplifiers w t low gain and high output power tend to make the optimal repeater spacing even larger. 162 Neal S. Bergano The total accumulated noise power depends on the amount of noise generated at each amplifier (Eq. 4.1) times the number of amplifiers in the system. As stated previously, the automatic control obtained through the amplifier’s gain saturation fixes the total output power. This total power contains both signal and accumulated noise. Because the total power is fixed and noise accumulates, then the signal’s power must decrease as the signal propagates down the amplifier chain.13As an example, consider a 6000-km-long amplifier chain with 120 amplifiers spaced every 50 km. From Eq. 4.1 we know that a 10-dB net gain amplifier with nsp = 1.4 will generate about 16 p.W of noise power over the 40-nm C-band. Thus, at the end of the amplifier chain there is a total noise power of -2 mW (+3 dBm) of noise power. A simple but useful estimate of the signal-to-noise ratio at the output of a chain of N amplifiers is: SNR, = PL NF ghu B,N ’ (4.3) where PL is the average optical power launched into the transmission spans, NF is noise figure, and g is the amplifier’s gain, and Bo is optical bandwidth in Hertz. Equation 4.3 assumes that all the amplifiers are identical, and that there is no signal decay from noise accumulation. Wide-band EDFAs are not ideal amplifiers that have a simple gain shape. The actual output power over the amplifier’s optical bandwidth is somewhat dependent on the spectral distribution of input power, which is known as spectral h~le-burning’~ (SHB). The result of SHB is that the output power at any individual wavelength will be influenced by other signals within some characteristic spectral range, or “hole width.” The influence of SHB can be useful for gain equalization. Unfortunately, it can also limit the ability to perform transmitter preemphasis15to correct for unequal SNR in a WDM system. Figure 4.9 shows a measurement of SHB on an amplifier chain and the effect it can have on 01 . -301 1544 1548 1552 1556 Wavelength (nm) -40 1544 1548 1552 1556 Wavelength (nm) Fig. 4.9 The influence of SHB is observed in the optical spectra at the output of a 6200-km amplifier chain. When pre-emphasis powers are changed, the SNR is altered for neighboring channels, but not for channels far away. 4. Undersea Communication Systems 163 the output signal-to-noiseratio in a WDM system. The left-hand side of the figure shows the spectrum of a signal before and after a single channel was turned on. The background ASE noise shows that the gain in the vicinity of the channel is diminished. The right-hand side of the figure shows two optical spectra after 6200 km; one with 32 channels, and the second with 28 channels. If the system had ideal, homogeneous amplifiers with simple gain shapes, we would expect the SNR of each channel to increase by 0.6 dB = 10 log(32/28) when the channel count is decreased from 32 to 28. However, SHB has a strong influence on the SNR before and after channels 1-3 and 5-7 were turned off. For example, the SNR of channel 4 increased by 3.6 dB, while the SNR of the long wavelength channels are unchanged. 4.4 Dispersion and Nonlinearity Management Chromatic dispersion causes different wavelengths to travel at different group velocities in single-mode transmission fiber.16 For those with an electrical engineering background, the chromatic dispersion can be considered similar to an “all-pass” filter with a nonlinear phase characteristic. High-speed optical data signals require low end-to-end dispersion to ensure good waveform fidelity. The approximatedispersion limit for a nonreturn-to-zero (NRZ) signal produced from a chirp-free external modulator is given by: D(ps/nm) = 104,000’ B2 ~ (4.4) where D is chromatic dispersion in ps/nm and B is bit rate in Gb/s.17 Thus, 10Gb/s NRZ operation requires dispersion values less than about 1000ps/nm to ensure low dispersion penalty, corresponding to -60 km of conventional single-mode fiber. This value is even smaller for signals with larger optical bandwidths, such as a return-to-zero (RZ) signal, or a directly modulated distributed feedback (DFB) laser. The chromatic dispersion and strength of the nonlinear index of the transmission fiber can limit the system’s performance in terms of the bit error ratio (BER) of a single channel, as well as the ultimate capacity that can be transmitted using WDM techniques. The important manifestations of the fiber’s nonlinear index include self-phase modulation, cross-phase modulation, and four-wave mixing.’* These terms are used to describe how intensity fluctuations can modify the signal‘s optical phase and the intermodulation between channels. The nonlinear index of a single-mode fiber is expressed as: 164 Neal S Bergano . where no is the linear part of the refractive index, N2 is the nonlinear coefficient (-2.5 x cm2/Watt),*P is the optical power in the fiber, and A@ is the effective area over which the power is distrib~ted.‘~ the strength of Since optical nonlinearity on a local level is quite small, the deleteriouseffects of the nonlinear index occur over many tens to hundreds of kilometers. This means that the nonlinear interaction lengths are long and that the local chromatic dispersion is an important factor in the system’s performance. Signals at the fiber’s zero dispersion wavelength (or wavelengths placed symmetricallyabout the zero dispersion wavelength) travel at the same group velocity. Hence a signal at A0 and another signal (or ASE noise) will always be exactly phase matched with one of its mixing products, which will be symmetrically on the other side of ho. Under these conditions, both the signal and noise waves have long interaction lengths during which they can exchange energy, which will degrade the performance of the signal. On the other hand, large local dispersion can reduce phase matching or the propagation distance over which closely spaced wavelengths overlap, and can reduce the amount of interaction through the nonlinear index in the fiber. Thus, in a long undersea system, signal distortions caused by the fiber’snonlinear index and the dispersion can be managed by tailoring the accumulated dispersion so that the phase-matching lengths are short, and the end-to-end dispersion is small. This technique, known as dispersion amounts to constructing the amplifier chain by concatenating optical fibers with specific lengths and opposite signs of dispersion. This is shown in Fig. 4.10 for a single “dispersion period” where the majority of the fiber used in an amplifier chain has a dispersion value of -2ps/nm-km, and is compensated by a conventional single mode fiber with 17ps/nm-km. This dispersion-mapping technique satisfies the engineering trade-off in the amplifier chain to have both large local dispersion and small end-to-end dispersion. The dispersion map shown in Fig. 4.11 is effective at improving the performance of channelslocated close to the amplifierchain’s averagezero dispersion wavelength. Unfortunately, optical channels located far away from the zero dispersionwavelength necessarily accumulate significantdispersion. The accumulated dispersion of these channels requires an additional compensation at the terminals. Figure 4.1 1 shows the accumulated dispersion for the amplifier chain in the previous figure duplicated over many dispersion periods. Here the outer channels accumulate in excess of 10,000pdnm for a transmission distance of 9000 km. Thus, the pulses at these outer wavelengths experience the greatest amount of dispersion along the transmission line and will broaden and overlap with many neighboring time-slots. The degree to which simple dispersion compensation in the terminal can recover the original pulse shape is inversely related to the accumulated nonlinear phase shift (i.e., the more + * The approximate sign is used since the value of nz is slightly fiber-design dependent. 4. Undersea Communication Systems Long wavelength channel 165 -500 -1 000 0 100 200 300 Length (krn) 400 500 -2ps/krn-nm +17ps/km-nrn -2pslkm-nm Fig. 4.10 The accumulated dispersion along a 500-km amplifier chain. The “dispersion map” consists of a dispersion value of -2 ps/nm-km for the majority of the fiber, compensated by a fiber with +17ps/nm-km. I 10,000 0 .t $2 a , 5,000- I a n 6 m U a , c 0 -- .. . 2 2 -5,000- 2 -1 0,000 - Terminal’s Dispersion Compensation I 0 2000 dmap6 4000 6000 Distance (km) 8000 10000 Fig. 4.11 Accumulated dispersion as a function of distance for several channels of a WDM transmission system. The average slope of the dispersion is 0.075 ps/km-nm*. The minimum and maximum wavelengths are f l 5 n m from the zero dispersion wavelength (ho). linear the system, the more accumulated dispersion that can be tolerated in the transmission line). In practice, it has been found that the performance of long transmission lines can be improved by placing half of the needed dispersion at the transmit end and half at the receiver21(sometimes known as 50/50 compensation). 166 Neal S. Bergano Trans. #3 ps Trans. #N-I Trans. #2 / Trans. #N Fig. 4.12 Block diagram of a WDM transmitter using “pair-wise” orthogonal polarization launch. The simplest method of reducing the deleterious effects of the fiber’s nonlinear index is simply to reduce the operating power of the WDM channels.22 Unfortunately, reduced power leads to poorer SNR and potentially, degraded bit-error ratios. Poor bit-error ratio performance can be greatly reduced by using forward error correction codes in the terminals (see Section 4.7). These effects can also be reduced by wisely choosing the transmission format used at the transmitter (see next section on transmission formats). For example, synchronous phase modulation added at the transmitter allows operation at greater launch power, particularly for channels located far away from the amplifier chains zero dispersion wavelength. One very effective method of reducing the interactions between WDM channels is to launch the even numbered channels orthogonally polarized with respect to the odd numbered channels (Fig. 4.12). This “pair-wise” orthogonal method23greatly reduces nonlinear crosstalk from four-wave mixing and linear crosstalk such as nonideal extinction in demultiplexingdevices. Figure 4.13 shows an example of two-tone four-wave mixing through a 500-km amplifier chain.24When two CW tones were launched in the same polarization, the two interacted to form many mixing products. However, the mixing was greatly reduced when the two were launched orthogonally p ~ l a r i z e d . ~ ~ 4.5 Modulation Formats The choice of modulation format has an important performancekconomic trade-off. From a performance standpoint, the best modulation format is dictated by many system parameters, such as system length, the fiber type, the dispersion management, and the optical bandwidth. For example, the chirped 4. Undersea Communication Systems Parallel Launch Orthogonal Launch 167 9 A ._ 4-4 1 Af-5GHz - 3 - S u) c a Frequency Separation Frequency Separation Fig. 4.13 Two-tone four-wave mixing in a 500-km dispersion managed amplifier chain. Two CW tones are launched with a frequency separation of 5 GHz. return-to-zero format is useful for lO-Gb/s transoceanic length WDM systems operating with the dispersion maps shown in Fig. 4.10 and Fig. 4.1 1. Although lightwave systems are known to be at the cutting edge of telecommunication technology, the basic signaling format is quite simple. Binary ones and zeros are sent by the presence or the absence of light pulses, much as one would use a flashlight with an on/off switch to convey binary data. Of course, the “hi-tech’’ aspect is that trillions (10l2) of these pulses can be transmitted per second through a single optical path that is mega-meters in length. One of the key challenges of the system design is to transmit a pulse shape that will survive the long transmission distance in the presence of dispersion, fiber nonlinearity, and the added optical noise from the amplifiers. Transmission based on this simple on/off pulse scheme is referred to a unipolar pulse system.26 When the shape of the light pulse used is a rectangular pulse that occupies the entire bit period, the format is referred to as a Non-Return to Zero format, or NRZ for short (Fig. 4.14). The name NRZ attempts to describe the waveform’s constant value characteristic when consecutive binary ones are sent (Fig. 4.15). A string of binary data with optical pulses that do not occupy the entire bit period are described generically as RZ, or Return-to-Zero. Figure 4.16 shows the power spectral density of a binary data stream formed from a unipolar signaling pulse, similar to the top of Fig. 4.15. The R F spectrum of the data signal P ( f ) is given in Eq. 4.6, which assumes that the data pattern is composed of random, uncorrelated bits.27The R F spectrum is formed from the Fourier transform of the signal pulse H ( o ) in accordance to the following: 168 Neal S. Bergano T=1IB Different unipolar binary C O ~ J signaling pulses. NRZ, nonreturn to zero; Fig. ’ RZ, return to zero. NRZ I I I Rz t Fig. 4.15 Data waveforms for different pulse formats. 0 B 2B 38 Fig. 4.16 Spectral content of a rectangular pulse and the RF spectrum of an NRZ signal. where T is the bit period andp is the probability of a “1 ” in the original binary sequence. The bottom part of the figure shows an actual measurement of an NRZ signal on an RF spectrum analyzer. Here we note that there is no RF power at integer multiples of the bit-rate frequency for the NRZ format. 4. Undersea Communication Systems 169 A robust transmission format that can propagate in the presence of dispersion, fiber nonlinearity, and accumulated noise is the chirped return-to-zero (CRZ) pulse shape.24Figure 4.17 shows a block diagram of a CRZ transmitter and associated eye diagrams. CW laser light is modulated by NRZ data stream at the required bit rate and is shaped by bit-synchronous amplitude modulator. At the 100%amplitude modulation level, the pulses usually occupy about 33% of the bit slot. Prechirping is accomplished using a bit-synchronous phase modulator, with an adjustable peak-to-peak level and phase relative to the center of the bit. Mathematically the complex amplitude of CRZ pulses is given by: A = &cos ( a n sin ( n ~ texp (ibcos ( 2 n ~ t , ) )) ) (4.7) where Pp& is the “ones” peak power, a is the level of amplitude modulation, b is the phase modulation index in radians, and F is the bit rate. The typical pulse width at 100% amplitude modulation level for a lO-Gb/s signal is 32 ps, and the peak power is 5.4 times higher than the time average power. The added bandwidth of the CRZ pulse together with the local dispersion and nonlinear phase shift in the amplifier chain determines the temporal evolution of the pulse. CRZ pulses periodically expand and contract as they traverse the different signs of local dispersion of the dispersion map. A single pulse might spread by several bit periods; thus making the actual data pattern “seem” to disappear at certain points in the system. For example, Fig. 4.18 shows the result of a calculation for peak intensity and pulse width of a CRZ pulse as it travels down an amplifier chain.28 The pulse width experiences Amplitude Amplitude Amplitude LASER AMPLITUDE OPTICAL OUTPUT DATA SOURCE 4 CLOCK Fig. 4.17 CRZ transmitter and associated waveforms. 170 Neal S. Bergano .0 c .Y w 6 $ 4 n g 2 0 g 8 5 8 ? l 4 0 0 2000 4000 Length (km) 6000 8000 Fig. 4.18 CRZ pulse width and peak intensity along the transmission path calculated for a channel operating43nm above the averagezero dispersion wavelength. The values of intensity and width are normalized to the input values. large changes, and for the channels operating away from the zero dispersion wavelength, the pulses completely overlap with their neighbors in time. The dispersion map for a transmission line is designed so that the pulses will have minimum temporal distortion at the receive terminal, after dispersion compensation. When the system is configured with the zero dispersion wavelength near the center channel, the side channels accumulate significant amounts of dispersion. Placing an appropriate amount of pre- and postdispersion at the transmitter and receiver individually optimizes the transmission performance of these channels (Fig. 4.1 1). In addition, the phase modulation index can be optimized to achieve the best performance. Figure 4.19 shows a calculated eye opening for a lO,OOO-km-long, 16-channel WDM transmission operating at 10.7 Gb/s, with 0.6 nm channel separation. The eye opening is given as a function of the phase modulation index for different values of accumulated dispersion (i.e., the amount of dispersion compensation placed at the terminals). The calculations indicate that the added phase modulation improves the eye opening by several dB. The improvement in performance by the CRZ format has also been demonstrated in transmission experiments. Figure 4.20 shows the measured Q-factor versus distance for CRZ, RZ, and NRZ modulation formats in a 64 x 12.3-Gb/s WDM transmission experiment up to 9000 km long.29For the CRZ case, one RAD phase modulation is used and the pre- and postdispersion compensation are optimized for each case. As shown in Fig. 4.20a, CRZ and RZ provide similar performance at distances up to about 5000 km. However, as the distance and with that the accumulated dispersion and nonlinear IS1 start to increase, CRZ eventually becomes superior to RZ, resulting in 1.5 dB higher Q-factor at 9000 km. Whereas phase modulation reduces the 4. Undersea Communication Systems 0v 171 5-0 p .c -2- 0" W a , a, w > .e r Y ;6-8 I w A - A 0 =-6nm I I I +compensation to 0 pslnmlkm +compensation to -1 00 pslnmlkm t compensation to 100 pslnmlkm Fig. 4.19 Relative eye opening versus phase modulation index for a channel located 6 nm lower than the system's average zero dispersion wavelength. The values of residual end-to-end dispersion are given in the legend. 22 v D m 18 14 10 1000 3000 5000 7000 9000 11.5 I (b) 0.5 1 1.5 2 Distance (km) Phase Modulation Index (RAD) Fig. 4.20 (a) Q-factor versus distance for channel 2 of 64 with different modulation formats (CRZ with one RAD phasemodulation). (b) Q-factorversus phase modulation index for channel 2 at 7900 km measured for different phase modulations. nonlinear ISI, it also increases the signal spectral width and with that the interchannel linear cross-talk. The optimum amount of phase modulation for CRZ is thus a compromise between the signal power and the accumulated dispersion on one side and the channel spacing on the other. As an example, Fig. 4.20b shows Q-factor versus phase modulation for channel 2 at 7900 km. As is clear in this figure, the optimum phase modulation for this case is in fact 1.5RAD, resulting in 1.3 dB higher Q-factor compared to RZ. However, for an increased signal power and/or channel spacing, the optimum amount of phase modulation and the corresponding improvement will also increase. Many in the optics and physics communities often wonder if optical solitons are used in undersea cable systems. The optical soliton is another pulse waveform that has been widely studied for long-distancedata transmis~ion.~~ There 172 Neal S Bergano . was an active debate in the early 1990s between using NRZ or solitons for the first single channel EDFA-based transmission systems. NRZ had the advantage of compatibility with existing systems, and solitons were thought to have the advantage of higher single channel bit rates. At that time it was decided to use the NRZ format for the first systems then switch to the “higher” capacity format for the next generation systems. When WDM techniques became available, new dispersion maps were invented that strongly reduced the four-wave mixing between NRZ channels, thus eliminating one of the big advantages of solitons. This made adding NRZ wavelength channels the preferred path for greatly increasing capacity, rather than making incremental improvements by increasing the bit rate per channel. In the intervening years a few eventschanged the debate. The understanding of the basic physics of optical propagation has increased, driving the evolution of both the NRZ and soliton transmission formats. The NRZ format evolved into RZ and CRZ, and soliton transmission evolved into dispersion-managed solitons.31The modulation format debate changed to a question of, on the one hand, purposely using the fiber’s nonlinearity to help guide data pulses, or on the other hand, to reduce the system’s nonlinear behavior and operate the system in a quasi-linear region. Clearly, designers working at multiples of 10 Gbls have chosen the quasi-linear approach of managing the fiber’s nonlinearity. Many variants of the simple NRZ and RZ formats have appeared, such as CRZ, alternating-phase RZ, duo-binary, vestigial side band, etc. All of these formats attempt to optimize some aspect of the transmission system. For example, some of the formats attempt to optimize spectral efficiency by transmitting a small optical bandwidth. 4.6 Measures of System Performance The performance of a digital lightwave system is specifiedusing the Q-fa~tor.~* The Q-factor (adapted from Personick’s work on calculating the performance of receivers in lightwave links33)is the electrical signal-to-noise ratio at the input of the decision circuit in the receiver’s terminal. This is shown schematically in Fig. 4.21 using a typical RZ eye diagram. For the purpose of calculation, the signal level is interpreted as the difference in the mean values, and the noise level is the sum of the standard deviations. The Q-factor is formed by the following ratio: 4 lPl - Pol a1+a0) where POand p1 are the mean values of the “zeros” and the “ones,” and a 0 and a are their standard deviations at the sampling time. 1 4. Undersea Communication Systems 173 Eye Diagram ............ ... = .. gJ Sampling time a, ” P 1 Decision Level ’ s I3 w .. Po I Q 5 Fig. 4.21 A typical received RZ eye diagram for a lightwave system. A voltage histogram is schematically shown to indicate the parameters that are included in the definition of Q-factor. The Q-factor is related to the system’s bit error ratio through the complementary error function, given by:* where M or in terms of the more standard error function erf( . ): (4.10) where The Q-factor given in Eq. 4.8 is a unitless quantity expressed as a linear ratio, or it can be expressed in decibels as 20log(q). The factor of 20 (or 10log (q2))is used to maintain consistency with the linear noise accumulation model. For example, a 3-dB increase in the average launch power in all of the spans results in a 3-dB increase in Q-factor (assuming signal-spontaneous beat noise dominates and ignoring signal decay and fiber nonlinearity). The *Here I use the definition of erfC(n) as given in MATLAB@ rather than the definition originally given in reference [32]. (MATLAB@is a product of The Mathworks Inc.) 174 Neal S. Bergano 1 Q-Factor (linear) 2 3 4 5 6 7 0- -4Log (BER) -8-10- -1 2-_ 0 5 10 Q-Factor (dB) 15 Fig. 4.22 Bit-error ratio as a function of Q-factor. relationship between the Q-factor and bit-error ratio is shown in Fig. 4.22. A convenient relationship to bear in mind is that a BER of requires a Qfactor of 15.6 dB (or a linear ratio of 6). A useful approximation for converting BER back into Q - f a ~ t o is~ ~ by: r given Let t = J-2 log, (BER) [ 2.307 + 0.2706t System margin is the amount that the Q-factor (measured in dB) exceeds the required value for a given bit-error ratio. In long-haul lightwave systems the BER is set by a combination of the electrical signal-to-noise ratio of the data signal at the decision circuit and any distortions in the data’s waveform. Optical noise, fiber chromatic dispersion, polarization mode dispersion, fiber nonlinearities, and nonideal settings in the transmitter and receiver degrade the BER. Also, the BER can fluctuate with time due to polarization effects in the transmission fiber and the amplifier’s components (see Section 1.8). The most accurate methods of measuring margin are based on bit-error ratio measurements. When practical, the simplest method is to measure the BER on the ampliiied line, convert it to a Q-factor using Eq. 4.1 1, and state the margin as the difference between the measured Q-factor and the requirement. This technique is possible in some systems that use forward error correcting codes in the terminals (see next section). However, if the line bit-error ratio is below the practical measurement limit (about 10-13), then other methods are needed. For systems operating in this “error free” region, the most accurate method is the decision-circuit method of measuring the Q - f a c t ~ r . ~ ~ This measurement technique includes the intersymbol interference present in the regenerator’s linear channel, as well as that generated in the system from dispersion and fiber nonlinearity. 4. Undersea Communication Systems 175 * v -0.4 -0.2 00 . 0.2 0.4 Decision Level (volts) Fig. 4.23 Typical Q-factor measurement for 5-Gb/s, 9000-kmoperation. The data shows the bit-error ratio vs the decision threshold. The solid lines show the fit of Eq. 4.12 to the data. The decision-circuit method of measuring Q-factor involves three steps. First, the system’s BER is measured as a function of the decision circuit’s threshold voltage (this voltage is shown on the vertical axis in Fig. 4.21). Figure 4.23 shows data for a typical Q-factor measurement for a 9OOO-km, 5-Gb/s transmission system operating at a Q-factor 7.2: 1 (linear ratio) or 17.2dB. Second, the measured data is fit to the ideal curve of BER as a function of threshold voltage, as given by: The form of Eq. 4.12 assumes Gaussian noise statistics, and the curve-fitting 00, operation results in calculating values for PO,ply and q.In the third and final step, the Q-factor is formed using the fitted values for p and CJ in Eq. 4.8. It is well known that the electrical noise at the decision circuit is not exactly Gaussian,35however, the Gaussian approximation can lead to close BER estimates36Figure 4.24 shows the measured voltage histogram of a detected optical signal emerging from a long lightwave system operating at 5 Gb/s. For this measurement, 1 million voltage samples were recorded for a zero bit and a one bit in a 27-1data pattern. The non-Gaussian probability density function is apparent when the actual density is compared with a best-fit Gaussian. The measurement of Q-factor as described previously measures only a subset of the distributions located near “inside” rails of the received eye or the voltages that are close to the decision circuit. Thus, the insides of the edges of the eye are fitted with an equivalent Gaussian function, and the underlying SNR is extrapolated from the fit. Mazurczyk and Duff have identified the inability of the decision circuit Q-factor measurement to measure large margins using long data Pattern-dependent effects cause the Q-factor measurement to underestimate 176 Neal S Bergano . a, P -1 I -3 r : t -4 -1 a/+ Gaussian Fit * 1.o .o -0.5 0 Voltage 0.5 Fig. 4.24 Typical voltage histograms of a 5-Gb/s NRZ data signal for the “ones” and “zeros” rails. One million voltage samples are recorded in each bit using a digital oscilloscope. Eye Diagram -5 ’ ‘.., -1 0 % Decision Point Fig. 4.25 An expanded view of the upper rail of an eye diagram showing ISI. The resulting BER vs decision level can have a slope change causing the Q-factor measurement to underestimate the actual value. the actual Q-factor for long pseudorandom data patterns with large margins. The root cause of the effect is shown schematicallyin Fig. 4.25. In the figure, the upper part of the received eye diagram is expanded to show the intersymbo1 interference (ISI). Here the pattern dependenceof the data causes different bits to have different mean voltages at the decision circuit’s timing point. The resulting BER vs decision voltage does not followa simple Gaussian characteristic, rather it follows the rules of total probability given each bit’s probability density function. For large margins, the resulting curve can exhibit a slope 4. Undersea Communication Systems 177 change at BERs less than what is practical to measure, and the extrapolated BER (and Q-factor) is then underestimated. In practice, this is not a serious limitation to characterizinga working system, because typical values of beginning of life optical margins are less than 5 or 6 dB. A practical engineering fix to this problem is to measure the Q-factor for a series of word lengths. It is often useful to know the ideal Q-factor as a starting place for system calculations. Marcuse describedthe ideal Q - f a ~ t oconsideringonly accumur~~ lated noise impairments in terms of the optical SNR. This formalism can be embellished to include other effects such as finite extinction ratio in the transmitter and other pulse shapes. For example, assuminga NRZ data format with extinction ratio (I), the ideal Q-factor is given as: where y= [d], 2(1 - where& andBE are the optical and electricalbandwidths in the receiver. Later in the section on system design (Section 4.9) we will use Eq. 4.13 to calculate the expected Q-factor using the SNR given by Eq. 4.3 as a starting point for the impairment budget. 4.7 Error Correcting Codes Up to this point our discussion has focused on the generation of optical signals, propagating them over fiber cables, and detecting them at the far end; that is to say the physics of getting optical data bits across the system. The topic of forward error correction (FEC) codes approaches the subject of data transmission from a more classical communications channel perspective, where information is transmitted over a nonideal noisy channel. FEC adds redundancy or extra information to the original input data before it is converted to an optical signal (Fig. 4.26). The decoder in the receiver uses this redundant information to identify and correct bit errors caused by the transmission channel. The result of this added information is that the actual transmitted bit rate is larger than the rate of the input data. For example, a 10-Gb/s system employing a 23% FEC overhead has a transmitted data rate of 12.3Gb/s. 178 Neal S. Bergano Data In 4 FEC Encoder Transmit Terminal Transiitter ,-I ......................................................................... Channel: Added Noise Chromatic Dispersion Fiber's nonlinear index , ......................................................................... Data Out : I ........................................................................ Receive Terminal ; Fig. 4.26 A transmission system using FEC codes in the terminals. Redundancy or extra bits are added at the transmit end before the data is transmitted into the system. The FEC-enabled receiver identifies and corrects bit errors. FEC codes can dramatically improve the performance of lightwave transmission systems by adding system margin.39 For example, a BER of lo-" requires a Q-factor of 17dB (calculated from Eq. 4.1 1). Figure 4.27 shows the output bit-error ratio as a function of input Q-factor for 7% and 23% FEC codesa For the 23% code shown in the figure, an output BER of 1O-I' is achieved with input Q-factor of about 8.4 dB, or a gross coding gain of 8.6 dB (i.e., 17dB - 8.4 dB)! The net coding gain will be reduced because the bit rate of the system increased. We can estimate the penalty of increasing the bit rate assuming an increased noise bandwidth of the receiver by 10 log (1.23)' or about 0.9 dB (assuming that the penalty scales linearly with the bit rate). This gives a net coding gain of about 7.7 dB. The FEC coding gain allows the target Q-factor (or line bit-error ratio) to be greatly relaxed, which can be used to improve the transmission system in several ways. For example, the system can be made more linear by operating the WDM channels at a lower average power. The resulting degraded error ratio (caused by the lower SNR) can be removed with the FEC. This more linear system could be used to transmit higher capacity by placing WDM channels closer together and/or using wavelengths that are farther away from the fiber's zero dispersion wavelength. Other benefits could be increased repeater spacing, longer transmission distances, or a relaxed tolerance on component specifications. The solid lines in Fig. 4.27 are theoretical calculations of the FEC code's performance assuming additive white Gaussian noise and ideal data streams without any intersymbol interference. Although these assumptions are not completely true, the measured data points are in good agreement with the theory. Figure 4.28 shows the results of a study that was performed to test the In ~ validity of the ideal c a l ~ u l a t i o n s ~this study the error correction capability 4. Undersea Communication Systems Input BER 179 2.7e-002 I 2.7e-003 3.6e-005 I IO-‘ I 0-3 Output BER I 0-5 I 0-7 I 0-9 IO-” 6 7 8 9 10 Input Q-Factor (dB) 11 12 Fig. 4.27 Output bit-error ratio as a function of input Q-factor for three cases: (1) No FEC, (2) 7% single-stage Read-Solomon code, and (3) 23% concatenated Reed-Solomon code. The solid lines are theoretical calculations and the symbols are measured points. 6 7 8 9 10 Input Q (dB) 11 12 6 7 8 9 10 Input Q (dB) 11 12 Fig. 4.28 Output BER vs input Q-factor for two different forms of distortion: (A) BER is degraded by added noise, (B) BER is degraded by waveform distortion caused by the fiber’s nonlinear index. Eye diagrams are shown in the inserts. for a 14% Reed-Solomon code was measured under two different conditions. In the first, data were collected for a noise-loaded system, and as expected, the measured data points fit the theoretical prediction. In the second experiment, waveform distortion arising from chromatic dispersion and the nonlinear behavior in the transmission line degraded the input BER. Even in this case the measured data points are in agreement with the simple theory. FEC codes have the added benefit of simplifyingthe measurement of margin in an operating system. Many FEC decoders can report how many errors have 180 Neal S. Bergano been corrected. This can give an accurate measurement of the actual bit-error ratio on the line. As stated in the previous section, the most accurate way of measuring margin is to know the BER on the line (thus knowing the received Q-factor). Alternatively, if the FEC decoder reports error-free operation on the line, then it is known with a high degree of confidence that the system is operating with a minimum margin equal to the FEC coding gain. 4.8 Polarization Effects Several polarization effects in lightwave systems can combine to degrade the performance of long-haul lightwave systems42(see Table 4.1). These effects can both reduce the mean received SNR43344 cause the S N R to fluctuate and with time.45Standard telecommunication optical fibers do not maintain the state-of-polarization of the transmitted signal. Random perturbations along the fiber’s length can couple the transmitted signal between the two polarization modes and give rise to the time-varying state-of-p~larization.~~ The unstable state of polarization interacting with the polarization dependence in the transmission line can lead to a fluctuating Q-factor at the receive-terminal. To accommodate this fluctuating performance, additional margin needs to be designed into the system (see Section 4.9 on system design). Figure 4.29 gives a graphical representation of how polarization dependence can result in a time-varying SNR. Consider an amplifier chain where each amplifier has some PDL. During a favorable time, the states of polarization at the inputs to the amplifiers could drift to coincide with a majority of the low-loss axes of the PDL in the amplifiers. At these times the SNR will be high. Alternatively, at unfavorable times, a majority of the input polarizations could coincide with the high-loss axes, producing lower SNR. The same type of effect is true for polarization mode dispersion in the transmission fiber Table 4.1 Important Polarization Effects Found in Lightwave Systems ~ ~~~~~~~ SOP (State Of Polarization) drift PDL (Polarization Dependent Loss) PMD (Polarization Mode Dispersion) PHB (Polarization Hole-Burning) The state of polarization evolves over time, caused by temperature and stress changes of the transmission fiber. A component’s attenuation has a small dependence on the signal‘s polarization. The group delay through the transmission fiber or component is polarization dependent. The amplifier’s saturated output power is slightly larger for light in the orthogonal polarization from the saturating signal. 4. Undersea Communication Systems Time Varying Birefringence (PMD) 181 : : + Input I ~~ ' F u L H H SNR Low Signal A rl' H T A - .,. , - L ~ow-Loss~xis of PDL Fig. 4.29 A transmission line containing polarization-dependent elements, such as polarization-dependent loss and polarization mode dispersion. The unstable state of polarization caused the received SNR to fluctuate with time. I I 1 I I I 0 1 2 3 Time (hours) 4 5 Fig. 4.30 Q-factor vs time for a 5-Gb/s signal transmitted over 7200 km. The insert shows a histogram of the Q-factor data. and the amplifier's components, only in this case we are more concerned with waveform distortions than SNR. Figure 4.30 shows a measurement of the Q-factor fluctuations in a WDM transmission experiment for 1 of 16 5-Gb/s channels after 7200 km. Polarization hole-burning results from an anisotropic saturation created when a polarized saturating signal is launched into the erbium doped fiber. The PHB effect was first observed as an excess noise accumulation in a chain 182 Neal S Bergano . of saturated E D F A s , ~ ~ was later isolated in a single amplifier and idenand tified as PHB.48The gain difference caused by PHB is quite small in a single amplifier, with a typical value of about 0.07 dB for an amplifier with 3 dB of gain compression. Although PHB is a very small effect in a single EDFA, its effect on the overall performance of an optical amplifier transmission line can be several dBs in received Q-factor. PHB can cause the amplified spontaneous emission noise to accumulate in the polarization orthogonal to the signal faster than along the parallel axis (Fig. 4.31). Noise accumulates faster than would be predicted by simple noise accumulation theory, and as a result, the signal decays at the expense of the noise. Fortunately, the deleterious effects of PHB can be avoided by depolarizing the total signal propagating in the amplifier chain at a rate faster than the EDFA’s gain recovery time. When the signal’s degree of polarization is low, there is no preferred polarization axis for the gain to be depleted, and the transmission performance returns to the expected value. Depolarizing the total signal in a WDM system can be performed passively by allowingchannels to take on random SOPSor actively by modulating the channel’s polarizations. In a WDM system with many optical channels, the unavoidable PMD in the amplified line causes the different channels to disperse in polarization, which leads to a natural decrease in the degree of polarization. This process can be accelerated by purposely launching the channels in a “pair-wise” orthogonal manner.23 Actual Signal Decay with Distance (km) Rg. 4.31 PHB causes the noise in the orthogonal polarization to have an excess gain. If left unchecked, this could lead to noise accumulation faster than would be predicted with simple noise accumulationtheory. 4. Undersea Communication Systems 183 Alternatively,the state of polarization can be activelymodulated or “scrambled” at the transmit end of the system. Because the dynamics of the EDFA gain are relatively polarization scrambling the signal at a rate faster than the EDFA can respond to eliminates any excess noise accumulation caused by PHB. The characteristic time constant associated with PHB is similar to the time contents that govern the large signal response of the EDFA, or about 130-200 ~ s e c . ~ O reduce the negative effects of PHB on transmissionsystems, To the SOP of the transmitted optical data signal should be scrambled at a rate that is high compared to the amplifier’sresponse time. Therefore, polarization scrambling should interchange the optical signal between orthogonal polarizations at a frequency higher than 1/ 130 sec, or about 7 kHz. Polarization scrambling techniques were particularly important for the first single-channel optical amplifier systems where the degree of polarization launched into the system was potentially large. Performance improvements have been reported for slow-speed ~crambling~lg~~ (i.e., much lower than the bit rate), synchronous scrambling53 (equal to the bit rate), and high-speed ~ c r a m b l i n g(faster than ~~.~~ the bit rate). 4.9 System Design Thus far, we have reviewed several aspects of optical amplifier transmission technology used in undersea cable systems. This section attempts to put the pieces together by reviewing the design of a 32-channel by lO-Gb/s transatlantic 6000-km system. The goal of the system design is to have adequate end-of-life margin considering many of the factors presented thus far, such as degradations caused by optical noise, waveform distortions, Q-factor fluctuations, and system aging. Key design parameters are the repeater spacing, launch power, and the dispersion management of the amplified line (Table 4.2). A 6000-km system will require about 120repeaters spaced every 50 km. The term repeater is taken from the nomenclature of analog transmission systems. For our purposes, a repeater is the pressure vessel that houses the erbiumdoped fiber amplifiers. In ow example the repeater’s EDFA will have a net gain of 10 dB, assuming an average cable attenuation of 0.2 dB/km. A total launch power of about 11dBm (-4 dBm per channel) is required to produce enough margin. A systembandwidth of about 19nm is required for 32 channels spaced every 75 GHz (-0.6nm at 1550nm). The dispersion map shown in Fig. 4.10 is used to limit the negative effects of the fiber’s nonlinear index, and each transmitter will use the chirped return to zero format. The impairment budget (Table 4.3) is a design tool used to account for all of the expected impairments over the system’s lifetime. The starting point is the ideal Q-factor that is calculated considering only the received SNR, calculated for example using Eq. 4.3 and Eq. 4.13. From this starting point the 184 Neal S. Bergano Table 4.2 Key Design Parameters Parameter Value 6000 km 50 km Length Repeater spacing Repeater count Repeater gain Total launch power Repeater noise figure Channel spacing Amplifier bandwidth Dispersion period dispersion map (see Fig. 4.10) Dispersion slope Line rate (23% FEC overhead) Transmitter extinction ratio Receiver optical bandwidth Receiver electrical bandwidth 120 10dB 11dBm 4.5 dB 75 GHz 19nm 500 km 0.075 ps/km-nm2 12.3Gbls 13.7dB 50 GHZ 8.6GHz Table 4.3 Impairment Budget for a 32-Channel, 10-Gb/s, 6000-kmSystem ~ ~~~ Line 1 2 3 Parameter dB Value 17.8 4.3 2.0 1.o 1.o 9.5 8.5 1.o 4 5 6 7 8 Mean Q value (from simple SNR calculation) Propagation impairment Manufacturing variations and impairments Q-factor time variations Aging End-of-life Q-factor (1 - 2 - 3 - 4 - 5) Required Q-factor End-of-lifemargin (6 - 7) values of all expected degradations are subtracted. For example, the 4.3-dB value in line 2 includes effects arising from propagation impairments such as the fiber’s nonlinear index, added noise from optical reflections, and nonideal dispersion compensation. This value is obtained using very detailed computer modeling of the optical propagation in a transmission system56(which unfortunately is beyond the scope of this chapter). Line 3 gives a 2-dB allotment for manufacturing variations, which covers all of the realistic population distributions of the components, and the imperfect system assembly process. Line 4 gives 1dB for Q-factor fluctuation, and line 5 allots 1dB for system aging. 4. Undersea CommunicationSystems 185 The Q-factor target of 8.5 dB represents the FEC threshold value shown in Fig. 4.27. The values in the table were recalculated for different launch powers and span lengths until the 1 dB end-of-life figure was reached. Much of the terminal transmission equipment for undersea systems differs significantly from equipment for terrestrial applications because of the large system-length differences. The transmission terminals include equipment to condition digital data for transmission undersea, power feed equipment to provide DC power to the undersea equipment,and line monitoring equipment to diagnose the location of undersea cable cuts and other undersea faults. An important part of the undersea cable network's terminal is the power feed equipment used to supply electricalpower to the optical amplifierslocated in the undersea repeaters. The active components (such as pump lasers) are powered by running a DC current through the copper conductor in the cable. The power feed equipment at the shore terminals supply a constant current of about 1 Ampere at -lO,OOOvolts, where one side of the cable is biased with positive voltage and the other with negative voltage. Interestingly, having a cable conductor across the ocean allows one to determinethe ground potential difference between continents, which is typically tens of volts, but can increase significantly during electrical or solar storms. Figure 4.32 shows a diagram of a typical amplifier pair that is located in an undersea system. A maintenance system is used to identify the location of faults andor degraded components by monitoring the undersea equipment from the shore terminals. An optical monitoring signal is coupled back into the fiber in the reverse direction at a low optical power. The signal-to-noise ratio of this low-level signal is enhanced using signal correlation techniques to provide data on repeater gain, gain tilt, and span attenuation. This approach Erbium Doped Fiber WDM Isolator U ' b 186 Neal S Bergano . also is synergistic with the use of COTDR (Coherent Optical Time Domain Reflectometer) techniques to identify the location of a cable cut or other fault b t e n repeaters. ewe 4.10 Transmission Experiments Most long-haul transmission experiments using optical amplifiers fall into one test bed^,^^ and special measurements of three categories: circulating performed on installed systems.59 Circulating-loop transmission measurements are by far the most important experimentaltechnique. Circulating-loop techniques applied to an amplifier chain of modest length can provide an experimentalplatform to study a broad range of transmission phenomena for EDFA-based transmission systems.60 A loop experiment attempts to simulate the transmission performance of a multithousand-kilometer-long system by making multiple passes through an amplifier chain of modest length (Le., hundreds of kilometers). The loop transmission experiment (Fig. 4.33) contains most of the elements found in conventionalexperiments, such as an opticaldata transmitterlregeneratorpair, a chain of amplifiedfibersections, and diagnosticequipment such as a bit-error ratio test set (BERTS). In the loop experiment, optical switching is added to allow data to flow into the loop (the load state) and then to circulate (the loop state, Fig. 4.34). The data circuIates for a specified time, after which the state of the experiment toggles, and the Ioacfnoop cycle is repeated. Load Switch 3dB Coupler BERTS Clock Gate Load J aitchL-1 State , :> , I : I + -,. /round -2.4 msec trip time Measurement Gate hl , Jnl , ,) Fig. 4.33 Top: Block diagram for a circulating-loop transmission experiment. Bottom: Timing diagram for the experiment showing the optical switch states and the time gate for making measurements. 4. Undersea Communication Systems A) Load 187 B) Loop Fig. 4.34 Simplified block diagram of a loop transmission experiment, showing: (A) the load state and (B) the loop state. The basic unit of time for the loop experiment is the time of flight for an optical signal around the closed loop, which is about 4 8 wsec per kilometer .9 of fiber. With reference to the timing diagram of Fig. 4.33, the experiment starts with the load switch on (or transmitting light) and the loop switch off (or blocking light). The two switchesare held in this load condition (Fig. 4.34a) for at least one loop time to fill the loop with the optical data signal. Once the loop is loaded with data, the switches change state to the loop configuration (Fig. 4.34b), and the data is allowed to circulate around the loop for some specified number of revolutions.A portion of the data signal is coupled to the receiver or other diagnostic equipment for analysis. The data signal is received and retimed by the regenerator and compared to the transmitted signal in the BERTS for error detection. The measurement continues, switchingbetween the load and the loop states so that errors can be accumulated over long intervals of time. Since errors are counted only during the measurement gate period, the effective bit rate for the experimentis diminished by the duty cycle of the gate signal; thus, the real time for demonstrating a particular BER might be increased by 50 or 100 times over conventional measurements. In addition to bit-error ratio, many other measurements are possible, such as optical spectra, eye diagrams, and Q-factor. For example, Fig. 4.35 shows the optical spectra of a 16-channel WDM experiment as a function of distance. One of the advantages of a circulating loop is that length dependencemeasurements are easily made. From this measurement, the nonideal gain equalization of the amplifier chain is clearly observed; the inner channels gain power, and the outer channels lose power as they propagate into the system. The length of the amplifier chain used in the loop experiment is an engineering tradeoff between cost and performance. To perform meaningful experiments, many in the lightwave community have settled on a minimum amplifier chain of about 500 km. The benefits of having a long amplifier chain are: 0 As the loop length is increased, the round-trip time becomes long compared to the optical amplifier’s recovery time. 188 Neal S. Bergano - Intensity (dB) i 62 1558 Distance (krn) 0 ‘ 1552 i556 1554 Wavelength (nrn) Fig. 4.35 Optical spectrum of 16, 5-Gbls channels measured for different transmission distances in a circulating loop. 0 0 0 Long amplifier chains have more accurate dispersion maps and/or more map periods. The statistics of the performance fluctuations become more realistic. Any attenuation that is added by the “loop specific” equipment becomes less significant. The benefits of having a short amplifier chain are: 0 0 0 Having a shorter amplifier chain reduces the cost of the experiment. Flexibility. For example, a direct comparison of two amplifier chains are more easily performed with a short amplifier chain. When performing experiments with new technologies, there might be a limited set of components available. Circulating-loop experimentshave been used to demonstrate massive transmission capacity over long distances. For example, Fig. 4.36 and Fig. 4.37 show the results of a 2400-Gb/s transmission experiment, where 120 channels, each carrying 20 Gb/s, were transmitted over 6200 km.61This experiment used many of the techniques described in the previous sections, such as gain equalization, dispersion management, FEC, RZ pulses, and orthogonal polarization launch. This massive capacity and high spectral efficiencywas achieved by using an optimum FEC code, a carefully engineered dispersion map with ultra-low dispersion slope, and full C-band EDFAs. 4. Undersea Communication Systems 189 -20 y 1525 I I I I 1535 1545 1555 1565 Wavelength (nrn) Fig. 4.36 Optical spectrum of 120 channels, each carrying 20 Gb/s, measured in a circulating loop after propagating over 6200 km. The channel spacing was 42 GHz (AA. x 0.33 nm). Note that the channels near 1535nm were purposely omitted because of insufficient BER performance caused by low SNR. 12 8 1525 1535 1545 1555 1565 8 Wavelength [nm] Fig. 4.37 Q-factor of 120 channels, each carrying 20 Gb/s, measured in a circulating loop after propagating over 6200 km. 4.11 Future Trends in Long-Haul Optical Transmission Systems The capacity of undersea fiber-optic systems will increase by using more optical bandwidth and by using the available bandwidth more efficiently. The conventional pass-band of the EDFA (C-band) is about 40nm wide, in the wavelength range of roughly 1526 to 1566nm, corresponding to optical frequencies of 196.5 to 191.4THz. Thus, the conventional erbium band has about 5 THz of bandwidth available for data transmission. The ultimate digital capacity that can be “fit” into the EDFA’s C-band will depend on how efficientlythis bandwidth can be used for data transmission. This spectral efficiency, expressed in (Bits/second)/Hz,is defined as the system’s average digital 190 Neal S. Bergano capacity divided by the average optical bandwidth of the system. The bestreported spectral efficiencies in WDM transmission range from 1.Obits/sec/Hz for very short (-100 km) distances, to roughly 0.5 bits/sec/Hz for transoceanic distance (Fig. 4.38). For example, the data shown in Fig. 4.36 represents a spectral efficiency of about 0.48 bits/sec/Hz. Assuming that this spectral efficiency could be achieved (with realistic margin) gives an upper limit on the C-band capacity of about 2.5TB/s on a single fiber. Table 4.4 gives some representative values for the optical bandwidth required to achieve a total transmission capacity for different spectral efficiencies. " ' I .j. 0000 0. 9. :.. e 3;' 0) *.4 ',0 .f0 --. . *. 9. Total Capacity in GBIs *'. -9.. 0.. d- q o o 0.. .... a. 0. *e.. * . *. 9. m 0 v) g i : 0.2 1 **.. =.. .3200 *. -. .. I *-a..po ***-*w.@oo # -3FEO ,:,ooo 180( *. With FEC "'-.,. I I000 -9. 0 Without FEC I 6 1000 . 8 I ' I I I 100 200 500 1,000 2,000 5,000 10,000 Transmission Distance (km) Fig. 4.38 Spectral efficiency of recently published transmission experimentsas a function of transmission distance. The data points list the total capacities. Experiments that ih used FEC are displayed w t a different symbol. Table 4.4 The Relationship between Total Capacity, Spectral Efficiency, and Required Optical Bandwidth (The required optical bandwidth is calculated assuming a center wavelength of 1545 nm. The missing entries calculate to a bandwidth that is much larger than 80 nm.) 640 Gb/s 1280 Gb/s 2560 Gb/s 5120 Gb/s - 0.1 (B/s)/Hz 0.3 (B/s)/Hz 0.5 (B/s)/Hz 1.0 (B/s)/Hz 50.9nm 16.7nm 10.2nm 5.1 nm 102nm 34.2nm 20.4nm 10.2nm 67.6nm 40.7nm 20.3nm 81.5nm 40.7nm 4. Undersea Communication Systems Cable “Repeater” Cable “Repeater” 191 + Fig. 4.39 An amplifier chain with both EDFAs and Raman gain. The performance of the C-band could also be improved by using a combination of Raman gain with the EDFA as shown in Fig. 4.39. Here Raman gain in the transmission fiber is used to “assist” the EDFA gain, which lowers the accumulated noise, thus increasing the SNR. Alternatively, the same SNR could be achieved while lowering the signal power, thus reducing the nonlinear effects.62 The system’stotal capacity could also be improved by increasing the number of optical fibers in the cable. This becomes an engineering challenge to make the optical amplifiers more efficient in physical space, given the limited amount of space in the pressure vessels, and require less electrical power, given the practical limits of electrical power transmission in the cable. We can continue to use this bandwidthhpectral efficiency idea to estimate the ultimate capacity of a transoceanic-length system (practicality notwithstanding). The low attenuation window of typical telecommunications-grade optical fibers is about 120 nm wide and extends from approximately 1500 to 1620 nm, corresponding to -1 5 THz. Assuming the same 0.5 bits/sec/Hz spectral efficiency yields a potential capacity of about 7.5 TB/s. Erbium amplifiers can cover about 2/3 of this bandwidth by using both the C-band and the newer “Long” wavelength band (or L-band) in the wavelength range of about 1570 to 1610nm. The leading optical amplifier candidate for the remaining short wavelength band (S-band) is stimulated Raman gain, which would be accomplished by pumping the transmission fiber at 1430 nm. Commensurate with the required wide-band optical amplifier, is the need for wide-band transmission fibers that have a “flattened” chromatic dispersion characteristic. Such fibers have been reported recently that extend the concept of dispersion mapping by alternating both the sign and the slope of the d i s p e r s i ~ n .The ,resulting fiber spans have relatively constant disper~~ ~ sion value over a broad bandwidth (Fig. 4.40). Ultimately, one could envision using the entire pass-band of the transmission fiber from 1300to 1700 nm, corresponding to 55 THz. This would pose many challenges to fiber and system designers. For example, a very broadband optical amplifier would be needed (or combinations of amplifiers), and the added attenuation of the fiber at the shorter wavelengths would decrease the signal-to-noise ratios for WDM channels in that region. - 192 Neal S. Bergano D+ D- D+ D- D+ D- ... ... ... .................... b Fig. 4.40 An amplifier chain using “dispersion matched” fiber spans. The dispersion characteristics in the two fiber types are similar in magnitude and opposite in sign over a wide optical bandwidth. 0 -2 -4 Log(BER) -6 0.8 -8 0.7 R = l .O -1 0 -1 2 \ 0 2 4 6 8 10 12 14 16 18 Input Q-Factor (dB) Fig. 4.41 Theoretical calculation for the bit-error ratio vs Q-factor. The parameter is the FEC code rate, which is the reciprocal of the overhead. As stated previously, the transmission performance of a lightwave system can be improved using FEC coding. Figure 4.41 shows a calculation for bit error ratio as a function of input Q-factor for different FEC code rates.65This calculation recasts Shannon’s capacity limit66,67 terms of a lightwave system in assuming a binary asymmetric channel. For example, the theoretical BER 4. Undersea Communication Systems 193 threshold for a code rate of 0.8 (or a 20% FEC overhead) is approximately 5 dB. This represents an improvement of 3.5 dB over the performance shown in Fig. 4.27 for a 23% FEC overhead. Thus, there is room for improvement in terms of the quality of FEC encoders and decoders. Some of the promising techniques to improving FEC performance are iterative concatenated codesY6* turbo product codes,69and low-densityparity check codes.70 4.12 Summary We have come a long way since the 1980s and the first undersea fiber-optic cables that revolutionized international telecommunications. Optical fiber cable networks now provide the bulk of the long-haul telecommunicationsfor voice and data over land and across seas. Today, transoceanic cable networks are being built with multi-Terabitcapacities.Ultimately, another order of magnitude increase in the data transmission capacity of single-mode fiber will occur given wider bandwidth amplifiers and improvements in spectral efficiency. These improvementswill foster unprecedented capacity improvements for international telecommunications. References Neal S. Bergano, “Undersea Fiberoptic Cable Systems: High-Tech Telecomunications Tempered By a Century of Ocean Cable Experience,” Optics and Photonics News Magazine, Vol. 11, No. 3, March 2000. Neal S. Bergano and Howard Kidorf, “Global Undersea Cable Networks,” Optics and Photonics News Magazine, Vol. 22, No. 3, March 2001. F R. Trischitta and W. C. Marra, “Global Undersea CommunicationsNetworks,” ! IEEE CommunicationsMagazine, Vol. 34, No. 2, February 1996. Bern Dibner, The Atlantic Cable, Burndy Library, Norwalk, CT, 1959. R. D. Ehrbar, “Undersea Cables for Telephony,” Chapter 1 in Undersea Lightwave Communications, edited by Peter K. Runge and Patrick R. Trischitta, IEEE Press, New York, 1986. P. K. Runge and P. R. Trischitta, “The SL Undersea Lightwave System,” Chapter 4 in Undersea Lightwave Communications, edited by Peter K. Runge and Patrick R. Trischitta, IEEE Press, New York, 1986. P. Trischitta, et al., “The TAT-12/13 Cable Network,” IEEE Communications Magazine, Vol. 34, No. 2, p. 24, February 1996. T. Li, “The Impact of Optical Amplifiers on Long-Distance Lightwave Telecommunications,’’Proceedings ofthe IEEE, Vol. 18, No. 11, p. 1568, 1993. A. M. Vengsarkar, et al., “Long-Period Fiber-GratingBased Gain Equalizers,” Opt. Lett., Vol. 21, No. 5, pp. 336-338, 1996. P. C. Becker, N. A. Olsson, and J. R. Simpson, Erbium-Doped Fiber Amplijiers Fundamentals and Technology,p. 206, Academic Press, Boston, 1999. * lo 194 l1 Neal S. Bergano l2 l3 l4 l5 l6 l7 ’* l9 2o 21 22 23 z4 25 26 J. P. Gordon and L. F. Mollenauer, “Effects on Fiber Nonlinearities and Amplifier Spacing on Ultra-Long Distance Transmission,” Journal of Lightwave Communication, Vol. 9, No. 2, p. 170, 1991. E. Lichtman, “Optimal Amplifier Spacing in Ultra-Long Lightwave Systems,” Electronics Letters, Vol. 29, p. 2058, 1993. C. R. a l e s and E. Desurvire, “Propagation of Signal and Noise in Concatenated Erbium-Doped Fiber Amplifiers,” Journal of Lightwave Technology, Vol. 9, No. 2, p. 147, 1991. A. K. Srivastava, et al., “Room Temperature Spectral Hole-Burning in ErbiumDoped Fiber Amplifiers,” in Proc. Optical Fiber Con$, p. 33, San Jose, CA, 1996. A. R. Chraplyvy, J. A. Nagel, and R. W. Tkach, “Equalization in Amplified WDM Lightwave Transmission Systems,” IEEE Photonics Technology Letters, Vol. 4, No. 8, August 1992. G. P. Agrawal, “Group-Velocity Dispersion,” Chapter 3 in NonlinearFiber Optics, edited by Ivan P. Kaminow and Thomas L. Koch, Academic Press, Boston, 1989. A. H. Gnauck and R. M. Jopson, “Dispersion Compensation for Optical Fiber Systems,” Chapter 7 in Optical Fiber Telecommunications IIU, edited by Ivan F ? Kaminow and Thomas L. Koch, Academic Press, Boston, 1997. G. P. Agrawal, NonlinearFiber Optics, Academic Press, Boston, 1989. D. Marcuse, A. R. Chraplyvy, and R. W Tkach, “Effects of Fiber Nonlinearity . on Long-Distance Transmission,” Journal of Lightwave Technology, Vol. 9, No. 1, pp. 121-128,1991. F Forghieri, R. W Tkach, and A. R. Chraplyvy, “Fiber Nonlinearities and Their . . Impact on Transmission Systems,” Chapter 8 in Optical Fiber Telecommunications IIIA, edited by Ivan P. Kaminow and Thomas L. Koch, Academic Press, Boston, 1997. T. Naito, T. Terahara, T. Chikama, and M. Suyama, “Four 5-Gbith WDM Transmission Over 4760-km Straight-Line Using Pre- and Post-Dispersion Compensation and FWM Cross-Talk Reduction,” Optical Fiber Communications,OFC ’96, pp. 182-183, 1996. A. Puc, F W Kerfoot, A. Simons, and D. L. Wilson, “Concatenated FEC Exper. iment Over 5000-km-long Straight Line WDM Test Bed,” OFC ’99 Paper ThQ6, San Diego, CA. Neal S. Bergano and C. R. Davidson, “Method and Apparatus for Improving Spectral Efficiency in Wavelength Division Multiplexed Transmission Systems,” United States Patent 6,134,033, issued October 17,2000. Neal S. Bergano, et al., “320 Gb/s WDM Transmission (64 x 5 Gb/s) over 7,200 km using Large Mode Fiber Spans and Chirped Return-to-Zero Signals,” OFC ’98, paper PD12, San Jose, CA, February 1998. E. A. Golovchenko, Neal S. Bergano, and C. R. Davidson, “Four-WaveMixing in Multispan Dispersion-Managed Transmission Links,” IEEE Photonics Technology Letters, Vol. 10, No. 10, October 1998. Bell Telephone Laboratories, Transmission Systems for Communications, Fifth Edition, Chapter 30, p. 741, Bell Telephone Laboratories, Inc., 1982. 4. Undersea Communication Systems 27 28 195 29 30 31 32 33 34 35 36 37 38 39 41 Bell Telephone Laboratories, Transmission Systems for Communications, Fifth Edition, Chapter 30, p. 741, Bell Telephone Laboratories, Inc., 1982. Ekaterina A. Golovchenko, Alexei N. Pilipetskii, and Neal S. Bergano, “Transmission Properties of Chirped Return-to-Zero Pulses and Nonlinear Intersymbol Interference in 10 Gbls WDM Transmission,” OFC 2000, paper FC3, Baltimore, MD, March 2000. B. Bakhshi, M. Vaa, E. A. Golovchenko, W. W Patterson, R. L. Maybach, and . Neal S. Bergano, “Comparison of CRZ, RZ, and NRZ Modulation Formats in a 64 x 12.3 Gbls WDM Transmission Experiment Over 9000 km,”OFC 2001, paper WF4, Anaheim, CA, March 2001. L. F. Mollenauer, J. P. Gordon, and P. V, Mamyshev, “Solitons in High Bit-Rate Long-DistanceTransmission,”Chapter 12in Optical Fiber TelecommunicationsIIIA, edited by Ivan P. Kaminow and T. L. Koch, Academic Press, Boston, 1997. M. I. Suzuki, et al., “Reduction of Gordon-Haus Timing Jitter by Periodic Dispersion Compensation in Soliton Transmission,”Electronics Letters, Vol. 31, No. 23, p. 2027, 1995. Neal S. Bergano, E W. Kerfoot, and C. R. Davidson, “Margin Measurements in Optical Amplifier Systems,” IEEE Photonics TechnologyLetters, Vol. 5, No. 3, March 1993. S. D. Personick, “Receiver Design for Digital Fiber Optic Communications Systems,” Bell System TechnicalJournal, Vol. 52, No. 6, pp. 843-886, 1973. Cecil Hastings, Jr., Approximations for Digital Computers, Princeton University Press, Princeton, p. 191, 1955. D. Marcuse, “Derivation of Analytical Expressions for the Bit-Error Probability in Lightwave Systems with Optical Amplifiers,” Journal o Lightwave Technology, f Vol. 8, pp. 18161823, 1990. P. A. Humblet and M. Azizoglu, “On the Bit Error Rate of Lightwave Systems with Optical Amplifiers,” JournaZ of Lightwave Technology,Vol. 9, pp. 15761582, 1991. V, J. Mamczyk and D. G. Duf, “Effect of Intersymbol Interference on Signal-toNoise Measurements,’’ Conference on Optical Fiber Communications, paper WQ1, 1995. D. Marcuse, “Derivation of Analytical Expressions for the Bit-Error Probability in Lightwave Systems with Optical Amplifiers,’’ Journal of Lightwave Technology, Vol. 8, pp. 18161823, December 1990. N. Ramanujam, et al., “Forward Error Correction (FEC) Techniques in LongHaul Optical Transmission Systems,” Paper WE1 presented at the LEOS Annual Meeting, Vol. 2, p. 405,2000. C. R. Davidson, et al., “1800Gbls Transmission of One Hundred and Eighty 10 Gbls WDM Channels over 7,000 km using the Full EDFA C-Band,” Paper PD25 at the conference on Optical Fiber CommunicationsOFC 2000, March 2000. H. Kidorf, et al., “Performance Improvement in High-Capacity, Ultra-Long Distance, WDM Systems using Forward Error Correction Codes,” Paper ThS3 presented at the conference on Optical Fiber CommunicationsOFC 2000, p. 274, March 2000. 196 42 Neal S. Bergano 43 45 46 47 48 49 50 51 52 53 54 55 56 C. D. Poole and J. Nagel, “Polarization Effects in Lightwave Systems,” Chapter 6 in OpticalFiber TelecommunicationsIIIa, edited by I. Kaminowand T. Koch, Academic Press, Boston, 1997. E. Lichtmann, “Performance Degradation Due to Polarization Dependent Gain and Loss in Lightwave Systems with Optical Amplifiers,” Electronics Letters, Vol. 29, NO.22, pp. 1971-1972,1993. F. Bruyere and 0. Audouin, “Penalties in Long-Haul Optical Amplifier Systems Due to Polarization Dependent Loss and Gain,” IEEE Photon. Tech. Lett., Vol. 6, No. 5, pp. 654-656, 1994. S. Yamamoto, N. Edagawa, H. Taga, Y. Yoshida, and H. Wakabayashi, “Observation of BER Degradation Due to Fading in Long-Distance Optical Amplifier System,” Electronics Letters, Vol. 29, No. 2, pp. 209-210, 1993. R. E. Wagner, C. D. Poole, H. J. Schulte, N. S. Bergano, V. P Nathu, J. M. Amon, . R. L. Rosenberg, and R. C. Alferness, “Polarization Measurements on a 147-km Lightwave Undersea Cable,” in Technical Digest o OFC ’86,Paper PDP7, Atlanta, f GA, February 24-26,1986. M. G. Taylor, “Observation of New Polarization Dependence Effect in Long-Haul Optically AmpUied System,” OFC ’93, Post-deadline paper, PDS, San Jose, CA, 1993. V. J. Mazurczyk and J. L. Zyskind, “Polarization Hole-Burning in Erbium-Doped Fiber Amplifiers,” CLEO ’93, Post-deadline paper, CPD26, Baltimore, MD, 1993. E. Desurvire, C. R. Giles, and J. R. Simpson, “Gain Saturation Effects in HighSpeed, Multichannel Erbium-doped Fiber Amplifiers at h = 1.53 pm,” Journal of Lightwave Technology, Vol. 7, No. 7, pp. 2095-2104, December 12,1989. Neal S. Bergano, “The Time Dynamics of Polarization Hole Burning in ErbiumDoped Fiber Amplifiers,” OFC ’94, San Jose, CA, 1994. Neal S. Bergano, V. J. Mazurczyk, and C. R. Davidson, “Polarization Scrambling Improves SNR Performance in a Chain of EDFAs,” OFC ’94, San Jose, CA, 1994. M. G. Taylor, “Improvement in Q with Low-Frequency Polarization Modulation on Transoceanic EDFA Link,” IEEE Photonics Technology Letters, Vol. 6, No. 7, pp. 860-862, July 1994. Neal S. Bergano, C. R. Davidson, and F Heismann, “Bit-SynchronousPolarization . and Phase Modulation Scheme for Improving the Transmission Performance of Optical Amplifier Transmission System,” Electronic Letters, Vol. 32, No. 1, pp. 52-54, January 4,1996. M. G. Taylor and S. J. Penticost, “Improvement in Performance of Long Haul EDFA Link Using High Frequency Polarization Modulation,” Electronics Letters, Vol. 30, No. 10, pp. 805-806, 1994. Y. Fukada, T. Imai, and A. Mamoru, “BER Fluctuation Suppression in Optical In-Line Amplifier Systems Using Polarization Scrambling Technique,” Electronics Letters, Vol. 30, No. 5, pp. 432433, 1994. E. A. Golovchenko, et al., “Modeling of Transoceanic Fiber-optic WDM Communications Systems,” IEEE Journal o Selected Topics in Quantum Electronics, Vol. 6, f No. 2, pp. 337-347, March/April2000. 4. Undersea Communication Systems 57 197 58 59 6o 61 62 63 64 65 66 67 69 70 Neal S. Bergano, Jennifer Aspell, C. R. Davidson, P. R. Trischitta, B. M. Nyman, and F. W Kerfoot, “A 9000-km 5 Gb/s and 21,000-km 2.4 Gb/s Feasibility Demon. stration of Transoceanic EDFA Systems Using a Circulating Loop,” Optical Fiber Communications Conference, PD-13, San Diego, CA, February 18-22,1991. H. Taga, N. Edagawa, H. Tanaka, M. Suzuki, S. Yamamoto, H. Wakabayashi, N. Bergano, C. Davidson, G. Homsey, D. Kalmus, P. Trischitta, D. Gray, and R. Maybach, “10-Gb/s, 9,000-km IM-DD Transmission Experiments Using 274 Er-Doped Fiber Amplifier Repeaters,” Post-deadlinePaper, OFC ’93. J. C. Feggeler, et al., “10-Gb/s WDM Transmission Measurements on an Installed Optical Amplifier Undersea Cable System,” Electronics Letters, Vol. 31, No. 19, p. 1676, September 14,1995. Neal S. Bergano and C. R. Davidson, “CirculatingLoop Transmission Experiments for the Study of Long-Haul Transmission Systems Using Erbium-Doped FiberAmplifiers,” IEEEJournalofLightwave Technology,Vol. 13, No. 5, p. 879, May 1995. J.-X. Cai, M. Nissov, A. N. Pilipetskii, A. J. Lucero, C. R. Davidson, D. Foursa, H. Kidorf, M. A. Mills, R. Menges, P. C. Corbett, D. Sutton, and N. S. Bergano, “2.4 Tb/s (120 x 20 Gb/s) Transmission over Transoceanic Distance with Optimum FEC Overhead and 48% Spectral Efficiency,” OFC 2001, PD-20, March 2001. Balslev C. Clausen, et al., “Modeling and Experiments of Raman Assisted Ultra Long-haul Terrestrial Transmission Over 7500 km,” Paper We.F. 1.2 presented at the 27th European Conference on Optical Communications, Amsterdam, The Netherlands, 2001. Stig Nissen Knudsen and Torben Veng, “Large Effective Area Dispersion Compensating Fiber for Cabled Compensation of Standard Single Mode Fiber,” OFC 2000, Paper TUGS,Baltimore, MD, March 2000. M. Tsukitani, et al., “LOW-Loss Dispersion-Flattened Hybrid Transmission Lines Consisting of Low-NonlinearityPure Silica Core Fibres and Dispersion Compensating Fibers,” Electronics Letters, Vol. 36, No. 1,2000. Y Cai, et al., “Performance Limit of Forward Error Correction Codes in Optical Fiber Communications,” Paper TuF2, presented at the Optical Fiber Communications Conference, Anaheim, CA, 2001. C. E. Shannon, “A mathematical theory of communication,” Bell System Technical Journal, Vol. 27, pp. 379423 and 623-656, July and October, 1948. W Weaver and C. E. Shannon, n e Mathematical Theory of Communication, University of Illinois Press, Urbana, Illinois: 1949, republished in paperback, 1963. 0. Ait Sab, “FEC Techniques in Submarine Transmission Systems,” Paper TuF1, presented at the Optical Fiber Communications Conference, Anaheim, CA, 2001. R. M. Pyndiah, “Near-optimum decoding of product codes: block turbo codes,” IEEE Trans. on Communications,Vol. 46, No. 8, pp. 1003-1010, August 1998. D. J. C. MacKay and R. M. Neal, “Near Shannon limit performance of low density parity check codes,” Electronics Letters, Vol. 32, No. 18, pp. 1645-1646, August 1996. Chapter 5 High-Capacity, Ultra-Long-Haul Networks John Zyskind, Rick Barry, Graeme Pendock, Michael Cahill, and Jinendra Ranka Sycamore NetworkF, Chelmsford,Massachusetts I. Introduction The advent of optically amplified transmission and of Dense Wavelength Division Multiplexing (DWDM) technology has transformed the technology and also the economics of optical network deployments. In less than 10 years, the capacity of a single optical fiber equipped with commercial transmission equipment has increased from a single OC-48 signal, transmitting at a rate of 2.488 Gb/s, to 160 OC-192s signals, totaling 1600Gb/s, a factor of close to 1000. The economics of DWDM are driving the development and deployment of a new generation of ultra-long-haul DWDM systems for terrestrial networks that can carry these high-capacity data streams over thousands of kilometers. During the same period, driven by the growth of the Internet and other data-based services, the demand for new capacity has exploded and the requirements for the public network have changed dramatically. This chapter will discuss the challenges of high-capacity, ultra-long-haul terrestrial transmission systems, the advanced technologies required for such systems, and the architectures of optical networks based on ultra-long-haul transmission capability designed to meet these new demands. In conventional time-division multiplexed (TDM) regenerated transmission, prevalent until the mid-l990s, one signal was transmitted over its own fiber and, because of the attenuation of the fiber, the signal had to be optoelectronicallyregenerated approximately every 50 km by a dedicated, 13 IO-nm optoelectronic regenerator, comprising expensive, complex, and bit-rate specific high-speed optical and electronic components, the bandwidth of which limited the capacity of the TDM signal. On the other hand, modern day DWDM systems can carry simultaneously on a single fiber numerous signals, each at the same bit rate as the aforementionedTDM signal, and each carried on a distinct optical wavelength. A single optical amplifier, which amplifies all the signal wavelengths simultaneously,is used periodically to overcome the fiber attenuation in place of the multitude of more complicated regenerators that would be required, one for each signal, in a regenerated TDM system. These technologies dramatically reduce the cost of long-haul transmission capacity, and this dramatic cost reduction has driven the development and 198 OPTICAL FIBER TELECOMMUNICATIONS, VOLUME N B Copyright 8 2002, Elsevier Science (USA). All rights of repmductionin any form reserved. ISBN 0-12-395173-9 5. High-Capacity, Ultra-Long-Haul Networks 199 widespread deployment of DWDM as the technology of choice for long-haul telecommunicationsnetworks. Since the introduction of the first commercial DWDM systems in 1995, the capacity of such systems has grown explosively from 8 channels each carrying a DWDM OC-48 signal in the first systems to 160 DWDM channels or more each carrying an OC-192 channel in some recently announced systems. Until recently, these systemswere typically able to carry the DWDM signals over distances of 300-600 km without optoelectronic regeneration. As the capacity of such systems has exploded, the cost of terminals and regenerators has become an ever larger fraction of total system cost. Minimizingthe number and the cost of regenerators is now a major economic driver in the design of new equipment and the design of carriers’ fiber networks. These economic factors are driving increased channel bit rates from OC-48 (2.488 Gb/s) to OC-192 (9.953 Gb/s) and, in the near future, to OC-768 (39.813 Gb/s) to minimize the number of regenerators, transmitters, and receivers. The demand to reduce system cost is also driving the demand for, and development of, ultra-long-haul terrestrial DWDM systems with reach between regenerators exceeding 2000 km. A hypothetical national-scale fiber network connecting major urban centers of the United States is shown in Fig. 5.1. A backbone network would connect such centers, and much of the traffic arriving at these network nodes would pass through destined for other nodes. The cost benefit lies in avoiding the necessity for expensive optoelectronic regenerators between these nodes and permitting optical pass through of express traffic destined for another node. Fig. 5.1 Hypothetical national-scale, backbone long-haul network with nodes situated at major urban centers of the United States. 200 John Zyskind et a . l Managing the large bandwith and the large number of channels carried by high capacity DWDM systems represents an unprecedented challenge both for equipment suppliers and for carriers. Applying the conventional technology of switchingand routing low-bit-ratetributaries, for example, at the STS-1 speed, requires very large and expensive switches, and the optical-to-electronicto-optical (OEO) conversions required for such electronic switching is also increasingly expensive. Optical networking promises a solution to the challenge of managing the immense bandwidth carried by such DWDM systems. The fact that the different signals are encoded on distinct optical wavelengths opens the possibility of optical manipulation, switching, and routing of each individual signal channel using optical filtering and switching technologies to which optics is naturally well suited. For such optical networking to be useful, it will be necessary to extend the unregenerated reach of DWDM optical transmission to support the extended optical path lengths that will result. Transmission of high data rate channels (presently at lOGb/s as in the future at 40 Gbls) over unregenerated links with lengths of 1000 to 5000 km poses major challenges that cannot be met with conventional DWDM technology. Foremost among these problems are the accumulation of optical noise and spectral gain nonuniformity arising from optical amplification, as well as distortion due to transmission effects (including chromatic dispersion, polarization mode dispersion, and optical nonlinearities). These challenges were first addressed for undersea systems in which both fiber and optical amplifiers are placed under the water while terminals and regenerators are restricted to the shores at the ends of transoceanic links spanning thousands of kilometers (see, for example, Bergano 1997). The undersea elements of transoceanic systems must meet much more stringent reliability requirements than terrestrial systems because of the great expense that deep water ship repairs entail; the range of technologies that can be deployed is thereby strictly limited. However, each undersea deployment is a “green field” system in which the fiber spans, fiber type, and optical amplifier spacing can be specially tailored to the link length and designed capacity of that particular system. As a result, there is a great deal of latitude for optimization of each system’s design to meet the challenges of ultra-long-haul transmission. In terrestrial systems, on the other hand, the fiber network is typically already installed, often with a very different system in mind, well before the ultra-long-haul system designer begins. The fiber type is usually already defined and is one of a number of widely deployed, distinct fiber types. The locations of optical amplifiers are predetermined by the locations of hut sites, selected based on the availability of rights of way and the economic incentive to support as few amplifier sites as possible, which as we shall see, is not conducive to overcoming the challenges of ultra-long-haul transmission. While terrestrial ultra-long-haul systems certainly have strong similarities to undersea systems, the differences are significant, and the solutions to the problems 5. High-Capacity, Ultra-Long-Haul Networks FEC encoder OC192 Tx’s Deriodic channel Dower manaoement & optional dispersion slope management dispersion 201 FEC decoder Raman pump Raman pump Fig. 5.2 Schematic of an ultra-long-haultransmission system illustratingthe use of Raman amplification, FEC, and periodic dispersion slope compensation and power ih management wt a reconfigurable gain-flattening filter. of ultra-long-haul transmission are quite distinct. Figure 5.2 depicts some of the key features of a high-capacity, ultra-long-haul transmission system which will be discussed in this chapter. 11. Noise and Optical Amplification A. NOISE IN OPTICALLYAMPLIFIED ULTRA-LONG-HA SYSTEMS UL The management of optical amplifier noise and the management of transmission distortions of the high-speed optical signals are the two most important considerations in the design of high-capacity, ultra-long-haul transmission systems. This section will focus on the management of amplifier noise, and in Section I11 we will turn to sources of distortion. The advent of practical optical amplifiers capable of simultaneously amplifying multiple signal wavelengths that occupy an appreciable range of the optical spectrum was the key technological advance that ushered in the DWDM revolution. Optical amplifiers are used at the end of each fiber span to boost the power of the DWDM signal channels to compensate for fiber attenuation in the span. EDFAs designed to operate with high inversion provide gain over a spectral range about 30 nm in width, from about 1530 nm to about 1560nm. This spectral range can support roughly 40 DWDM signal channels with a separation of 100GHz and 80 channels with a separation of 50 GHz, corresponding to 400 or 800 Gb/s, respectively, for 10 Gb/s OC-192 or STM-64 channels, and in the future, with 40-Gb/s channels, capacities of 1.6Tb/s (1600 Gb/s) for 100-GHz spaced channels. EDFAs designed to operate with lower inversion can provide gain over an even wider spectral range, including the so-called L-band, starting at about 1570nm, thus offering the opportunity to double the capacity on a single fiber through addition of an L-band EDFA. For a system transmitting over both the C- and L-bands, the 202 John Zyskind et a. l capacity can reach 1.6 Tb/s or more for lO-Gb/schannels spaced at 50 GHz or 3.2 Tb/s or more for 40-Gb/s channels spaced at 100GHz. Unfortunately, optical amplification is not possible without the generation of amplified spontaneous emission (ASE), and the noise resulting from this ASE constitutes perhaps the most severe impairment that limits the reach and capacity of such systems. Each optical ampljjier contributes ASE, and these contributions add cumulatively along the amplifier chain. This accumulated ASE gives rise to signal-spontaneousbeat noise at the receiver, which is the fundamental noise limit in an optically amplified transmission system. Each EDFA contributes an amount of ASE: where PME is the ASE power in an optical bandwidth Av, h is Planck’s constant, v is the optical frequency, nsp is the spontaneous emission factor, and G is the optical amplifier gain. The spontaneous emission factor, nsp, is determined by the inversion of the amplifier’s Er ions. The contribution of each amplifier’s ASE to the accumulated ASE is characterized by the amplifier’s noise figure, which at high gain is well approximated by NF r 2nsp. z The signal-spontaneousnoise impairment can be characterizedin terms of the Optical Signal to Noise Ratio (OSNR), defined as the ratio of the signal channel power to the power of the ASE in a specified optical bandwidth, usually taken by convention to be 0.1 nm. This OSNR target must be sufficient to achieve the required system performance, which for commercial systems is today most often a bit-error rate (BER) of Le., effectivelyerror free. The OSNR target must include s a c i e n t margin to provide for any impairments that may be encountered. These include transmission impairmentsarising, for example, from chromatic dispersion, nonlinearities, and PMD discussed in the following sections; distortions introduced by the transmitter and receiver; amplser gain ripple; manufacturingmargin to provide for variances in performance of parts such as transmitters and receivers produced in a commercial manufacturing environment; and aging both of the system equipment and fiber plant during the expected life of the system. The target OSNR must theoretically increase by 6 dB for each factor of 4 increase in the channel bit rate in order to maintain equivalent noise performance. The actual increase in required OSNR with channel bit rate may be greater than this value due to the greater difficulty in achieving comparable transmitter and receiver performance at higher bit rates, and because of the greater severity of transmission impairments at higher bit rates, especially for rates as high as 40 Gb/s. As the length of a system increases, and the number of amplifiers contributing ASE increases, the OSNR at the end of the system decreases. The maximum unregenerated reach of an optically amplified system is the length of the system at which the OSNR at its end equals the target OSNR for acceptable system performance. However, the realization of this maximum length is contingent 5. High-Capacity, Ultra-Long-Haul Networks 203 on successful management of transmission impairments that generate signal distortion. The length of the system that results in this OSNR is determined by the characteristics of the fiber network and of the optical amplifiers. For a system consisting of NaW fiber spans, each of loss LsPM(in dB) followed by an optical amplifier with output power Pout(in dBm) per channel launched into the span and noise figure NF (in dB), the OSNR (in dB) of a signal channel at the end of the system is approximately (Zyskind et al. 1997): OSNR (in dB) = 58 + Pout - LsPw - NF - 10log (Namp). (5.2) Although the fiber spans of actual commercial fiber networks are typically not uniform in length, Eq. 5.2 can be used to illustrate some of the constraints placed on ultra-long-haul system design as a result of amplifier noise. The first thing to note is that if the amplifier spacing is fixed, for each dB that the available OSNR is increased (or the target OSNR can be reduced), the unregenerated reach of the system can be increased by about 25% (i.e., 1 dB). If the OSNR is increased by 3 dB, the length of the system can be doubled. The OSNR can be increased dB for dB by increasing Pout,by decreasing noise figure, or by decreasing span loss. The OSNR can also be increased by reducing the number of spans, but the dependence is much weaker. The system , reach (in dB of loss) can be represented as the product of the span loss, L in dB, which is proportional to span length in kilometers, and the number of spans, Namp.Equation 5.2 shows that, if the system reach Lspan. Nampis kept constant, the OSNR increases as the span length is reduced, and the number of spans is increased in the same proportion, because the OSNR depends only logarithmically on the number of spans. Figure 5.3 shows the OSNR as a 0 500 1000 1500 Aggregate Fiber Loss (dB) 2000 Fig. 5.3 OSNR of systems with the indicated span losses (15, 20, 25, and 30dB) as a function of the aggregate fiber loss, which is the span loss multiplied by the number of spans The amplifier noise figure is taken to be 5 dB, and the launched power per channel is 0 dBm. 204 John Zyskind et al. function of system reach for systems with span losses of 15,20,25, and 30 dB, correspondingto spans of approximately60,80,100, and 120km, respectively, in practical fiber networks. A noise figure of 5 dB, typical of an EDFA, and launched channel power of 0 dBm per channel have been assumed. No margin has been allowed that would shorten the reach of a practical system. These parameters are typical of systems capable of transmitting DWDM signals several hundred km without regeneration. For longer systems and higher bit rates, the availableOSNR must be increased and/or the target OSNR must be reduced. An ultra-long-haul system with an unregenerated reach of several thousand kilometers, for example, would require an OSNR increase on the order of 10dB. For ultra-long-haul systems with 4O-Gb/s DWDM channels, the OSNR would need to increase by at least an additional 6 dB to achieve the same system reach. The OSNR could be improved by reducing the span loss, Lspan,and this is done in commercial undersea systems where span lengths for transoceanic systems can be 50 km or less, corresponding to span losses of about 10dB. However, for terrestrial systems,reducing the span loss by reducing the separation between amplifier sitesis expensive and commerciallyunattractivebecause more optical amplifiers are required and additional amplifier sites are needed to accommodate them. In addition, the amplifier sites in terrestrial systems must often be placed in pre-existing equipment huts, the locations of which cannot be changed. The OSNR could also be improved by increasing Pout. Increasing Pouris possible only to a certain extent, because as Pourincreases, impairmentsarising from optical nonlinearitiesbecome more severe, especially for very long transmission distances. The remaining alternativeis to reduce the noise figure of the optical amplifiers. For each 1dB decrease in the noise figure, the accumulated ASE will be reduced by 1dB. For high-gain EDFAs, the noise figure is in principle limited to values above 3 dB. For practical designs this is generally a few dB higher, and there is little opportunity to reduce the noise figure of EDFAs. B. DISTRIBUTED RAlMAN AMPLIFICATION Distributed Raman ampliiication is a new technology that offers the promise of effective noise figures that break the 3-dB barrier; this w l increase sysi l tem OSNR and enable extended transmission distances (Hansen et al. 1997). Raman amplification makes use of high power laser light, or Raman pump light traveling in the transmission fiber, as illustrated in Fig. 5 . 3 to produce ~ amplification in the transmission fiber over an appreciable distance due to the stimulated Raman scattering effect. The Raman pump typically has a wavelength approximately 100nm shorter than that of the signals to be amplified. Raman pumping requires relatively high pump powers; approximately a few hundred milliwatts are needed to provide gains of 10-15dB in commonly deployed transmission fibers. Early Raman pump units were based 5. High-Capacity, Ultra-Long-Haul Networks 205 on high-power, double-clad fiber lasers pumping cascaded Raman fiber resonators. Due to their cost, these units were used primarily for specialized applications such as repeaterless transmission. However, the recent availability of pumps with adequate power has made the deployment of Raman amplification in commercial transmission systems possible. The units most likely to see widespread deployment will use multiplexed, single-transverse mode 14xxs-nm semiconductor pump diodes, i.e., diodes with various wavelengths within several 10s of nm of 1450nm depending on the range of signal wavelengths to be amplified. The technology for 14xx-nm diode lasers used in such pump modules is similar to that for 1480-nmpumps used to pump erbium-doped fiber amplifiers, and there has been dramatic progress in the technology of such pumps during the last decade. Presently 14xx-pumpdiodes are available with pump powers exceeding200 mW of fiberpigtailed power, and vendors are working on development of diodes with even higher power. These diodes are suitable for use in modules that employ several such diodes multiplexed in both polarization and wavelength. Multiplexing multiple pumps delivers greater Raman pump powers than possible from a single diode. In addition, multiplexingin polarization minimizes polarizationdependent Raman gain, whilst using pumps at different wavelengths broadens and flattens the Raman spectral-gain profile (Emori and Namiki 1999). For Raman-enhanced systemswith very wide optical bandwidth, for examplethose employing both the C- and L-bands, the Raman pumps must employ multiple pump wavelengths to deliver flat gain over a bandwidth of 70nm or more, and the various pump wavelengths and powers must be carefully selected to take into account not only the Raman gain spectra produced by the various wavelengths,but also the Raman interactions among the various Raman pump wavelengths as they propagate down the fiber. The distributed Raman gain induced in the fiber can dramatically improve the OSNR. This is because the distributed Raman amplification overcomes the attenuation in the latter part of the span and the minimum signal power is increased roughly by the loss of the fiber over that portion of the span where the Raman amplification exceeds the fiber attenuation as shown in Fig. 5.3. This improvement in performance is typically represented by an “equivalent”noise figure,which is the noise figure of a hypothetical lumped amplifier located at the end of the span that would produce the same gain and the same contribution to the accumulated ASE. The equivalent noise figure for a counter-pumped distributed Raman pump is: %- 2 (5.3) In GR where NFeqis the equivalent noise figure in linear units, GRis the Raman gain in linear units at the signal wavelength, asis the fiber attenuation at the signal 206 John Zyskind et al. wavelength, and olP is the attenuation at the Raman pump wavelength. The > approximation for NF,, holds for large gain G > 1 and in the approximation that as x 9. Figure 5.4b shows the measured gain and effective NF from a Raman pump with a single pump wavelength. The figure also shows the a m rate agreement obtained with simulated performance using a model of pump propagation and distributed Raman gain. Wavelength multiplexing pumps of different wavelengths are often used in commercial practice to produce a broader and flatter gain profile. The theoretical limit, often termed the quantum limit, to the noise figure of a discrete optical amplifier located at the end of the span is 3 dB, and the noise figures of commercialEDFAs are typically a few dB higher. If the gain of a distributed Raman amplifier is, for example, 15dB or approximately 30 in linear units, then the equivalent noise figure is approximately 0.7, or -1.5 dB, an improvementof 6 dB or more over a discrete EDFA. This is possible because the equivalent noise figure is referenced to the end of the span. However, the distributed Raman amplifier is not actually located at the end of the span, but provides distributed amplification over an appreciable part of the preceding fiber span. Thus distributed Raman amplification delivers a substantial improvement in OSNR as illustrated in Fig. 5.5. At gains above 20dB, the improvement in noise figure is significantly degraded by the onset of Rayleigh scattering, which places an upper limit on the usable Raman gain (Hansen et al. 1997). Because of the logarithmic dependence of NF,, on GR, the Raman noise figure is not significantly impacted by this limit. However, it does mean that the gain available from a distributed Raman amplifier is insufficient to compensate fully the loss of a typical terrestrial transmission span. This is all the more so when the extra loss of dispersion compensation and possible adddrop multiplexing are included. Raman On I a distance [km] 3 Raman Pump Raman-Signal WDM IO' . . . 1550 1560 12 - 1530 1540 wavelength [nm] Fig. 5.4 (a) Raman amplification showing evolution of signal gain. The distributed gain provides improvement in OSNR. (b) Spectral gain and equivalent noise figure (NF) obtained with Raman pump. Experimental data (solid line) agrees closely with simulated results (dotted). 5. High-Capacity, Ultra-Long-Haul Networks 207 0 5 10 15 Distributed Raman Gain (dB) 20 Fig. 5.5 Discrete EDFA quantum-limited noise figure compared to equivalent noise figure for distributed Raman amplification. Thus in commercial systems, it is most common to follow a backward-pumped distributed Raman amplifier with a relatively low-gain EDFA. The noise figure of the hybrid Raman-EDFA combination is determined primarily by that of the distributed Raman amplifier because of the higher power it delivers to the input of the EDFA, hence the additional noise is negligible. The improvement in OSNR performance offered by Raman amplification can be used to improve systemperformance in a number of ways. First, Raman amplification can be used to extend the reach of an unregenerated link. Referring to Fig. 5.3, if the span losses and channel launch powers are kept constant and the link length is limited by ASE accumulation, and if the Raman amplification delivers 6 dB improvement in OSNR, it will be possible to quadruple the length of the link. Alternatively,if the number of spans is kept constant, it will be possible to increase the span loss by about 6 dB, which would correspond to about 25-30% of a typical terrestrial span. For fiber networks with relatively short hut spacings, distributed Raman amplification may make it possible to skip huts and totally eliminate amplifier sites that would be necessary for systems relying only on EDFAs or other discrete amplifiers. This represents a substantial savings to carriers, both in terms of equipment costs and in terms of the costs associated with maintaining the site where the amplifier would have been located. Raman amplification can also assist systems that employ channels closely spaced in wavelength. With closely spaced channels, the nonlinear interactions among the channels, particularly four-wave mixing and cross-phase modulation, become more severe. With distributed Raman amplification it is 208 John Zyskind et al. possible to reduce launched channel powers in order to mitigate the nonlinear interactions while maintaining the link‘s OSNR performance. CounterpropagatingRaman pumping is preferred to copropagatingRaman pumping, as this reduces the transfer of pump noise to the signal, as well as pump mediated cross-talk between the signals. Further improvements in OSNR may also be possible through the use of copropagating Raman amplification, and the development of Raman pump sources suitable for copumped distributed Raman amplifiers is an area of active research. The use of second-orderRaman pumping in conjunction with first-order counterpumping (Rottwitt et al. 2000; Dominic et al. 2001) has been proposed, as has the use of low-noise pump sources with a low degree of polarization (Dominic et al. 2001b). C FORWARD ERROR CORRECTION . An additional way of improving the system bit error rate without requiring an increase in the OSNR, is by making use of forward error correcting (FEC) codes. With FEC, extra bits are appended to the data by the FEC encoder at the transmitter. These extra bits help the FEC decoder at the receiver to detect and correct bits that become corrupted through transmission. Consequently, FEC enables the system to operate at a far lower received OSNR than would be possible without FEC, whilst maintaining an acceptable BER. The system’s target OSNR can be correspondingly reduced, which makes extended transmission distances possible. The strength of the FEC in correctingerrors is characterized in terms of the coding gain, which is the difference in the OSNR at which the system operates with a specified bit error rate (BER) without and with FEC. The coding gain is usually defined at the system’s target BER, for example for many 10-Gb/s-based and 40-Gbh-based terrestrial systems. The serial addition of the extra bits with FEC increases the bit rate. There are penalties associated with the expanded serial bit rate of FEC-encoded signals. To maintain the same noise performance, the required OSNR increases by the ratio of the rate expansion, for example a 7% rate expansion requires a 0.3 dB increase in OSNR. Thus the coding gain is often quoted as a Net Equivalent Coding Gain, which is obtained by subtracting the linear noise penalty associated with the expanded serial rate from the raw coding gain. In addition, higher transmission rates may, depending on the channel bit rate and the system design, entail greater transmission penalties from nonlinearities, dispersion, and PMD, and from limitations in transmitters and receivers. The higher bandwidth components required for the expanded rate may also be more expensive. Typically, for a given type of error correcting code, the stronger the FEC coding gain, the higher this overhead will be, but the better the FEC will be at correcting a severely corrupted signal. The most advantageousFEC encoding is that which requires the least rate expansion and delivers the greatest coding gain. 5. High-Capacity, Ultra-Long-Haul Networks 209 In fact, coding schemes have been found and implemented commercially that deliver very substantial coding gains with acceptable overhead. FEC is typically implemented on one or more integrated circuits specially designed for this purpose. Beyond the rate expansion and the coding gain, the complexity of the encoding algorithm and the feasibility of implementing it on an integrated circuit (IC) are critical considerations in designing a FEC code for ultra-long-haul systems. Some commercial optical communications systems placed the extra bits in the SONET overhead. This has the advantage that the aggregate bit rate transmitted is not increased, but the available overhead rate is relatively small and the FEC coding gain is correspondingly weak. In so-called “out of band of band” FEC, the serial bit rate carried on a wavelength is expanded above the rate required to carry the data. The most widely used FEC code is the ITU G.975 standard (ITU 1999), which is a Reed-Solomon (255,239). The RS(255,239) code increases the bit rate by 7%, from 9.95 Gb/s to 10.66Gb/s, but is able to correct a BER of IO-’ down to a BER below lo-’’, corresponding to a coding gain of approximately 6 dB. This code was first adopted for commercial undersea systems where it was widely used. Single-chip codecs capable of providing G.975 FEC for terrestrial systems are now offered for OC-192 lO-Gb/s transmission by a number of commercial vendors of telecommunications ICs, and codecs for 4O-Gb/s OC-768 transmission are currently under development. The ability to reduce OSNR requirements by up to 6 dB has a dramatic impact on system capabilities similar to that delivered by the OSNR improvement achieved with distributed Raman amplification. In a system limited by noise accumulation and not by transmission impairments, 6 dB of coding gain results in a quadrupling in the length of an unregenerated link. Advanced FEC schemes that deliver even greater coding gain will be critically important for future ultra-long-haul systems, and are the object of a great deal of work by equipment manufacturers and companies specializing in integrated circuits for the telecommunications industry. The most straightforward improvement would be to use a stronger Reed-Solomon code, and Kidorf et al. (2000) have reported a further 1.2dB increase in coding gain for a RS (255,223) code, but at the cost of increasing the rate expansion from 7% for the G.975 RS code to 14% for the RS(255,223) code. More powerful FEC schemes can be designed by using other coding approaches. Concatenated FEC codes use two FEC codes, an inner code and an outer code, and at the transmitter the data is sequentially encoded with the outer code and then the inner code, and at the receiver sequentially decoded with the inner code and then the outer code. Concatenated codes can significantly increase the coding gain, but at the cost of greater complexity and, in many cases, greater rate expansion to support the FEC overhead. Ait et al. (1999) proposed concatenation of the RS(255,223) with the RS(255,239) code, which provides an additional 2 dB of coding gain relative to the RS(255,239) code alone with a rate expansion of about 25%. This approach, based as it is 210 John Zyskind et a. l on concatenation of the already commerciallydeployed Reed-Solomoncodes, appears very attractive. PUC al. (1999) reported use of a Reed Solomon (255,239) code concateet nated with a soft-decision Viterbi convolutionalcode to produce a net coding gain of 10dB. However, this coding scheme requires an overhead of 113% or an expanded rate 2.13 times greater than the rate of the payload data. For high-speed fiber-optics systems where rate expansion entails not only proportionatelygreater OSNR requirements for an ideal receiver, but also more severe nonlinear transmission impairments and bandwidth limitations of optoelectronic components, such dramatic rate expansion is likely not practical. Sab and Lemaire (2001) have reported results of the performance calculated for a block turbo-code, which is an interative, soft-decision code. This code should deliver a net coding gain of 10dB with a more feasible rate expansion of 28%, but its implementation is likely to be complex. Keeton et al. (2001) have proposed the use of BCH both in conjunction with Reed-Solomoncodes in a concatenated scheme and for even greater coding gain in a two-dimensional product code. In the product code the data are encoded with the BCH code in each dimension of a two-dimensional array. The product code can be decoded iteratively by iterating alternately on the product code in each dimension, resulting in higher coding gain while materially increasing the overhead. Simulated results for these schemes indicate that they offer high net coding gain with modest rate expansion. Figure 5.6 shows 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Q factor per information bit (dB) Fig. 5.6 Simulated coding gains for three coding schemes: a RS(255,239) code, a BCH(239,223)-RS(255,239)concatenated code, and a BCH(255,239) product code. In each case the raw coding gain (dotted curve) and the Net Equivalent Coding Gain (solid curve) are shown. (From Keeton et al. 2001.) 5. High-Capacity, Ultra-Long-Haul Networks 211 the simulated results for three coding schemes, the widely used G.975 ReedSolomon (255,239) coding scheme, concatenated FEC with a BCH(239,223) code and an RS(255,239) code, and finally a BCH(255,239) product code. In each case the raw coding gain and the Net Equivalent Coding Gain are shown as a function of the Q-factor per information bit, which scales dB for dB as the OSNR in a noise limited system. The RS(255,239) code has a raw gain and a NECG of 6.1 dB with a rate expansion of of 6.4dB at a BER of 6.7%. The BCH(239,223)-RS(225,239) concatenated code has a raw coding gain of 8.5dB and a NECG of 7.9dB with a rate expansion of 14.3%. The BCH(255,239) product code has a rate expansion only slightly greater, 14.7%, but has a raw coding gain of 10.1 dB and NECG of 9.5 dB. For 4O-Gb/s transmission, FEC will be even more critical because of the extremelyhigh OSNR that would otherwise be required at the receiver. Because the transmission penalties and transceiver penalties increase dramatically with bit rate at this very high transmission rate, there will be much more pressure to deliver greater coding gain with less rate expansion. FEC with 7% overhead will be used in long-haul and ultra-long-haul applications, but any additional rate expansion may not be attractive at 40 Gb/s. D. P O n R MANAGEMENT A major limitation in long transmission systems is the deviation in power amongst the channels that results from the accumulation of optical amplifier gain nonuniformities. This impacts the system in two ways: those channels that decrease in power sufTer a penalty from reduced OSNR, while those channels that increase in power may degrade due to fiber nonlinearity. Figure 5 . 7 ~ input Spectrum -6 O@ut spectfum and OSNR 11 3 (a) without pre-emphasis -5 5 -5 x5 b 4 E 3 ‘1530 1540 1550 1560 ..- s4 b E3 $ 30 5 I z I 2 29 1530 1540 1550 1560 Ei5 I (b) with pre-emphasis E5 0 3 0 : E . E 4 . s $3 ‘1530 1540 1550 1580 wavelength [nm] s 4 E3 2 E 2 9 1530 1540 1550 1560 wavelength [nm] I Fig. 5.7 Simulated impact of accumulated EDFA gain ripple on channel powers and their OSNRs (a) without pre-emphasis and (b) with pre-emphasis.be-emphasis of the input spectrum is used to equalize the OSNRs for a l channels at the output. l 212 John Zyskind et al. illustrates the impact that a 0.8 dB EDFA gain ripple has on a flat input spectrum after only five spans. There is significant variation in both the output power and OSNR of the channels. Consequently, there will be substantial variation in error rate among the channels. For systems with sufliciently few concatenated optical amplifiers(or sufficientlyflat amplifiers)resultingin accumulated ripple less than about 10 dB, the effect of accumulated gain ripple can be successfully managed by pre-emphasizing the input spectrum to obtain an equal OSNR for all the channels at the output (Chraplyvy et al. 1993), as illustrated in Fig. 5.7b. This is done by redistributing the power at the booster amplifier among the various channels so that the total launched output power is still constant, but the OSNR at the end of the system is uniform among the channels. This improves the OSNR of the worst channels and helps to achieve uniform performance across the band. As the system must deliver error-free performance in all channels, it is the channel with the lowest OSNR that will impose the noise limit on the system’s engineering rules. The case illustrated, with only five spans and fairly flat amplifiers, is mild compared to a conventional long-haul system with six to eight spans and with greater gain ripple per amplifier. As the number of cascaded amplifiers increases, the required preemphasis needs to be much stronger; the OSNR before pre-emphasis of the weakest channel will be much lower than the average. The OSNR of the preemphasized channels will then represent a much more dramatic improvement in overall system performance. Due to the large number of cascaded amplifiers present in ultra-long-haul systems, the accumulation of gain ripple generally will be so great that it cannot be adequately managed with pre-emphasis. It then becomes necessary to reset the channel powers to desired levels at periodic sites along the link, as illustrated in Fig. 5.2. This can be done by demultiplexing the channels and adjusting the power of each individually and then multiplexing them again to continue on their way. Alternatively,the wavelengthscould be divided in bands, and the powers of the various bands could be balanced. This approach would be less expensive and more compact, but for high-capacity systems with close channel spacing, it will be necessaryto sacrifice some wavelengthchannels, and the systemcapacity will be correspondinglyreduced because filtering the bands with adequate cross-talk rejection in the demultiplexing and remultiplexing filters requires guard bands between the signal bands. Furthermore, if the bands are too wide, then the gain ripple within a band may exceed the range that can be corrected by transmitter pre-emphasis, and system performance would suffer. Alternatively, if the bands are too narrow, much of the available bandwidth would be eaten up by the guard bands, resulting in significantly reduced system capacity. An ideal device for maintaining the power balance among channels in ultra-long-haul systems would be a reconfigurablegain flatteningfilter (GFF). A reconfigurable GFF can be adjusted to compensate for the accumulated gain nonuniformity from the previous spans. Several different technologies 5. High-Capacity, Ultra-Long-Haul Networks 213 are currently being investigated and developed for implementing reconfigurable GFFs, including silica waveguide arrays (Doerr et al. 1999), MEMs (Ford et al. 1998), liquid crystal spatial light modulators and acousto-optic tunable filters (Kim et al. 1998). Ultimately, simple and low-cost reconfigurable GFFs may find a place in every optical amplifier. This would ensure optimum amplifier gain flatness and assist the manufacturer in eliminating the wide range of fixed GFFs that become necessary for producing different amplifier designs. While the cost and size of the first devices preclude their use in every amplifier, they are attractive for periodic use in ultra-long-haul systems. For DWDM systems with high channel counts and broad optical bandwidth, Stimulated Raman Scattering (SRS), an optical nonlinear interaction among the signal channels, transfers power from short-wavelength channels to long-wavelength channels, thereby inducing a tilt in the power spectrum. Additional tilt is induced in each span, and the tilt grows cumulatively with the number of spans. The magnitude of the tilt induced in each span is proportional to the number of channels, to their launched power and to the optical bandwidth that they occupy (Forghieri et QZ. 1999). As with other nonlinearities, the effects of SRS are reduced if the fiber’s effective area is larger and if the launched signal channel powers are lower. The effects of SRS tilt may be managed along with other sources of power ripple and power tilt by the use of pre-emphasis and periodic filtering. But for high-capacity systems with large total power and wide optical bandwidth, it may be necessary to combat the accumulation of SRS tilt by filtering at each amplifier site. In considering how such systems will actually be deployed in the field, it is important to develop automated procedures to ensure quick and accurate balancing of the channels of the DWDM channels, i.e., to adjust the powers of the individual transmitters and the spectral characteristics of the optical amplifiers and the power equalization sites. Otherwise, turning up the system or adding additional waves to an already operational system will be very time consuming and will also be susceptibleto errors that would degrade the performance of the system. As a 1 dB additional penalty will shorten the reach of a 2500-h, noise-limited system by 500 km,accurate and efficient pre-emphasis and power balancing are essential. 111. Transmission Impairments In order to realize the reach and the capacity made possible through the OSNR enhancementsmentioned in this chapter, distortions arising from transmission impairments must be limited so that the associated penalties are modest. These penalties are typically accounted for during the system design phase by budgeting an allowance for the associated penalties when the target OSNR is set, and the smaller these penalties are, the further the system reach and/or the 214 John Zyskind et al. higher its capacity. One of the objects in the design of high-capacity, ultralong-haul systems is to minimize the impact of these penalties on the target OSNR and to ensure that the penalties for these impairments do not exceed their allocated penalties. A. CHROMATICDISPERSION AND OPTICAL NONLINEARITIES Chromatic dispersion results from the dependence of the optical fiber’s index of refraction on optical wavelength and is the most important source of distortion of high-speed signals. As a result of chromatic dispersion, different frequencies of light travel at different speeds. For on-off keyed data transmission, where data 1s and Os are represented by the presence and absence of light, respectively, the pulses representing 1s contain a range of frequencies, and chromaticdispersion causes the pulses to spread as they propagate. Signal pulses correspondingto 1swill spread into the time slots for adjacent bits leading to the generation of bit errors when the distorted data trains are detected after transmission. The dispersion length LD, corresponding to the distance after which a pulse has broadened by one bit interval, is: 1 whereB is the bit rate, D is the dispersion, and AA is the spectralwidth of a pulse (Gnauck 1997).This length, which provides an estimate of the limit chromatic dispersion imposes on the length signals can be transmitted, is shorter when the bit rate is higher, when the dispersion is greater, or when the spectral width of the signal is greater. For high bit-rate long-haul transmission, external modulation of continuous wave diode lasers is used in preference to direct modulation of diode lasers because of the narrower spectrum that results. For signals produced by external modulation, the spectral width approximatesthe bit rate, B. The dispersion limit is then: 105 LD X D*B2’ (5.5) where LD is in km, is in ps/nm. km, B is Gb/s. The precise limit depends D and on the details of the modulation format and the design of the receiver circuitry, but Eq. 5.5 provides a reasonable approximation. The dispersion limit for externally modulated signals is inversely proportional to the square of the bit rate; for lO-Gb/s OC-192 signals on standard single mode fiber (SMF) with a dispersion of 17ps/nm km,it is about 60 km, corresponding to a residual dispersion of about 1000ps/nm, and for 4O-Gb/s OC-768 signals, it is less than 4 km, correspondingto about 60 pdnm. These lengths are a great deal shorter than the link lengths permitted by noise accumulation, and techniques to compensate and manage the dispersion are essential for high-capacity, ultralong-haul transmission. - 5. High-Capacity, Ultra-Long-Haul Networks 215 Two complementaryapproaches are used to manage the fiber chromaticdispersion: design of transmission fiber to reduce dispersion in the signal bands in the 1500-1600nm spectral region, and the use of dispersion compensation to provide negative dispersion to compensate the accumulated positive dispersion of transmission fiber. In a system with no nonlinear effects, transmission fibers designed to have no chromatic dispersion at about 1550nm (so-called dispersion shifted fiber, or DSF) could be used, and the small residual dispersion for channels whose wavelengths do not coincide with the zero dispersion wavelength could be compensated at any point in the link, as long as the final residual dispersion at the receiver is less than the dispersion limit. However, in typical high-capacity, ultra-long-haul-systems, it is desirable to launch the highest signal powers possible in order to maximize the OSNR (see Eq. 5.5), but as the launched power increases the impairments resulting from optical nonlinearities become more severe. The optimum launched signal power is therefore determined by the tradeoff between maximizing the launched power to maximize the OSNR and reducing the launch power to mitigate nonlinearities. At the optimal launch power, nonlinearities are sigdicant but not unduly severe, and the BER at the end of the system as a function of launch power is a minimum. In order to design the system with optimized launch power and thus the best system performance, it is not suEcient merely to control the residual dispersion at the end of the system. The local dispersion and the accumulated dispersion at each point along the length of the system are also important. Where dispersion compensation is used, both the amount of dispersion compensation and its placement are also important; if the dispersion map is not properly designed, impairments from nonlinearities will be severe, resulting in stricter limits on the launched channel power. Forghieri et al. (1997) have provided a comprehensivereview of optical nonlinearities and dispersion management. In this chapter we shall focus on some of the aspects that are of particular importance for high-capacity, ultra-longhaul systems. The effects of cross-phasemodulation and self-phasemodulation can be converted into amplitude modulation by chromatic dispersion if the accumulated dispersion is allowed to grow too large before compensation. The dispersion map must be designed so that either the dispersion of the transmission fiber is very low or its dispersion is compensated with sufficient frequency along the route to keep the accumulated dispersion sufficiently low at all points along the link. Against this need for keeping the accumulated dispersion low is the need for local dispersion to be suf€iciently large in order to minimize nonlinear interactions among the channels. Four-wave mixing and cross-phase modulation are especially severe when the local dispersion is low. Only fibers with substantial local dispersion are suitable for most DWDM applications. DSF with zero dispersion near 1550nm is not suitable. Standard single mode fiber (SSMF), which has zero dispersion at 1310nm and a dispersion of 17ps/nm .km at 1550nm (sometimes also called nondispersion shifted fiber, 216 John Zyskind et a. l or NDSF), is well-suited to DWDM transmission in both the C- and L-bands but residual dispersion accumulates very quickly. Nonzero dispersion shifted fibers (NZDSF) have been designed to support transmission of high-data-rate DWDM channels. The dispersion of NZDSF in the signal band is designed to be large enough to avoid significant impairments from four-wave mixing for practical systems with channel spacing of 50 GHz (about 0.4 nm) or greater (typically about 4 ps/nm .km)but small enough that dispersion accumulates much more slowly with propagation distance than for NDSF. However, even for NZDSF, the dispersion limit is only about 240 km for lO-Gb/s signals at the center of the C-band and about 16km for 40-Gbh signals. In fact, because the dispersion increases with wavelength, for transmission fibers the limit is even lower at the red end of the C-band and in the L-band. Thus, both NDSF and NZDSF require dispersion compensation for both 10-Gb/s and for 4O-Gb/s signal channels, but for NDSF much more is needed. For transmission over more than a few spans in systems where the channels cover a broad spectral range, careful attention must be given to matching the dispersion slope, defined as the derivative of dispersion D with respect to the wavelength, so that an acceptable dispersion map can be provided for all channels across the channel band. This requires matching the relative dispersion slope (i.e., RDS, the ratio of the dispersion slope to the dispersion) of the transmission fiber and the dispersion compensation. A mismatch in the relative dispersion slope between the transmission fiber and dispersion compensationwill cause a walk-off in the accumulated dispersion that varies across the channel band and increaseswith distance. Consequently, a small group of channels may have good transmission, whereas channels in the remainder of the band perform poorly. Dispersion compensating fiber (DCF) is single mode fiber designed to have a large dispersion opposite in sign to that of the transmission fiber to be compensated, and is the technology most widely used to compensatedispersion. It is generallylocated between the stages of optical amplifiers, which are designed with two amplifyingstagesand accessto the mid-stageregion to accommodate the dispersion compensation and optical add/drop filters. DCF typically has a far lower RDS than transmission fibers This disparity is especially severe for the NZDSF fibers, which have been widely deployed for high-speed DWDM applications. NZDSF fiber designs tend to have a large RDS because of their low absolute dispersion. In recent years there has been a drive to address this problem by reducing the dispersion slope of transmission fibers and increasing it for DCF. The walk-off in accumulated dispersion across the C-band with transmission distance for two different fiber and DCF combinations is illustrated in Fig. 5.8. In these figures the accumulated dispersion is plotted for three wavelengths across the C-band (1530, 1546, and 1562nm) for a system comprising twenty 100-km fiber spans with equal amount of dispersion compensation applied at the end of every span. Figure 5 . 8 ~ shows the accumulated dispersion for TrueWave Classic fiber, an early NZDSF, that has a 5. High-Capacity, Ultra-Long-Haul Networks 217 2500 r---- 1000 -2000 -1 500 -~~~~~ 1530nm 500 1000 distance [km] 1500 0 2000 (b) p 1500 1000 500 o 1 1562 nm / J 5 a c 8 9 7J 1546 nm f- I -500 -1 500 4 -1000 -2000 0 ' 1530 n 500 1000 1500 2000 Fig. 5.8 Dispersion maps illustrating the walk-off in accumulated dispersion across the band due to residual dispersion slope. The walk-off is large in case of (a) older TW-classic fiber that had a large slope, but is substantially smaller when using (b) TrueWave Reduced Slope fiber in conjunction with newer DCFs with greater slope. The far smaller walk-off allows the eyes to be recovered across the band. high RDS (where D = 2.7 ps/nm/km at 1550nm and S = 0.07 ps/nm2/km) compensated by an older DCF with negligible dispersion slope. The walk-off in accumulated dispersion across the band over 2000 km is 4000 pshm. It is tempting to try and use dispersion management at the end of the link on each channel independently to bring its accumulated dispersion to the optimum. Unfortunately,this works only for channels near the center of the band. This is illustratedby the simulatedeye diagrams for the two channels at the edges of the band that, in addition to the common dispersion compensation at each optical amplifier, have been passed through a dispersion compensating fiber located before the receiver, the length of which is adjusted for optimum performance of that individual channel. The eyes cannot be recovered because the impact from fiber nonlinearity on the signals while they are substantially dispersed 218 John Zyskind et al. cannot be undone using dispersion compensation at the end. This result is expected as the effects of dispersion and nonlinearity are not commutative. Figure 5.8b plots the accumulated dispersion for a similar system with TrueWave RS fiber that has a much lower dispersion slope (where D = 4.4ps/nm/km at 1550nm and S = 0.045ps/nm2/km) compensated by a newer DCF having RDS S / D = 0.0067 ps/nm. Here the walk-off in accumulated dispersion has been reduced to around 1100ps/nm. In this case, by using appropriate per-channel dispersionadjustment of the channels at the end of the link, the eyes can be recovered across the band. With further improvements in slope-matched dispersion compensation, the need for end-of-the-system, per-channel dispersion adjustment can be minimized or even avoided. The design of single mode DCF with sufficiently high dispersion slope to match that of NZDSFs is the object of intense work and is progressing rapidly. Single mode DCFs have been reported for two of the most widely deployed NZDSF fiber designs (Srikant 2001; Quang Le et al. 2001). However, the design and manufacturing of single mode DCFs with sufficientlyhigh relative dispersion slope to match the NZDSFs with the highest relative dispersion is quite challenging and, at this writing, commercial single mode DCF solutions are not yet generally available for some of the widely deployed varieties of NZDSF with high dispersion slope. Because of the importance of slope-matched dispersion compensation for ultra-long-haul systems at 10 Gb/s, and even more so at 40 Gb/s, other technologies to provide slope-matched dispersion compensation for NZDSF have been proposed. One possible alternative technology that can provide high negative dispersion slopes is higher-order mode dispersion compensation, first proposed in 1993 (Poole et al.) and presently the subject of revived interest to meet the need for slope-matched dispersion compensation for NZDSF (Gnauck et al. 2000; Ramachandran 2000). These are wideband devices that work across the C- or L-band. Signalsare passed through amode converter and transformed into a higher mode that propagates through a length of specially designed higher-order mode fiber that has negative dispersion and high RDS for this transmitted mode. A second converter at the output end transforms the signals back to single mode to continue down the link. In addition to the possibility of higher relative dispersion slope, higher-order mode dispersion compensation also offers the possibility of lower losses and, because of the larger effective area of the higher-order mode fiber, reduced nonlinear impairments compared to single mode DCFs. The Virtually Imaged Phased Array (VIPA), another potential technology for slope-matched dispersion compensation for NZDSFs with high relative dispersion slope, is a resonant device. It works for a prescribed comb of wavelengths, and is thus more limited for use in wide optical bandwidth applications than single mode DCF or higher-order mode dispersion compensators (Ishikawa and Ooi 1998; Shirasaki and Cao 2001). Because it is based on bulk optics, the VIPA will avoid the nonlinear effects that are produced in single mode DCFs. 5. High-Capacity, Ultra-Long-Haul Networks 219 Without dispersion compensating modules that adequately match the dispersion slope of transmission fiber, or to the extent that slope matching is imperfect, it may be necessary to perform dispersion slope equalization at periodic sites along the link (Zyskind et al. 2000), as shown in Fig. 5.2. This can be done either on a per-channel basis, which is expensive, bulky, and difficult to manage, or it can be done on groups of channels or bands (Haxell et al. 2000). As with power balancing by bands, the use of bands is less attractive than broadband slope-matched dispersion compensation because elaborate filtering arrangements are required for compensation by band. Such band filtering entails a reduction in system capacity because of the necessity for dead bands in which no channels can be supported and because, compared to use of broadband slope-matched dispersion compensation, it is more expensive, complex, and bulky. Nondispersion shifted single mode fiber has larger dispersion than NZDSF, and thus requires longer lengths of dispersion compensating fiber, which have higher loss. But the RDS of NDSF is smaller than that of NZDSF, and NDSF is the transmission fiber for which commercially available DCFs provide the best slope compensation. NDSF also has the largest effective area, which permits higher launched signal powers and the largest local dispersion, which tends to minimize interchannel nonlinear effects, particularly four-wave mixing. For 40-Gbls transmission where residual dispersion must be much smaller than for 10-Gbls, additional trimming will be required to compensate for imperfect dispersion slope compensation and for variations of dispersion arising from temperature variations experienced by transmissionfiber over the link (Kato 2000). Tunable dispersion compensation on a per-channel basis will be able to provide the required slope trimming, as well as adjust for temporal variations. Tunable dispersion compensationmay also be required for dynamically reconfigurable networks to compensate for differences in cumulative dispersion a wavelength channel will experience when its path through the network is changed. Component suppliers are now working on developing such devices (see Eggleton 2001 for a review of work on tunable dispersion compensation) using a variety of approaches. Tunable dispersion compensation has been reported for dispersion compensating chirped-fiber Bragg gratings controlled by temperature or strain tuning (Eggleton 1999; Eggleton 2000; Willner 1999; Fells 2000); integrated all-pass filters (Madsen 1999; Horst 2000), which are planar deviGes based on ring resonators; and the virtually imaged phased array devices (Shirasaki 2000), which as described previously, are bulk optic devices based on resonant multipath reflections. The ideal fiber span would have high local dispersion to mitigate nonlinearities, but would have accumulated dispersion that is small and equal to the value for optimal transmission performance (depending on the modulation format). It has been proposed to create such spans by combining fibers having large positive dispersion with fibers having large negative dispersion (Reverse 220 John Zyskind et al. Dispersion Fibers, RDF) in the same span. The positive dispersion fibers typically have a large effective area-wbich reduces optical nonlinearities-and small RDS, so as to facilitate matching of dispersion slope of the positive dispersion and reverse dispersion fibers. Such fiber spans would obviate the need for additional dispersion compensation at the amplifier sites In addition to the direct cost savings of dispensing with dispersion compensation, such fiber spans would be more suitable than currently deployed fiber networks for all Raman systems in which only distributed Raman amplifiers are used. Without the need to compensate the loss of dispersion compensation at amplifier sites, it would be more practical to compensate the fiber span loss with distributed Raman amplification, which would improve OSNR performance, reduce costs, as well as reduce the power transients that accompany changes in channel loading in EDFAs. Such fiber spans would also be well adapted to support dispersion-managedsoliton transmission. Such dispersion-managed spans are currently used for undersea systems where the fiber spans and the system equipment are designed together, and each fiber span is the same length. For terrestrial systems with nonuniform spacings between-amplifierhuts, it would be necessary to tailor the lengths of the two fiber types for each individual span to the total length required for that span. Before such networks are deployed it will be necessary to meet these practical engineering challenges in the deployment of such fiber networks. B. POLARIZATION EFFECTS Polarization effects arise from three phenomena: polarization-mode dispersion; polarization-dependent loss and polarization-dependent gain; or polarization hole-burning. Polarization-dependent losses and polarization hole-burningare effectsthat become important in systems where signals propagate over long distances through many optical amplifiers and other components. Polarization-mode dispersion (PMD) becomes increasingly important for higher data rates and the effectgrows with the squareroot of the link length. Polarization-dependent loss (PDL) arises from the fact that the optical components (such as isolators, filters, optical amplifiers,etc.) through which the signals pass have insertion loss that depends on the incident polarization. When the light impinges on a component in a polarization state with relatively less loss, the input power to subsequent optical amplifiers is raised and the OSNR improves. Conversely, when the light impinges on a component with a polarization state with relatively greater loss, the input power to subsequent optical amplifier is lowered and the OSNR is degraded. PDL manifests itself as a statistical impairment because of the random and variable evolution of the polarization state as light propagates over long distances in fiber; the PDL arising from each of many components is a random variable that varies with time. PDL is controlled in long systems by requiring that the PDL is small for 5. High-Capacity, Ultra-Long-Haul Networks 221 all components in the signal path. Components with sufficientlylow PDL for terrestrial ultra-long-haul applications are commercially available. Polarization-dependentgain (PDG) or polarization hole-burning(PHB) in the optical amplifiers arises from inhomogeneity in the saturation characteristics of the gain. Polarization hole-burning arises from the greater saturation experienced by erbium ions oriented so as to be preferentially saturated by light with the polarization of the signal. The result in a system with a long chain of optical amplifiers is that a strong saturating signal experiences more severe gain saturation than ASE with the orthogonal polarization because the signal always interacts most strongly with precisely those ions that will not be as deeply saturated for the polarization state that is orthogonal to that of the signal. This effect, unlike PDL, is deterministic, and the orthogonal ASE steals power from the signal at each optical amplifier as the signals propagate down the amplifier chain. But PHB is very weak, therefore it is only a problem for very long chains of amplifiers, such as are used in submarine systems. In addition, for DWDM systems with multiple channels, the polarization states of the differentwavelengthchannels are independent and their PHB contributions cancel each other. PDG also arises from the relative orientation of the pump light and the signal light. The system behavior of pump-induced PDG is similar to PDL, but, like PHB, it is very weak and in multistage EDFAs with multiple pumps the PDG tends to get washed out. Polarization-mode dispersion (PMD) is the most important polarization effect for high-capacity, ultra-long-haul systems with high bit-rate channels. PMD arises from the birefringence in the fiber that gives rise to differential group delay between the two principal states of polarization. PMD is manifest as a time varying and statistical pulse broadening and pulse distortion because the perturbations to the fiber symmetry that give rise to the birefringence vary randomly in orientation along the fiber and are also dependent on environmental variations, particularly temperature. For lengths of fiber longer than the correlation length for the birefringence, which is generally the case for a fiber span, PMD is characterized in terms of the differential group delay (DGD) between the two principal states of polarization after a given length of fiber. Because of the statistical nature of PMD, the differential group delay increases with the square root of the length of the fiber and is expressed in units of p s / G . The PMD of a fiber span is typically specified in terms of a mean PMD, which is the average over time of its net DGD. The statistical distribution of DGD about this mean is determined by the physics of PMD and follows a Maxwellian distribution. Because of its statistical nature, the possibility of errors arising from PMD can never be totally eliminated, but the probability of an outage, defined as the probability of a penalty greater than the OSNR margin assigned to PMD, can be calculated from the mean PMD and its statistical distribution. In practice, a given amount of margin is allocated to PMD impairments, and the probability of an outage is defined as the probability that the instantaneous 222 John Zyskind et al. DGD will exceed that value that induces a penalty equal to this margin allocation. When the DGD exceeds this value, the possibility of PMD-induced bit errors at the end of the system cannot be excluded. For modest DGD, the penalty due to PMD goes up quadratically with both the bit rate and with the DGD (Poole and Nagel 1997). This means that the acceptable mean PMD is inversely related to the bit rate, but that as the instantaneous PMD increases above the acceptable level, the penalty increases rapidly. It follows that the acceptable mean differential group delay is proportional to the bit period and is generally of the order of 10-15% of a bit period depending on the modulation format and other detailsof the systemdesign and the permitted outage probability (Poole and Nagel 1997). In addition to DGD, or first-order PMD, second-order PMD must be considered when the PMD varies over the bandwidth of the source. Components of higher-order PMD tends to increase as first-order PMD increases (Shtengel et al. 2001), therefore higher-order PMD becomes relatively more important when the PMD is larger, and in fact becomes dominant for 10Gb/s per second above about 20 ps of mean PMD (Taga et al. 1998). For recently manufacturedfiber with a mean PMD of 0.125 ps/& or less (Noutsias and Poirier 2001), PMD is not an obstacle to ultra-long-haultransmission at 10Gb/s. However, for older fiber where PMD can be significantly larger or for 4O-Gb/s ultra-long-haultransmission,PMD can limit the reach of a system. For IO-Gb/s ultra-long-haul transmission on older vintage fiber and for 40-Gb/s ultra-long-haul transmission over a wide range of fibers, PMD compensation will be necessary. Because of the importance of higher-order PMD for larger PMD where compensation is needed, it is likely compensation of second-order PMD, in addition to fist-order PMD, will be necessary in a useful compensator. The development of optical and electrical components to counteract the PMD is an area of active research and commercial development. See Penninckx and Lanne (2001) for a recent review. C MODULATIONFOR2MAT . The modulation format is also an important consideration in the design of ultra-long-haul systems. The most commonly used format in long haul optical communications is Nonreturn to Zero (NRZ) modulation, in which 1s are represented as rectangular pulses occupying the full bit period and Os by the absence of a pulse. N R Z pulses are normally formed either by directly modulating a semiconductor laser (for DWDM transmission generally a single longitudinal mode Distributed Feedback Laser) to turn its power on for a data “1” and off for a “0,” or by using a continuous wave semiconductor laser followed by an external modulator (or sometimes by an integrated modulator on the same semiconductor chip with the diode laser) passing the laser’s optical power for a “1” and blocking it for a “0.” 5. High-Capacity, Ultra-Long-Haul Networks 223 In high bit rate (2.5 Gb/s per channel or greater), long-haul transmission, it is necessary to use external modulation. Direct modulation of semiconductor lasers induces chirping, and the associated spectral broadening entails unacceptable dispersion penalties, as can be seen by reference to Eq. 5.4. External modulators for IO-Gb/s applications are commercially available and widely deployed. The two most widespread technologies are electro-optically controlled Mach-Zehnder interferometers fabricated in LiNbO3 and electroabsorption modulators fabricated in semiconductor-based devices. In fact, a modest amount of chirp of the proper sign (a chup parameter of approximately -0.7) results in improvement in performance because dispersion initially acts to narrow the c h q e d pulse (Agrawal 1992). LiNb03-based Mach-Zehnder modulators are commercially available that can provide properly prechirped NRZpulses. Return to Zero (RZ) modulation uses pulses that are substantially narrower than a bit period to represent “ls,” so even for consecutive “1s” the power level returns to zero between successive pulses. For practical receiver designs, RZ modulation results in receiver performance superior to that for NRZ modulation by 1 to 2 dB (Boivin and Pendock 1999). With RZ modulation, systems can also be designed to be more robust against impairments such as self-phase modulation and polarization mode dispersion (Taga et al. 1998; Sunnerud et al. 2001). In fact, if the dispersion map and signal powers are appropriately managed and the input pulse is properly shaped to produce solitons (Mollenauer 1997) or dispersion managed solitons (Suzuki et al. 1995; Smith et al. 1997; Cao and Yu 2001) the effects of dispersion and self-phase modulation can be held in balance. In this case, pulse spreading induced by dispersion is balanced by pulse narrowing induced by self-phase modulation so that, in the case of classical solitons, the pulse shape is maintained for transmission over arbitrary distance, or so that, in the case of dispersion managed solitons, it returns to the same shape at the end of each span for an arbitrary number of spans. Carrier-suppressed RZ (Miyamoto et al. 1999; 2001) and chirped RZ (Bergano et al. 1997) modulation have also been proposed as modulation formats to further mitigate nonlinear interactions. Although RZ modulation offers improved performance, transmitters for RZ modulation are more complex and expensive than those for NRZ modulation. Generally, two modulation stages are required, one to form the pulses and a second to modulate the pulses to imprint the data on the signals. For example, whereas NRZ modulation can be implemented with a single LiNbO3 Mach-Zehnder modulator, RZ modulation requires either two separatemodulators or a two-stagemodulator. For dispersion-managedsoliton transmission, the dispersion map must generally be controlled more tightly than for NRZ transmission, which can pose a practical challenge for commercial systems deployed on carriers’ actual fiber networks with highly variable amplifier-hut separations. 224 John Zyskind et a. l IV. Optical Networking A. CHANGING NETWORK NEEDS The 1990s marked tremendous advances in optical networking. In particular, the 1990s saw the widespread deployment of SONET systems. SONET, o Synchronous Optical Networking, was a tremendous step forward for r the public telecommunications network infrastructure over the preceding plesiosynchronoussystem, enabling networks to be built that provided switching granularity (down to a voice call), scalability (up to 40 Gb/s or beyond), and high availability though the use of automatic protection switching (APS) and ring-based restoration. In addition, the 1990ssaw the widespread deployment of point-to-point DWDM systems used for fiber multiplication. By the end of the 199Os, most major public telecommunicationsnetworks were primarily built by stacking SONET rings on top of one another through the use of DWDM. The 1990s also saw an explosion in data traffic driven by both corporate and public demands. IP, driven by the Internet, corporate requirements, and the World Wide Web, became the dominant data networking technology by the end of the 1990s. The dominatingtrend of IP is continuing, with most technologists agreeing that IP will within the next 10 years become the technology of choice to carry all traffic, including voice. As IP traffic grew, IP routers grew in size, speed, and complexity, which drove fundamental changes in the way backbone IP and optical networks were constructed. At first, data growth helped spur the deployment of WDM as more and more SONET rings were stacked. However, as the speed of the backbone router ports increased, the core network evolved from a highly layered one (e.g., IP over Frame Relay over ATM over SONET over DWDM) to a flat IP-over-wavelength architecture. The latter architecture, also known as an IP-over-glass architecture,was motivated, enabled, and in a large sense required when IP router ports started to run at the speed of a wavelength (first OC-48c, now OC-l92c, and moving to OC-768c). So dramatic was the failure of SONET rings to respond to the high-speed service requirements on IP, that many backbone networks had bifurcated by early 2000, with the SONET network providing voice and lower-speed services, and the WDM network providing wavelengths to the IP layer, SONET layer, ATM layer, and for sale to external customers as “transparentunprotectedwavelength services.”These customers in turn used these wavelengths to construct their IP, SONET, and ATM networks. In summary, the core of the public network infrastructure is changing for three fundamental reasons. First, the service requirements are changing from voice to data, electrical to optical, and static to dynamic. Second, the existing SONET-over-WDMinfrastructure is insdlicient to meet those changing requirements. Third, new technologies and architecturesare available to meet the new requirements in a far more economical and scalable fashion. For these 5. High-Capacity, Ultra-Long-Haul Networks 225 reasons, the first decade of the new millennium will see the maturation of the second phase of optical networking, the intelligent optical network, beginning with IP over wavelengths as the first step. The rest of this section will discuss the use of ULH transmission as a key enabler of this new architecture. B. THE VALUE OF ULTRA-LONG-E4UL TRANSMISSION Data networks are not the same as voice networks. There are many fundamental differences, including the use of packet switching, statistical multiplexing, and the use of both connectionless data transfer and logical connectionoriented data transfer in data networks versus the use of circuit switching, time division multiplexing, and physical-connection-oriented bit transfers in voice networks. There are also fundamentally different traffic and service requirements with data traffic being characteristically distance insensitive, dynamic, and unpredictable, and voice tr&c being characterized as more local, steady, and predictable. For these and other reasons, the SONET ring networks optimized for the voice network are no longer optimal, or even adequate, to continue to build the public data infrastructure. Because of the different traffic requirements, the different underlying technologies, and the different historical evolution of the two technologies (data being an unregulated and relatively new technology), data networks are not planned, engineered, or constructed the same way as voice networks. Two key attributes of an IP backbone are: (1) that it is typically an irregular mesh topology, often evolving in an organic and hard-to-predict fashion; and (2) the The trunks of the mesh run at wavelength speeds (OC-48c/OC-192c/OC-768~). design and optimization of an IP network is a complicated process, but ideally the trunks should be determined from the traffic requirements and not the physical layout of the fiber backbone. In fact, this is currently done with wavelength services criss-crossingthe country interconnectingdistant routers. This enables the construction of flatter IP networks whose topologies are better optimized to the traffic. Such a design has the benefits of using fewer IP router ports (which reduces costs and keeps the routers smaller) and reducing latency. In fact, the construction of long trunks is absolutely critical in the construction of a scalable optical Internet in that the IP layer traffic is essentially off-loaded to the more scalable optical layer. Today, these trunks are constructed with the concatenation of shorterreach WDM systems. For instance, a 5000-km, cross-continental trunk, for example, from New York to Los Angeles, might be required to cross 8-10 or more DWDM systems with a corresponding number of costly optical-toelectrical-to-optical (OEO) conversions along the way. It is the reduction of these intermediate OEO conversions by using optical bypass that saves money, space, and power and leads to a far more economical and manageable network. Optical bypass and long trunks are not only useful in constructing IP networks. In fact, one of the driving forces behind ULH transmission is the move away from an interconnected SONET ring-based architecture to an optical 226 John Zyskind et al. mesh network. Similar to an IP network, the optical mesh network topology is optimized by considering the traffic demands, not the physical fiber topology. Whether for express wavelengths for an IP-over-glass architecture, the construction of an optical mesh network, or for more narrow applications, ultra-long-haul transmission can greatly reduce costs by reducing the amount of required OEO conversions in the network. In general, the higher the bandwidth-distance requirements on the traffic, the more motivation there is for optical bypass and ULH transmission. C ALL-OPTICAL NETWORKS' . With the advent of ultra-long-haul transmission, it is now possible to construct IP and optical mesh networks with long express trunks with little or no intermediate OEO conversion. Such a technology has significant architectural implications on building the network, as has already been alluded to. One of these implicationsis the ability to construct all-optical networks. An all-optical network is one in which the signal remains in the optical domain from the source to destination without any conversion to electronics within the network. The primary motivation for an all-optical network is that optics, and not electronics, is the most cost effective way to tap the multi-Tb/s capacity of the optical fiber, i.e., through the use of optical bypass and optical switching nodes. Another motivation is that fiber transparency provides an element of future proofing the network against advances in technology. Two other architectural implications were already discussed: that long express wavelengths enable the construction of flatter IP and optical mesh networks optimized to meet the traffic demands rather than on the physical layout of the fiber. With automation in the all-optical layer, these express trunks can be quickly brought into service and/or reconfigured to meet changing tr&c demands. All-optical networks have been commercialfor some time now, albeit in very limited form; linear and ring systems with intermediate wavelength-selective adddrop are widely deployed in long-haul as well as metropolitan networks. By slotting in a transponder card of the appropriate wavelength, the carrier can route the signal from the source node to the destination node without any conversion to electronics. The technologies that have enabled these static wavelength routing networks are well known and include DFB lasers, LiNbOs im modulators, thin f l filters, fiber Bragg gratings, etc. Tunablelasers and reconfigurable adddrop technologies promise to enable configurable versions of these systems. ULH transmission enables the construction of larger networks, to the point that now linear or ring network topologies are insufficient to realize the full benefit of the transmission capabilities, i.e., the reach exceeds the normal distance between the major backbone junction nodes (a node with 5. High-Capacity, Ultra-Long-Haul Networks 227 a fiber-route degree greater than 2; typically 3 4 in a continental U.S. network, or sometimes even higher). Thus, having surpassed the unregenerated reach required to connect junction nodes without intermediate regeneration, carriers can now further reduce OEO conversion cost by building a ULH mesh network optically interconnecting many or all of these junction nodes. These junction nodes may be manually patch-paneled or may contain alloptical switches. The choice of a manual versus automatically configurable node depends on the trade-off between the cost of the optical switches and the benefits of wavelength configuration. In fact, there are more detailed cost trade-offs involving levels of network configuration/automation dependent on such things as the tuning range of the transmitters, and in some cases the tunability of the dispersion compensation. Such trade-offs are complicated and time and business sensitive depending upon such considerations as the dynamic nature of the IP trunks, OXC trunks, or wavelength services, the predictability of such traffic, the potential lost opportunity costs, as well as the operational cost savings of the automatic node over the manual node, to name more than a few. D. ALL-OPTICAL ISLANDS ULH and all-optical mesh networks can greatly reduce the OEO conversion cost in the network. However, these cost savingsdo not come without potential drawbacks. First, because of the lack of standards for optical midspan meet, the larger the mesh, the larger the portion of the network built from one vendor. Unlike the o/e/o intelligent optical networking interoperability standards that have been demonstrated several times and that are continuing to be addressed at various industry fora (IETF GMPLS, OIF UNI, ITU G.ASON,ODSI UNI), there is no current or expected activity to standardize the interfaces necessary to support an optical midspan meet. Thus, for the foreseeablefuture, all-optical networks will have to be interconnected through OEO. Second, for a given technology at a certain point in time, the capacity (per lit fiber) of the optical mesh will decrease as the reach of the wavelengths is increased. If the supported capacity is sufficient, then the reduced OEO cost will result in reduced network cost. However, if more capacity is required, extra fibers will have to be lit, increasing the amplifier costs. Thus, infinitely extending reach (toward the goal of one optical hop across the core backbone) at the expense of capacity may or may not be the most cost effective way to build the network. Third, although optical switching is maturing, the level of configuration/ automation in an all-optical junction node is less than in its OEO counterpart. For example, the OEO nodes typically support STS-1 grooming, fully automatic circuit setup, various protection and restoration mechanisms, logical dissociation between the client-side interfaces and the network-side interface/wavelength, client-side APS, rate adaptation (e.g., 10G to 40G), 228 John Zyskind et al. conversion of transmission formats such as from NRZ to RZ or from FEC to SuperFEC, wavelength changing, etc. Thus, there is a balance to be made in the core backbone architecture between functionality and cost, and a carl rier may choose to make some nodes all OEO, others al or mostly OEO, and others hybrid. For these and other reasons, larger backbone networks will continue to be built from an interconnectionof all-opticalnetworks, or islands. These islands are surrounded by OEO performing regeneration, wavelength changing, rate adaptation, format conversions, etc. These islands not only encompass the traditional static linear point-to-point systems, the emerging ULH meshes, but also newer, short fat-pipe OC-768 systems and future islands using as yet undeveloped technologies. Over time these islands will become larger, support more capacity, and become more sophisticated in their functionality. Future islandswill likely support optical regeneration as well as wavelength changing. Because of the OEO around the islands, multivendor and multitechnology networks are easily constructed. It also allows the use of different technologies in different parts of the network, for example, high-capacity, short fat-pipe systems in shorter distance, high density areas and longer skinnier ultralong-haul pipes in more widely spaced, sparse parts of the network. The use of OEO switches ties this bandwidth together into a complete multivendor multitechnology mesh network. V. Conclusions Some of the most exciting technological advances in optical communications have made possible dramatically increased reach for high-capacity DWDM systems. The ability to extend high-capacity optical paths to thousands of kilometers between regenerators makes possible dramatic reductions in network cost as well as scalable network architectures, based on increased optical functionality, to support the rapid growth of data-based servies. Acknowledgments The authors would like to acknowledge useful conversations and fruitful collaborations with our colleagues at Sycamore Networks. References 0. Ait Sab, “FEC Techniques in Submarine Transmission Systems,” Paper TuF1, Proceedings OFC 2001, OSA, Anaheim, California, 2001. 0. Ait Sab and V. Lemaire, “Block turbo code performances for long-haul DWDM optical transmission systems,” Proceedings of OFC 2000, Baltimore, MD, paper ThS5, pp. 280-282,2000. 5. High-Capacity, Ultra-Long-Haul Networks 229 N. Bergano, “Undersea Lightwave Systems Design,” in Optical Fiber Communications IIIA, I. P. Kaminow and T. L. Koch, eds., pp. 302- 335, Academic Press, 1997. N. Bergano, et al., Proceedings OFC 1997, Paper PD16, 1997. L. Boivin and G. Pendock, “Receiver Sensitivity for Optically Amplified R Z Signals with Arbitrary Duty Cycle,” in OpticalAmpliJers and TheirApplications, 1999, Nara, Japan: Paper ThB4, pp. 106-109 (Optical Society of America). X. Cao and Y . Yu, “Ultra Long-Haul DWDM Transmission via Nonlinearity Management,” Optical Amplifiers and Their Applications, OSA Trends in Optics and Photonics, vol. 44, A. Mecozzi, M. Shimizu, and J. L. Zyskind eds., Optical Society of America, Washington DC, 2001, pp. 203-210. A. R. Chraplyvy, J. A. Nagel, and R. W Tkach, “Equalization in Amplified WDM . Lightwave Transmission Systems,” IEEE Photon. Technol. Lett., vol. 4, p. 920, 1992. V. Dominic, A. Mathur, and M. Ziari, “Second-order Distributed Raman Amplification with a High-Power 1370-nm Laser Diode,” in O p t i d Amplijiers and Their Applications, 2001, Stresa, Italy, OMC6 (Optical Society of America). V. Dominc, E. Mao, J. Zhang, B. Fidric, S. Sanders, and D. Mehuys, “Distributed Raman Amplification with Co-PropagatingPump Light,” in O p t i d AmpliJers and Their Applications, 2001, Stresa, Italy, OMC5 (Optical Society of America). C. R. Doerr, et al., “DynamicWavelength Equalizer in Silica Using the Single-FilteredArm Interferometer,”IEEE Photon. Technol.Lett., vol. 11, pp. 581-583, 1999. B. J. Eggleton, J. A. Rogers, P. B. Westbrook, and T. A. Strasser, “Electrically Tunable Power Efficient Dispersion Compensating Fiber Bragg Grating,” IEEE Photon. Technol. Lett., vol. 11, pp. 854-856, 1999. B. J. Eggleton, A. Ahuja, P S. Westbrook, J. A. Rogers, P Kuo, T. N. Nielsen, and . . B. Mikkelsen, “Integrated Per-ChannelDispersion Compensating Bragg Gratings,” Journal of Lightwave Technology, v01. 18,2000. B. J. Eggleton, “Dynamic Dispersion Compensation Devices for High-speed Transmission Systems,” Paper WH1, OFC 2001, Anaheim, California (Optical Society of America). J. A. J. Fells, et al., “Twin Fiber Grating Adjustable Dispersion Compensator for 40 Gbith,” ECOC 2000, Munich, Germany, Postdeadlinepaper 2.4,2000. Y . Emori and S. Namiki, “100-nm Bandwidth Flat Gain Raman Amplifiers Pumped and Gain-Equalized by 12-wavelength-channel WDM High-Power Laser Diodes,” OFC 1999, San Diego, CA, Postdeadline paper PD19. J. E. Ford and J. A. Walker, “DynamicSpectralPower EqualizationUsing Micro-OptoMechanics,” IEEE Photon. Technol. Lett., vol. 10, pp. 1440-1442, 1998. F. Forghieri, R. W. Tkach, and A. R. Chraplyvy, “Fiber Nonlinearities and Their Impact on Transmission Systems,” in Opical Fiber Communications IIIA, Ivan P . Kaminow and Thomas L. Koch, eds, pp. 196264, AcademicPress, San Diego, 1997. L. D. Garrett, et al., “Demonstration of VIPA Device for Tunable Dispersion Compensation in 16 x 10-Gb/s WDM Transmission over 480 km Standard Fiber,” OFC 2000, Baltimore, Maryland (Optical Society of America). A. H. Gnauck and R. M. Jopson, “Dispersion Compensation for Optical Fiber Systems,” in Opical Fiber Communications IIIA, Ivan P =now . and Thomas L. Koch, eds., pp. 162-195, Academic Press, San Diego, 1997. 230 John Zyskind e t a . l A. H. Gnauck, L. D. Garrett, Y. Danziger, U. Levy, and M. Shur, “Dispersion and Dispersion Slope Compensation of NZ-DSF for 40 Gbps Operation over the Entire C-band,” Technical Digest of OFC 2000, San Diego, CA, PD-8,2000. P. B. Hansen, et al., “Capacity Upgrades of Transmission Systems by Raman Amplification,” IEEE Photon. Technol. Lett., vol. 9, pp. 262-264, 1997. I. Haxell, et al., “2410 km All-Optical Network Field Trial wt 10 Gb/s DWDM ih Transmission,” OFC 2000, Postdeadline paper PD41. I. Haxell, M. Ding, A. Akhtar, H. Wang, and P. Farmgia, “52 x 12.3 Gbit/s DWDM Transmission over 3600 km of True Wave Fiber with 100km Amplifier Spans,” Optical Amplifiers and Their Applications, OSA Trends in Optics and Photonics, vol. 44, A. Mecozzi, M. Shimizu, and J. L. Zyskind, eds., Optical Society of America, Washington, DC, 2001, pp. 217-219. F. Horst, “Tunable Ring Resonator Dispersion Compensators Realized in HighRefractive-IndexContrast SiON Technology,” ECOC, Munich, Germany, Postdeadline paper PD2.2,2000. G. Ishikawa and H. Ooi, “Demonstration of Automatic Dispersion Equalization in 40 Gbps OTDM Transmission,” ECOC 1998, WdC-6, pp. 519-520,1998. ITU-T G.975, November 1999, “Forward Error Correction for Submarine Applications.” T. Kato, Y. Koyano, and M. Nishimura, “Temperature Dependence of Chromatic Dispersion in Various Types of Optical Fiber,” Optics Letters, vol. 25, pp. 115 6 1 158, 2000. S. Keeton, S. Sridharan, and M. Jarchit, “EnablingNext Generation Optical Networks with Forward Error Correction,” NFOEC 2001 Proceedings, pp. 54-59, Baltimore, Maryland, 2001. H. Kidorf, et al., “PerformanceImprovement in High-Capacity, Ultra-Long-Distance, WDM Systems Using Forward Error Correction Codes,” OFC 2000, San Diego, CA, THS3, pp. 274-276. H. S. Kim, et al., “Actively Gain Flattened Erbium-Doped Amplifier over 35 nm by Using All-Fiber Acousto-OpticTunable Filters,” IEEE Photon. Technol.Lett., vol. 10, pp. 790-702,1998. C. K. Madsen, et al., “Integrated All-Pass Filters for Tunable Dispersion and Dispersion Slope Compensation,” IEEE Photon. Technol. Lett., vol. 11, pp. 1623-1625, 1999. Y. Miyamoto, et ab, Electronics Letters, vol. 35, no. 23, pp. 2041-2042, 1999. Y. Miyamoto, S. Kuwahara, A. Hirano, Y. Tada, Y. Yamane, and H. Miyazawa, “Reduction of nonlinear crosstalk of carrier-suppressed RZ format for 100GHzspaced Nx43-Gbith WDM in non-zero shifted band,” Proceedings of ECOC 2001, Paper Th.B.3,2001. R. E. Neuhauser, P. M. Krummn’ch, H. Bock, and C. Glingener, “Impact of Nonlinear Pump Interactions on Broadband Distributed Raman Amplification,” Paper MA4-1, OFC 2001, OSA, Anaheim, California, 2001. C. D. Poole, et al., “Elliptical-CoreDual Mode Fiber Dispersion Compensator,” IEEE Photon. Technol. Lett., vol. 5 , pp. 194-197, 1993. 5. High-Capacity, Ultra-Long-Haul Networks 231 P. Noutsios and S. Poirier, “PMD Assessment of Installed Fiber Plant for 40 Gb/s Transmission,” NFOEC 2001 Proceedings, pp. 1342-1 347, Baltimore, Maryland, 2001. D. Penninckx and S. Lanne, “Reducing PMD Impairments,” Paper TuP1, OFC 2001, OSA, Anaheim, California, 2001. A. Puc, E Kerfoot, A. Simons, and D. Wilson, OFC 1999, ThQ6, pp. 255-258. N. T. Quang Le, T. Vng, and L. Gruner-Nielsen, “New Dispersion Compensating Module for Compensation of Dispersion and Dispersion Slope of Non-Zero Dispersion Fibres in the C-band,” Paper TuH5, OFC 2001, OSA, Anaheim, California, 2001. S. Ramachandran, B. Mikkelsen, L. C. Cowsar, M. F Yan, G. Raybon, L. Boivin, . M. Fishteyn, W. A. Reed, P. Wisk, and D. Brownlow, “All-fiber, Grating-Based, Higher-Order-Mode Dispersion Compensator for Broadband Compensation and 1000-km Transmission at 40 Gbps,” ECOC 2000, PD-2.5,2000. K. Rottwitt, A. Stentz, T. Nielsen, P. Hansen, K. Feder, and K. Walker, “Transparent 80-km Bidirectionally Pumped Distributed Raman Amplifier with Second-Order Pumping,” in European Conference on Optical Communications 2000, Nice, France, 11-14. M. Shirasaki, et al., “Variable Dispersion Compensator Using the Virtually Imaged Phase Array (VIPA) for 40-Gbh WDM Transmission Systems,” ECOC 2000, Munich, Germany, Postdeadline paper 2.3,2000. M. Shirasaki and S. Cao, “Compensation of Chromatic Dispersion and Dispersion Slope Using a Virtually Imaged Phased Array,” Paper TuS1, OFC 2001, OSA, Anaheim, California, 2001. G. Shtengel, E. Ibragimov, M. Rivera, and S. Suh, “Statistical Dependence Between First and Second-Order PMD,” Paper MO-3, OFC 2001, OSA, Anaheim, California, 2001. N. J. Smith, N. J. Doran, W Forysiak, and E M. Knox, “Soliton Transmission Using Periodic Dispersion Compensation,” Journal ofLightwave Technology,vol. 15, p. 1808, 1997. V. Srikant, “Broadband Dispersion and Dispersion Slope Compensation in High Bit Rate and Ultra Long Haul Systems,” Paper TuHl, OFC 2001, OSA, Anaheim, California, 2001. H. Sunnerud: M. Karlsson, and P. A. Andrekson, “A Comparison Between NRZ and RZ Data Formats with Respect to PMD-induced System Degradation,” Paper WT3, OFC 2001, OSA, Anaheim, California, 2001. M. Suzuki: I. Morita, N. Edagawa, S. Yamamoto, H. Taga, and S. Akiba, Electronics Letters, vol. 31, p. 2027, 1995. H. Taga, et al., “Polarization Mode Dispersion Tolerance of 10-Gbitls NRZ and RZ Optical Signals,” Electronics Letters, vol. 34, pp. 2098-2100, 1998. J. L. Zyskind, J. Nagel, and H. D. Kidorf, “Erbium-Doped Fiber Amplifiers for Optical Communications: in Optical Fiber Communications IIIB, I. P. Kaminow and T. L. Koch, eds., pp. 13-68, Academic Press, San Diego, 1997. J. L. Zyskind, G. J. Pendock, M. J. L. Cahill, G. D. Bartolini, J. K. Ranka, and S. Y. Park, “High Capacity, Ultra-Long-Haul Transmission,”Pmc. o NFOEC 2000, f Denver, CO, 2000. Chapter 6 Pseudo-Linear Transmission of High-speed TDM Signals: 40 and 160 Gb/s Renk-Jean Essiambre, Gregory Raybon, and Benny Mikkelsen* Bell Laboratories, Lucent Technologies. Holmdel, New Jersey 1. Introduction High-capacity fiber-optic communication systems transport bits of information (optical pulses) by having them first time-division multiplexed (TDM) to form a channel centered at a given wavelength. Many channels at different wavelengths are then wavelength-division multiplexed (WDM) together and launched in an optical fiber for transport. A given capacity can be implemented through a large number of low-speed TDM channels or a reduced number of high-speed TDM channels. By July 2001, the highest bit rate per channel in state-of-the-art installed commercial WDM systems is 10 Gb/s. The next anticipated higher standard bit rates for the synchronous optical network (SONET) and synchronous digital hierarchy (SDH) standards are 40Gb/s and 160 Gb/s. The deployment of transmission systems based on 40 Gb/s per channel and above (from here referred to as high-speed TDM systems) can be advantageous in many ways. Benefits include a reduced number of opto-electronic components used for transport such as lasers, modulators, and receivers. The size of optical networking components used for routing and switching, such as optical add-drops and cross-connect, is also reduced dramatically for low channel count. Such reductions in component count, complexity, and size generally lead to a decrease in overall system size, cost, electrical power consumption, and channel sparing. Besides the advantages in hardware mentioned above, provisioning, operation, administration, and maintenance (POAM) are also simplified by the deployment of high-speed transport as the number of paths to monitor and restore in case of hardware failure or malfunction is reduced. Also, it is cheaper to spare a single high-speed transponder than several lowspeed WDM transponders. Furthermore, as networks predominantly carry Internet protocol (IP) traffic, it is important to take into account that a few high-data-rate links perform better than many low-data-rate links in a packetswitched network. While offering many advantages, high-speed systems have not been deployed because they faced numerous challenges. * Author’s present address: Mintera Corporation,Lowell, Massachusetts. 232 OPTICAL FIBER TELECOMMUNICATIONS, VOLUME IVB Copyright 0 2002, Elsevier Science (USA). All rights of reproduction in any form reserved. ISBN 0-12-395173-9 6. Pseudo-Linear Transmission of High-speed TDM Signals 233 A first challenge is related to the development of commercial-grade broadband high-speed electronics for high-speed transmitters and receivers (see chapters found elsewhere in this book). Depending on the availability of such components, it may become necessary to introduce optical multiplexing and/or demultiplexingtechniques, especially if speeds higher than 40 Gb/s are considered. Transmission of high-speed signals over optical fibers by itself brings an important set of challenges. Higher-speed signals become increasingly demanding in dispersion accuracy. For instance, in the absence of fiber nonlinearity and dispersion slope, a 40-Gbls-based WDM system has 16 times less dispersion margin than a 10-Gb/s-based WDM system for a given modulation format. Such high-speed systems may exhibit sensitivity to dispersion variations induced by temperature variations in the transmission fiber [l, 21, whereas 10-Gb/s-basedsystems are less critically affectedby these environmental changes. The effects of polarization-mode dispersion (PMD) also become an important factor for high-speed systems as the PMD values of commercial transmission fibers and other optical components in the transmission path may start producing distortions for medium-haul (300 to 1000km), long-haul m, (1000 to 3000 k ) and ultra-long-haul (>3000 km) high-speed TDM systems. Metropolitan systems (e300km) are fairly immune to intrinsic PMD ifstateof-the-art low-PMD transmission fibers and optical components are used. Schemes and devices for PMD compensation (PMDC) are being developed (see chapters found elsewhere in this book) to reduce the impact of PMD on transmission and detection. Despite important technical challenges, the dispersion accuracy and PMD characteristics of transmission fibers and optical components have steadily improved over time. On the other hand, improvement of the nonlinear characteristics of transmission fibers have only been minimal (nonlinear coefficient has been reduced by -1 dB). These minimal improvements in nonlinearity characteristicsexacerbate the problem of transmitting high-speed signals, because the higher density of bits in high-speed signalsrequires proportionally higher power per channel to preserve the energy per bit. Until recently, it was believed that the distortions induced by fiber nonlinearity in high-speed transmission were too large to allow the energy per bit of high-speed TDM signals to become comparable to the energy per bit of low-speed signals for comparable signal distortion. However, with the uncovering of pseudo-linear transmission [3-91, there has been a renewed interest in high-speed TDM transmission as it opens the possibility of having efficient transmission of information with high-speed TDM signals. Pseudo-lineartransmissionis a regime for transmission of high-speedTDM signals where fast variations of each channel waveform with cumulative dispersion (see Fig. 6.1) allow important averaging of the intrachannel effects of fiber nonlinearity. As a result of the redistribution over many bits of the effects of fiber nonlinearity, pulse distortions are minimized. Additionally, a partial cancellation of some intrachannel effects can be achieved through appropriate dispersion mapping. Nonlinear interactions between WDM channels 234 RedJean Essiambre et al. h 460 8 0 E 40 20 -300 _-_--- -200 /-- ...----100 0 Time @s) -_-------300 Time @s) Fig. 6 1 The fast waveform evolution of rapidly dispersing pulses provide a redistribu. tion of the effects of fiber nonlinearity that is the basis for pseudo-linear transmission. The upper part of the figure shows the location of a 40-ps t h e window in the pulse train made of 5-ps pulses. The average power of the full train is 8 a.u. The lower part shows the waveform evolution in that time window after the pulses have been dispersed by 100 to 120p s / m of cumulative dispersion by step of 5 pdnm. After such dispersion, pulses strongly overlap and any trace of the intensity proiile of individual pulses is lost. Note that each step of 5 pslnm corresponds to the cumulative dispersion of about 300 m of STD unshifted fiber. (interchannel interactions) are generally much weaker than intrachannel nonlinear interactions for pseudo-linear transmission. The evolution of the signal during propagation is generally characterized by important pulse overlap. In this regime, the optimum transmissionis obtained when the net residual dispersion at the end of the system is nearly zero, a characteristiccommon with transmission when fiber nonlinearity is negligible, i.e., when transmission is linear. The origin of the term “pseudo-linear” transmission (pseudo means false, spurious, etc.) can be understood as follows The ultimate performance of a transport system is measured as the energy per bit it can transport at fixed signal distortion from fiber nonlinearity for a given spectral efficiency S (bits/s/Hz). It can be shown that the maximum energy per bit of a highspeed TDM signal transmitted in the pseudo-linear regime is on the order of the energy per bit of systems using lower bit rates (lOGb/s per channel and below) for high but practical spectral efficienciesof intensity-modulatedsignals (S = 0.2 to 0.4 bits/s/Hz). It follows that the transmission in the pseudo-linear regime is highly nonlinear since the energy per bit in this regime does not 236 RenB-Jean Essiambre et al. wavelengths are multiplexed together to form a WDM signal that is transmitted and demultiplexed before being detected by receivers. A number of spans made of transmission fibers are linked with discrete amplifiersthat are used to periodically amplify the signal. In some instances, the transmissionfiber can be Raman pumped to reduce the generation of amplified spontaneous emission (ASE) noise in the system. Typically dispersion compensation is applied at the amplification sites and is included within two stages of a multistage amplifier. This positioning inside an amplifier reduces the impact on ASE noise generation associated with the introduction of a lossy element in the transmission line. Amplifiers linking two spans are referred to as in-line amplifiers, and dispersion compensation at these amplifier sites is referred to as in-line compensation. The amplifier following the transmitter is called a postamplifier (the prefix “post” is relative to the transmitter as the amplification is generally considered as an extension of the transmitter). The dispersion compensation at this amplification site is referred to as precompensation (the prefix “pre” is relative to the transmission line as the choice of dispersion compensation is dictated mainly by the transmission line ahead). Similarly, the amplifier just before the receiver is the preamplifier, and the associated dispersion compensation for this amplifier is the postcompensation. One should note that additional lossy elements might be present in the in-line amplifiers to perform other functions such as gain equalization, channel adddrop, channel crossconnect, performance monitoring, optical regeneration, polarization-mode dispersion (PMD) compensation, etc. Additional amplification stages may be inserted in the amplifiers to accommodate these additional elements. A typical cumulative dispersion map is displayed in Fig. 6.3. It is generally desirable to have identical spans so as to minimize system complexity and cost. For such systems, the three parameters, pre-, in-line, and postcompensation, uniquely define the map. When convenient, in-line compensation is sometimes replaced by the residual dispersion per span and postcompensation by the net residual dispersion at the end of the link. Points of zero cumulative dispersion are labeled ZO. In the absence of nonlinear effects and residual dispersion slope, the pulses at these points are identical to the pulses at the output of the transmitter. For pseudo-linear transmission, the pulses at zo are nearly transform-limited like those in linear transmission. Positions corresponding to launch points to the transmission fibers are labeled zin. 21 . NOISE ACCUMULATION AND OPTICAL SIGNAL-TO-NOISE RATIO REQUIREMENTS In optically amplified systems, the main source of noise is the accumulation of ASE of the optical amplifiers. The noise accumulation can easily be calculated for passive transmission fibers. The ASE noise generated in both polarization statesby an individualamplifier (composed of multiple stages or not) measured 6. Pseudo-Linear Transmission of High-speed TDM Signals (Undo the kcompensation) 237 I, Postcompensation Net Residual Dispersion Zin Dispersion per Span Cumulative Dispersion at Input of First Span Fig. 6.3 Cumulative dispersion maps and nomenclature of dispersion mapping. The locations in the map where the net dispersion is zero are labeled z0 and the fiber input zi,. The precompensation is undone prior to postcompensation to make preand postcompensation independent quantities. in a bandwidth A u at its output is given by [lo]: where nsp is the spontaneous emission factor of the amplifier, G its gain, u the optical frequency, hu the energy of a photon of frequency, and Au the optical bandwidth considered. The value of nspis generallygiven through the amplifier noise figure N F = 2 nsp(l - 1/G) 1/G knowing the amplifier gain. In a chain of amplifiers, the total noise generated is the sum of the noise generated by each amplifier. It is useful to calculatethe optical signal-to-noiseratio (OSNR) of a channel after propagation over a chain of Nap identical amplifiers. The OSNR (in dB) is given by [lo, 111: + where P i n is the launch averaged power per channel (at the input of the transmission fiber) in dBm; N F is the noise figure of the amplifiers and Lsp is the span loss, both in dB. The reference bandwidth, Au, for the OSNR calculation is 0.1 nm (12.5 GHz at 1550nm). Note that Eq. 6.2 shows that, assuming similar signal distortion from fiber nonlinearity at the end of the transmission line, each dB gained in permissible launch power, P i n , is translated into a 1 dB gain in OSNR. Figure 6.4 displays the OSNR evolution with an increase in the number of amplifiers, Namp, in a link. The launch power, P i n , is OdBm, and the noise figure of each amplifier, NF, is 5 dB. The transmission fiber loss varies from 13 to 28 dB in steps of 5 dB from the upper to the lower curve. 238 Renk-Jean Essiambre et al. I " " I " " l ~ " " ' " ' - 13 dg SpanLoss ....... 18 dB span Loss --.-.-. 23 dB Span Loss 28 dB Span Loss ---------_ 0 1 0 20 30 .................. .................. __ -.-.-.-.-._._._._,_. 40 50 Number of Amplifiers Fig. 6.4 OSNR evolution as a function of the number of amplifiersNmp. The launch power per channel is P, = 0 dBm and the noise figure of each amplifier is NF = 5 dB. i At 40 Gbk, a typical OSNR requirement for bit-error-rate is 23 dB. The required OSNR to achieve a given bit-error-rate (BER) depends in general on the nature of noise limiting detection, the type and distortions of the waveform being detected, and the receiver design. One can, however, derive an approximateexpression for the required OSNR, assumingthat the main source of noise results from the beating between the signal and the ASE noise and for large duty cycle intensity-modulated formats. Under these approximations, one can express the required OSNR, OSNRR,as [12], Q2Be l + r OSNRR = Bo ( 1 - a 2 ' where Bo is the optical reference bandwidth of 12.5GHz and Be is the electrical Nter's bandwidth of the receiver. The transmitter extinction ratio, I, is where is the defined as r = Izems/Iones, Io, (Izems) instantaneous current at the bit. sampling instant for a "1" ("0") The parameter Q is given by [13]: Q = aones + azeros where cones (azems) is the standard deviation of Iones (Izems). the optiAt mum decision threshold, the parameter Q is related to the BER through the relation [131, ~ X (-Q2/2) P BER=-er$c Q2/2;7 ' Iones - Izems ($)% where erfc(x) = 2/fiLm exp (-u2)du is the complementary error function. A BER of corresponds to a value of Q = 6. The parameter Q is often quoted in dB by using 10 loge', giving 15.6dB for Q = 6. According to Eq. 6.3, the corresponding required OSNR is 19.4dB for an electrical filter bandwidth of Be = 30 GHz and an infinite extinction ratio (r = 0). 6. Pseudo-Linear Transmission of High-speed TDM Signals 239 Alternatively, one can estimate the OSNR requirement if the receiver sensitivity Prec known using: is OSNRR = 58 + P m - NF, (6.6) where NF is the noise figure of the optically preamplified receiver. For such receivers the best (quantum-noise-limited) receiver sensitivity for an intensity-modulated format and direct detection corresponds to 38 photons per bit (NF = 3 dB), which gives an ideal OSNR requirement of 17.8 dB at 40Gb/s (infinite extinction ratio and ideal electrical filtering). This ideal OSNR requirement is 1.6dB lower than the value given by the approximate expression of Eq. 6.3. For realistic transmitters and receivers with nonideal responses, the required OSNR is generally higher by more than 3 dB from the quantum limit. A direct measurement of receiver sensitivitycan determine the required OSNR for a given transmitter-receiver pair by using Eq. 6.6. One should also mention that 40-Gb/s systems using Reed-Solomon (239,255) forward-error correction (FEC) operate at a higher bit rate of 42.68 Gb/s and have a slightly higher OSNR requirement than systems operating without FEC (see chapters found elsewhere in this book). One can estimate the impact on the noise figure of an in-line amplifier from the presence of an additional amplikation stage required to compensate for the loss of a dispersion-compensatingfiber (DCF). Assuming the same noise of figure NF for all amplification stages, the noise figure NFDC an amplifier that includes a dispersion compensation module (DCM) of loss r] is approximately given by where Pk is the average power at the input of the amplifier and PDCFis the averagepower at the input of the DCM (see Fig. 6.2). Let’s consider an example where the presence of a DCM significantly impacts the amplifier’snoise figure. For a 12-dBDCM loss ( r ] = 12dB), a PDCF of -2 dBm to avoid nonlinearities in the DCF, and a Pk of -16 dBm (launch power Pi,= 4 dBm and 20 dB span loss), the excess noise figure contribution is 2.1 dB. The impact of the presence of dispersion compensation can be minimized by using dispersion compensation devices that have higher-power thresholds for nonlinearities or lower loss such as the higher-order mode (HOM) DCF described in Section 3.1.1. 2.2 INTENSITTY-MODULATED FORMATS AND SPECTRAL EFFICIENCY Intensity-modulated direct-detection (IMDD) formats are the most commonly used transmission formats in high-capacity fiber-opticcommunication systems (for alternative formats see chapters found elsewhere in this book). 240 RedJean Essiambre et al. These formats generally lead to the simplest designs for transmitters and receivers. Such simplicityis desirablesince the fabricationof commercial-grade high-speed receivers and transmitters is a challenge by itself. The IMDD formats are defined by the duty cycle and the pulse shape. The duty cycle d is defined as: d = -T ,P (6.8) T B where Tp is the pulse full-width at half maximum (FWHM) and TB the bit period. The IMDD formats consideredhere are the non-return-to-zero(NRZ) format and the return-to-zero(RZ) format with Gaussianpulse shapes. Besides simplicityin transmitter and receiver designs, there are two important properties determining the choice of optimum format for WDM transmission. The first is related to the spectral efficiency S defined as: B S=-, Vch (6.9) where B is the bit rate and Vch is the channel spacing. The spectral efficiency correspondsto the spectraldensity of information and is expressedin bits/s/Hz. The spectra of various modulation formats differ in their bandwidths and shapes. Moreover, the intersymbol interference created by narrow optical or electrical filtering varies widely according to the modulation format. As a result, the ability to achieve high spectral efficiency by closely spacing WDM channels generally depends on the modulation format. Figure 6.5 shows an exampleof the differencesbetween formats for demultiplexinga 40-Gbls signal from a WDM field with 100-GHzspacing. The fist two columns are the speo tra of an isolated channel and the WDM field, respectively. Third and fourth columns are the eye diagrams after demultiplexingfor an isolated channel and from an arbitrary channel from the WDM field, respectively. For both cases, optical demultiplexingis performed using a 80-GHz FWHM Gaussian optical filter and a 28-GHz, 3-dB electrical Bessel filter included in the receiver. The upper row is the NRZ format, whereas from second to fourth row, the format is RZ (Gaussianpulses) with duty cycles of 50,33, and 20%, respectively. Singlechannel eye diagrams clearly show minimal distortions, whereas eye diagrams of the demultiplexed WDM channel are distorted from coherent cross-talk. Coherent cross-talk originates from the overlap of the spectra of neighboring channels with the demultiplexed channel. The coherent cross-talk increases as the duty cycle decreases because the broader spectra of short pulses lead to greater spectral overlap. As a result, one should expect low-duty-cycleformats to be limited to lower spectral efficienciesthan formats with larger duty cycles. Nonetheless, it is noticeable that for duty cycles as low as 20%, the eye diagrams still show significant opening for the spectral efficiency of 0.4 bits/s/Hz of Fig. 6.5. 6. Pseudo-Linear Transmission of High-speed TDM Signals Isolated Channel 241 speceum WDM spectrum DemultiplexedIsolated DemultiplexedWDM Channel Eye Diagram ChannelEye D a r m iga 2 1.5 1 0.5 0 2 -1 v z !.I22 1.5 go I 0.5 0 3 -30 40 do -100 -50 0 50 100 -100 do 0 50 loo Frequency (GHz) Frequency (GHz) 0 5 10 15 20 250 5 10 15 20 25 Time (ps) Time (ps) Fig. 6.5 Coherent cross-talk and eye diagrams in dense WDM systems. From fmt to fourth column: single-channel spectra, WDM spectra, single-channeleye diagrams, and eye diagrams of center channel. For both single channel and WDM cases, eye diagrams are obtained using an optical demultiplexer consisting of a Gaussian filter of bandwidth 80 GHz and an electrical Bessel filter of 28 GHz. The relative time delay between channels is a random number of bits, including fractional bit delays. The bit rate is 40 Gbls and channel spacing 100 GHz (0.4 bitslslHz spectral efficiency). The displayedoptical spectra were generated by passing the signal through a 0.02-nm optical filter. Degradation of the eye diagrams as the duty cycle decreases comes from spectral overlap between channels causing coherent cross-talk. 23 DISPERSION . The equation of evolution of a field A(z,t) representing a modulated signal propagating in a lossy/amplifyinglinear dispersive medium is given by, aZ aA i + -/32(z)- a2A + a(z) -A 2 at2 2 = 0, (6.10) where BZ(Z) is the group velocity dispersion (GVD) representing the dispersion of group velocity with angular frequency. This is obtained from the propagation constant B(w) using /32 = [d2B/d~2]0=00, wo is the anguwhere lar frequency. The fiber loss/gainis accountedfor through a(z),which describes the signal power evolution, for passive transmission fiber a(z), equal to the fiber loss coefficienta over the length of the fiber. Solving Eq. 6.10 for a pulse 0 labeled n (n = 1,2, . ..) with the initial condition of an unchirped Gaussian 242 RenLTean Essiambre et al. pulse at location zo (where the cumulative dispersion is zero), gives, (6.11) where the characteristicpulse width T is given by: , the chirp C,(z) by: CGVD(Z) C,(Z) = , T2 o the cumulative GVD, CGVD(Z), point zo to z by: from (6.13) (6.14) and finally the complex amplitude b,(z) is given by: (6.15) where the cumulativelosdgain factor Z(z) is given by: ZZ = () Jc,l .(z’)h’. (6.16) The pulse position t,(z), frequency w,(z), and phase &(z) are not affected by linear dispersive propagation and keep their initial values to, WO, and 00, respectively. The parameter TOis the characteristic pulse width and is related to the pulse full width at half maximum Tp at the transform-limited points zo by T 3 TP(zo) = 2 m TO. is the initial pulse amplitude at z = ZO. p A0 The pulse bandwidth does not change with dispersive propagation. Note that the pulse characteristicbandwidth is given by (2nTo)-l, whereas the pulse full bandwidth at half maximum Av, is given by: (6.17) and the root-mean-square (RMS) bandwidth is given by ( 2 n a To)-’. One can associate to the pulse temporal broadening induced by dispersion a characteristic length referred to as dispersion length LD defined as: LD= 1 -. 0 ‘F2 1821 (6.18) 6. Pseudo-Linear Transmission of High-speed TDM Signals 243 This length is indicative of when dispersive effects start to impact a pulse and corresponds to the broadening of a Gaussian pulse by a factor of &. It is often useful to express the evolution of the pulse width Tp(z)in more commonly used quantities as (6.19) where c is the speed of light in vacuum and cumulative dispersion C&) from to z, 20 (6.20) and dispersion D, 2RC D(z) = -- B ~ ( z ) . h2 (6.21) Figure 6.6 shows the pulse width evolution with cumulative dispersion of initially transform-limitedGaussian pulses. Significantpulse overlap in a pulse train occurs when the pulse width approaches the bit period TB. At 40 Gb/s, TB = 25 ps, and pulses of width Tp = 2.5,5,8.3, and 1 2 . 5 ~broaden to 25 ps s after 18, 35, 56, and 77ps/nm of cumulative dispersion (see Fig. 6.6), respectively. The shorter the pulse the faster it broadens and the faster neighboring pulses in a train overlap. This range of cumulative dispersion corresponds, for instance, to the cumulative dispersion of 1 to 4.5 km of standard (STD) unshifted fiber [D = 17ps/(kmnm)]. Such small tolerance on cumulative dispersion makes it impossible to prevent pulse overlap during propagation within one span (50-100 km) made of either nonzero dispersion-shifted fibers [NZDSFs, 4ps/(kmnm) < I1 < 8ps/(kmnm)] [14] or STD unshifted fibers. D 0 20 40 60 80 100 Cumulative Dispersion (pdnm) Fig. 6.6 Pulse broadening with cumulative dispersion for Gaussian pulses. Short pulses reach the boundary of the bit slot (25 ps at 40 Gb/s) faster than long pulses. 244 RenC-Jean Essiambre et a. l However, the use of dispersion-shifted fibers [DSFs, 11 < Zps/(kmnm)] 0 makes it possible to prevent pulse broadening for long pulses (Tp > lops). Alternatively, dispersion compensation on a short scale (a few kilometers) using segments of NZDSF or STD unshifted fibers can prevent the a m mulation of dispersion. The impact of nonlinearity on high-speed TDM transmission using these low cumulative dispersion maps can be quite different from the impact of nonlinearity in the pseudo-linear regime and wl il not be covered in this chapter (even though a glimpse of the effects of nonlinearity on dispersion-shifted fibers (DSFs) for high-speed TDM signals can be seen in the upper part of Fig. 6.24). Accumulating dispersion not only leads to pulse broadening, but also to pulse chirping. This can be understood by considering that dispersion leads to a spread in delays between the different frequency components of a pulse. As a result, in a medium having normal dispersion(negativeD, positive &), the leading edge of a dispersed pulse becomes composed of the lowest-frequency components (“red”) of the pulse, whereas the trailing edge contains the highestfrequency components (“blue”). The opposite situation occurs for a medium having anomalous dispersion (positive D,negative 82). Because the phase of these frequency components evolve at different speeds, a chirp is created across the pulse. The waveform evolution of a 40-Gbh pulse train made of 5-ps Gaussian pulses is shown in Fig. 6.7. As dispersion accumulates, the initially transformlimited pulses (Fig. 6.7a) broaden (Fig. 6.7b) until significant pulse overlap 001001001 10110101 101 60 60 (d) 30pdnm 40 40 20 n 2 0 n 40 20 0 0 100 200 300 400 500 0 100 200 300 400 500 Time (ps) Time (ps) Fig. 6.7 Waveform evolution of a 5-ps Gaussian-pulse train at 40 Gb/s with cumulative dispersion. The transform-limited pulses in (a) gradually broaden in (b) u t l there ni is significant overlap between nearest neighbors in (c)-(f). At large values of cumulative dispersion (100 ps/nm and above) a large number of pulses overlap in time and individual pulses are no longer identifiable. 6. Pseudo-Linear Transmission of High-speed TDM Signals 245 10 0 5 10 15 20 2t 5 10 15 20 25 Time (ps) Time (ps) Fig. 6.8 Electrical eye diagrams for the waveforms of Fig. 6.7. A 28-GHz Bessel electrical filter is used in the receiver. occurs between neighboring pulses (Figs. 6.7c-e). The oscillations between pulses seen in Figs. 6.7c-e are the result of the beating between the dispersed pulses, “blue” and “red” frequency components. For large cumulative dispersion (Fig. 6.7f), the large number of pulses interfering at a given location in time and the wide distribution of phases that originates from pulse chirping creates the appearance of a “random field” (but with the limited bandwidth content of individual pulse spectra). Because the waveform is determined by interference of fields (the chirped pulses) as opposed to addition of powers (if the pulses were chirp-free), the waveform evolves very rapidly in a dispersive medium as any small change in phase due to dispersion affects greatly the interference between the dispersed pulses. This rapid waveform evolution leads to a generally beneficial redistribution of nonlinear phase distortions among pulses and is one of the bases of the pseudo-linear regime. The eye diagrams for the dispersing pulses of Fig. 6.7 are presented in Fig. 6.8. There is no optical demultiplexer and a 28-GHz electrical Bessel filter is included in the receiver. One notices that the fast oscillations observed between pulses in Figs. 6.7c-e that result from the beating of the high-frequency components of the pulses are reduced as a result of electrical filtering. 231 .. Eye Closure Penalty To compare the transmission performance of various modulation formats it is important to be able to isolate the deterministic effects of distortion (dispersion, nonlinearity, and coherent cross-talk)in the eye diagrams from stochastic effects(such as the effectof amplifier noise). To be able to do a meaningful comparison between various transmission formats (without having to do extensive simulation to achieve statistical averaging), we excluded the effect of optical amplifier noise in the simulations presented in this chapter. Under such conditions, the shape of an eye diagram becomes a reliable indicator of the effect 246 RenC-Jean Essiambre et al. 40 35 + m (u 30 - 25 20 15 10 5 n 0 5 10 15 20 25 30 35 Time (ps) Fig. 6.9 Example of the determination of the eye diagram closure using a box of width 20% of the bit duration. of transmission. One measure of distortions of an eye diagram is given by the eye closure penalty. It is calculated by first determining the height PR of the highest rectangle that can be fitted inside the eye opening. The rectangle width is 20% of the bit period TB.This width is chosen to include the effect of clock jitter on the decision sampling instant. The eye closure penalty (expressed in dB) is then given by: Ceye= -1Olog ~ (2ppXe)’ (6.22) where Pave the signal average power. is Note that twice the average power is equivalent to the height of the rectangle for an unfiltered NRZ sequence having the same number of “zeros” and “ones.” A negative eye closure Ceyerepresents an eye more opened than the reference (the unfiltered NRZ signal), whereas a positive Ceye represents an eye more closed than the reference. Figure 6.9 shows a typical example of the positioning of the box for determining PR used in the eye closure calculation of Eq. 6.22. Note that the relation between eye closure penalty and receiver sensitivity penalty or BER penalty is not straightforward, as it depends on the nature of the noise and the type of waveform distortion. 2.3.2 Dispersion Margin and Modulation Formats High-speed TDM systems based on intensity-modulated formats use closelyspaced short pulses that rapidly broaden and overlap (as seen in Fig. 6.7), causing intersymbol interference (ISI) [15]. This leads to much tighter requirements on dispersion margins for high-speed TDM systems than for lowerspeed systems. Dispersion margins are dependent on the modulation format. Figure 6.10 shows the eye closure as a function of the cumulative dispersion applied to transform-limited RZ and NRZ formats. One first notices that RZ formats have negative eye closure at the transform-limited point (0 pshm) that represents the fact that the eye opening is larger for RZ than unfiltered NRZ, 6. Pseudo-Linear Transmission of High-speed TDM Signals I ' ' ' . 247 ' I ' ' ' ' I ' ." ' I' y ' ' ' ' I..' ' ' ' I 'y ' 1 ....... 12.5 ps I-NRZ 1 120 0 20 40 60 80 Dispersion (ps/nm) 100 Fig. 6.10 Eye closure penalties at 40 Gb/s for the RZ format w t duty cycles of 0.2, ih 0.33, and 0.5 and the NRZ format. Low duty cycles have reduced dispersion margins. The eyes diagram are obtained after electrical filtering with a 28-GHz Bessel filter. No optical filtering is applied. the reference. However, as cumulative dispersion CD increases, the larger eye opening for RZ decreases faster than NRZ. This is because the broader spectrum of RZ as compared to NRZ makes the shorter pulses broaden faster (see Fig. 6.6). The lower the duty cycle (shorter pulses) the faster the eye closes with cumulative dispersion. Consequently, in systems with negligible fiber nonlinearity, the lower the duty cycle the more sensitive the transmission becomes to offsets of dispersion. 2.4 FIBER NONLINEARITY 2.4.1 Introduction The evolution of the optical field A(z, t ) experiencing Ken- nonlinearity in optical fibers has been derived in Ref [16]. Assuming that the slowly varying envelope approximation (SVEA) holds, the equation of evolution for the field A(z, t ) can be written as: The coefficient83 accounts for the change of the GVD ( 8 2 ) with angular frequency Cg3 [dp2/dw],=,) and is referred to as the third-order dispersion (TOD) parameter. When fiber losdgain and TOD are neglected and 82(z)is a constant independent of z, Eq. 6.23 is known as the nonlinear Schrodinger equation (NSE), which has soliton solutions when dispersion is anomalous (negative /32 or positive 0) Ref [17] and Chapter 11 of Ref [ 8 ) When (see 1]. &(z) is a periodic function with suf€iciently low path-averaged value, Eq. 6.23 can have an approximate solution called dispersion-compensated (DC) or dispersion-managed soliton (see chapters found elsewherein this book). When 248 RenC-Jean Essiambre et al. any additional terms are included, Eq. 6.23 is referred to as a generalized nonlinear Schrodinger equation (GNSE). The coefficient 83 of Eq. 6.23 is related to the more straightforwardly measurable quantity dispersion slope S through the following relation, S(2) = - = -p dh h3 dD 4nc (6.24) The coefficient y in Eq. 6.23 represents the effect of the Kerr nonlinearity and is defined as: y = - n 2 wo (6.25) cAeff ’ where 122 is the nonlinear refractive index coefficient and A& is the effective mode area. Typical fiber parameter values for some common fibers are given in Table 6.1. One can associate different length scales that are specific to nonlinear transmission. A first length scale is related to power only and is defined as: 1 LNL=-, YPP (6.26) where LNL is the fiber length required to produce nonlinear phase rotation of one radian at a power Pp. When Pp is interpreted as a pulse peak power, LNLis related to the effect of self-phase modulation (SPM, see Section 2.4.3). A second length scale is related to the power evolution in fibers; for a passive fiber this is known as the effective length L,ff and is given by: Leff = 1 - exp (-a0 L) Y (6.27) a 0 where L is the length of the fiber segment considered. The effective length L,ff gives the length over which the power decreases by a factor of e in passive fibers. This length is related to most nonlinear effects but is perhaps most critical for SPM and cross-phase modulation (XPM) (see Chapter 8 of Ref. [18] and chapters found elsewhere in this book). A third length scale is the walk-off length and is defined as: L w = - TD DAV ’ (6.28) where Av is the spacing between the two spectral components of interest and TO is a time delay. In WDM systems experiencing XPM (see Fig. 6.1 l), the relevant delay corresponds to 2 Tp (see Chapter 8 of Ref [18]), the delay necessary for two pulses from channels separated by Av to fully walk through each other. The time delay TD may have different values depending of the nonlinear interaction of interest. Note that the dispersion length LD defined in Table 6.1 Some Nominal Values of Fiber Parameters at 1550 nm of Different Commercial Fiber Brands. The Ratio Dispersion to Slope (RDS) Is Defined as S/D D Fiber TrueWaveTM RS LEAFW TeralightTM Ultra STD unshifted fiber Submarine DeeplightTM TeralightTM Metro DCF WB-DCF HS-DCF Manufacturer Lucent Corning Alcatel Lucent, Coming, Furukawa Lucent Pire11i Alcatel Lucent Lucent Lucent ps/(km nm) 4.5 4.2 8 16.9 -3.1 -2.2 8 S ps/(ltm nm2) RDS nm-’ (110 Aefi pm2 PMD ps/& dBhm 0.22 0.22 t0.22 0.23 0.045 0.09 0.052 0.01 0.021 0.0065 0.0033 55 12 63 87 50 70 63 t o .1 tO.l 0.055 0.05 (0.04 t o .1 -0.016 >-0.055 0.0073 0.0022 0.0035 0.0067 t0.12 0.058 -0.22 -0.33 -0.67 0.215 t0.23 t0.25 t o .1 to. 1 t0.08 t0.25 t0.25 t0.25 -100 -95 0.5 0.5 0.68 -100 20 19 15 A W N 250 RenkJean Essiambre et al. Eq. 6.18 can be interpreted as a walk-off length within a pulse where the delay To = TO,the characteristic pulse width, and the spectral separation Au = -sgnm(B$t2/(ncTo). The various length scales defined in Eq. 6.18 and Eqs. 6.26-6.28 characterize different length scalesfor the effects of nonlinearity. In general, these length scales depend on channel bit rate, channel spacing, modulation format, fiber types, input powers, power evolution, dispersion mapping, amplifier spacings, system length, etc. It is difficult to determine a general rule that would determine which scale is the most important for a given set of system parameters. Nonetheless, it is still instructive to define these length scales and, whenever possible, we will point out the scale relevant to nonlinear interactions specific to pseudo-linear transmission. Equation 6.23 describes the evolution of the full field (which may include many WDM channels) with distance. In general, nonlinear interactions for the WDM field can be decomposed into more basic nonlinear interactions (see Fig. 6.1 1). These interactions are single-channel self-phase modulation, multiple-channelcross-phase modulation (XPM), and multiple-channel fourwave mixing (FWM). Cross-phase modulation and four-wave mixing are interchannel interactions that are the strongest for moderate- (-10 Gb/s) and low-speed (c10Gb/s) signals. Cross-phasemodulation and four-wave mixing SingleChannel Modulation Self-Phase Modulation Instability (MI) Self-Phase Nonlinear Modulation (SPM) Intersymbol (Soli:ons, etc.) Intefierence Pulse Distortion - \ ---Basic Nonlinear Interactions Single Channel v Coherent Cross-talk Multiple Channels (WDM) Four-Wave Cross-Phase Mixing {FWM) Modulation JXPM) v Timing Jitter and Pulse Distortion Intrachannel Intrachannel Cross-Phase Four-Wave Modulation Mixing v Timing Jitter Amplitude Jirrer U m a I I F W M ) /\ 10 Gbls and Above ... 10 Gb/s and Below Fig. 6.11 Distribution of inter- and intrachannel nonlinear impairment in a WDM system for different bit rates per channel. For high-speed TDM systems, the dominant nonlinear interactions are self-phase modulation (SPM), intrachannel cross-phase modulation (IXPM), and intrachannel four-wave mixing (IFWM). 6. Pseudo-Linear Transmission of High-speed TDM Signals 251 have been extensively studied (see Ref [ 161and references therein and chapters found elsewherein this book). For high-speed systems operating in the pseudolinear regime of transmission, XPM and FWM are usually much weaker than the nonlinear interactions within each channel (intrachannel interactions). This can be evidenced by numerical simulations or system experiments where single-channel transmission is compared to WDM transmission. For pseudo-linear transmission, one observes that when adding WDM channels, in the worst case, only a moderate increase in waveform distortions compared to single-channel transmission is observed. Moreover, assuming small spectral overlap between channels, adding WDM channels has rather small impact (if any) on the choice of the optimum schemes of transmission, such as dispersion mapping for instance. In a way similar to the decomposition of nonlinear interactions among channels in WDM systems, it is possible, when operating in the pseudo-linear regime, to separate the various nonlinear interactions among bits of the same channel. These intrachannel interactions are single-pulse self-phase modulation or simply self-phase modulation (SPM), cross-phase modulation among pulses or intrachannel cross-phase modulation (IXPM), and four-wave mixing among pulses or intrachannel four-wave mixing (IFWM). The field of a single channel can be represented as a sum of the fields of individual pulses, A = A , where A , is the field representing the mth of M pulses centered at tm.By replacing this sum in Eq. 6.23 we obtain, zt=, M =i y m= I A,A;A~. m,n,p=l (6.29) The nonlinear terms on the right-hand side (RHS) of Eq. 6.29 can be identified as follows: when m = n = p we have SPM, when m = n # p or m # n = p it is IXPM, and when m # n # p or m = p # n it is IFWM. This separation between nonlinear interactions is meaningful only if all Am’s fields in Eq. 6.29 can be separated in time, Le., that the pulse width (at the transform-limited point) Tp is smaller than the bit period TB (d < 1). This condition is similar to the condition of separability between nonlinear interactions in the analysis of WDM transmission where the channels are assumed to be separated in frequency by more than their bandwidth. In the pseudolinear regime of transmission where pulses disperse rapidly and extensively (Lo< LNL), the location of the nonlinear interaction is given approximately < by tm + tp - tn. This relation is analogous to the phase-matching condition used to determine the frequency location of a wave generated by four-wave mixing (FWM). By considering only three interacting pulses, A , , A2, and A3 and assuming pseudo-linear transmission, one can obtain from Eq. 6.29 separate equations 252 RedJean Essiambre et aL for the evolution of each pulse, SPM IXPM IFWM aA2 i a 2 ~ 2 1 a3~ -+ -82(z)- -83az 2 at2 6 at3 3~~ aZ 2 + a(z) -A2 2 = iylAzI2Az + 2iy(lA1I2 + IA312)A2 +2iyAlAtA3, (6.31) i 1 ~ 3 3 a(z) + -82(z)- aat2* -~-83-aat3 ~ + -A3 2 6 2 = iyIA3I2A3 +2iy(lAiI2 + IA212243 +iyA;Az, (6.32) where Eqs. (6.30) through (6.32) give the evolution of the field envelope for pulses 1 through 3, respectively. For the equation of evolution of each pulse, only nonlinear terms leading to a perturbation at this pulse location have been retained. We notice that the evolution of each field in Eqs. (6.30H6.32) is affected by three types of nonlinear terms (RHS). The first term on the RHS leads to the modulation of a pulse phase by itself (i.e., SPM). The second and third nonlinear terms represent the modulation of the phase of a given pulse by the neighboring pulses (i.e., IXPM). The last terms on the RHS of Eqs. (6.30H6.32) lead to nonlinear mixing between pulses (ie., IFWM). 2.4.2 Energy Exchange It is well known that the nonlinear Schrodinger equation (NSE) has an infinite number of so-called conservation laws [17]. The fist two have direct physical interpretations, energy and momentum conservation during propagation. In the case of the generalized nonlinear Schrodinger equation (GNSE), - aA, az i + 2/32(z)-@A, at2 -- 1 a3A, 837 6 at az + -A, 2 =N , (6.33) describing the evolution of pulse A, under an arbitrary nonlinear effect N , the ” evolution of the pulse energy E, = f ,IA,I2 dt is given by: E, = Ej,, exp[-Z(z)] + 2 exp[-l(z)] where Eh is the pulse energy at the input of the transmission fiber and %{A,(z, t)* N(z, t ) ] stands for the real part of A,(z, t)* N(z, t). The first term on the RHS of Eq. 6.34 represents simply the variation of the pulse energy due to fiber loss or gain common to all pulses. The second term on the RHS represents the variations in the relative pulse energies due to fiber nonlinearity. Clearly when A,(z, t)*N(z, t ) is an imaginary number the second term 6. Pseudo-Linear Transmission of High-speed TDM Signals 253 on the RHS of Eq. 6.34 vanishes and all pulses undergo only the common fiber losdgain. It results in the absence of energy exchange among pulses during propagation. All the terms describing the nonlinear interactions among M pulses are included when writingN = i y CEn,=, , A: Ap (RHS of Eq. 6.29). A For pseudo-lineartransmission, the nonlinear terms that overlap with the pulse located at tq are the ones having tq = t, tp - tn. It can be easily verified that t>* considering only the SPM and IXPM terms makes %{A4(z, N ( z , t)}vanish. As a result, no energy exchange among pulses results from these two nonlinear interactions. On the other hand, the IFWM terms generally lead to a nonvanishing %{A,(z, t)* N(z, t)},resulting in pulse energy variations and exchange of energy between bit slots. Using this property, it was possible to identify that the formation of shadow pulses in pseudo-linear transmission originated from IFWM [6]. + 2.4.3 Self-Phase Modulation The effect of nonlinearity on the propagation of an isolated pulse is referred to as self-phase modulation (SPM) and can take many forms. One form, thoroughly studied, is the optical fiber soliton (see Ref. [16], Chapter 12 in Ref. [18], Ref [19], and chapters found elsewhere in this book). The soliton in fibers is characterized by a strict balance between the effect of fiber nonlinearity and the dispersive effects for isolated pulse propagation. It requires which happens to be a medium having an anomalous dispersion (positive D), the sign of dispersion of fused silica at wavelengths longer than -1300nm. Thus, the most straightforwardlymanufacturable fibers have positive D in the third communication window (-1 550 nm) where fiber loss is minimum. The requirement of an exact balance between dispersion and nonlinearity for solitons is not always necessary to achieve acceptable isolated pulse transmission. In many instances, an average compensation of dispersion by nonlinearity leads to adequate isolated pulse transmission. An example of intricate pulse evolution that still produces low distortion is given by chirpedRZ transmission [20, 211. Another one is related to transmission of NRZ signals where some compensation of dispersion by nonlinearity can occur despite the fact that the NRZ pulse shape is quite dif€erent from a soliton (see optimum dispersion for NRZ in Fig. 6.24 for instance). For high-speed TDM systems, two forms of solitons are of particular interest. In its first form, SPM can compensatecontinuously for the local dispersion, and the corresponding pulses are referred to as local solitons, adiabatic solitons, or simply solitons. In a second form, SPM compensates for the residual dispersion in a periodically dispersion-compensated system according to a precise prescription of the dispersion map, and the corresponding pulses are referred to as dispersion-compensated(DC) or dispersion-managed solitons. It is difficult to use local solitons in high-speed systems with constantdispersion fibers (fibers with constant dispersion along its length). This is due 254 Red-Jean Essiambre et a . l to the large range of signal power experiencedduring propagation in the transmission fiber. This range is approximately 10to 25 dB for passive transmission fibers. Self-phase modulation (SPM), the nonlinear effect that compensates the effects of dispersion also experiences the same range as the power evolution. To preserve the local soliton, the width of each local soliton would also experience the same range as the power evolution, and consequently, the spectral bandwidth of the local soliton would vary by 10 to 25dB [22-251. Such large variations in spectral bandwidth prevent efficient use of WDM techniques. It is possible, however, to preserve the balance of nonlinearity and dispersion required by local solitons without the spectral broadening by tailoring the fiber dispersion along its length. For passive fibers, the dispersion should decrease exponentially along the length, and such fibers are referred to as dispersion-decreasing fibers (DDFs) [26-281. Even though interesting and manufacturable from a single fiber draw [29-331, DDFs impose some important constraints on system designs by forcing fixed values of powers (the local soliton power) for the signal evolution, as well as unidirectionality for individual fibers. Moreover, the wide dispersion range necessary to accommodate the span loss (>20 dB) for large amplifier spacings (100 km) requires large values of dispersion at the input end of the fiber. Such a high dispersion value increases significantly the path-averaged dispersion and can lead to large timing jitter [34-371. The other possibility for using the soliton effect in high-speed TDM systems is to use DC solitons. The design of DC soliton links allows some pulse broadening but limited to the bit slot duration to prevent pulse overlap during transmission. As shown in Fig. 6.7, at 40 Gb/s the pulse overlap becomes significant when the cumulative dispersion reaches approximately f10ps/nm. For large amplifier spacings (-100 km), this requirement restricts the disperD sion of the transmission fiber to a range of I1 c 2 ps/(km nm), assuming a precompensation of half the span cumulative dispersion. Dispersion-shifted fibers (DSFs) meet the requirement of low-dispersion values (over a limited bandwidth of -60nm however) and can support DC solitons [38-44]. Such low, local dispersion, however, makes WDM difficult due to four-wave mixing (FWM) [45] and limits the use of DSFs in high-capacity transport. An alternative to DSFs is to use dispersion compensation on a short length scale (on the order of a few kilometers) [46-48]. Such short-scale dispersion compensation allows one to use fiber segments with relatively high dispersion and still have a low value of average dispersion and a low excursion of cumulative dispersion. As for the fabrication of DDFs, one can fabricate continuously dispersion compensating fibers (CDCFs) from a single fiber draw [49]. However, the scale of dispersion compensation cannot be arbitrarily small because for too frequent dispersion compensation, the fiber starts to behave like a constant-dispersion fiber with a low average value of dispersion. As for DSF, such CDCF would suffer from FWM in WDM transmission [46]. As a result, an optimum scale of dispersion compensation may exist that will correspond 6. Pseudo-Linear Transmissionof High-speed TDM Signals 255 to a tradeoff between the effects of pulse broadening and FWM. One should note that the design of such a CDCF is a priori bit rate and modulation format specific. Since the use of DSFs is likely to be limited by FWM, it suggests that the effectiveuse of DC solitons in high-speed systems may require deployment of special transmission fibers. In pseudo-linear transmission, the propagation of short pulses (Tp 10ps) over dispersive fibers [ D > 2 ps/(km nm)] is dominated by dispersion I1 (LD < LNL),and each pulse can broaden by up to several orders of magni< tude. As a result, the length scale over which the soliton dynamics start to play a significant role can increase considerably, reducing the importance of SPM in pseudo-linear transmission [50, 511. The effect of SPM is best seen by considering the bandwidth evolution of a short pulse transmitted in the pseudo-linear regime as depicted in Fig. 6.12 (from Ref. [SO]). The transformlimited pulse width is 5 ps, and each span is made of 100km of large effective area STD unshifted fiber fully dispersion compensated at each amplifier. The parameters of the transmission fiber differ slightly from Table 6.1 and are = -2Ops2/km [D= 15.75ps/(km nm)], (11= 0.25 dB/km, A,ff = 108 pm2 and n2 = 3.2 x cm2/W. Three dispersion maps are represented. Curve (a) in Fig. 6.12 shows the bandwidth evolution when no precompensation is Pulse RMS bandwidth evolution as a function of distance for transmission of-5-ps Gaussian pulse over multiple spans. Full dispersion compensation at each span is applied. The amplifier spacing is 100 km and the transmission fiber is a large effective area fiber with the followingparameters, #32 = -20 ps2/km, ct = 0.25 dB/km, Aen = 108 pm2 and n2 = 3.2 x cm2/W. In case (a) there is no precompensation, (b) a precompensation of 2000 ps2/km (D = - 1575ps/nm), equivalent to one span, and (c) a precompensation of 100 ps2 [D= -79 ps/(km nm)]. The input peak powers Pp are 35,20, and 50mW for (a), (b), and (c), respectively. From Ref. [50]. 256 Renk-Jean Essiambre et al. applied. Curve (b) is when a precompensation compensating for a full span is used (CpE = -1575ps/nm), whereas curve (c) is when 5% of the span dispersion is precompensated (CpE = -79 ps/nm). When no precompensation is present [curve (a)], a sharp bandwidth decrease occurs where the pulse peaks up (at the beginning of each span), with negligible bandwidth evolution for the remainder of each span. When the precompensation is set to compensate a full span dispersion [curve (b)], the pulse peaks up at the very end of the span where the power is low and the effect of SPM is neghgible. For a moderate precompensation of 5% of the span [curve (c)], the pulse bandwidth experiences a rapid breathing just around the point of zero cumulative dispersion (points zo in Fig. 6.3) where the pulse is nearly transform-limited.This evolution consists of a rapid spectral broadeningjust before zo followed by a rapid spectral narrowing (of nearly equal magnitude as the spectralbroadening)just after zo. The net result is the generation of a small residual spectralbroadening every time the pulse goes through a point zo in the transmission fiber. For a transmission fiber of opposite sign of dispersion, a net spectral narrowing is observed instead of a net spectral broadening. Figure 6.12 also showsthat the magnitude of the spectralnarrowing after 13 spans is about 10%for the case of no precompensation [curve(a)], 2% when the precompensation is equal to 5% of the span [curve(c)], and virtually no change in pulse bandwidth when the precompensation is equal to the fidl span [curve (b)]. Because the chirp caused by SPM is a smooth function, the broadening or narrowing of the spectra observed in Fig. 6.12 is virtually equivalent to change in duty cycle of the pulses. The effect of SPM is identical for each pulse of a data sequence. The net result of SPM is to produce a train of identical pulses with a slightly different duty cycle than the original train. For the example of curve (c) of Fig. 6.12 that has a 2% spectral broadening after 13 spans, no significant degradation in the eye diagram and reception results from SPM in single-channel transmission. Even for larger spectral broadening (like a doubling of the spectrum bandwidth), no significant degradation is expected from SPM other than the difference in eye openings between various duty cycles (see Section 2.3.1). In the case of a 5% precompensation [curve (c) of Fig. 6.121, the spectralbandwidthdoubles only after approximately40,000 km. The most likely detrimental effect of SPM in the pseudo-linear transmission regime is to broaden the spectrum so much as to create coherent cross-talk from overlapping spectra in dense WDM. It is clear that for pseudo-linear transmission of high-speed TDM signals, the effect of nonlinearity on isolated pulse transmission (the effect of SPM) is greatly reduced relative to lower bit rate systems. However, in pseudo-linear transmission, a large number of pulses belonging to the same channel simultaneously overlap and interact through fiber nonlinearity. As mentioned in Section 1, these intrachannel interactions among neighboring pulses are the main source of waveform distortions in pseudo-linear transmission. 6. Pseudo-Linear mansmission of High-speed TDM Signals 257 2.4.4 Intrachannel Cross-Phase Modulation A first source of waveform distortion in pseudo-linear transmission comes from the modulation of a pulse phase by nonlinear interactions with its neighboring pulses from the same channel. This intrachannel cross-phase modulation (IXPM) generates a frequency shift on each pulse that depends on the pulse pattern and, through dispersion, it leads to generation of timing jitter. This is well illustrated in Fig. 6.13, which displays part of a pulse train before and after transmission along with the eye diagram after transmission showing significant timing jitter. The simplest way to demonstrate the effect of IXPM is by considering propagation of a single pair of pulses. The equations of propagation describing the effect of SPM and IXPM on each pulse of a pulse pair can be obtained by using Eqs. (6.30) and (6.31) and keeping only the first two terms on the RHS and neglecting TOD. We can then rewrite the simplified equations as, (6.35) 0 16 14 12 10 50 100 150 200 Time @s) v s 0 8 $ 6 4 2 0 0 10 20 30 40 50 Time (ps) Fig. 6.13 Waveform and eye diagram after transmission of a PRBS sequence of 128 bits consisting of 5-ps Gaussian pulses. Transmissionis over 80 km ofTrueWaveTM fiber (D= 4 ps/nm). The bit rate is 40 Gb/s, launch power 18dBm, fiber loss 0.21 d B h , and precompensation - 17 ps/nm. The degradation in the eye diagram is from timing jitter caused by IXPM. 258 RenHean Essiambre et a. l where we renormalized the amplitude A , to B = A n exp[W/21, n (6.36) where n = 1,2 and s(z) = exp[-Z(z)] is the power evolution (normalized to the power at z = ZO). second term on the RHS of Eq. 6.35 represents the effect The of cross-phase modulation of one pulse on another and is the MPM effect. A few methods have been used to evaluate analyt~cally effects of MPM the [52-61]. One approach consists of using the variational formulation [62]. The variational approach is particularly useful in nonlinear optics for describing the propagation of chirped pulses in nonlinear media. It was first used to describe pulse propagation in nonlinear diffractive media [63] and then in nonlinear dispersive media [64] such as optical fibers. Even though perturbative in nature, the variational method can be quite accurate even when the effect of nonlinearityis large [65,66]. The accuracy essentiallydepends on how well the ansatz chosen represents the evolution of the field experiencing nonlinearity. It is known through numerical simulations that the pulse integrity is essentially preserved after nonlinear interactions [3-61 in the pseudo-linear regime and that dispersion dominates the waveform evolution. Let’s consider the following ansatz, (6.37) where the parameters B,, C,, T,, t,, w,, and e assumed real, are the nth pulse , renormalized amplitude, chirp, characteristic width, position, frequency, and phase, respectively. All these parameters are assumed to be slowly varying functions of the distance z. Applying the variational method to Eq. 6.35 with the ansatz (6.37) gives [52, 541, (6.38) (6.39) daw dAt -= E dz y s(z)(Ei + E2)P2T exp (-T2/2) (6.41) (6.42) -= - B ~ ( z ) A ~ , dz 6. Pseudo-Linear Transmission of High-speed TDM Signals 259 where n = 1,2, Aw = 01 - w2 is the relative angular frequency between pulses 1 and 2, and At = tl - tz the separation between the two pulses. The dummy parameters P = and T = P A t have been introduced for convenience. The equation of evolution of the phase 6, has been omitted here. From Eqs. 6.38 and 6.39, one can easily show that the renormalized , energy per bit, E1,2 = f i B ; , z T ~ , zis conserved (as one would expect for IXPM, as discussed in Section 2.4.2). From Eq. 6.36, the actual energy in the fiber, E,(z), is given by E&) = E~exp[-Z(z)], where EO is the energy -( of a pulse at z = zo or E n ( Z ) = Eh exp [ Z z - in)], where Ein = f i B i T n = f i A : ( z = Zin)Tn(z = Zin) is the energy per pulse at the input of the span located at z = zh. Equations 6.38-6.42 represent the evolution of the pulse parameters with distance. One can easily verify that in the absence of nonlinearity [ y = 01 one recovers the evolution of a pulse in a dispersive medium (Eqs. 6.1 1-6.15) except for the pulse amplitude B, that is given by: &/dm B, = Bo/[l + C,(Z)~]~’~ = Ibnl. (6.43) The latter equation means that the phase of the complex amplitude b, associated with a linearly dispersive pulse is not included in the choice of ansatz (Eq. 6.37) used in the variational method. It is obvious from Eqs. 6.38-6.42 that IXPM affects all pulse parameters, B,, C,, T,, t,, and 0,. The chup is affected by both SPM (second term on the RHS of Eq. 6.40) and IXPM (third term on the RHS of Eq. 6.40). Chirping of a transform-limited pulse results in spectral broadening [16]. As discussed in Section 2.4.3, the spectral broadening caused by chirp induced by SPM is generally small in the pseudo-linear regime, especially when a precompensation equivalent to a few kilometers of transmission fibers is used (see Fig. 6.12). Moreover, the pulse chirping induced by SPM is identical for all pulses and does not lead to significantclosure of the eye, as all pulses have the same degree of broadening at a given point in the dispersion map. The pulse chirping effect of IXPM, however, depends on the presence or absence of neighboring pulses and creates a differential in pulse parameters that depends on the bit pattern. This effect of IXPM will create not only a differential in pulse bandwidth as the pulses propagate, but also in the value of net residual dispersion at which the pulses are transform-limited at the end of transmission. As a result, pulse chirping induced from IXPM can be responsible for some degree of amplitude jitter as pulses with different degrees of temporal broadening will be observed at a given point in the dispersion map. However, a potentially more dramatic effect of IXPM is the generation of timing jitter through the generation of frequency shifts (Eq. 6.41) that are converted into time shifts by dispersion (Eq. 6.42). For the case of a single pulse pair, the relative time shift induced by the nonlinear interaction between two pulses can be obtained by simultaneously solving Eqs. 6.41 and 6.42. For the purpose of solvingEqs. 6.41 and 6.42, 260 Renk-Jean Essiambre et a. l the evolution of the characteristicpulsewidth T,, and chirp C,, are assumed not to be affected by nonlinearity and are given by Eqs. 6.12 and 6.13. Assuming identical pulses, Eq. 6.41 can then be rewritten as: where Au = Aw/(2n) is the relative frequency between the pulse pair. From Eq. 6.44 it is clear that the relative frequency shift Au due to IXPM can only grow monotonically (the RHS of Eq. 6.44 does not change sign with propagation). Figure 6.14 (from Ref. [60]) shows an example of the frequency shift Au induced by IXPM in a simplified system where fiber loss is neglected. The dispersion map of length LM = lOOkm is composed of two fiber segments. The fiber lengths L1 and LZ are identical and equal to 50km. The disper1 sions D and D2 are of opposite signs and equal to 10 and -1Ops/(kmnm), respectively. The nonlinearity y = 2 W-' km-' for both fiber segments. The launched energy for each pulse Ein = 100 fJ, the characteristic pulse widths TO= 3 ps (Tp = 5 ps), and pulse separation TB= At(0) = 25 ps corresponding to the separation between two pulses in a 4O-Gb/s pulse train. The two launch positions considered in Fig. 6.14 are at the input of each fiber type. The frequency shifts for both launch positions are virtually identical when using the variational method (Eq. 6.44), and almost identical when the frequency shift is calculated by numerically solving the NSE (Eq. 6.29) with a pulse pair as the input. From Fig. 6.14, one can estimate the accuracy of the variational formulation. As mentioned earlier, the net time shift between two pulses, A(z) 3 At(z)-TB,experienced after one dispersion map period depends on the launch position in the dispersion map. Figure 6.15 (from Ref. [60]) displays the net time shift as a function of the launch position in the anomalous fiber segment for the dispersion map described above. It shows that a launch point in the middle of the anomalous fiber segment [Dl = lOps/(kmnm)] leads to zero net time shift. Similarly,a launch point in the middle of the normal dispersion fiber segment [D2 = -10 ps/(km nm)] also suppresses the net time shift. In the absence of fiber loss, it is possible to find a dispersion mapping that makes the pulse separation after one span At(L) come back to its initial value TBif one neglects variations of At(z) on Au(z) in Eq. 6.44 and assumes that the evolution of the characteristicpulse width T,(z) originates from dispersion alone. One can make the net time shift A vanish if the RHS of Eq. 6.42 is , an antisymmetric function around an arbitrary point of symmetry, 2. This requires 82(z - z,) and Aw(z - z,) to have opposite symmetry properties (one symmetric and the other antisymmetric) at the location z,. From Eq. 6.44, the symmetry of the frequency shift Au(z)is determinedby the symmetry of pulse width evolution T,,(z).From Eq. 6.12, T,(z) can only be a symmetricfunction 6. Pseudo-Linear Wansmission of High-speed TDM Signals 1 261 0 1 0 20 30 40 50 60 70 80 90 100 Distance (km) Fig. 6.14 Frequency shifts Au as a function of distance for a lossless system for two launch positions (transmission fiber parameters described in the text). The first launch position is at the input of the normal fiber [Ll = 50 km and D = - 10ps/(km nm)]; 1 solid curve from variational approximation (Eq. 6.44) and asterisks from solving the NSE (Eq. 6.23) directly. The second launch position is at the input of the anomalous fiber [Lz = 50 km and 4 = IOps/(km nm)]; dashed curve from variational approximation (on top of solid curve) and open circles from solving the NSE. From Ref. f601- I 0.5 ' 1 0.4 .rl ' -. 01 b -0.2 + , Q -0.3 -0.4 -0.5 0 5 10 15 20 25 30 35 40 45 50 Launch Position (km) Fig. 6.15 Net t h e shift A after one dispersion map as a function of launch position in the anomalous fiber [Dl = -lOps/(kmnm)]. For a launch point at midpoint in the fiber, the net time shift vanishes. From Ref. [60]. 262 RenUean Essiambre et al. if the cumulative dispersion map is either symmetric or antisymmetricaround zo, i.e., ICGW(Z-zo)l = ~CGW(ZO where zo is the point of zero cumulative -).I dispersion. Such a map makes the RHS of Eq. 6.44 a symmetric function and, after integration overz, makes Au(z) an antisymmetricfunction (see Fig. 6.14). The net time shift A(z) will then vanish when the dispersion map &(z - ZO) is a symmetric function. This condition is fulfilled when zo coincides with the middle of a fiber segment. Realistic systems differ in a few ways from the rather idealized symmetric system considered above. First the dispersion map is generally made of a long monotype lossy transmissionfiber. Fiber nonlinearity cannot be avoided in this fiber as the launch power Pinshould be maximum to maximize the link OSNR (see Eq. 6.2). The dispersion compensation is generally performed using a DCF having a much larger magnitude of dispersion than the transmission fiber so as to make it short enough to ease its insertion inside the in-line amplifiers and minimize its loss. The maximum power launched in the DCF is generally set to minimize any additional nonlinear signal distortions to the distortions already present from propagation in the transmission fiber. The degradation of the N F of the in-line amplifier from the presence of the DCF is set by this maximum power value. Depending on the type of transmission fiber, the type of DCF and whether or not these fibers are active (by using Raman pumping for instance), the impact of the presence of the DCF on the N F of the in-line amplifiers (see Eq. 6.7) and the link budget (see Eq. 6.2) can vary greatly. = Let's consider a system with the following parameters: pSpan -5 ps2/km, ~ D C F= 20 ps2/km, L,,, = 100km, LDCF= 25 km (the length LM of the 0 dispersion map is 125km), a = 0.21 dB/km in the transmission fiber with a nonlinearity coefficient y = 2 W-I km-' while the DCF is considered as having no significant nonlinearity for simplicity. Note that the relatively low value for the dispersion of the DCF has been chosen to facilitate visualization in Fig. 6.17. The cumulative dispersion of the transmission fiber Cspan - 5 0 0 ~ s ~ . = Figure 6.16 shows the frequency and net time shifts after one dispersion map period LM as a function of precompensation (see Fig. 6.2). Even though it is clear from Fig. 6.16 that there is always a frequency shift associated with IXPM, there is a precompensation length L,, that leads to zero net time shift. From Fig. 6.16 this value of L, = 3 km or a cumulative dispersion C,, = 60 psz. The mechanism allowing a zero net time shift for the optimum launch point (LpE = 3 km) can be understood by considering the evolution of the pulse parameters with distance as displayed in Fig. 6.17. The launch point of transform-limited pulses is at z = zo = 0 km. Besides the launch point, the pulses go through only one other point of zero net cumulative dispersion zo at z = 15km during the first map period. As the pulses propagate they broaden and recompress (Fig. 6.17d) according to the cumulative GVD (Fig. 6.17b). The pulses initially have their relative frequency Au(z) = 0 at z = 0 km.As 6. Pseudo-Linear Transmission of High-speed TDM Signals 263 F Y 2 0 I - j -2oof . i 3 ., 5 , 1 0 15 , , . , 0 , , 20 , ,,I 25 Precompensation (km) Fig. 6.16 Frequency shift Au and net time shift A after one dispersion period in a lossy system (described in the text) as a function of the length of DCF before the transmission fiber. Despite a nonvanishing frequency shift, it is possible to induce a zero net time shift with proper precompensation. In this exmple, the precompensation that minimizes the net time shift corresponds to 3 km of DCF. the pulses enter the transmission fiber at zin = 3 km, their relative frequency gradually shifts (3 km < z < 10 km) as the pulses start to compress as they approach zo = 15 km. No significant frequency shift occurs when pulses are close to the point of zero cumulative GVD (10 km < z 20 km) because of their negligible temporal overlap near ZO.An additional frequency shift occurs when pulses have passed the point z = 20 km where they start to overlap again. This process leaves a nonvanishing frequency shift at the end of the transmission fiber and a net time shift from the continuous conversion of the frequency shift Au(z) by the dispersion of the transmission fiber. Then, the remainder of the dispersion compensation (-Cspan - Cp, = 440 ps’) is applied using 22 km of DCF (103 km < z < 125km) to bring back the cumulative GVD to zero (pseudo-linear transmission). It is worth remembering that there is no additional frequency shift generated in the DCF because nonlinearity is neglected in this fiber. The accumulated frequency shift at the end of the transmission fiber translates into a time shift in the DCF that is opposite in sign to the one generated in the transmission fiber (because of the opposite sign of dispersion of the DCF). Under certain conditions these two net time shifts can cancel exactly. Using the relation A(z = L,) = 2n s,”- Bz(z)Au(z)dz, the requirement for timing cancellation can be written, 264 RenkJean Essiambre et a. l Distance (km) Fig. 6.17 Evolution of various parameters as a function of distance for the lossy system described in the text. The parameters monitored are: (a) GVD; @) cumulative GVD; (c) average power, Pm; (d) pulse peak power, Pp, normalized to the average power, P (e) frequency shift Au; and (f) net time shift A. The launch point in the , ; dispersion map is 3 km inside the DCF or 60 ps2of cumulative GVD. It has been chosen to make the net time shift after one dispersion map cancel out. where Au(Lpre Lspan)is the frequency shift at the end of the transmission fiber and Lcomp= (CPE Cspm)/j?spanthe length of fiber between the point is of zero cumulative dispersion zo that is located inside the transmission fiber and the end of the transmission fiber. The previous discussion considered a pulse pair. Note that inverting the pulse positions in the pulse pair, Le., At + - At in Eq. 6.44, reverses the frequency shift Au(z). Consequently, when considering a pulse train, a pulse located symmetricallyin a pulse pattern will experienceno net frequency shift as all frequency shifts cancel out. Conversely, the pulses experiencingthe maximum frequency shift are the trailing and leading pulses of an isolated long sequence of pulses where all nearest neighbors pulses are on the same side of the pulse pulling the pulse frequency in the same direction. However, as discussed earlier, such frequency shift does not necessarily immediately lead to a time shift, as it depends on the dispersion map. + + 6. Pseudo-Linear Transmission of High-speed TDM Signals 265 2.4.5 Intrachannel Four-Wave Mixing The second type of nonlinear interaction occurring among rapidly dispersing pulses in the pseudo-linear regime is intrachannel four-wave mixing (IFWM) [6, 71. This nonlinear interaction among pulses leads to generation of small pulses (shadow pulses) that can have positive or negative delays with respect to the original pulses. Observation of shadow pulses due to Kerr nonlinearity in fibers has been reported as early as 1992 [67] in the context of an experiment on an ultrashort (-90 fs) pair of pulses. IFWM in fibers has similarities with the phenomena of photon echoes in inhomogeneously broadened media and stimulated photon echoes in transient four-wave mixing [68]. One way to understand the mechanism of IFWM is as follows. Dispersed pulses that experiencenonlinearity can see a small portion of their field shifted by a discrete frequency value [69] as a result of FWM occurring between different spectral components of overlapping pulses. After propagation in a dispersive medium followed by full dispersion compensation, the discrete frequency shift is translated in a discrete time shift that, for sufficiently high dispersion, localizes the shifted field near the middle of a neighboring bit slot. If the bit slot is empty, the shifted field appears as a pulse of small amplitude referred to as “shadowy’ pulse. If there is a pulse in the bit slot where the field localizes, the two fields beat together, creating variations in main pulse amplitude. Figure 6.18 shows an example of how IFWM affects transmission. The eye diagram shows that the eye closure comes from amplitude jitter and shadow pulse generation. Pseudo-linear transmission usually operates at sufficiently high values of dispersion that result in having the shadow pulses localized near the middle of a bit slot. Let us consider the nonlinear interaction between two pulses propagating in a highly dispersivemedium shown in Fig. 6.19. Four power levels separated by 3 dB are displayed. The difference in power between the first-order shadow pulses (seen on each side of the pulse pair) is 9 dB. This power difference is proportional to the product of three times the original pulse power and is identical to the power scaling of sidebands generation in well-known FWM between channels. The second-order shadow pulses located further away from the pulse pair are separated by 15dB, consistent with FWM scaling for the dominant IFWM product (twice the pulse pair power and one time the firstorder shadow pulse) that leads to the second-order shadow pulse. A variety of analytic methods have been used to describe IFWM [5541,70, 711. They can be divided into three groups: small field perturbations [55, 56, 61, 711, nonlinear evolution in a dispersive frame [58, 591, and the variational method [57, 701. These different approaches have different advantages and drawbacks. The small-field-perturbationmethod has the advantage that it is not limited to a particular type of interaction between pulses, as it encompasses all types of nonlinear interactions. However, it is not clear how accurate the analysis becomes for large perturbation, especially when the timing jitter 266 RenC-Jean Essiambre et al. 1000 v F loo E 5 10 5 a 1 500 600 700 800 Time (ps) + v 4 $ 3 0 $ 2 1 0 0 IO 20 30 40 50 Time (ps) Fig. 6.18 Waveform and eye diagram after transmission for the same system considered in Fig. 6.13 except that TrueWaveTMfiber has been replaced by STD unshifted fiber (D= 17 ps/nm) and the precompensation has been changed to -527 ps/nm. The degradation in the eye diagram is from amplitude jitter and shadow pulse generation caused by IFWM. Time (ps) Fig. 6.19 Shadow pulse generation after transmission of 1.25-ps Gaussian pulse pair separated by TB = 6.25 ps (160 Gb/s separation) over 80 km of TrueWaveTMfiber (D = 4ps/nm). The pulse pair peak powers are 75, 150, 300, and 600mW, corresponding to average powers of 18, 15, 12, 9dBm for a pulse train assuming the same number of "zeros" and "ones." First- and second-order shadow pulse generation can be seen on both sides of the pulse pair. The separation in powers is 9dB between first-order shadow pulses and 15 dB between second-order shadow pulses. 6. Pseudo-Linear Transmission of High-speed TDM Signals 267 becomes on the order of the pulse width. Nevertheless, considerable insights can be gained by using the small-field theory. As for the study of IXPM, let us study the interaction in a pulse pair. Assuming that the optical field of the pulse pair after transmission B(z, t ) can be written as B(z, t ) = B,(z, t ) AB(z, t ) where Bn(z,t ) is the dispersive pulse evolution given by Eqs. 6.11-6.15 and AB(z,t) is the perturbed field (assuming ]AB1 < IB1,21) from the nonlinear interaction, the equation of < evolution of AB can be written as [55,61], + where s(z) is the power evolution along the line normalized to the transmission fiber input power. After full dispersion compensation, Eq. 6.46 can be rewritten as [55, 611, (6.47) where Bin = & is the renormalized pulse peak amplitude at the input of the transmission fiber, C(z) = B~(z')dz'/Tiis the chirp due to dispersion as defined in Eqs. 6.13 and 6.14 and z a point of zero cumulative dispersion. The o phases 0, with q = m,n,p are the phases of the individual pulses labeled m,n, andp. Notice that to simplify the expression in Eq. 6.47 the time is shifted to TFrt where Tpert = tm tp - tn is the location of each perturbation for pseudo-linear (highly dispersive) transmission. For a pulse pair there are eight sets of indices [m,n,p] in the sum of Eq. 6.46 but only six lead to different terms. The perturbations representing SPM are the combinations [ 1111 and [222]. The effect of IXPM is given by [221], [122], [112], and [211], whereas the combinations for IFWM are [121] and [212]. The type and location of the perturbations generated by a pair of pulses located at 25 and 50ps are summarized in Table 6.2. The amplitude and instantaneous frequency of the six perturbations AB,,,n, as given by Eq. 6.47 are plotted in Fig. 6.20. The p parameters used are /32 = -5ps2/km, T = 5ps (TO= 3ps), Pp = 20mW, y = 1.84 (Wkm)-', L = 100km, and zo = 0 (no precompensation). As seen in Fig. 6.20 and discussed in Section 2.4.3, the perturbations from SPM are identical for all pulses and do not generally contribute significantly to the degradation in transmission. Figure 6.20 shows that perturbations created by + 268 RenUean Essiambre et al. Table 6.2 Classification and Location of the Perturbations Generated by the Interaction of a Pulse Pair (Located at t = TB and t = 2 TB)Transmitted in the Pseudo-Linear Regime Location Type of Nonlineariq 0 TE 2T E 3T E SPM IXPM IFWM [1111 [122] and [221] [112] and [211] 11211 W I 0.01 5 E & B ; 0.001 0.m1 -25 0 25 50 75 100 -25 0 25 50 15 100 Time @s) Fig. 6.20 Perturbation AB of the transmitted field for the six different interactions generated by a pulse pair. Perturbationgenerated by SPM and IXPM fall on top of the original pulse pair located at tl = 25 ps and t2 = 50 ps, whereas IFWM perturbations are generated at t = 0 and 75 ps, respectively. IXPM are shifted in frequency by about 25 GHz. As discussed in Section 2.4.4, this frequency shift leads to timing jitter and some degree of pulse chirping. The contributions from IFWM create first-order shadow pulses located near t = 0 and t = 75 ps as seen in Figs. 6.19 and 6.20. When in the pseudolinear regime of transmission, shadow pulses are located near the center of the bit slots on either side of the generating pulse pair. One also observes in Fig. 6.20b that the shadow pulses are shifted in frequency by about 12 GHz and have some degree of chirp. Such shift in frequency means that the relative 6. Pseudo-Linear Transmission of High-speed TDM Signals 269 position of the shadowpulses changes with dispersion. Interestingly,for conditions such as Lo LNL(for short dispersion map periods or for low-dispersion fibers), the shadow pulses are located closer to the pulse pair [71]. However, this condition generally falls outside the scope of pseudo-linear transmission. Equation 6.47 can help to predict the effect of dispersion mapping on IFWM and how it impacts amplitude jitter. As mentioned before, the system impact of IFWM is to create shadowpulses in the “zeros” and amplitude variations in the “ones” as a result of the beating between the shadow pulses and the “ones.” The amplitude variations in the “ones” are maximum when shadow pulses and “ones” are in phase (in-phase cross-talk). This produces the maximum power variations (azlB,(L, t)lABmrp(L,t)l) in the “ones.” Conversely, when the shadow pulses and the “ones” are in quadrature (quadrature crosstalk) only to small amplitude variations (alAB,,np(L, t)I2)appear. When the = the difference of “ones” from the pulse train have identical phases t ) and the pulse B&, t ) at the end of phase between the perturbation transmission is given by the phase of the integral on the RHS of Eq. 6.47. When the power distribution is symmetric [s(z) = s(-z)], it is easily verified that the t ) (the real part of ABm,ng(~, in-phase cross-talk) vanishes for a symmetric dispersion map (zo = L/2) whereas it is not possible to make the imaginary part of AB,,, vanish with a specificdispersion mapping. A symmetric map will thus make the in-phase cross-talk disappear but keep the quadrature component. Since the amplitude jitter created by the in-phase cross-talk is the dominant source of distortion introduced by IFWM, a symmetric dispersionmap is very efficient to reduce the effect of IFWM on the “ones” for a symmetric power distribution. Figure 6.21 (from Ref. [61]) illustratesthe strong reduction of the effects IFWM for a symmetric dispersion map compared to a nonsymmetric map. The dispersion map is shown in the inset. The only difference between the dispersion maps of Figs. 6.21a and 6.21b is that the first occurrence of dispersion compensation is applied halfway in the transmission in Fig. 6.21b instead of as a precompensation in Fig. 6.21a. The small fluctuations in the “zeros” and beating in the “ones” in Fig. 6.21b come from the nonvanishing quadrature component of the perturbed field AB that does not disappear for a symmetric dispersion map. - 2.4.6 Dispersion Mapping When the effects of fiber nonlinearity are negligible (at low power levels), the waveform distortions observed at the end of a link depend only on the residual dispersion and dispersion slope at the center wavelength of each channel. However, to maximize the OSNR, operation at the highest possible power producing minimal waveform distortions becomes necessary. Unlike transmission at low powers, transmission at high powers suffersfrom the effects of fiber nonlinearity that strongly depend on the details of the dispersion maps. We present 270 RedJean Essiambre et al. s 4 (a) 3 2 ~ 1 0 0 to 10 f: 0 10 20 time [PSI 20 time [PSI 30 30 Fig. 6.21 Eye diagrams after transmission over 1600km in a lossless fiber for (a) a symmetric dispersion map and (b) a nonsymmetric dispersion map. The strong reduction of amplitude jitter originates from the cancellation of the in-phase component of the perturbed field AB for a symmetric dispersion map. The system parameters are 40 Gb/s, Pin = 1 dBm, Tp = 2.5 ps, y = 1.2W-I km-' ,D = 17ps/(km nmz)(from Ref. [61]), @ 2001 IEEE. in this section the effects of dispersion mapping on nonlinear transmission of high-speed signals. Finding the optimum dispersion map, modulation format, and fiber type for transmission is a multidimensionalproblem that suggests the use of a suitable multidimensional representation. Figure 6.22 shows a matrix of plots arranged in columns of different intensity-modulated formats and rows of different powers. Each plot in the matrix shows the eye closure penalty Ceye after transmission of a single 40-Gbh channel over 80 km of Ti-ueWaveTM fiber [D= 4 ps/(lun nm) and other fiber parameters given in Table 6.11 as a function of pre- and postcompensation. Full dispersion and dispersion slope compensations are applied prior to postcompensation. The nonlinear effects in the DCFs are neglected for simplicity. The eye closure C,, (see Section 2.3.1) is expressed in dB and color-coded according to the color bar seen in the figure. White areas represent regions of eye closure penalties above the maximum closure indicated on the color bar (3 dB in Fig. 6.22). One can see that transmission fidelity improves slightly when reducing the duty cycle with significant improvement for a low duty cycle of 10%. Figure 6.23 shows the same results but for STD unshifted fiber (see Table 6.1). Transmission fidelity for large duty cycle (33% and above) is poorer for transmission over STD fiber than over TmeWaveTM gradually improves to become comparable to but Ti-ueWaveTM low duty cycles ( ~ 2 0 % ) . for Transmission over multiple spans (8 spans of 80km) is presented in Fig. 6.24. The average launch power at each span input is 12dBm and the fiber dispersionD has a low value of 2 ps/(km nm). Dashed lines represent points of zero net residual dispersion. Two regimes of transmission can clearly be identified in Fig. 6.24. The first regime (solitonicregime), mostly for large duty cycle formats, shows optimum transmission at a net positive residual dispersion. The secohd regime (pseudo-linear regime), mostly for low duty cycle formats, 6. Pseudo-Linear Transmission of High-speed TDM Signals 271 Duty Cycle lOO%(NRZ) 50% 33% 20% 10% preeornp (pdnm) Rccornp (pslnrn) precomp (pslnm) Rccornp (pdnrn) Pnecornp (pshrn) Fig. 6.22 Eye closure penalties as a function of modulation formats and launch (average) power for 40-Gb/s single-channel transmission over 80 km of TrueWaveTM fiber [TrueWaveTMparameters are given in Table 6.1 except for the value of D = 4ps/(km nm) here]. Full dispersion and dispersion slope compensation are assumed before the postcompensation. Each plot in the matrix of plots shows the color-coded eye closure penalties Ceye a function of pre- and postcompensation. as Only a small improvement in eye opening (essentially the back-to-back difference in eye opening) can be seen when decreasing the duty cycle. Only when the duty cycle is reduced to a value as low as 10% is a significant improvement in transmission observed. Amplifier noise is not included. See also Plate 1. shows optimum transmission at zero net residual dispersion. For some formats and residual dispersion per span (for instance 33% duty cycle and 16 p s h m residual dispersion per span), both regimes are present and simultaneously produce moderate penalties. In the solitonic regime, there is compensation of dispersion by nonlinearity (even for NRZ!) through the solitonic effect that results on having optimum transmission at positive net residual dispersion. As discussed in Section 2.4.3, for pseudo-linear transmission, the effect of dispersion on the single-pulse transmission dynamics is so large as to reduce the compensation of nonlinearity by dispersion to a practically almost negligible value. This results in an optimal dispersion compensation in the pseudo-linear regime along the dashed line of zero net residual dispersion. For a higher value of dispersion of D = 4ps/(kmnm) (see Fig. 6.25), the solitonic regime starts to have reduced performance relative to pseudolinear transmission. For D = 8 ps/(km nm) shown in Fig. 6.26, transmission 272 Red-Jean Essiambre et al. Duty Cycle 100% 50% 33% 20% 10% D = 17 ps/(km nm) -5w -250 n -sw -250 n s w 250 o sw 250 n -5w 250 o Precomp (pdnm) Precomp (pslnm) Precomp (pdnm) Precamp (pslnm) Precomp (p”nm) Fig. 6.23 Identical to Fig. 6.22 except for STD unshifted fiber (parameters given in Table 6.1). For large duty cycles (NRZ and 50% duty cycle), transmission is limited to lower powers than TrueWaveTM fiber but rapidly increases as the duty cycle decreases. See also Plate 2. performance is clearly better for pseudo-linear transmission. Finally, for STD unshifted fiber (Fig. 6.27), pseudo-linear transmission with low duty cycle is the only regime that allows transmission fidelity. It is clear from Figs. 6.246.27 that optimum transmission is achieved by proper choice of precompensation. For the solitonic regime (Fig. 6.24) over low-dispersion fibers, there is a range of optimal values of precompensation that can be quite large, especially for the NRZ format. For pseudo-linear transmission, a more precise value of precompensation leads to optimum transmission, as evident from Figs. 6.24-6.27. Observations of the type of distortions for nonoptimum precompensation reveals that the eye diagram closes mostly due to timing jitter in these cases. As described in Section 2.4.4, the main source of timing jitter originates from IXPM. In the absence of fiber loss, the timing jitter can be minimized choosing a precompensation that makes the dispersion map symmetric (see Figs 6.14 and 6.15 for the single-span case). For a single span, the precompensation where Lspanis the span length. that makes the map symmetric is -Lspan Dspan/2 When many spans are concatenated, one should add half of the sum of the residual dispersion per span, or -N Cre,/2 where C,,, is the net residual dispersion per span. When fiber losses are included, the single-span precompensation should be shorter since the power distribution is concentrated at the beginning of the span (see Fig. 6.16) [6, 72, 731. The optimal precompensation can be 6. Pseudo-Linear Transmission of High-speed TDM Signals Residual Dispersion per Span 0 8 oslnm 16 vslnm 273 24 uslnm 40 Gb/s Single channel m 12 dBm 8 spans of 80 km TrueWaveTM/DSF s m m with D = 2 p s / ( h nm) -I 0 I 2 3 4 Eye Closure Penalty (dB) Fig. 6.24 Eye closure penalties after 8 spans of 80 km as a function of residual dispersion per span and modulation format. The transmission fiber has the same parameters as TrueWaveTM (Table 6.1) except we use here D = 2 ps/(km nm). The dashed lines are the points of zero net residual dispersion. Two regimes of transmission are present. A first regime (solitonic regime) has optimum transmission with a net positive residual dispersion. In this regime the solitonic effect is at play and is responsible for the compensation of dispersion by nonlinearity (even for NRZ!). The second regime (pseudo-linear regime) has its optimum transmission at zero net residual dispersion. See also Plate 3. expressed by the following semi-empirical relation [74], It is easily verified that this relationship approximately describes the optimum value of precompensation for pseudo-linear transmission limited by IXPM in Figs. 6.24-6.27. For systems operating at higher powers per channel, such as those in Figs. 6.22 and 6.23, IFWM can become the dominant nonlinear interaction. This can be easily understood as an increase in the “ones” pulses, peak power Pp followed by a much faster increase of the shadow pulses, peak power (c( P j , see Fig. 6.19. In contrast, the frequency shift associated to IXPM increases linearly with Pp (E1 and E2 in Eq. 6.41 are linearly proportional to Pp). As 274 Red-Jean Essiambre et al. Residual Dispersion per Span 0 8 pshm 40 Gb/s Single channel 12 dBm 8 spans of 80 km TrueWaveTM with D = 4 ps/(km nm) - 1 0 1 2 3 4 Eye Closure Penalty (dB) -200 -100 0 -200 -100 0 Preeomp (psinm) k o m p (pdnm) 0 -100 0 -200 -100 0 Precomp (pdnm) Precomp (psinm) Fig. 6.25 Same as Fig. 6.24 except D = 4ps/(km nm). See also Plate 4. Residual Dispersion per Span 40 Gb/s Single channel 12 dBm 8 spans of 80 km TrueWaveTM with D = 8 ps/(km nm) 1 0 1 2 3 4 Eye Closure Penalty (dB) Precomp. (pdnm) Precomp. (pdnm) Reeomp. (pslnm) Precomp. (pdnm) Fig. 6.26 Same as Fig. 6.24 except D = 8 ps/(km nm). See also Plate 5. 6. Pseudo-Linear Transmission of High-speed TDM Signals Residual Dispersion per Span 0 e 200 275 8 pslnm 16 pslnm 24 pshm E 200 Single channel ........ ....... ........ ....... Z"'Oo 0 12 dBm ........ ....... ........ ....... p -100 -200 8 spans of 80 km 5; m 5 g 0. - 2w 100 STD Unshifted Fiber ~~-L*-.--+--<-%. . . . .. . =___ ;. with D = 17 ps/(km nm) .... ... - . r.... ? -... 4 ... 0 100 -200 200 IO0 - - s2 s K s 2 n W E - 1 0 1 2 3 4 U O ? O r-4 B I 0 0 200 -200 100 0 Eye Closure Penalty (dB) .. .. 200 -100 0 -200 -100 0 -200 -100 0 Precomp. (pslnm) Precomp (pdnm) Precomp. (pdnm) Precomp. (pslnm) Fig. 6.27 Same as Fig. 6.24 except for STD unshifted fiber D = 17 ps/(km nm) and A,E = 80 km2. See also Plate 6. a result, the importance of IFWM relative to IXPM grows as the power per channel increases. For transmission limited by IFWM, multiple optimum values of precompensation can be found, as is obvious in Fig. 6.23, for low duty cycle and high powers. These multiple optimal values of precompensation originate from the recurrence of FWM that results from periodic generation and reabsorption of FWM products. Such phenomenon in the case of FWM between channels in WDM systems has been described in Ref. [75]. A similar phenomenon occurs with creation and reabsorption of shadow pulses for IFWM. The number of optimal values increases with the channel bit rate and the transmission-fiber dispersion. 3. High-speed TDM Pseudo-Linear Transmission Experiments Commercial systems operating at 40 Gb/s are expected to be installed by the beginning of 2002. Although high-capacity (3.08 Tb/s) long-span (100 km) 40-Gbk-based long-haul transmission over 1200km has been demonstrated [76] using the NRZ format [with the added advantage of forward error 276 RedJean Essiambre et al. correction (FEC) coding and distributed backward Raman amplification], pseudo-linear transmission using the RZ format generally enables transmission at higher powers (see Figs. 6.22-6.27). Higher powers provide larger power budget margin that can be used to extend the reach of the system. In this section we focus our attention on pseudo-linear transmission using the RZ format and describe experimentalhesearch systems operating at 40 Gb/s and 160 Gb/s per channel. The details of experimental transmission systems are examined, including the pulse sources in the transmitter, the demultiplexer, and clock recovery in the receiver. The 4O-Gb/s systems presented here are based on electrical time-division multiplexing (ETDM). However, the highest bit-rate systems will most likely be based on optical time-division multiplexing (OTDM) because the need for high bit-rate optical channels may develop before high-speed electronicsbecome commerciallyavailable.The components that make up a 16O-Gb/s system are examined carefully with emphasis on the potential for the systems to be practical and reliable in the near future. 3.1 40-Gbh SYSTEMS Pseudo-lineartransmission of 4O-Gb/s RZ signals is a very robust transmission technique and, as mentioned earlier, can reduce the effects of fiber nonlinearity even with simplified dispersion maps (with no precompensation for instance). In Section 2.4.6, transmission simulations examined various fiber types, duty cycles, and dispersion maps at 40 Gb/s. Sometimes it is difficult to construct the best dispersion map in a research laboratory and early experimentalresults used simplified dispersion maps. Ludwig et al. examined a single span of NZDSF with either 100% pre- or 100% postcompensation [4]. The span consisted of 150km of Ti-ueWaveTMfiber, and the measured system penalty as a function of input optical power is displayed in Fig. 6.28. The figure shows both experimental and simulated results. Clearly, higher optical power can be launched into a postcompensated span than a precompensated span for minimum system penalty. At a launch power slightly above 12dBm, a 1dB system penalty is observed for a postcompensated span. The highest power launched into the precompensated span is limited to 4dBm for a 1dB BER penalty. The ability to launch higher optical powers at constant waveform distortions from fiber nonlinearities increases the OSNR and power margin. In the experiment of Ludwig et al., a power margin of 8 dB is observed. This range is limited by nonlinearity for high powers and optical noise for low powers. Because terrestrial systems (see Chapter 9 of Ref [18]) can extend up to a few thousand kilometers, it is necessary to periodically reamplify the signal to prevent excess fading of the signal that will make it undetectable (see Fig. 6.2). Typical span length for terrestrial systems ranges from 80 to 120km, and dispersion compensation is applied at each in-line amplifier (in-line compensation). As mentioned in the Introduction, the main feature of pseudo-linear transmission is the fast waveform evolution that reduces or averages out the I I I - - - . -G--l-polt.ooap. -O--.-pR&mp. - , ; I , I I 1 0 -2 0 2 4 6 8 10 12 fiber-input power [am] 1476 Fig. 6.28 System penalty as a function of launched optical power into NZDSF fiber for pre- and postdispersion Compensation (from Ref [4]), @ 1998 IEEE. 1600 .E D -g 1200 800 400 O 6 16000 : : 9 ._ 12000 s - 8000 0 0 200 400 600 800 101 I Transmission Distance (km) Fig. 6.29 Dispersion maps for 800-km transmission system based on STD unshifted fiber. (a) Typical per span compensation. (b) All dispersion compensation is at the end of line. effect of intrachannel nonlinearities. Because of this property, it is possible to place all the dispersion compensation at the end of the line (end-of-line compensation) and still expect reasonably good transmission performance. The two types of dispersion compensation schemes are depicted in Fig. 6.29 for a 800-km transmission line. Figure 6.29a shows the dispersion map when there is full dispersion compensation at each in-line amplifier, whereas Fig. 6.2913 shows end-of-line dispersion compensation. The latter approach was demonstrated experimentally by Gnauck et al. at 40 Gb/s using 80-km amplifier spacing and 2.5-ps pulses [9]. In this experiment, two wavelengths spaced by 3.2 nm were transmitted over 800 km of STD unshifted fiber using typical erbium-doped fiber amplifiers (EDFAs) every 80 km. All of the dispersion compensation, which consisted of 10 DCF modules that had an average cumulative dispersion of -1360ps/nm per module, was placed at the receiver end. Here the average power per channel launched into the fiber span and DCF span were +4.0 dBm and +1 .O dBm, respectively. 278 Renk-Jean Essiambre et al. The measured Q-factor on the two wavelength channels was 18.0dB for an optimized launched state of polarization and the transmission penalty compared back-to-back ranged from 2.5 to 3.5dB. The measured OSNR (in a 1.0-nm-resolution bandwidth) was 13.0 dB (or 23 dB in the reference 0.1-nmresolution bandwidth). Longer amplifier spacing was also investigated [77]. Figure 6.30 shows the measured Q in dB vs launched optical power for different transmission distances (360, 480, 600, 720 km) using 120-km spans. The maximum Q decreases with increasing transmission distance due to OSNR degradation. The best Q-value in each system occurs between +6dBm and +8 dBm launch power, and the launch power into the DCF at the end is kept at +1 dBm. The low Q-values at low launch powers are due to poor OSNR, while at high launch powers distortions from nonlinearities are responsible for the degradation in Q. Although the Q-value decreases to 15.6dB (corresponding to an error rate floor of after 720 km, the experiment demonstrates that long amplifier spacing, long distance, and end-of-linedispersioncompensation are possible using RZ pseudo-linear propagation. Placing the dispersion compensation at the end of the line may offer some advantages for a point-to-point link. One potential advantage is to allow a “skip” of dispersion compensation at amplification sites where insertion of DCF may be difficult. A second potential advantage is to dramatically reduce the number of sites where dispersioncompensationis performed to one (or two if precompensation is used). On the other hand, a potential drawback of endof-line compensation arises in the context of optical networking where add and drop points are present in the line. For such an optical network, large dispersion compensation is necessary at adddrop sites to bring back the signal to a point of zero net dispersion when end-of-line dispersion compensation is 20 -v-3360km -0-480km -B-600km . - . 14 2 4 6 8 10 12 14 Power into SMF (dBm) Fig. 6.30 Q measurements vs launched power into STD unshifted fiber spans for three, four, five, and six 120-km span systems (from Ref [77]), @ 2000 IEEE. 6. Pseudo-Linear Transmission of High-speed TDM Signals 279 used. In contrast, for in-line compensation, a more modest cumulative dispersion is necessary at adddrop sites to bring the signal to a value of dispersion where it can be detected. 3.1.1 ETDM Long-Haul Pseudo-Linear Transmission Maintaining high OSNR is important for any communicationsystem, but it is especially criticalfor 40-Gb/s-based systems because of the high density of bits. Today’s receivers require typically at least 24 dB OSNR (measured in 0. l-nmresolution bandwidth) in order to achieve bit error rate (BER) performance better than lop9. For noise accumulation purposes, the transmission line can be considered as a chain of amplifiers with lossy elements inserted between them. Because one wants to minimize the number of in-lineamplifiersfor a given transmission line, the span loss is generally the most lossy element in the line. As a result of limitation of launch power to the transmission fiber because of fiber nonlinearity, the amplifiers in the chain receiving the lowest input powers are the in-line amplifiers. Consequently, the degradation of OSNR arises primarily from the in-line amplifiers and is related to their NE However, the presence of DCF inside the in-line amplifiersmay also add to the degradationof OSNR by increasing the NF of the in-line ampuer (see Section 2.4.4). The degradation of NF from the presence of DCF can be decreased by reducing the DCF loss and by increasing the input power to the DCF. However, the input power to the DCF is limited by its small effective area of 15-22 vm2 (see Table 6.1). To improve the performance, alternative methods of providing dispersion compensation are being examined. Recently, Ramachandran et al. described a dispersion compensation module that is based on higher-order mode (HOM) fiber that has an effective area, A,r of 65 pm2 [78]. Such effective area is much larger than conventional DCFs and reduces the nonlinear effects in the fiber. A schematic of the module is shown in Fig. 6.3 1. The device consists of a 2-km spool of HOM fiber spliced to two long-period gratings (LPGs) written using ultraviolet light. The LPG converts the linearly polarized mode, LPol, into the LP02 mode at the input and then back to the LPol at the output. Up to 99% mode conversion over a bandwidth of 40 nm has been demonstrated. 2 km EOM-JXF LPG lull \/ LPG 11111 11 4 u 1 + Lp,, LP, LO Pl Fig. 6.31 Schematic of higher-order-mode dispersion compensation module (HOM-DCM). Block arrows represent propagation of dominant mode. LPG- long period grating, X fusion splice. 280 Renuean Essiambre et al. Single-wavelength 4O-Gb/s transmission over 1700km of TWRS has been demonstrated in a recirculating-loop experiment with the HOM dispersion-compensationmodule [79]. The loop consisted of 100km of TWRS fiber followed by the HOM module used in a per-span postcompensation scheme. The 100-km span loss is 22 dB, and backward Raman amplification with an on-off gain of 15dB is used in the transmission fiber. The launched signal power into the transmission fiber is 0 dBm, whereas the signal power into the DCM module could be varied from 0 to +5 dBm. The back-to-back performance of the transmitter and receiver in this system showed a receiver sensitivity of -30 dBm, which corresponds to a required OSNR of 24 dB according to Eq. 6.6. The BER penalty at a BER = after transmission is 5 dB. The penalty is attributed in part to reaching the OSNR limit and to some degree to PMD. The average differential group delay (DGD) of the TWRS due to PMD is 0.05 ps/& (2.1 ps after 1700km),whereas the PMD contribution from the 17 cascaded HOM-DCMs is 4.1 ps. This module has the potential to significantlyimprove the performance of in-line amplifiers that have degradation in NF from the presence of the DCF. 3.1.2 Alternating Polarization Format To increase transmission capability further, the state of polarization of the last two tributaries to be time-division multiplexed can be set orthogonal [80,81]. This techniqueproduces alternatepolarization (AP) between adjacent bit slots. It is well known [61, 821 that FWM between orthogonal waves propagating in a nonlinear medium of low birefringence is reduced substantially. Similarly, the orthogonal polarization between adjacent bits reduces the effects of IFWM and enables higher launch powers at a fixed level of signal distortion. Figure 6.32 (from Ref. [83]) shows the BER performance as a function of -5 0 5 10 15 20 input power,dBm Fig. 6 3 Bit-error-rate (a) and reflected power (b) vs input power for single polar.2 ization (open symbols) and alternating polarization format (closed symbols) (from Ref [83]). 6. Pseudo-Linear Transmission of High-speed TDM Signals 281 launched optical power for a 40-Gb/s transmission experiment over a single span of 240 km of STD unshifted fiber. Two curves are plotted; one for the AP format and one for single polarization (SP). By going from SP to AP,an additional 2 dB of launch power is possible, extendingthe ultimate span length of the system. Furthermore, the AP format reduces the effect of stimulated Brillouin scattering in the fiber (the Brillouin gain is polarization dependent), thus enablinghigher launch powers to be achieved. This technique is also used in later sections to achieve long-distance transmission of 160G b k 32 IdO-Gb/s OTDMSYSTEMS . To keep pace with the increasingdemand for capacity in the backbonenetwork, fiber transmission systems with high TDM bit rates are becoming increasingly important. Long-haul system evolution has historically occurred in four-fold increases in bit rate because of the SONET and SDH standards. Economics have determined that the cost per bit can almost be cut in half if the bit rate increases by a factor of four. High TDM bit rates are attractive because a single wavelength can accommodate a higher-capacitypayload. Already there are applications that require a high serial data rate. One such applicationinvolves the aggregation of data from a high resolution radio telescope that consists of 13,000dipole antenna elements [84]. Each antenna produces digitized data at 2.O-Gb/s rates and the antennas are grouped into stations of 78 antennas, resulting in an aggregate data rate of 160Gb/s for each group. Consequently, the interest for the next generation TDM bit rate of 160Gb/s is emerging rapidly [85-871. However, the speed of electronic circuits currently limits the possible TDM data rate to -40 Gb/s. Therefore, OTDM techniqueshave to be applied today to construct serial data rates of 160Gb/s. This section examines the progress toward TDM rates of 160Gb/s based on OTDM. In particular, the focus is on practical implementation, fiber transmission limitations at 160Gb/s, and the possible channel spacing for WDM systems. 3.2.1 16O-Gb/s OTDM Transmitter and Receiver Optical time-division multiplexing (OTDM) is a well-known technique that allows very high serial data rates to be obtained while using lower bit-rate electronics. This multiplexingtechnique is often used in research laboratories to explore high bit-rate transmission before electronic components are available [88]. In order to realize a 160-Gbh TDM system, optical multiplexing and demultiplexing of lower-rate tributaries must be used at the transmitter and receiver ends. Figure 6.33 shows a schematicof a 16O-Gb/sOTDM system including the most important building blocks. To minimize the complexity of the transmitter, the highest possible tributary/ electronic data rate should be used, Le., 40 Gb/s at present. The optical multiplexing can be performed by either bit-interleavingfour 4O-Gb/s RZ signals with the SP or with AP where adjacent bits have orthogonal polarization. 282 RenCJean Essiambre et al. 160 Gbls Transmitter: Ootical MUX 160 Gbls Receiver: Ootical DeMUX Gbls RZ 40 Gb/s ETDM 4 x 40 Gbls ETDM 40 GH7 Fig. 6.33 Schematic of 160-Gb/s OTDM single wavelength transmitter and receiver based on 40-Gbit/s tributaries. TDM-SP TDM with same polarization. TDM-AP TDM with alternating polarization (AP) format. CR: clock recovery. It should be noted that the interleaved tributaries in general have a random relative carrier phase. This is also the case when a single pulse source is used, unless the four data modulators are integrated on a single chip. As a consequence, to optically multiplex from 40 Gb/s to a single polarization 16O-Gb/s signal (TDM-SP), 2-ps pulses with more than 30 dB extinction ratio (suppression of pedestal) are needed to avoid coherent beat-noise between adjacent bits. Multiplexing by alternating polarization (TDM-AP) considerably relaxes the requirements to the RZ pulse-width and extinction ratio. Clearly, a key technology at the transmitter is a stable and reliable pulse source that generates transform-limited narrow pulses. In addition to high extinction ratio and low insertion loss, the pulse source must provide high SNR, well-controlled repetition frequency,and wavelength. Several techniques have been considered such as mode-locked lasers [89-931 (both fiber and semiconductor types), gain-switched lasers [94], as well as gatingkarving devices like electro-absorption modulators (EAMs) [95-981 and Mach-Zehnder (MZ) modulators [99]. Mode locking provides the shortest pulses (< 1ps is possible), but stability is a critical issue as well as control of repetition frequency, in particular for semiconductor devices. The carving technique, on the other hand, provides very stable and versatile pulse sources. Although EAMs have intrinsic carving loss, they are very attractive since transform-limited 3-4 ps pulses with more than 20 dB extinction ratio can be realized at 40 GHz with a sinusoidal drive voltage of only 5 V peak-to-peak. To further improve the extinction ratio and to reduce the pulse width, EAMs can be cascaded or be followed by an optical compression and/or optical regenerator stage [98, 1001. A short length of DSF and an optical filter can very efficientlyrealize the latter, but various kinds of nonlinear optical loop mirrors (NOLMs) have also been used as regenerators [loll. At the receiver, optical demultiplexing is necessary to go from the serial rate of 160 Gb/s to a data rate that can be handled by electronics. Several optical demultiplexing techniques have been investigated, such as electronically controlled gating devices such as EAMs [98, 100, 1021and MZ modulators [103], 6. Pseudo-Linear Transmission of High-speed TDM Signals 283 or optically controlled gates based on semiconductor optical amplifiers (SOAs) [94, 104-1061 or NOLMs [107, 1081. Also, optical coherent demultiplexers based on, e.g., FWM in SOAs or optical fibers, have been considered [109, 1lo]. The most practical solutions, however, are the modulator-based techniques. The simplicity of use and general availability allows for reliable system implementation. Of course these solutions have drawbacks. Lithium niobate (LiNb03) modulators, although very mature, reliable, and low loss, are sensitive to polarization and do not provide enough extinction ratio in a single stage. EAMs on the other hand, provide high extinction ratio (>20 dB) and low polarization sensitivity but are wavelength dependent and fairly high loss at present. The semiconductor approach lends itself to potential integration into more complex circuitry. As shown in Fig. 6.33, the OTDM receiver requires parallel demultiplexers. Naturally, it is preferable to reduce costs through integration, and this architecture lends itself to larger-scale photonic integration. One example of a step in this direction is the tandem EAMs with integrated semiconductor optical amplifiers that have been described operating as transmitters at 40 Gb/s [l 1 1, 1121. This configuration is also well suited for application in the demultiplexer and clock recovery. All of the above demultiplexing techniques require a clock signal that is either electronic or optical. The clock from the 160-Gb/s TDM signal can be recovered by all-optical techniques such as self-pulsating lasers [ 1131 or by electro-optic clock-recovery schemes where the phase detection is performed optically [114, 1151 or by injection locked electro-optic oscillators [98, 1161. Even conventional electronic clock recovery circuits are possible since narrow-band 160-GHz electronics are sufficient to retrieve the clock frequency. The electro-optic scheme and injection locking have been demonstrated at 160 Gb/s, whereas the conventional all-electronic phase-locked-loop has been demonstrated at 100 Gb/s [1171. Figure 6.34 shows a semiconductor-based transmitter and receiver that operates at 160 Gb/s. Using two cascaded EAMs driven at 40 GHz, 3-ps pulses with 30dB extinction ratio are generated. The pulses allow multiplexing to 160 Gb/s with the TDM-AP format. Typical laboratory experiments are performed using a single data modulator as is shown. A real transmitter would be Optical MUX Fig. 6.34 Experimental 160-Gb/s transmitter realized using electroabsorption modulators (EAMs) and passive delay-line interleaving. 284 Red-Jean Essiambre et al. implemented using at least four modulators, as shown previously in Fig. 6.33. To achieve appropriate decorrelation of the 40-Gb/s tributaries, several meters of fiber are used in each arm of the multiplexer, providing sufficient delay. To generate pulses that can be multiplexed to 160 Gb/s with the TDM-SP format, further reduction of pulse width and improved extinction ratio are necessary. A simple fiber regenerator based on self-phase modulation in optical fiber can achieve sufficient improvement to multiplex in single polarization [118]. In this case, less than 2-ps pulses with higher than 30 dB extinction ratio are generated. The optically preamplified receiver shown in Fig. 6.35 consists of two concatenated EAMs, one being driven at 40 GHz, the other at 10 GHz [119]. Consequently, optical demultiplexing from 160 to 10 Gb/s is performed. This allows BER testing to be performed at 10 Gb/s, where electronic testing equipment is widely available. Future implementations would most likely be based on 40-Gb/s tributaries to reduce the number of demultiplexers required. The clock recovery is realized by an injection locking technique, using components very similar to the demultiplexer. The clock recovery is also a series of concatenated EAMs that perform TDM demultiplexing from 160 to 10 Gb/s. This tributary is detected and then filtered using an electronic filter with a very high Q (900-1000). The oscillator is formed by amplifying the 10-GHz component and driving the EAMs with this signal. Electronic phase adjustment can be used to control which time slot or tributary is demultiplexed. This approach can also be used to provide add/drop functionality. One advantage of this approach is that the reconfiguration of the demultiplexer can be achieved on a very fast time scale. Tong et al. demonstrated 4-11s reconfiguration of a similar EAM-based demultiplexer [ 1201. Taking advantage of optical demultiplexing to 10 Gb/s, the bandwidth required at the receiver is less than 10 GHz. For most of the experimental results described, back-to-back receiver sensitivity at 160 Gb/s is approximately -26 dBm, which corresponds to a required OSNR of 28 dB according to Eq. 6.6. Receiver sensitivity is measured before the optical demultiplexer. 160 Gb/s Rz -lOGb/s ,@Phase shifter CR Fig. 6.35 Experimental 160-Gb/s receiver with electro-absorption-modulator-based demultiplexing and electro-optic clock recovery. 6. Pseudo-Linear Transmission of High-speed TDM Signals 285 3.2.2 Single-Channel Transmission A transmission experiment using a single channel at 160Gb/s was demonstrated. The 160Gb/s transmitter and receiver are similar to those described in Figs. 6.34 and 6.35 except that the baseband modulation rate is at 20 Gb/s, which requires an extra passive optical multiplexer. In addition, the 16O-Gb/s signal was multiplexed in a single polarization. The link is assembled similar to the typical system layout shown in Fig. 6.3. It consists of four, 100-km spans of TmeWaveTMreduced slope (TWRS) with a measured dispersion of 3.5 ps/(km nm) and an average loss of 0.2 dB/km. The dispersion is compensated at each amplifier site using DCF to nearly 100% compensation. PMD measurements on the fiber show PMD values of less than 0.06 ps/d&. The launch power into the transmission span is +8 dBm while the launch power into the DCF is +3 dBm. Figure 6.36 shows the measured BER vs received power for 0 and 300 km. Bit error rate is plotted for two arbitrarily chosen 10-Gb/s tributaries. The receiver sensitivity for the 160-Gb/s transmission, measured at a BER = is -21.5 dBm after 300 km. Compared to backto-back (0 km), the BER penalty is 3.5 dB. The penalty is mainly attributed to PMD and IXPM, which occurs in NZDSF as described in Section 2.4.4. To extend the reach further, an experiment with an improved system, displayed in Fig. 6.37 has been performed [ 1211. In this system, Raman amplification is used to compensatefor the span loss of 22 dB and no in-line amplifiers are required. This enables an improvement in OSNR to be achieved through both lower NF of the Raman amplifiers and by removing EDFAs and DCF throughout the span. As described earlier in Section 3.1, placing the DCMs -4 -5 E-6 B -0 0 , -7 -8 -9 -1 0. 3 2 3 0 -28 -26 -24-22 -20 -18 Received Power (dBm) -' 6 Fig. 6.36 BER vs received power for two arbitrarily chosen lO-Gb/s tributaries after 300-kmtransmission compared to back-to-back. 6. Pseudo-Linear Transmission of High-speed TDM Signals 287 High-speed TDM rates of 160 Gb/s will impose stringent requirements to PMD of the transmission fiber and any in-line optical components. Although no experimental investigation has been reported, a mean PMD of less than 0.75 ps is expected to be required at this data rate. Hence, optical PMD compensation techniques [1231will be imperative at 160Gb/s. Moreover, tunable dispersion compensators are likely to be needed at the receiver to trim the optimum dispersion. The dispersion tolerance at 160 Gb/s is approximately f 2 ps/nm for single-polarization transmission (TDM-SP) and f 6 ps/nm for the case where adjacent bits have alternating polarization (TDM-AP). Tunable dispersion compensation has been demonstrated over a tuning range of 50 pshm at 160Gb/s by a fully packaged, electrically tunable fiber Bragg grating [ 1241. 3.2.3 Wavelength-Division Multiplexing (WDM) The short RZ pulses necessary to form a 160-Gb/s high-speed TDM signal will increase the spectral bandwidth per channel at 160Gb/s to values above the bandwidth of a 4O-Gb/s RZ signal and well above the bandwidth of a 4O-Gb/s NRZ signal. Even though 16O-Gb/s signals have larger bandwidths per channel, they also have more information per channel. The presence of trade-offs between single-channelbit rate and spectral efficiency in WDM systems presents the classic argument for moving toward high-speed TDM signals or not. The first 3-Tb/s system was demonstrated using 19 wavelengths each modulated at a bit rate of 160Gb/s [ 4 . channelswere spaced by 480 GHz, 4 ] The giving a spectral efficiency of 0.33 bits/s/Hz. More recently, 16O-Gb/s WDM transmission has been demonstrated with a channel spacing of 300 GHz, corresponding to a spectral efficiency of 0.53 bit/s/Hz [125]. A schematic of the transmission experiment over 400 km of NZDSF fiber is shown in Fig. 6.39. At the transmitter, 4O-Gb/s RZ channels with a channel spacing of 300 GHz are sliced from a 40-nm-wide spectrum produced by SPM in 2km of DSF MUX and D m 6 Ch., 300 GHz TDM Transmitter ~ p : Pumps Fig. 639 WDM at 160 Gb/s per channel transmission experiment. Six wavelengths spaced by 300 GHz created through spectral slicing of SPM spectrum are transmitted. 288 RedJean Essiambre et al. by a waveguide-grating-router-based(WGR) demultiplexer. The WGR has a Gaussian-shaped bandpass characteristic with a 3 dB bandwidth of 1.35 nm and an insertion loss of -4.5 dB. Only six channels are selected. This is considered sutlicient to capture all transmission distortions. The pulse width of the 4O-Gb/s RZ signals is 3 ps after the demultiplexer as determined by the bandwidth of the WGR. Next, the six wavelength channels are delayed by different fiber lengths to decorrelate them before they are multiplexed by a second WGR with characteristics similar to the first one. After being multiplexed onto the same fiber, each wavelength channel is opticallytime multiplexed to a 160-Gb/s TDM-AP format. Figure 6.40 shows the measured optical spectrum at (a) the transmitter, (b) after transmission, and (c) after WDM demultiplexing. The transmission span is the same as just described for the single-channel400-km experiment. Since the channel spacing at 160Gb/s will be several hundred GHz, the transmission limitationseven on NZDS fibers are solely single-channeleffects. Therefore, the optimum dispersion map for WDM transmission is similar to the map for single-channel transmission. Note, that WDM transmission with the TDM-AP format is more effective than TDM-SP with polarizationinterleaved WDM channels, because transmission limitations are caused by intrachannel effects at 160Gb/s. TDM-AP also permits wavelength channels f -0 3 -50 1535 240 1545 1555 1565 1575 Wavelength (nm) Fig. 6.40 Optical spectrum of six 16O-Gbivs channels at (a) transmitter output, (b) after transmission, and (c) after wavelength demultiplexing. 6. Pseudo-Linear Transmission of High-speed TDM Signals -4- 289 * . -5- fi 4. g m P 0 ch 4 pd-1,400 km WDM W ch4 pol-2 4M)lrmWDM A alngie chainel pol.i,m B A aingie channel pol-2.400 kn 0 ch 3, po1-l14Wkm WDM 0 ch3,pol-2.400kmWDM back-tpbmk w U -6-7- e :* W - 0 a @ -8-9 ' * B B ' ' ' -10-1 1 ' ' .' ' ' . ' ' ..' ' ' ' to be dropped and added without the need for polarization tracking. Furthermore, TDM-AP allows for a high spectral efficiency because this format tolerates broader R Z pulses. The measured BER performance after transmission is shown in Fig. 6.41. For comparison, back-to-back BER (no transmission fiber) is also shown for wavelength channel three (1556nm) for the case where only channel three is being multiplexed (squares). The difference in sensitivity between the squares and circles gives the combined penalty for allocating 160Gb/s on a 300GHz grid and transmitting the six channels over 400 km over fiber. As seen, the penalty is less than 0.2 dB, confirming that the cross-talk from adjacent wavelength channels is very small. Furthermore, the change in measured receiver sensitivity is less than 0.5 dB when varying the launched power per channel between -1.4 and +7.6 dBm (power limited by EDFA), demonstrating the large power margin of the pseudo-linear transmission regime. 4. Conclusions A description of high-speed TDM systems operating at 40 and 160 Gbls per channel has been presented. We first discussed and described a new regime of transmission, the pseudo-linear regime. Such a regime allows transmission of high-power, high-speed TDM signals over fibers with relatively high disI1 persion [ D > 2ps/(kmnm)], the type of fiber composing the vast majority of the installed and currently deployed networks. The pseudo-linear regime of transmission is characterized by a rapid pulse broadening that results in a dramatic reduction of the solitonic effect (compensation of dispersion by 290 RenCJean Essiambre et al. nonlinearity) on each pulse. As a result, full dispersion compensation is used in this regime. We described two new forms of nonlinear interactions between rapidly dispersing pulses, intrachannel cross-phase modulation (IXPM) and intrachannel four-wave mixing (IFWM). These two intrachannel effects are the most important nonlinear interactions in pseudo-linear transmission and determine the dispersion mapping even for wavelength-division multiplexed (WDM) systems. Both effects limit the maximum power that can be transmitted. It was shown that the most important effect of IXPM is the generation of timingjitter, whereas for IFWM it is the generation of amplitudejitter and the creation of shadow pulses. The effects of dispersion mapping on the reduction of the effects of DLPM and IFWM has been discussed. In the second part of the chapter, the semiconductor-based technologies enabling the development of stable and reliable high-speed transmitters and receivers have been described. Long-distance error-free transmission at 40 and 160Gb/s per channel has been demonstrated with the return-to-zero (RZ) format. Both electronic time-division multiplexing(ETDM) and optical time-divisionmultiplexing (OTDM) systemshave been demonstrated. Finally, transmission of a WDM signal based on 160 Gb/s per channel with 300-GHz channel spacing (spectral efficiency of 0.53 bits/s/Hz) has been demonstrated. Acknowledgments We would like to thank Andrew Chraplyvy, Daniel Fishman, Lisa Wickham, Taras Lakoba, Alan Gnauck, Peter Winzer, Vadim Zharninsky, Diego Grosz, Bob Jopson, Jau Tang, and Colin McInstrie for interestingdiscussions and/or their comments on the manuscript. List of Symbols Symbol Meaning Fiber loss at the signal wavelength Gadloss coefficient at the signal wavelength Constant of propagation Group-velocity dispersion (GVD) parameter Third-order dispersion (TOD) parameter Temporal separation between two pulses relative to the bit period Spectral width Spectral separation Pulse full-spectralwidth at half maximum Loss of a dispersion-compensatingmodule Nonlinear coefficient Av Av AP V rl Y 6. Pseudo-Linear Transmission of High-speed TDM Signals 291 D Dspan E O Ein G h Iones L Lcotn, Signal wavelength Signal optical frequency Channel spacing in WDM systems Angular frequency of the nth pulse Pulse initial angular frequency Standard deviation of Iones Standard deviation of Izeroh Phase of the n* pulse Pulse initial phase Complex pulse amplitude of the nthpulse Effective area of the fiber transverse mode Field amplitude of the nth pulse Bit rate 3-dB electrical filter bandwidth Renormalized electric field of the n* pulse Full width at half maximum optical filter bandwidth Speed of light in vacuum Pulse chirp Initial pulse chirp Cumulative dispersion Eye closure penalty Cumulative group-velocitydispersion Cumulative dispersion precompensation Net residual dispersion per span Dispersion Dispersion of the fiber making a span Energy of pulse at a point zo Pulse energy at the input of a span Amplifier gain Planck constant Current at the sampling instant in a “one” bit Current at the sampling instant in a “zero” bit Cumulative losslgain factor Length of the communication link Fiber length from the point zo inside the transmission fiber to the fiber end Length of the nth fiber segment Length of fiber used as precompensation Soliton period (= n L D / 2 ) Length ofthe transmission fiber making a span span loss Dispersion length (= T / j 2 ) ;l31 Dispersion map period 292 RenBJean Essiambre et al. Nonlinear refractive index Spontaneous emission factor Number of amplifiers Noise figure of an amplifier Noise figure of an optical amplifier with mid-stage dispersion compensation Required OSNR to achieve a given bit error rate Noise power in a given bandwidth Au Average signal power Average power at the input of the dispersion compensating fiber Average power at the input of a transmission fiber Average power at the input of the first stage of a multistage amplifier Peak power of a pulse Height of the rectangle fitting inside the eye diagram Receiver sensitivity Q-factor Extinction ratio between “zeros” and “ones” Power evolution normalized to the power at the fiber input Spectral efficiency Dispersion slope Time Timing of the nth pulse Bit period Characteristic width of a transform-limited pulse Full width at half maximum of the nth pulse Full width at half maximum of a pulse Evolution of Tp with cumulative dispersion Distance where the net cumulative dispersion is zero Distance correspondingto the input of the transmission fiber Point@)of symmetry of a dispersion map Propagation distance List of Useful Relations PASE= 2ns,(G - 1)hu Au OSNR = 58 + Pi,- NF - Lsp - 1010gNmp OSNRR = Iones Q2Be l f r B o (1 -a2 Q = nones + ozeros - Leros 6. Pseudo-Linear Transmissionof High-speed TDM Signals 293 OSNRR = 58 +Prec- NF D=--p2 2KC A= 2Jrc 83 8 = 3 (%) ( : D + S ) 2 Ti LD = 1821 Lm=Leff 1 = YPP 1 - exp (-aoL) a0 TD Lw= DAv 294 RenCJean Essiambre et al. List of Acronyms AP APT ASE BER CDCF DC DCF DCM DDF DGD DSF EAM ETDM EDFA FEC FWHM FWM GNSE GVD HOM IFWM IMDD IP IS1 IXPM LEAF LHS LPG MZ NF NRZ NSE NZDSF NOLM OSNR OTDM PMD PMDC POAM PRBS RDS RZ Alternate polarization Adiabatic perturbation theory Amplified spontaneous emission Bit error rate Continuously dispersion-compensatingfiber Dispersion-compensated or dispersion compensation Dispersion-compensatingfiber Dispersion compensation module Dispersion-decreasingfiber Differential group delay Dispersion-shifted fiber Electro-absorptionmodulator Electrical time-division multiplexing Erbium-doped fiber amplifier Forward error correction Full width at half maximum Four-wave mixing Generalized nonlinear Schrodinger equation Group-velocity dispersion Higher-order mode Intrachannel four-wavemiXing Intensity-modulateddirect-detection Internet protocol Intersymbol interference Intrachannel cross-phase modulation Large effective area fiber Left-hand side Long-period grating Mach-Zehnder Noise figure Nonreturn-to-zero Nonlinear Schrodinger equation Nonzero dispersion-shifted fiber Nonlinear optical loop mirror Optical signal-to-noise ratio Optical time-divisionmultiplexing Polarization-modedispersion Polarization-modedispersion compensation Provisioning, operation, administration, and maintenance Pseudo-random bit sequence Ratio dispersion to slope Return-to-zero 6. Pseudo-Linear Transmission of High-speed TDM Signals 295 RHS SDH SNR SOA SONET SP SPM STD SVEA TDM TWRS TOD WDM WGR XPM Right-hand side Synchronous digital hierarchy Signal-to-noiseratio Semiconductor optical amplifier Synchronous optical network Single polarization Self-phasemodulation Standard unshifted fiber Slowly-varying-envelopeapproximation Time-division multiplexing TrueWavem reduced slope Third-order dispersion Wavelength-divisionmultiplexing Waveguide grating router Cross-phase modulation References [1] K. S. Kim and M. E. Lines, “Temperature Dependence of ChromaticDispersion in Dispersion-Shifted Fibers: Experiment and Analysis,” Appl. Phys. Left. 73, pp. 2069-2074 (1993). [2] K. Yonenaga, A. Hirano, S. Kuwahara, Y Miyamoto, H. Toba, K. Sato, and H. Miyazawa, “Temperature-independent 8O-Gbit/s OTDM Transmission Experiment Using Zero-Dispersion-Flattened Transmission Line,” Electron. Lett. 36,pp. 343-345 (2000). [3] D. Breuer and K. Petermann,“Comparison of NRZ- and RZ-Modulation Format for 4O-Gbls TDM Standard-Fiber Systems,” ZEEE Photon. Technol. Lett. 9, pp. 398-400 (1997). [4] D. Breuer, H. J. Ehrke, F. Kiippers, R. Ludwig, K. Petermann, H. G. Weber, and K. Weich, “Unrepeatered40-Gb/s RZ Single-ChannelTransmission at 1.55 pm Using Various Fiber Types,” ZEEE Photon. Technol. Lett. 10,pp. 822-824 (1998). [5] K. Ennser, R. I. Laming, and M. N. Zervas, “Analysis of 40-Gb/s TDMTransmissionover Embedded Standard Fiber EmployingChirped Fiber Grating Dispersion Compensators,” J. Lightwave Technol. 16,pp. 807-81 1 (1998). [6] R.-J. Essiambre, B. Mikkelsen, and G. Raybon, “Intrachannel Cross-Phase Modulation and Four-Wave Mixing in High-speed TDM Systems,” Electron. Left. 35,pp. 1576-1578 (1999). [7] P V. Mamyshev and N. A. Mamysheva, “Pulse-Overlapped Dispersion. Managed Data Transmission and Intrachannel Four-Wave Mixing,’’ Opt. Left. 24, pp. 1454-1456 (1999). [8] C. M. Weinert, R. Ludwig, W. Pieper, H. G. Weber, D. Breuer, K. Petermann, and E Kiippers, “40Gb/s and 4 x 40Gb/s TDM/WDM Standard Fiber Transmission,”J. Lightwave Technol. 17,pp. 2276-2284 (1999). 296 ReneJean Essiambre et al. [9] A. H. Gnauck, Sang-Gyu Park, J. M. Wiesenfeld, and L. D. Garrett, “Higbly Dispersed Pulses for 2 x 40 Gbit/s Transmission over 800 km of Conventional Single-Mode Fiber,” Electron. Lett. 35, pp. 2218-2219 (1999). [lo] I?. C. Becker, N. A. Olsson, and J. R. Simpson, Erbium-Doped Fiber Amplij?ers, Fundamentab and Technology (Academic Press, San Diego, 1999). [ll] I. Kaminow and T. Koch (Editors), Optical Fiber Telecommunications IIIB (Academic Press, San Diego, 1997), Chap. 2. [12] N. A. Olsson, “Lightwave Systems With Optical Amplifiers,” J. Lightwave Technol. 7 , pp. 1071-1082 (1989). [13] G. I Agrawal, Fiber-optic Communication Systems, 2nd Edition (AcademicPress, ? San Diego, 1997). [14] A. R. Chraplyvy, R. W. Tkach, and K. L. Walker, “Optical Fiber for Wavelength Division Multiplexing,” U.S. Patent 5,327,516, July 5, 1994. [15] H. Meyr, M. Moeneclaey, and S. A. Fechtel, Digital Communication Receivers: Synchronization, Channel Estimation, and Signal Processing (Wiley, New York, 1997). [16] G. P Agrawal, Nonlinear Fiber Optics, 3rd Edition (AcademicPress, San Diego, . 2001). [17] P. G. Drain and R. S. Johnson, Solitons: An Introduction (CambridgeUniversity Press, UK, 1989). [18] I. Kaminow and T. Koch (Editors), Optical Fiber Telecommunications IIIA (Academic Press, San Diego, 1997). [19] R.-J. Essiambre and G. I?. Agrawal, “Soliton Communication Systems” in P r o p s s in Optics, Vol. XXXVII (North Holland, Amsterdam, 1997). [20] E. Golovchenko, A. Pilippetskii, and N. S. Bergano, “Transmission Properties of Chirped Return-to-Zero Pulses and Nonlinear Intersymbol Interference in 10 Gb/s WDM Transmission,” OFC’OO, Baltimore, USA, paper FC3, pp. 38-40 (2000). [21] A. Pilippetskii, E. Golovchenko, and S. G. EvangelidesJr., “Nonlinear Interaction of Signal and Noise in 10-GblsLong-DistanceRZ Transmission,”OFC‘OO, Baltimore, USA, paper FC1, pp. 41-43 (2000). [22] T. C. Damen, P. B. Hansen, H. A. Haus, and R. H. Stolen, “Applications of Solitonsin Transmission Systems EmployingHigh Launch Powers,” U.S. Patent 5,737,460, April 7, 1998. [23] M. L. Dennis, T. E Carruthers, W. I. Kaechele, R. B. Jenkins, J. U. Kang, and I. N. Duling 111, “Long Span Repeaterless Transmission Using Adiabatic Solitons,” IEE Photon. TechnoZ.Lett. 11,pp. 478-480 (1999). [24] M. L. Dennis, W. I. Kaechele, L. Goldberg, T. E Carruthers, and I. N. Duling 1 1 1, “Wavelength-Division Multiplexing of Adiabatic Solitons for Repeaterless Transmission,” Electron. Lett. 35, pp. 1003-1004 (1999). [25] M. L. Dennis, W I. Kaechele, L. Goldberg, T. E Carruthers, and I. N. Duling 111, . “Wavelength-DivisionMultiplexed 4 x 10 Gb/s Adiabatic Soliton Transmission Over 235km,” IEEE Photon. Technol. Lett. 11,pp. 1680-1682 (1999). 6. Pseudo-Linear Transmission of High-speed TDM Signals 297 [26] K. Tajima, “Compensation of Soliton Broadening in Nonlinear Optical Fibers with Loss,” Opt.Le#. 12, pp. 54-56 (1987). [27] H. H. Kuehl, “Solitons on an axially nonuniform optical fiber,” J. Opt. SOC. Am. B. 5,pp. 709-713 (1988). [28] Q. Ren and H. Hsu, “Solitons in Dispersion-Compensated Optical Fiber with Loss,” IEEE J. Quantum Electron. 24, pp. 2059-2062 (1988). [29] V. A. Bogatyrev, M. M. Bubnov, E. M. Dianov, A. S. Kurkov, I? V. Mamyshev, A. M. Prokhorov, S. D. Rumyantsev, V. A. Semenov, A. A. Sysoliatin, S. V. Chernikov, A. N. Gur’yanov, G. G. Devyatykh, and S. I. Miroshnichenko, “A Single-Mode Fiber with Chromatic Dispersion Varying Along the Length,” J. Lightwave Technol. 9, pp. 561-566 (1991). [30] A. J. Stentz, R. W Boyd, and A. F. Evans, “Dramatically Improved Transmis. sion of Ultrashort Solitons Through 40 km of Dispersion-Decreasing Fiber,” Opt. Lett. 20, pp. 1770-1772 (1995). [31 V. A, Bogatyrjov, M. M. Bubnov, E. M. Diabov, and A. A. Sysoliatin, “Advanced 1 Fibres for Soliton Systems,” Pure Appl. Opt. 4, pp. 345-347 (1995). [32] D. J. Richardson, R. P. Chamberlin, L. Dong, and D. N. Payne, “High-Quality Soliton Loss-Compensation in 38-km Dispersion-Decreasing Fibre,” Electron. Lett. 31, pp. 1681-1682 (1995). [33] D. J. Richardson, L. Dong, R. I? Chamberlin, A. D. Ellis, T. Widdowson, and W A. Pender, ‘‘Periodically Amplified System Based on Loss Compensating . Dispersion-Decreasing Fibre,” Electron. Lett. 32, pp. 373-374 (1996). [34] J. D. Moores, W S. Wong, and H. A. Haus, “Stability and Timing Maintenance in Soliton Transmission and Storage Rings,” Opt. Commun. 113, pp. 153-175 (1994). [35] D.-M. Baboiu, D. Mihalache, and N.-C. Panoiu, “Combined InfluenceofAmplifier Noise and Intrapulse Raman Scattering on the Bit Rate Limit of Optical Fiber Communication Systems,” Opt.Le#. 20, pp. 2282-2284 (1995). [36] R.-J. Essiambre and G. P.Agrawal, “Timing Jitter Analysis for Optical Communication Systems Using Ultrashort Solitons and Dispersion-DecreasingFibers,” Opt. Commun. 131, pp. 274-278 (1996). [37l R.-J. Essiambre and G. I? Agrawal, “Timing Jitter of Ultrashort Solitons in High-speed Communication Systems. I. General Formulation and Application to Dispersion-Decreasing Fibers,” J. Opt.SOC. B 14, pp. 314-322 (1997). Am. 1381 T. Morioka, S. Kawanishi, H. Takara, and 0. Kamatani, “Penalty-Free, 100Gbith Optical Transmission of t 2 ps Supercontinuum Transform-Limited Pulses over 40km,” Electron. Lett. 31, pp, 124-125 (1995). [39] T. Morioka, S. Kawanishi, H. Takara, 0. Kamatani, M. Yamada, T. Kanamori, K. Uchiyama, and M. Saruwatari, “lOOGbit/s x 4ch, l 0 O h Repeaterless TDM-WDM Transmission Using a Single Supercontinuum Source,” Electron. Le#. 32, pp. 468-470 (1996). [40] S. Kawanishi, H. Takara, 0.Kamatani, T. Morioka, and M. Saruwatari, “100 Gbit/s, 560 km Optical Transmission Experiment with 80-km Amplifier 298 Renuean Essiambre et al. Spacing Employing Dispersion Management,” Electron. Lett. 32, pp. 470-471 (1996). [41] T. Morioka, H. Takara, S. Kawanishi, 0. Kamatani, K. Takiguchi, K. Uchiyama, M. Saruwatari, H. Takahashi, M. Yamada, T. Kanamori, and H. Ono, “1 Tbitls (100 Gbit/s x 10 channel) OTDMWDM Transmission Using a Single SupercontinuumWDM Source,” Electron. Lett. 32, pp. 906-907 (1996). [42] S. Kawanishi, H. Takara, K. Uchiyama, I. Shake, 0. Kamatani, and M. Saruwatari, “ 1.4 Tbitls (200 Gbit/s x 7 ch) 50 km Optical Transmission Experiment,” Electron. Lett. 33, pp. 1716-1717 (1997). [43] J. Hansryd, B. Bakhshi, B. E. Olsson, P. A. Andrekson, J. Brentel, and E. Kolltveit, “80 Gbitls Single Wavelength Soliton Transmission over 172km Installed Fibre,” Electron. Lett. 35, pp. 313-315 (1999). [44] S. Kawanishi, H. Takara, K. Uchiyama, I. Shake, and K. Mori, ‘‘3TbitIs (160 Gbit/s x 19 channel) Optical TDM and WDM TransmissionExperiment,” Electron. Lett. 35, pp. 826827 (1999). [45] A. R. Chraplyvy, “Limitations on Lightwave Communications Imposed by Optical-Fiber Nonlinearities,” J. Lightwave Technol. 8, pp. 1548-1 557 (1990). [46] R. W Tkach, A. R. Chraplyvy, E Forghieri, A. M. Gnauck, and R. M. Derosier, . “Four-Photon Mixing and High-speed WDM Systems,” J. Lightwave Technol. 13, pp. 841-849 (1995). [47l D. Marcuse and C. Menyuk, “Simulation of Single-ChannelOptical Systems at 100Gb/s,” J. Lightwave Technol. 17, pp. 564-569 (1999). [48] S. K. Turitsyn, M. P. Fedoruk, and A. Gornakova, “Reduced-Power Optical Solitons in Fiber Lines with Short-Scale Dispersion Management,” Opt. Lett. 24, pp. 869-871 (1999). [49] H. Anis, G. Berkey, G. Bordogna, M. Cavallari, B. Charbonnier, A. Evans, I. Hardcastle, M. Jones, G. Pettitt, B. Shaw, V Srikant, and J. Wakefield, . “Continuous Dispersion-ManagedFiber for Very High Speed Soliton Systems,” ECOC’99, Nice, France, pp. 230-231 (1999). [SO] M. Zitelli, E Matera, and M. Settembre, “40-GbitlsTransmissionin DispersionManagementLinks with Step-IndexFiber and Linear Compensation,” Opt. Lett. 24, pp. 1169-1171 (1999). [Sl] M. Zitelli, F. Matera, and M. Settembre, “Single-Channel Transmission in Dispersion Management Links in Conditions of Very Strong Pulse Broadening: Application to 4O-Gb/s Signals on Step-IndexFibers,” J. Lightwave Technol. 17, pp. 2498-2505 (1999). [52] M. Matsumoto, “Analysis of Interaction Between Stretched Pulses Propagating in Dispersion-Managed Fibers,” IEEE Photon. Technol. Lett. 10, pp. 373-375 (1998). [53] S. Kumar, M. Wald, E Lederer, and A. Hasegawa, “Soliton Interaction in Strongly Dispersion-Managed Optical Fibers,” Opt. Lett. 23, pp. 1019-1021 (1998). 6. Pseudo-Linear Transmission of High-speed TDM Signals 299 I 4 T. Inoue, H. Sugahara, A. Maruta, and U. Kodama, “Interactions Between 51 Dispersion Managed Solitons in Optical-Time-Division-MultiplexedSystem,” ZEEE Photon. Technol.Lett. 12, pp. 299-301 (2000). [55] A. Mecozzi, C. B. Clausen, and M. Shtaif, “Analysis of Intrachannel Nonlinear Effects in Highly Dispersed Optical Pulse Transmission,”IEEE Photon. Technol. Lett. 12, pp. 392-394 (2000). 1561 A. Mecozzi, C. B. Clausen, and M. Shtaif, “System Impact of IntraChannel Nonlinear Effects in Highly Dispersed Optical Pulse Transmission,” IEEE Photon. Technol.Lett. 12, pp. 1633-1635 (2000). [571 T. Inoue and A. Maruta, “Prespread RZ Pulse Transmissionfor Reducing Intrachannel Nonlinear Interactions,” LEOS’OO,Puerto Rico, USA, Paper MJ3 (2000). [58] F. Merlaud and S. K. Turitsyn, “Intrachannel Four-Wave Mixing and Ghost Pulses ‘Generation:Time Domain Approach,” ECOC’OO, Munich, Germany, Paper 7.2.4, pp. 39-40 (2000). [59] M. J. Ablowitz and T. Hirooka, “Resonant Nonlinear Intrachannel Interactions in Strongly Dispesion-Managed Transmission Systems,” Opt. Lett. 25, pp. 1750-1752 (2000). [60] J. MBrtensson, A. Berntson, M. Westlund, A. Danielsson, P Johannisson, . D. Anderson, and M. Lisak, “Timing Jitter Owing to Intrachannel Pulse Interactions in Dispersion-Managed Transmission Systems,” Opt. Lett. 26, pp. 55-57 (2001). [61] A. Mecozzi, C. B. Clausen, M. Shtaif, S.-G. Park, and A. H. Gnauck, “Cancellation of Timing and AmplitudeJitter in SymmetricLinks Using Highly Dispersed Pulses,” ZEEE Photon. Technol. Left. 13, pp. 441-447 (2001). [62J H. Goldstein, ClassicalMechanics, 2nd Edition (Addison Wesley, Massachusetts, 1980). [63] W. J. Firth, “Propagation of Laser Beams Through Inhomogeneous Media,” Upt. Cornmun. 22, pp. 226-320 (1977). [ D. ] & Anderson, “VariationalApproach to Nonlinear Pulse Propagation in Optical I Fibers,” Phys. Rw. A 27, pp. 313.5-3145 (1983). [65l D. Anderson, M. Lis& and T. Reichel, “Approximate Analytical Approaches to Nonlinear Pulse Propagation in Optical Fibers: a Comparison,” Phys. Rev.A. 38, pp. 1618-1620 (1988). [66] M. Desaix, D. Anderson, and M. Lisak, “Accuracy of an Approximate Variational Solution Procedure for the Nonlinear Schrodinger Equation,” Phys. Rw. A. 40, pp. 2441-2445 (1989). [67] M. K. Jackson, G. R. Boyer, J. Paye, M. A. Franco, and A. Mysyrowicz, “Temporal Diffraction by Nonlinear Interaction in Optical Fibers,” Opt. Len. 17, pp. 1770-1772 (1992). [68] Y. R. Shen, The Principles of Nonlinear Optics (Wiley, New York, 1984), Chaps. 21.3 and 21.4. 300 Renk-Jean Essiambre et al. [69] I. Shake, H. Takara, K. Mori, S. Kawanishi, and Y. Yamabayashi, “Influence of Inter-Bit Four-Wave Mixing in Optical TDM Transmission,”Electron. Lett. 34, pp. 1600-1601 (1998). [70] T. Inoue and A. Maruta, “Reduction of Intra-Channel Four-Wave Mixing in StronglyDispersion-ManagedLine for High Speed OTDM system,”NLGWO1, Clearwater, USA, Paper MC59, pp. 211-213 (2001). [71] P. Johannisson, D. Anderson, A. Berntson, and J. Mfirtensson, “Perturbational Analysis of the Generation of Ghost Pulses in 40-Gbids Fibre-Optic Communication Systems,” NLGWO1, Clearwater, USA, paper MC74, pp. 257-259 (2001). [72] R.-J. Essiambre, E. Mikkelsen, and G. Raybon, “Modulation format with low sensitivity to fiber nonlinearity,” U.S. Patent Application, IDS No. 117529 July 1999). [73] J. Mhtensson, A. Berntson, and M. Westlund, “Intra-Channel Pulse Interactions in 40-Gbids Dispersion-Managed RZ Transmission system,” Electron. Lett. 36, pp. 244246 (2000). [74] R. I. Killey, H. J. Thiele, V. Mikhailov, and P Bayvel, “Reduction of Intrachan. ne1 Nonlinear Distortion in 40-Gbh-Based WDM Transmission over Standard Fiber,” IEEE Photon. Technol. Lett. 12, pp. 1624-1626 (2000). [75] D. Mamse, A. R. Chraplyvy, and R. W. Tkach, “Effect of Fiber Nonlinearity on Long-DistanceTransmission,”J. Lightwave Technol. 9, pp. 121-128 (1991). [76] B. Zhu, L. Leng, L. E. Nelson, Y. Qian, S. Stulz, C. Doerr, L. Stulz, S. Chandrasekar, S. Radic, D. Vengsarkar, Z. Chen, J. Park, K. Feder, H. Thiele, J. Bromage, L. Gruner-Nielsen, and S. Knudsen, “3.08 Tb/s (77 x 42.7 Gb/s) Transmission over 1200km of Nonzero Dispersion-Shifted Fiber with 100km Spans Using C- and L-band Distributed Raman Amplification: OFC‘O1, Anaheim, USA, paper PD-23, pp. 23-1-23-3 (2001). [77] Sang-Gyu Park, A. H. Gnauck, J. M. Wiesenfeld, and L. D. Garrett, “4O-Gb/s Transmission Over Multiple 120-km Spans of Conventional Single-ModeFiber Using Highly Dispersed Pulses,” IEEE Photon. Technol. Left. 12, pp. 1085-1087 (2000). [78] S. Ramachandran, B. Mikkelsen, L. C. Cowsar, M. E Yan, G. Raybon, L. Boivin, M. Fishteyn, W. A. Reed, P. Wisk, D. Brownlow, R. G. Huff, and L. Gruner-Nielsen, “All-Fiber Grating-Based Higher Order Mode Dispersion Compensator for Broad-Band Compensation and 1000-km Transmission at 40 Gb/s,” IEEE Photon. Technol. Lett. 13, pp. 632-634 (2001). [79] S. Ramachandran, G. Raybon, B. Mikkelsen, M. Yan, and L. Cowsar, “1700-km Transmission at 40 Gb/s with 100-km Amplifier-SpacingEnabled by Higher-Order-Mode Dispersion-Compensation,” ECOC‘O1 , Amsterdam, The Netherlands, paper We.F.2.2, pp. 282-283 (2001). [SO] E Matera, M. Settembre, M. Tamburrini, E Favre, D. Le Guen, T. Georges, M. Henry, G. Michaud, F? Franco, A. Schiffini, M. Romagnoli, M. Gugliehwci, and S. Cascelli, “Field Demonstration of 40-Gb/s Soliton Transmission with Alternate Polarizations,” J. Lightwave Technol. 17,pp. 2225-2234 (1999). 6. Pseudo-Linear Transmission of High-speed TDM Signals 301 [81] I. Morita, K. Tanaka, N. Edagawa, and M. Suzuki, “40-Gb/s Single-Channel Soliton Transmissioii over Transoceanic Distances by Reducing Gordon-Haus Timing Jitter and Soliton-Soliton Interaction,” J. Lightwave Technol. 17, pp. 2506-2511 (1999). [82] R. W. Boyd, Nonlinear Optics (Academic Press, San Diego, 1992). I831 K.-I. Suzuki, N. Ohkawa, M. Murakami, and K. Aida, “Unrepeatered40-Gbitls RZ SignalTransmission over 240-km Conventional SinglemodeFibre,” Electron. Lett. 34, pp. 799-800 (1998). [84] T. Koonen, H. Waardt, J. Jennen, J. Verhoosel, D. Kant, M. de Vos, A. van Ardenne, E. J. van Velduizen, “A Very High Capacity Optical Fibre Network for Large-ScaleAntenna Constellations: The RETINA Project,” European Conference on Networks and Optical Communications (NOC) Ipswitch, U.K.: A. Lord, D. W. Faulkner and D. W Smith (Eds.), 10s Press, pp. 165-172 (2001). . [85] S. Kawanishi, Y Miyamoto, H. Takara, M. Yoneyama, K. Uchiyama, I. Shake, and Y Yamabayashi, “120-Gbitls OTDM System Prototype,” ECOC’98, Madrid, Spain, Vol. 3, paper PD, pp. 41-46 (1998). [86] E. Mikkelsen, G. Raybon, R.-J. Essiambre, K. Dreyer, Y Su, L. E. Nelson, J. E. Johnson, G. Shtengel, A. Bond, D. G. Moodie, and A. D. Ellis, “160-Gbit/s Single-Channel Transmission over 300-km Nonzero-Dispersion Fiber with Semiconductor-BasedTransmitterand Demultiplexer,” ECOC’99, Nice, France, paper PD 2-3, pp. 28-29 (1999). [87] A. D. Ellis, J. K. Lucek, D. Pitcher, D. G. Moodie, and D. Cotter, “Full 10 x 10 Gbit/s OTDM Data Generation and Demultiplexing Using Electroabsorption Modulators,” Electron. Lett. 34, pp. 17661767 (1998). I881 R. S. Tucker, G. Eisenstein, and S. K. Korotky, “Optical Time-Division Multiplexing for Very High Bit-Rate Transmission,” J. Lightwave Technol. 6 , pp. 1737-1749 (1988). [89] G. Raybon, I? B. Hansen, R. C. Alferness, L. L. Buhl, U. Koren, B. I. Miller, M. G. Young, T. L. Koch, J.-M. Verdiell, and C. A. Burrus, “Wavelength-Tunable Actively Mode-Locked Monolithic Laser with an Integrated Vertical Coupler Filter,” Opt. Lett. 18, pp. 1335-1336 (1993). [90] R. Ludwig, S. Diez, A. Ehrhardt, L. Kller, W. Pieper. and H. G. Weber, “A Tunable Femtosecond Mode Locked Semiconductor Laser for Applications in OTDM Systems,” ZEZCE Trans.Electz E81, pp. 140-145 (1998). [91] M. C. Wu, Y K. Chen, T. Tanbun-Ek, R. A. Logan, M. A. Chin, and G. Raybon, “Transform-Limited 1.4-ps Optical Pulses from a Monolithic Colliding-Pulse Mode-Locked Quantum Well Laser,” Appl. Phjx Lett. 57, pp. 759-761 (1990). [92] T. F. Carruthers and I. N. Duling, “IO-GHz, 1.3-ps Erbium Fiber Laser Employing Soliton Pulse Shortening,” Opt. Lett. 21, pp. 1927-1929 (1996). [93] M. Nakazawa, E. Yoshida, and K. Tmura, “IO-GHz, 2-ps &generatively and Harmonically FM Mode Locked Erbium Fiber Ring Laser,” Electron. Lett. 32, pp. 1285-1286 (1996). 302 Renk-Jean Essiambre et al. [94] M. Vaa, B. Mikkelsen, K. S. Jepsen, K. E. Stubkjaer, M. Schilling, K. Daub, E. Lach, G. Laube, W. Idler, K. Wunstel, S Bouchoule, C. Kazmierski, and . D. Mathoorasing, “A Bit-Rate Flexible and Power Efficient All-Optical Demultiplexer Realised By Monolithically Integrated Michelson Interferometer,” ECOC‘96, Oslo, Norway, paper PD-ThB3.3, pp. 11-14 (1996). [95] E Devaux, N. Souli, A. Ougazzaden, F. Huet, and M. Carre, “High Speed Tandem of MQW Modulators for Coded Pulse Generation with 14-dB Fiber to Fiber Gain:’ IEEE Photon. Technol. Lett. 8, pp. 218-219 (1996). [96] M. Suzuki, H. Tanaka, K. Utaka, N. Edagawa, and Y Matsushima, “Transform-Limited 14-ps Optical Pulse Generation with 15GHz Repetition Rate by InGaAsP Electroabsorption Modulator,” Electron. Left. 28, pp. 1007-1008 (1992). [97] D. G. Moodie, P. J. Cannard, A. J. Dann, D. D. Marcenac, C. W. Ford, J. Reed, R. T. Moore, J. K. Lucek, and A. D. Ellis, “Low Polarisation Sensitivity Electroabsorption Modulators for 160-Gbitls Networks,” Electron. Lett. 33, pp. 2068-2070 (1997). [98] G. Raybon, B. Mikkelsen, R.-J. Essiambre, J. E. Johnson, K. Dreyer, and L. E. Nelson, “100-Gbitls Single-channel Transmission over 200-km TrueWave and 160-km Conventional Fiber with EAM Source and Demultiplexer,” ECOC‘99, Nice, France, paper WeC2.5, pp. 92-93 (1999). [99] J. J. Veselka, S. K. Korotky, P. V. Mamyshev, A. H. Gnauck, G. Raybon, and N. M. Froberg, “A Soliton Transmitter Using a CW Laser and an NRZ Driven Mach-Zehnder Modulator,” ZEEE Photon. Technol. Lett. 8, pp. 950-952 (1996). [loo] B. Mikkelsen, G. Raybon, R.-J. Essiambre, J. E. Johnson, K. Dreyer, and L. E Nelson, “Unrepeatered Transmission over 150k m of Nonzero-Dispersion Fibre at 100Gbitls with Semiconductor-BasedPulse Source, Demultiplexer,and Clock Recovery,” Elechon. Lett. 35, pp. 1866-1868 (1999). [loll I. Y Khrushchev, J. D. Bainbridge, J. E. A. Whiteaway, I. H. White, and R. V. Penty, “Multiwavelength Pulse Source for OTDMMDM Applications Based on Arrayed Waveguide Grating,” ZEEE Photon. Technol. Lett. 11, pp. 1659-1661 (1999). [lo21 D. D. Marcenac, A. D. Ellis, and D. G. Moodie, ‘%O-Gbit/sOTDM Using Electroabsorption Modulators,” Electron. Lett. 34, pp. 101-103 (1998). [lo31 M. L. Dennis et al., “Eight-to-One Demultiplexing of 100-Gbitls TDM Data Using LiNbOs Sagnac Interferometer Modulators,” OFC‘98, San Jose, CA, USA, pp. 110-112 (1998). [lo41 S. Diez, R. Ludwig, and H. G. Weber, “Gain-Transparent SOA-Switch for High-Bitrate OTDM Add/Drop Multiplexing,” ZEEE Photon. Technol. Lett. 11, pp. 60-62 (1999). [I051 M. Vaa, B. Mikkelsen, K. S. Jepsen, K. E. Stubkjaer, R. Hess, M. Duelk, W. Vogt, E. Gamper, E. Gini, P. A. Besse, H. Melchior, and S. R. Bouchoule, “Bit Error Rate Assessment of 80-Gbls All-Optical Demultiplexingby a Monolithically Integrated Mach-Zehnder Interferometer w t SemiconductorOptical ih Amplifiers,” ECOC‘97, Edinburgh, Scotland, Vol. 3, pp. 31-34 (1997). 6. Pseudo-Linear Transmission of High-speed TDM Signals 303 [lo61 S. Nakamura, Y Ueno, K. Tajima, J. Sasaki, T. Sugimoto, T. Kato, T. Shimoda, M. Itoh, H. Hatakeyama, T. Tamanuki, and T. Sasaki, “Demultiplexing of 168-GblsData Pulses w t a Hybrid-Integrated SymmetricMach-Zehnder Allih Optical Switch,” IEEE Photon. Echnol. Lett. 12, pp. 425427 (2000). [lo71 B. E. Olsson and P.A. Andrekson, “Polarization-IndependentDemultiplexingin a Polarization Diversity Nonlinear Optical Loop Mirror,” IEEE Photon. Techno!. Lett. 9, pp. 764-766 (1997). [lo81 T. Yamamoto, E. Yoshida, and M. Nakazawa, “Ultrafast Nonlinear Optical Loop Mirror for Demultiplexing 64O-Gbit/s TDM Signals,” Electron. Lett. 34, pp. 1013-1014 (1998). [lo91 R. Ludwig and G. Raybon, “All-Optical Demultiplexing Using Ultrafast FourWave-Mixing in a Semiconductor Laser Amplifier,” ECOC’93, Montreux, Switzerland, paper ThP12.2, pp. 57-60 (1993). [110] S. Kawanishi, K. Okamoto, M. Ishii, 0. Kamatani, H. Takara, and K. Uchiyama, “All-Optical Time-Division Multiplexing of lOO-Gbit/s signal based on Four-Wave-Mixing in a Traveling Wave Semiconductor Optical Amplifier,” Electron. Lett. 33, pp. 976-977 (1997). [l 1I] N. Souli, E Devaux, A. Ramdane, Ph. Kraus, A. Ougazzaden, F Huet, M. Carre, . Y S o d , J. E Kerdiles, M. Henry, G. Aubin, E. Jeanney, T. Montallan, J. Moulu, B. Nortier, and J. B. Thomine, “20-Gbit/s High-Performance Integrated MQW TANDEM Modulators and Amplifier for Soliton Generation and Coding,” IEEE Photon. Technol.Lett. 7 , pp. 629-631 (1995). [112] A. Ougazzaden, C. W Lentz, T. G. B. Mason, K. G. Glogovsky, C. L. Reynolds, . G. J. Przybylek, R. E. Leibenguth, T. L. Kercher, J. W Boardman, M. T. Rader, . J. M. Geary, F. S. Walters, L. J. Peticolas, J. M. Freund, S. N. G. Chu, A. Sirenko, R. J Jurchenko, M. S. Hybertsen, L. J. P. Ketelsen, and G. Raybon, “40-Gb/s . Tandem Electroabsorption Modulator,” OFC‘O1, Anaheim, CA, USA, paper PD-14 (2001). [113] C . Bornholdt, B. Sartorius, S. Schelbase, M. Mohrle, and S. Bauer, “SelfPulsating DFB Laser for All-Optical Clock Recovery at 40Gbit/s,” Electron. Lett. 36, pp. 327-328 (2000). [114] S. Kawanishi, T. Morioka, 0. Kamatani, H. Takara, and M. Saruwatari, “100Gbit/s, 200-km Optical Transmission Experiment Using Extremely Low Jitter PLL Timing Extraction and All-Optical Demultiplexing Based on Polarisation Insensitive Four-Wave Mixing,” Electron. Lett. 30, pp. 800-801 (1994). [115] D. T. K. Tong, Kung-Li Deng, B. Mikkelsen, G. Raybon, K. E Dreyer, and J. E. Johnson, “160-Gbith Clock Recovery Using ElectroabsorptionModulatorBased Phase-Locked Loop,” Electron. Lett. 36, pp. 1951-1952 (2000). [l lq E Cisternino, R. Girardi, S. Romisch, R. Calvani, E. Riccardi, and F Garino, ! “A Novel Approach to Prescaled Clock Recovery in OTDM Systems,” ECOC 1998, Vol. 1, Madrid, Spain, pp. 477-478 (1998). [117] I. D. Phillips, A. D. Ellis, T. Widdowson, D. Nesset, A. E. Kelly, and D. Trommer, “lOO-Gbit/s Optical Clock Recovery Using Electrical Phaselocked Loop 304 RenMean Essiambre et al. Consisting of Commercially Available Components,” Electron. Lett. 36, pp. 65M52 (2000). P. V.Mamyshev, “All-OpticalData Regeneration Based on Self-phaseModulation Effect,” ECOC‘98, Madrid, Spain, pp. 475476 (1998). B. Mikkelsen, G. Raybon, R.-J. Essiambre, A. J. Stentz, T. N. Nielsen, D. W Peckham, L. Hsu, L. Gruner-Nielsen, K. Dreyer, and J. E. Johnson, . “320-Gbls Single-ChannelPseudolinearTransmission over 200 km of NonzeroDispersion Fiber,” IEEE Photon. Technol.L e t . 12, pp. 1400-1402 (2000). K. L. Deng, D. T. K. Tong, C. K. Chan, K. Dreyer, and J. E. Johnson, “Rapidly Reconfigurable Optical Channel Selector Using RF Digital Phase Shifter for Ultra-Fast OTDM Networks,” Electron. Lett. 36,pp. 1724-1725 ( 0 0 . 20) G. Raybon, B. Mikkelsen, B. Zhu, R.-J. Essiambre, S. Stulz, A. Stentz, and L. Nelson, “160-Gbls TDM Transmission Over Record Length of 400km (4 x 100km) Using Distributed Raman Amplification Only,” OAA’OO, Qukbec City, Qukbec, Canada, PD-1, pp. 1-3 (2000). L. Griiner-Nielsen, S. N. Knudsen, B. Edvold, P. Kristensen, T. Veng, and D. Magnussen, “Dispersion Compensating Fibers and Perspectives for the Future,” ECOC‘OO, Munich, Germany, Vol. 1, paper 2.4.1, pp. 91-94 (2000). H. Bulow, “PMD Mitigation Techniques and Their Effectiveness in Installed Fiber,” OFC‘OO, Baltimore, MD, USA, Vol. 3, paper ThH1-1, pp. 110-112 (2000). B. J. Eggleton, B. Mikkelsen, G. Raybon, A. Ahuja, J. A. Rogers, P. S. Westbrook, T. N. Nielsen, S. Stulz, and K. Dreyer, “Tunable Dispersion Compensation in a 160-GblsTDM System by a Voltage Controlled Chirped Fiber Bragg Grating,” IEEE Photon. TechnoZ. Lett. 12, pp. 1022-1024 (2000). B. Mikkelsen, G. Raybon, B. Zhu, R.-J. Essiambre, P. G. Bernasconi, K. Dreyer, L. W Stulz, and S. N. Knudsen, “High Spectral Efficiency [0.53bit/s/Hz] . WDM Transmission of 160Gb/s Per Wavelength over 400 km of fiber,” OFC‘O1, Anaheim CA, USA, paper THFZl(2001). [118] [119] [120] [121] [122] [123] [124] [125] Chapter 7 Dispersion-Managed Solitons and Chirped Return to Zero: What Is the Difference? Curtis R. Menyuk Computer Science and Electrical Engineering Department, Universityof Maryland Baltimore County, Baltimore,Maryland and Photon& Corporation,Maynard, Massachusetts Gary M. Carter ComputerScience and Electrical EngineeringDepartment, Universityo Maryland Baltimore f Coun& Baltimore,Maryland and Laboratoryfor Physical Sciences, College Park, Maryland William L. Kath ComputerScience and Electrical Engineering Department, Universityo h4atyhnd Baltimore f County, Baltimore, Maryland and Applied MathematicsDepartment, Northwestern University, Evanston. Illinois Ruo-Mei Mu QCO Telecommunications, Eatontown, New Jersey I. Historical Overview Once upon a time, a long time ago-at least as time is counted in the telecommunications industry-there was a place called AT&T Bell Laboratories. At this remarkable place, it was possible for great scientists to do long-term research-research that might take many years to affect the development of telecommunications products. In 1973, Akira Hasegawa, who was then at AT&T Bell Laboratories,published a paper with Fred Tappert that contained three important contributions [l]. First, it showed that in an idealized optical fiber, with just second-order dispersion and the Kerr nonlinearity, the wave envelope of the light obeys the nonlinear Schrodinger equation, which may be written: i a u pii a2u 2 - -- + ylUl U = 0, az 2 at2 where U is the complex wave envelope,z is the distance along the optical fiber, is the second-order dispersion coefficient, and y is the nonlinear coefficient due to the Kerr effect. The quantity t = t - z/vg is referred to as retarded time, where t is physical time and vg is the group velocity. Second, this paper 305 OPTICAL FIBER TELECOMMlJKICATIONS VOLUME IVB Copyright 0 2002, Elrevier Science (USA). All rights of reproduction in any form reserved. ISBN 0-12-395173-9 306 Curtis R Menyuk et a. l showed that when p” c 0, Eq. 7.1 has a soliton solution, where T is an arbitrary parameter indicating the soliton pulse duration. Third, they showed that Eq. 7.1 can be solved using the split-step Fourier transform method. Equation 7.1 neglects the effects of higher-order dispersion, the Raman and Brillouin nonlinearities, device impairments along the transmission line, amplified spontaneous emission (ASE) noise, and fiber birefringence. However, in the succeeding quarter century, a large number of scientists have contributed to amend Eq. 7.1 to include the missing effects [2]. To this day, almost all modeling of optical fiber transmission is based on Eq. 7.1, its amendments, and its reductions. Moreover, almost al numerical solutions of Eq. 7.1 l and its amendments are based on the split-step method. At the time that Eq. 7.1 was first written down, the notion of using solitons in communications systems seemed farfetched. Optical fiber losses were still too large in the anomalous dispersion regime where p” < 0 for nonlinear propagation to be practical, and no sources that could produce the needed powers at these long wavelengths, h > 1.3 pm, existed [2]. By 1980, however, fibers with losses almost as low as 0.2 dB/km at h = 1.5 pm had been fabricated, and color center lasers that could produce short, high-intensity pulses at 1.5 w m had been perfected. In that year, a scientist at AT&T Bell Laboratories named Linn Mollenauer, along with his colleagues Roger Stolen and Jim Gordon, demonstrated that it was possible to transmit solitons in optical fibers [3]. While a large number of scientists have contributed to the development of optical fiber solitons since 1980, Hasegawa and Mollenauer have been tireless advocates of their use in optical fiber communications systems. The interest in using solitons stems from their remarkable propertyapparent in Eq. 7.2-that they do not spread in the time domain due to dispersion or in the frequency domain due to nonlinearity. A single soliton pulse in an ideal, lossless fiber is completely stable. Moreover, solitonsthat are in different wavelength channels do not change their shape after a collision. The central timesjust shift slightly [2,4]. At an early stage, it was thought that it might be possible to actually use the nonlinearity to compensate for dispersion [5]; however, it soon became apparent that it is more fruitful to view the dispersion as compensating for the nonlinearity. Any communications system is impaired by noise in the transmission line and the receiver. Optical fiber transmission systems are no exception. To ensure a reasonable signal-to-noise level, some nonlinearity is inevitable. However, it was already apparent in the early 1980s that the nonlinearity was too small to matter in the communications systems of the day in which the signal was typically regenerated every 40 km or less. Thus, soliton advocates began in the early 1980s to explore the possibility of replacing repeaters with amplifiers. Raman amplification 7. Dispersion-ManagedSolitons and Chirped Return to Zero 307 seemed particularly promising [6],and its feasibility was demonstrated in one of the earliest recirculating loop experiments with optical amplification [7]. However, the Raman effect was deemed too inefficient to be practical when amplifying 1 channel or as many as 8 channels-the maximum then being seriously discussed. The first practical optical amplifier was the erbium-doped fiber amplifier (EDFA), and it is still the dominant amplifier technology [8]. By 1991, it was apparent that they would be widely used in long-haul optical fiber communications systems. Systems based on EDFAs were far cheaper than systems based on optical regenerators. Moreover, they offered the immediate prospect of upgrading the base data rate from 622 Mbitsh (OC-12) to 2.5 Gbitsh (OC48), as well as the longer-term prospect of taking advantage of the broad gain bandwidth of erbium to transmit many wavelengths at once. Amplifiers remove the effect of attenuation, but they allow the other optical impairments to accumulate. At this point, systems designers were confronted by the necessity of working with systems in which the Kerr nonlinearity and chromatic dispersion interacted strongly enough to impact the pulse propagation and could seriously degrade the bit error rate (BER) if improperly handled [9, 101. Historically, long-haul optical fiber communicationssystems designershave almost always used a nonreturn to zero (NRZ) format, and it remains the predominant modulation format as of this writing. This format is also referred to as IM-DD (intensity modulation-direct detection) or OOK (on-off keying). As shown in Fig. 7.1, a mark in this format amounts to filling the bit window almost uniformly with optical energy, whereas a space amounts to putting as little energy as possible in the bit window. At an early stage, when optical NRZ Solitons Fig. 7.1 Schematic comparison of the nonreturn to zero (NRZ) and soliton formats. The boundaries of the bit windows are shown as dashed lines. The optical energy of the marks (1s) in the NRZ format fills the bit window evenly, whereas it is concentrated in the center of the bit window in the soliton format. 308 Curtis R Menyuk et al. . communications systems used multimode fibers and light emitting diode (LED) transmitters, this format became dominant because it required s h pler device technology than any of the alternatives [ll]. Later, when laser diodes replaced LEDs, the NRZ format also had the advantage of a smaller bandwidth than a return-to-zero (RZ) format at the same data rate [12]. Prior to the deployment of EDFAs, transmission impairments were dominated by attenuation and dispersion; therefore, it was important to reduce the impact of dispersion. As of 1990, practical NRZ sources based on distributed feedback lasers were available [13]. By contrast, solitons were generated by color-center lasers, which were too expensive and bulky for use in systems [14]. Moreover, receivers were optimized for the NRZ format [131. The challenge posed by nonlinearity to the NRZ format-particularly in transoceanic systems where signals are required to propagate all-opticallyfor thousands of kilometers-was quite serious. Unless the fiber’s dispersion is close to zero, the NRZ pulses spread unacceptably, leading to a large intersymbol interference at the receiver. If, however, the dispersion is zero, the ASE noise from the amplifier is pumped through a resonant four-wave mixing interaction and grows exponentially, leading to severe distortion of the signal [15]. The solution to this problem in undersea systems was dispersion management. In one example of this approach, one alternates sections of standard fiber that have a dispersion of 17ps/nm-km at 1.55pm with sections of dispersion-shifted fiber that have a dispersion of -2 ps/nm-km at 1.55 p,m [16]. Doing so, one keeps the average dispersion close to zero, which minimizes the spreading, while mitigating the resonant four-wave mixing. Dispersion management is not as critical in terrestrial systems, but as the single-channel rates have increased to 10 Gbits/s and distances between repeaters have increased, it has become increasingly important. A typical dispersion-managed configuration in a terrestrial setting alternates sections of standard fiber that have a dispersion of 17ps/nm-km at 1.55p,m and are already in the ground with sections of higher-dispersion fiber, at perhaps -85 ps/nm-km, that are placed in the amplifier huts or are part of the amplifier design [17,18]. After the invention of the EDFA, the next great step forward was the development of wavelength-division multiplexing (WDM) systems in which many wavelengths are transmitted at once. The earliest experiments came before 1990, and by 1993 there had been several important field trials [19]. As early as 1993, Chraplyvy et al. [20]had demonstrated the transmission of 8 channels in the NRZ format over distances that were relevant for terrestrial communications at 10 Gbits/s per channel. By 1996, Mollenauer et al. [21] had demonstrated the transmission of 6 and 7 channels of solitons using sliding-guiding filters over transoceanic distances at 10Gbitds per channel. Also by 1996, Bergano and Davidson [22] had demonstrated the transmission of 20 channels in the NRZ format, using bit-synchronous polarization and phase modulation, over transoceanic distances at 5Gbitds. At around this time, AT&T began to massively deploy WDM transmission systems in their networks. 7. Dispersion-Managed Solitons and Chirped Return to Zero 309 In 1996, the situation stood as follows: Soliton sources had finally been invented that used laser diodes and LiNbOs modulators, much like in the NRZ systems [23]. Testbed experiments had demonstrated the feasibility of long-haul WDM transmission at 10 Gbits/s using both the NRZ format and the soliton format. The NRZ experiments had used dispersion management with all the wavelength channels in the normal dispersion regime. The soliton experiments had used constant dispersion with all the wavelength channels in the anomalous dispersion regime. The NRZ systems avoided the bad effects of a sizable average dispersion by setting it close to zero, while accepting the residual nonlinear impairments. The soliton systems avoided the nonlinear impairments by using a sizable dispersion, but the dispersion combined with ASE noise led to a large timing jitter [24,25]. At this point, the NRZ and soliton formats appeared completely different. Experiments up to this time to study these formats were done on separate testbeds. Since the leading protagonists of the different formats did not work cooperativelywith each other, it was almost impossiblefor an outside observer to judge among the competing claims. But things were about to change! Soliton systems evolved dramaticallyin the next few years. Suzuki et al. [26] and Carter et al. [27] showed that it is advantageous to combine dispersion management with a soliton system. The nonlinearity of the system is effectively lowered by the spreading of the solitons, leading to a substantial reduction of the timing jitter [28, 291. While these dispersion-managed solitons are no longer stationary, they are at least periodically stationary in that they return to the same pulse shape after every period. By contrast, in the later experiments of Favre et al. [30] and Tsurutani et al. [31], this is no longer true. Here, the initial RZ pulses, which are simply raised-cosine pulses, have an initial chirp, which is related to the entire dispersion in the transmission line. While in the case of Favre et QI. [30] they continue to refer to their format as a disperslonmanaged soliton (DMS) format, Tsurutani et al. [31] refer to their format as simply an RZ format. At the same time, the NRZ format also evolved dramatically. Bergano et al. [32] first introduced phase modulation in order to reduce the degree of polarization of the signal stream and, consequently, the effect of polarization hole-burning. They observed that the optimal phase modulation led to pulse compression by the end of the propagation, so that the marks appeared as clearly distinguishable RZ pulses. They later observed that this effect could be enhanced and better performance obtained by using an initial amplitude modulation as well as an initial phase modulation [33-351. Using this approach, they successfully transmitted wavelength channels on both sides of the zero dispersion point over transoceanic distances. The initial pulses in this case are raised-cosine pulses, and they are referred to by Bergano et al. [34] as chirped pulses. return to zero (CRZ) It is our contention that the NRZ and soliton formats have effectively converged. The result is a new format, which some call DMS and others call 310 Curtis R. Menyuk et al. CRZ. One can certainly point to exceptions. Work continues on,more traditional soliton formats, in which solitons are at least periodically stationary, and the traditional NRZ format is still widely used in commercial systems. Nonetheless, both experiment and theory abundantly demonstrate that this new converged format is superior when the goal is to obtain the best possible performance over the longest possible distance in a system that contains many WDM channels. (We note, however, that this converged format is not necessarily optimal when spectral efficiency is more important than distance.) In the remainder of this chapter, we present the evidence that the DMS and NRZ formats have converged. We begin with single-channel systems with moderate dispersion,in which dispersion-managedsolitonswere first observed and this convergencefirst became apparent. We then move on to a comparison of two WDM systems-the CRZ system of Bergano et al. [35]and the DMS system of Le Guen et al. [36]. 11. Single-Channel Systems The modern era of converged formats began with the groundbreaking experiments of Suzuki et al. [26] and Morita et al. [37] in which they demonstrated that it was possible to successfully propagate DMS pulses at 20 Gbitsls over 9000 km in a dispersion-managed system with in-line filters. These authors noted that the Gordon-Haus jitter was reduced by a factor of three relative to standard solitons. Other early work in field tests [38] and fiber lasers [39] also played an important role. Later, Jacob et al. [40] demonstrated that it was possible to send periodic DMS pulses at 10 Gbitsls over 28,000 km, as shown in Fig. 7.2. Conceptually, it is possible to understand the reduction of rii.n.n.n 0 Time @) 500 10,000 km Fig. 7.2 Evolution of a train of solitons over 28,000 km. ms figure is modified from i Ref. 40.1 7. Dispersion-Managed Solitons and Chirped Return to Zero 311 e B m E I g o 1.4 1.0 Distance Ln Time Time Time Fig. 7.3 Evolution of the pulse duration in one period of the dispersion map. The duration increases and the amplitude decreases in the normal dispersion fiber, effectively decreasing the nonlinearity. The pulse returns to its original shape in the normal dispersion fiber. phis figure is modified from Ref. 40.1 To Receiver / InputJ4 Fig. 7.4 Schematicillustration of the recirculating loop. [This figure is modified from Ref. 40.1 the Gordon-Haus jitter as an effect of the periodic stretching of the pulses that effectively lowers the nonlinearity as shown in Fig. 7.3. The recirculating loop that was used in the Jacob et al. experiment is shown in Fig. 7.4. It consisted of 100km of fiber in the normal dispersion regime (SMF-LS) with D = - 1.2 ps/nm-km at 1.55pm, followed by approximately 7 km of fiber in the anomalous dispersion regime (SMF-28) with D = 16.5 ps/nm-km at 1.55 Fm. This loop also had a 1.2-nm optical bandpass filter. By changing the amount of anomalous dispersion fiber in the loop, it was possible to change the total average dispersion from anomalous to normal, allowing these authors to compare the NRZ format and the DMS format in the same loop [41]. It was only necessary to make slight changes in the loop when shifting from one format to the other. First, the transmitter included a grating filter when DMS signals were launched to reduce the initial pulse duration of the marks to 20 ps. 312 Curtis R. Menyuk et al. 200 0 9 E 9 0 m, E c t 0 v,h o o o O I DMS -200 6 $ r e, 20000 w" a , ._ L 4 0 NRZ 0 Distance (krn) 10000 Fig. 7.5 Final eye diagrams and amplitude margins as a function of distance for the DMS and NRZ formats. Note the difference in length scale in the plots of the amplitude margins. [This figure is modified from Ref. 41.] This grating filter was not included when NRZ signals were launched. Second, the length of the anomalous dispersion fiber was 7.5 km for DMS transmission and 6.5 km for NRZ transmission. In Fig. 7.5, we show the final eye diagrams for the DMS transmission and the NRZ transmission, along with the corresponding amplitude margins as a function of distance. The amplitude margins indicate the voltages at which the BER reached a threshold of and allowed the authors [41] to determine the major sources of impairments in this system. For DMS transmission, it was the growth of noise in the spaces, while for NRZ transmission, it was degradation of the marks. At about the same time, Bergano et al. [33] carried out a key WDM experiment in which they demonstrated that it was possible to transmit 32 wavelength channels at 5 Gbits/s over 9300 km, with channels in both anomalous and normal dispersion regimes. The optical pulses of the channels in the anomalous dispersion regime appeared to be more soliton-like, while the optical pulses of the channels in the normal dispersion regime appeared to be NRZ-like. In later work, Marcuse and Menyuk [42] compared the NRZ, RZ, and DMS formats in simulations of an unfiltered single-channel system with a data rate of 100Gbitsh. They simulated an amplifier spacing of 20 km, which allowed them to obtain propagation distances in excess of 1000km. They did not include standard solitons in the comparison because previous experimental and theoretical work had made the advantagesof dispersion-managedsolitons over standard solitons abundantly clear. DMS pulses are less susceptible than standard solitons to interpulse interactions if the dispersion management is not too strong [43, 441, and they are less susceptible to timing jitter [26-291. 7 Dispersion-Managed Solitons and Chirped Return to Zero . 313 They even perform better in WDM systems [45,46]. Marcuse and Menyuk [42] studied a canonical system for each of these three formats. For the NRZ and RZ formats, the average dispersion was zero; for the DMS system, the pathaveraged dispersion was 0.025 ps/nm-km. The canonical path-averaged peak power for the DMS pulses was 14mW, their FWHM pulse duration when maximally compressed was 2.56 ps, and their shape was approximately Gaussian. These values for the DMS pulses were chosen to minimize the interpulse interaction, which requires a minimum full width half maximum (FWHM) pulse duration that is a little more than one-fourth the pulse separation [43,44]. The actual pulse shape was determined using an algorithm that ensured that the pulse evolution was periodically stationary [47]. The RZ pulses were raisedcosine pulses with no initial chirp. In both cases, the peak path-averaged power was 1mW. In all cases, the canonical slope was dD/dh = 0.03 ps/nm2-km and the canonical polarization-mode dispersion (PMD) was &MD = 0. The DMS pulses vary periodically in the dispersion map, while the RZ and NRZ pulses distort continually. In Fig. 7.6, we show the margins for the power, chromatic dispersion, higher-order dispersion, and PMD. We see that in all cases, it was possible to reach distances of 2000 km, with comparable power margins but in different power ranges. We note that it was not possible to propagate the RZ pulses 2000 km at a path-averaged peak power of 1mW and a dispersion slope of 0.35ps/nm2-km. Either the peak power must be higher or the dispersion slope must be lower. For this reason, there is no contribution for the RZ pulses in the margin plots for &MD and the average dispersion (D). short line at The Fig. 7 6 Power, chromatic dispersion, higher-order dispersion, and PMD margins for . the DMS, RZ, and NRZ formats as a function of fiber length. phis figure is modified from Ref 42.1 314 Curtis R.Menyuk et al. 2000 km on the margin plot for Dm is a tic mark, not an indication of RZ p pulses. There is no fundamental distinction between the RZ pulses and the DMS pulses. As we raise the powers of the RZ pulses, we find that the pulses begin to oscillate in the dispersion map and that the overall dispersion must become anomalous. The NFU pulses do not evolve continually into DMS pulses unless they are modulated with a phase chirp. In that case, Bergano et aZ. [33] showed that they do evolve into pulses that are like DMS pulses. Schematically, a picture like the one that we show in Fig. 7.7 emerges from these considerations. It is possible to successfullytransmit pulses with a wide range of combinations of the path-averaged peak power and the average chromatic dispersion over a length L. The interior of the oval marked Lj indicates the combinationsthat are possible over that length. High-powerpulses require an average dispersion that is slightly anomalous and a shape that is soliton-likefor optimal performance. At lower powers, the pulses should be more like RZ pulses. At even lower powers, the pulses should be more like NRZ pulses. As the length L increases from L1 to LZ to L3 to L4, the range of possible parameter values decreases, and the power and chromatic dispersion become tightly coupled. However, a range of power values is still possible as long as one chooses the appropriate chromatic dispersion to match each power value. This convergencehas not been universally accepted to date by the strongest advocates of the soliton and NRZ formats. It is certainly possible to question whether DMS pulses are really solitons. After all, solitons were first defined as stationary solutions to partial differential equations that had the additional property of passing through each other unscathed in collisionsexcept for a time shift [48], as, for example, the solution to Eq. 7.1 given in 7.2. DMS pulses are far from stationary. They oscillate in amplitude due to spatially varying gain 1 8 ' RZ-like ~~ NRZ-like I 7. Dispersion-Managed Solitons and Chirped Return to Zero 315 and loss; they oscillate in pulse duration due to dispersion; and, finally, they do not have the standard hyperbolic-secant shape given by Eq. 7.2. One could respond that the DMS pulses are at least periodically stationary, and that is true if they are launched with a precisely correct combination of pulse duration, pulse power, and shape. Moreover, in filtered systems, a fairly broad set of initial pulses will ultimately evolve into periodically-stationary DMS pulses after a transient period. A periodically-stationaryDMS pulse will ultimately emerge even in unfiltered systems, but, unless the initially launched pulse shape is close to the required pulse shape, a large amount of continuum radiation is created as well, leading to an unacceptable level of intersymbol interference. That said, when NRZ pulses are launched with an initial chirp, Bergano and his coworkers showed that they can ultimately evolve into pulses that are at least reminiscent of solitons. Should they be called solitons?While this question is still actively debated in scientific meetings, it is one that in OUT view has little scientificcontent at this point and is largely a matter of semantics. It is apparent that the new formats that we have described here have evolved substantially from the standard soliton and NRZ formats. The advocates of both the soliton and NRZ formats have contributed significantly to this development. 111. Wavelength-Division Multiplexed Systems There is a serious problem with using the periodically stationary DMS modulation format in WDM systems-at least with currently available fibers and dispersion-compensationtechniques. To understand this problem, the reader should turn to Fig. 7.8, where we show the interaction length as a function of the dispersion map strength, y = 2(/3;’L1 - /3;Lz)/tO, where /3;’ and are the second-order dispersion in the first and second legs of the dispersion map, 50 0 0 Y = 2( p;L,-p;L,)/~,2 3 6 Fig. 7.8 Interaction length vs the map strength y. The larger curve with dots corresponds to a pulse separation of four times the maximally compressed FWHM, whereas the lower curve with squares corresponds to a separation of three times the maximally compressed FWHM. phis figure is modified from Ref. 43.1 316 Curtis R Menyuk et al. respectively, L1 and L2 are the corresponding lengths, and t is the FWHM o pulse duration at the point of minimum compressionfor the soliton. The interaction length is the length scale on which two neighboring solitons attract each other, leading to a transmission error. Therefore, a longer interaction length is better. Here, we normalize the interaction length to the soliton’s characteristic nonlinear length scale [2]. We see that as the map strength increases, the interaction length increasesup to a point, indicatingthat DMS pulses perform better than standard solitons, as noted earlier. However, beyond y = 3, the interaction length rapidly plunges and becomes worse than for standard solitons. This precipitousdegradation is related to the increased stretching of the solitons. When the solitonsstretch enough to begin to interact significantly,the mutual nonlinear interaction ultimately destroys them. Later work confirmed that periodically stationary DMS pulses cannot tolerate any significant interpulse interaction [49]. As a consequence, it is not possible to use periodically stationary DMS pulses in a WDM system unless a strong form of soliton control is used like frequency sliding [50] or active amplitude modulation [51, 521. The reason is that the third-order dispersion in today’s optical fibers implies that some channels in a WDM system that uses more than 10nm of bandwidth must experience large dispersion, so that the pulses in the channels with large dispersion will stretch by large factors. In modern-day systems, it is commonplace to have 30 nm of bandwidth, and systems have been demonstrated with more than 80nm [53]. Even when slope-compensatingfiber is used inside one map period, so that there is no large change in the average dispersion from channel to channel, the spread within one map period is typically large. Moreover, it is desirable to keep the residual dispersion large in order to minimize the nonlinear interchannelinteractions that occur due to cross-phasemodulation. As systems are upgraded from 10 Gbits/s to 40 Gbitsls, the stretchingwill become even larger. The response to this dilemma has been to lower powers in the soliton systems until the mutual interactions are at a tolerable level and to add an initial chirp [30,36]. In this case, the pulses are no longer periodically stationary. Instead, they change shape throughout their propagation along the entire transmission line. With a properly chosen initial chirp, their pulse durations at the end are smaller than at the beginning. This new kind of DMS system closely resembles the CRZ system of Bergano et al. [35]. We will show that these systems resemble each other far more closely than either resembles a single-channel, periodically stationary DMS system. We show a schematic illustration of the first system that we are studying in Fig. 7.9. The dispersion map has L1 = 160km and L2 = 20km, while B;’ = -2.125 ps/nm-km and l?; = 17ps/nm-km at 1.55 km, so that the average dispersion is zero at 1.55 bm. The dispersion slope is 0.075 ps/nm2-km, and the total propagation length is 5040 km.Each channel in the simulation has an average power of 0.3mW at the point of largest amplification and contains a 64-bit, pseudo-random stream with 32 marks and 32 spaces. The pulses have a raised-cosine pro€ile; so the FWHM pulse duration is 50ps. 7. Dispersion-Managed Solitons and Chirped Return to Zero Channel 1 317 &-chirp Post-compensation 1 & compensation n -1 n n -1 n . d L, L, Fig. 7.9 Schematic illustration of the simulated CRZ system. This system resembles the experimental system in Ref. 35. Channel Pre-chwp Post-compensation 1 n -1 n n -1 n Fig. 7.10 Schematicillustration of the simulated DMS system. This system resembles the experimental system i Ref. 36. n Each period of the dispersion map contains four amplifiers, spaced 45km apart. The pulses in each wavelength channel are prechirped, and the chromatic dispersion in each channel can be compensated at both the beginning and the end of the transmission. Symmetric compensation is superior to either just precompensation or just postcompensation [54, 551. This system resembles the experimental system of Bergano et al. [35], although these authors used alternating polarizations in neighboring channels, while we used a single polarization so that Eq. 7.1 applies, They also used large-effective-area fiber, and we did not. We will refer to this system as the CRZ system. We show a schematic illustration of the second system that we study in Fig. 7.10. In this case, the dispersion map has L1 = 102km and LZ = 17.3km, while B;’ = 16.4ps/nm-km and = -85ps/nm-km at 1.55km. The dispersion in the first leg corresponds to SMF (single-mode fiber) 318 Curtis R Menyuk et al. and in the second leg to DCF (dispersion-compensating fiber). The average dispersion at 1.55 pm is 0.25ps/nm-km. The average dispersion slope is p”’ = 0.03ps/nm2-km, which includes the combined effect of the slope of the SMF fiber, which is 0.075ps/nm2-km, and the DCF fiber, which is -0.2 ps/nm2-km. The loss in the first leg of the dispersion map is 0.21 dB/km, and the loss in the second leg of the dispersion map is 0.6 dB/km. There is one amplifier in each leg of the dispersion map. As with the previous example, the simulation uses a pseudo-random bit stream with 32 marks and 32 spaces in each wavelength channel. In this case, the average signal power at the point of largest amplification is 2.0 mW, and the amplitude of the marks has a raisedcosine shape so that the FWHM of the pulse intensity is about 36ps. The total propagation length is 1400km. The signal was prechirped using a 4.5km length of DCF fiber, and the average dispersion was compensated using a variable length of SMF at the end of the transmission. This system resembles the experiment of Le Guen et al. [36], although these authors used 20-Gbitds transmission with alternating polarizations in each wavelength channel. The simulated system that we present here corresponds to 10-Gbits/stransmission per channel, all in the same polarization, so that Eq. 7.1 applies. We will refer to this system as the DMS system. We note that the amplifier spacings and peak powers are typical for terrestrial systems. Comparing the parameters of the two systems, we find that the powers in the DMS system are approximately three times larger in the CRZ system, but the total propagation length is approximately three to four times smaller. Local dispersions are about six times larger in the DMS system, and the rate of ASE noise accumulation is also approximately six times larger. Thus, if we only look at the raw parameters of the two systems, they look quite different; however, when we rescale the critical scale lengths, like the dispersive and nonlinear scale lengths, in both systems by dividing by the system length, we find that they agree within a factor of two in all cases. It is remarkable that these scaled parameters are nearly the same almost regardless of the system design or whether the system is intended to model terrestrial or transoceanic systems. Because of the way in which the dispersion compensation is done in each leg of the dispersion map in the two systems, the intermediate evolution appears quite different. In the CRZ system, each channel is separately compensated at the beginning and at the end, so that for most channels the averaged dispersion in each dispersion map is quite large. In Fig. 7.1 1, we show the FWHM duration at the point of maximum compression in the dispersionmap for both the CRZ system and the DMS system for a single channel. In the CRZ system, this channel is shifted -4.8 nm from h = 1.55 pm, while the channel is not offset in the DMS system. The CRZ pulses broaden by a factor of almost four in the initial precompensating fiber. After that, the minimum pulse duration in each map decreases up to the middle of the propagation path, at which point the minimum pulse duration increases once again, reaching a maximum right before the postcompensating fiber. After postcompensation, the final 7. Dispersion-Managed Solitons and Chirped Return to Zero 319 5000 I i 0 0 1500 Distance (km) Fig. 7.11 Pulse duration as a function of propagation distance in the CRZ and DMS systems. Note the change in scale. pulse duration is approximately half the initial pulse duration. In the DMS system, the pulses are stretched as well as chirped in the initial prechirping fiber so that the pulse duration is nearly twice what it was originally. After that, the minimum pulse duration in each map slowly decreases, and the k a l postcompensating fiber brings the final pulse duration to approximately two-thirds the initial pulse duration. Despite the significant differences in the evolution, two salient points of similarity emerge. The first is that neither system is periodically stationary, in contrast to the single-channel DMS systems that we presented in the previous section. Moreover, both systems use a prechirp to obtain a final pulse duration that is smaller than the initial pulse duration. The pulse evolution in both the CRZ and the DMS systems is only weakly affected by the nonlinearity. To verify this point, one can calculate the pulse evolution from Eq. 7.1 both with and without the nonlinear term and compare the results. In Fig. 7.12, we show the ratio of the output FWHM pulse durato tion Tout the input FWHM pulse duration Ti, as a function of the initial chirp when the average dispersion (0) set to achieve optimal compression is that at the output. We also show the average dispersion (0) achieves the optimal compression as a function of the chirp. In the CRZ system, the chirp is included by introducing a phase variation in the wave envelope of the form q5 = A n cos (2nt/Tin),where t is the time measured from the center in one bit window. This chirp is similar to what a LiNbOs modulator would produce. We define the chirp C as the negative of the second derivative of the phase for at the peak of the initial pulse, so that C = An(27r))2/qi the CRZ pulses. For the DMS pulses, we determine the initial chirp by solving Eq. 7.1 in the prechirping fiber, which must be done nonlinearly. After that, we calculate the remainder of the evolution both linearly and nonlinearly to determine the importance of nonlinearity in the pulse evolution after the initial chirp has been introduced. 320 Curtis R Menyuk et al. 00 .7 s _a E 8 0 pq-/;pq % --* 0 C (GHzIps) 15 0 C (GHz /ps) 2 Fig. 7.12 Final pulse compression and required average dispersion as a function of initial pulse chirp. The solid curves indicate the linear results and the dashed curves indicate the nonlinear results. The curves show that the pulse evolution is dominated by linear, dispersive evolution, although arguably nonlinearity is somewhat more important in the DMS system than in the CRZ system. A key point is that the relationship between the initial chirp and the average dispersion is chosen to maximize the pulse compression at the output of the entire transmission line. In the periodic DMS systems that we presented in the last section, it may be useful to introduce a prechirp to decrease the initial transient, but there is no relationship between the initial prechirp and the final pulse durations once the transient oscillation have damped out. There is another way in which both the CRZ and DMS systems presented here behave like linear systems. In linear systems, the spread in the eye diagrams is dominated by signal-spontaneous beat noise, so that the ONES rail is more spread out than the ZEROS rail [56, 571. By contrast, the spread in the eye diagrams of the periodically stationary DMS pulses is dominated by exponential noise growth in the spaces, so that the ZEROS rail is as spread out as the ONES rail [58]. In Fig. 7.13, we show electrical eye diagrams for both systems, where we first excluded and then included the Kerr nonlinearity and ASE noise. We see that the eye diagrams in both systems are nearly identical and clearly indicate the dominance of spontaneous-signalbeat noise. Because the pulse evolution is dominated by the linear dispersion in both the CRZ and DMS systems, and because the spread in the eye diagrams is dominated by spontaneous-signal beat noise, it seems reasonable to refer to these systems as quasilinear. It is our contention that the quasilinear DMS system that we have studied in this section resembles the CRZ system far more closely than it 7. Dispersion-Managed Solitons and Chirped Return to Zero CRZ No nonlinearity No ASE 321 ,? - ? Nonlinearity No ASE v Q) c 2 No nonlineanty ASE e 8 Nonlinearity ASE 0 300 0 Time (ps) 300 Fig. 7.13 Electrical eye diagrams for the CRZ and DMS systems, with and without ASE noise and nonlinearity. resembles the periodically stationary DMS system that we introduced in the last section. The simulations that we have described thus far in this section are based on a single channel in order to focus on the behavior of individual pulses and to reduce the computational costs sufficiently to allow the study of a wide range of parameters. The channel was offset sufficiently from h = 1.55 km to ensure that the dispersive stretching is large, as is the case in most wavelength channels. Nonetheless, it is necessary to verify that the behavior that was observed still persists in full WDM systems. In simulations with up to 7 channels spaced 0.6 nm apart and centered about h = 1.55 km, it was found that there is no change in the quasilinear behavior previously described. However, both the ONES rail and the ZEROS rail broaden in the eye diagrams due to interchannel interactions. We show the eye diagrams for several channels in a seven-channel simulation and for both the CRZ and DMS systems in Fig. 7.14. We note that the spreading is particularly severe for channel 4 in the CRZ system, which is located at h = 1.55 km. At this wavelength, there is almost no net dispersion between the channels, which implies that pulses that initially overlap at high powers tend to repeat the same interactions periodically, enhancing the nonlinear interaction between this channel and its neighbors. We note that earlier studies indicate that 7 channels are sufficient to determine the evolution for this type of system [59]. 322 Curtis R.Menyuk et al. CRZ DMS ch.1 ch .I 0 300 0 300 Time (ps) Fig. 7.14 channels. Electrical eye diagrams for the CRZ and DMS system with seven WDM IV. Conclusions In 1996, five short years ago at the time that this chapter was written, all communications systems employed the traditional NRZ modulation format, and standard solitons were the only alternative being seriously considered. In the last 5 years, solitons evolved into periodically stationary, dispersion-managed solitons in single-channel systems and, ultimately, into quasilinear, dispersionmanaged solitons in WDM systems. During the same period, the traditional NRZ format first evolved into the phase- and amplitude-modulated “nearly” return to zero format. From there, this format evolved into the chirped return to zero format that is now being deployed in undersea systems. In singlechannel systems, the periodically stationary DMS format, the RZ format, and the NRZ format all form part of a continuum of formats. In modernday WDM systems, the CRZ format has a substantially longer reach than the traditional NRZ format, and the quasilinear DMS format appears to be the only soliton format without strong control that is compatible with today’s fibers and components. These two formats-CRZ and quasilinear DMS-are essentially the same. For many years, there was a debate over whether a soliton format or an NRZ format was preferable. Because of this public debate, we are often asked whether solitons “won” or “lost” and whether solitons will “ever be used.” In our view, the debate is over, and both sides won. There continues to be a debate over whether the quasilinear DMS/CRZ format should really be called a soliton format, but this debate is really about semantics, not science. The reality is that an intermediate format has emerged, and both camps arrived at 7. Dispersion-Managed Solitons and Chirped Return to Zero 323 this format almost simultaneously, starting from different points. This intermediate format is remarkably robust and effective. Like Juliet, when she is bemoaning the hostility between the Capulets and the Montagues, we can conclude, “What’s in a name? That which we call a rose, by any other name would smell as sweet.. . .” [60]. Acronyms ASE BER CRZ DMS EDFA FWHM IM-DD LED NRZ OOK PMD Amplified spontaneous emission Bit error rate Chirped return to zero Dispersion managed soliton Erbium-doped fiber amplifier Full-width half-maximum Intensity modulated-direct detection Light emitting diode Nonreturn to zero On-off keying Polarization-mode dispersion Return-to-zero Single mode fiber Wavelength-division multiplexing Rz SMF WDM References [1 A. Hasegawa and E Tappert, “Transmissionof stationarynonlinear optical pulses 1 in dispersive dielectricfibers. I. Anomalous dispersion,” Appl. Phys. Lett., vol. 23, pp. 142-144 (1973). [2] ’Ibo general references that contain copious further references to the literature on soliton communications are: A. Hasegawa and Y. Kodama, Solitons in Optical Communications, Clarendon: Oxford, UK (1995) and G. P. Agrawal, Nonlinear Fiber Optics, Academic Press: San Diego, CA (1995). [3] L. E Mollenauer, R. H. Stolen, and J. I? Gordon, “Experimental observation of picosecond pulse narrowing and solitons in optical fibers,” Phys. Rm. Lett., V O ~ .45, pp. 1095-1098 (1980). [4] The original paper that demonstrates this property for soliton solutions of the nonlinear Schrodinger equation is: V. E. Zakharov and A. E. Shabat, “Exact theory of two-dimensionalself-focusingand one-dimensionalself-modulation of waves in nonlinear media,” Sov. Phys.-JETP, vol. 34, pp. 62-69 (1972) [Original RussianZh. Ehp. Teol: Fiz., vol. 61, pp. 118-134 (1971)]. [5] A. Hasegawa and Y Kodama, “Signal transmission by optical solitons in monomode fiber,” Proc. IEEE, vol. 69, pp. 1145-1 150 (1981). 324 Curtis R Menyuk et a. l [6] A. Hasegawa and Y Kodama, “Amplificationand reshaping of optical solitons in a glass fiber-IV Use of the stimulated Raman process,” Opt. Lett., vol. 8, pp. 650-652 (1983). [7] L. E Mollenauer and K. Smith, “Demonstration of soliton transmission over more than 4000 km in fiber with loss periodically compensated by Raman gain,” Opt. Lett., vol. 13, pp. 675-677 (1988). [8] Historical discussions and discussions of systems applications may be found in: E. Desurvire, Erbium-DopedFiber AmpliJers: Principles and Applications,Wiley: New York, N Y (1994) and P. C. Becker, N. A. Olsson, and J. R. Simpson, Erbium-Doped Fiber Amplifzers: Fundamentals and Technology, Academic Press: San Diego, CA (1999). [9] E Forghieri, R. W Tkach, and A. R. Chraplyvy, “Fiber nonliiearities and . their impact on transmission systems,” in Optical Fiber TelecommunicationsIIIA, I. P. Kaminow and T. L. Koch, eds., Academic Press: San Diego, CA (1997). Contains an extensive set of references to the key literature. [lo] A. R. Chraplyvy, “Limitations on lightwave communications imposed by opticalfibernon-linearities,”J. Lightwave Technol.,vol. 8, pp. 1548-1 557 (1990). Contains a lucid review of the physical effects. [ll] See, e.g., the discussions in L. Kazovsky, S. Benedetto, and A. Wdner, Optical Fiber Communication Systems, Artech Boston, MA (1996), Chap. 3.1 and S. Gowar, Optical Communication Systems, Prentice-Hall New York, NY (1993), Chap. 7.3. [12] G. P. Agrawal, Fiber-optic Communication Systems, Wiley: New York, NY (1997), Chap. 1.2.3. See also the historical discussion at the beginning of E Heismann, S. K. Korotky, and J. J. Veselka, “Lithium niobate integrated optics: Selected contemporary devices and systems applications,” in Optical Fiber Telecommunications IIIB, I. P. Kaminow and T. L. Koch, eds., Academic Press: San Diego, CA (1997). [13] D. A. Fishman and B. S. Jackson, “Transmitter and receiver design for amplified lightwave systems,” in Optical Fiber TelecommunicationsIIIB, I. P. Kaminow and T. L. Koch, eds, Academic Press: San Diego, CA (1997). [14] K. Smith and L. E Mollenauer, “All-optical long-distance soliton-based transmission systems,” in Optical Solitons--Theory and Experiment, J. R. Taylor, ed., Cambridge: Cambridge, UK (1992). [15] D. Marcuse, “Single-channel operation in very long nonlinear fibers w t optical ih amplifiers at zero dispersion,” J. Lightwave Technol.,vol. 9, pp. 356-361 (1991). [16] N. S. Bergano, “Undersea amplified lightwave systems design,” in OpticaZ Fiber Telecommunications IIU, I. P. Kaminow and T. L. Koch, eds., Academic Press: San Diego, CA (1997). 1171 A. H. Gnauck and R. M. Jopson, “Dispersion compensation for optical fiber systems,” in OpticaIFiber TelecommunicationsIIIA, I. P Kaminow and T. L. Koch, . eds., Academic Press: San Diego, CA (1997). 7. Dispersion-Managed Solitons and Chirped Return to Zero 325 [I81 J. L.Zyskind, J. A. Nagel, and H. D. Kidorf, “Erbium-doped fiber amplifiers for optical communications,”in OpticalFiber TelecommunicationsIIIB, I. P. Kaminow and T. L. Koch, eda, Academic Press: San Diego, CA (1997). [19] R. Gi!es and T. Li, “Optical amplifiers transform long-distance lightwave telecommunication^" Proc. IEEE, vol. 84, pp. 87-83 (1996). [ZO] A. R. Chraplyvy, A. H. Gnauck, R. W. Tkach, and R. M. Derosier, “8 x 10 Gb/s transmission through 280 km of dispersion-managedfiber,” ZEEE Photon. Technol. Lett., vol. 5, pp. 1233-1235 (1993). [21] L. E Mollenauer, I?V. Mamyshev, and M. J. Neubelt, “Demonstration of soliton WDM transmission at 6 and 7 x 10Gbit/s, error-free over transoceanicdistances,” Electron. Lett., vol. 32, pp. 471-473 (1996). [22] N. S. Bergano and C. R. Davidson, “Wavelength division multiplexing in longhaul transmission systems,” J. Lightwave Technol., vol. 14, pp. 1299-1307 (1996). [23] P. V. Mamyshev, “Dual-wavelength source of high-repetition-rate, transformlimited optical pulses for soliton transmission,” Opt. Lett., vol. 19, pp. 2074-2076 (1994). [24] J. P. Gordon and H. A. Haus, “Random walk of coherently amplified solitons in optical fiber,” Opt. Lett., vol. 11, pp. 665-667 (1986). [25] H. A. Haus, “Quantum noise in a solitonlike repeater system,” J. Opt. SOC. B, Am. V O ~ .8, pp. 1122-1 126 (1991). [26] M. Suzuki, I. Morita, N. Edagawa, S. Yamamoto, H. Taga, and S. Akiba, “Reduction of Gordon-Haus timingjitter by periodic dispersion compensation in soliton transmission,” Electron. Lett., vol. 31, pp. 2027-2029 (1995). [27] G. M. Carter, J. M. Jacob, C. R. Menyuk, E. A. Golovchenko, and A. N. Pilipetskii, “Timing-jitter reduction for a dispersion-managed soliton system: Experimentalevidence,” Opt. Lett., vol. 22, pp. 513-515 (1997). [28] N. J. Smith, W. Forysiak, and N. J. Doran, “Reduced Gordon-Haus jitter due to enhanced power solitons in strongly dispersion managed systems,” Electron. Lett., vol. 32, pp. 2085-2086 (1996). [29] R.-M. Mu, V. S. Grigoryan, C. R. Menyuk, E. A. Golovchenko, and A. N. Pilipetskii, “Timing-jitter reduction in a dispersion-managed soliton system,” Opt Lett., vol. 23, pp. 936932 (1998). [30] E Favre, D. Le Guen, M. L. Moulinard, M. Henry, and T. Georges, “320 Gbit/s soliton WDM transmission over 1300km with 100km dispersion-compensated spans of standard fibre,” Electron Lett., vol. 33, pp. 2135-2136 (1997). [31] T. Tsurutani, K. Imai, N. Edagawa, and M. Suzuki, “340Gbit/s (32 x 10.66Gbit/s) WDM transmission over 6054 km using hybrid fibre spans of large core fibre and dispersion shifted fibre with low dispersion slope,” Electron. Lett., vol. 35, pp. 646-647 (1999). It is difficult to do justice to the wide scope of the work of Suzuki, Edagawa, and their colleagues at KDD with any single reference. In a series of publications that appeared mostly in Electronics Letters, they have examined all of the major formats that are being considered for use in transoceanic systems at different data rates and with different numbers of channels. The work just cited uses initial raised-cosinepulses with an initial chirp,just 326 Curtis R.Menyuk et a. l like the work by Favre et al. cited in Ref. [31]. In other work, H. Taga, K. Imai, N. Takeda, M. Suzuki, S. Yamamoto, and S. Akiba, “10 WDM 10 Gbitfs recirculating loop transmission experiments using dispersion slope compensator and non-soliton RZ pulse,” Electron. Lett., vol. 33, pp. 2058-2059 (1997) study an RZ format with unchirped raised-cosine pulses. An experiment based on more traditional soliton-like pulses is described in K. Tanaka, I. Morita, M. Suzuki, N. Edagawa, and S. Yamamoto, “400 Gbit/s (20 x 20 Gbit/s) denseWDM solitonbased RZ signal transmission using dispersion flattened fibre,” Electron. Lett., vol. 34, pp. 2257-2258 (1998). N. S. Bergano, C. R. Davidson, and F. Heismann, “Bit-synchronouspolarisation and phase modulation scheme for improving the performance of optical amplifier transmission systems,” Electron. Lett., vol. 32, pp. 52-54 (1996). N. S. Bergano, C. R. Davidson, M. A. Mills, P. Corbett, S. G. Evangelides, B. Pedersen, R. Menges, J. L. Zyskind, J. W. Sulhoff, A. K. Srivastava, C. Wolf, and J. Judkins, “Long-haul WDM transmission using optimal channel modulation: A 160Gb/s (32 x 5Gb/s) 9,500km demonstration,” OFC ’97 Technical Digest (Dallas, TX), Feb. 1997, postdeadline paper PD16. N. S. Bergano, C. R. Davidson, M. Ma, A. N. Pilipetskii, S. G. Evangelides, H. D. Kidorf, J. M. Darcie, E. A. Golovchenko, K. Rottwitt, P C. Corbett, R. . Menges, M. A. Mills, E. Pedersen, D. Peckham, A. A. Abramov, and A. M. Vengsarkar, “320 Gb/s WDM transmission (64 x 5 Gb/s) over 7,200 km using large mode fiber spans and chirped return-to-zero signals,” OFC ’98 Technical Digest (San Jose, CA), Feb. 1998, postdeadline paper PD12. N. S. Bergano, C. R. Davidson, C. J. Chen, B. Pedersen, M. A. Mills, N. Ramanujam, A. B. Kidorf, H. D. PUC,M. D. Levonas, and H. Abdelkader, “640 Gb/s transmission of sixty-four 10 Gb/s WDM channels over 7,200 km with 0.33 (bits/s)/hz spectral efficiency,” OFC ’99 Technical Digest (San Diego, CA), Feb. 1999, postdeadline paper PD2. D. Le Guen, S. D. Burgo, M. L. Moulinard, D. Grot, M. Henry, E Favre, and T. Georges, “Narrow band 1.02Tbit/s (51 x 20Gbit/s) soliton DWDM transmission over 1000km of standard fiber with 100km amplifier spans,” OFC ’99 Technical Digest (San Diego, CA), Feb. 1999, postdeadline paper PD4. I. Morita, M. Suzuki, N. Edagawa, S. Yamamoto, H. Taga, and S. Akiba, “20Gb/s single-channel soliton transmission over 9000 km without inline filters,” IEEE Photon. Technol.Lett., vol. 8, pp. 1573-1574 (1996). M. Nakazawa, Y. Kimura, K. Suzuki, H. Kubota, T. Komukai, E. Yamada, T. Sugawa, E. Yoshida, T. Yamamoto, T. Imai, A. Sahara, H. Nakazawa, 0. Yamauchi, and M. Umezawa, “Field demonstration of soliton transmission at 10 Gbitls over 2000 km in Tokyo metropolitan optical loop network,” Electron Lett., vol. 31, pp. 992-994 (1995). H. A. Haus, K. Tamura, L. E. Nelson, and E. P. Ippen, “Stretched-pulseadditive pulse mode-locking in fiber ring lasers: Theory and experiment,” J. Quantum Electron., vol. 31, pp. 591-598 (1995). [32] [33] [34] [35] [36] [37] [38] [39] 7. Dispersion-ManagedSolitons and Chirped Return to Zero 327 [40] J. M. Jacob, E. A. Golovchenko, A. N. Pilipetskii, G. M. Carter, and C. R. Menyuk, “Experimentaldemonstration of soliton transmission over 28 Mrn using mostly normal dispersion fiber,” IEEE Photon. Technol. Lett., vol. 9, pp. 130-132 (1997). [41] J. M. Jacob, E. A. Golovchenko, A. N. Pilipetskii, G. M. Carter, and C. R. Menyuk, “10Gb/s transmission of NRZ over 10,000krn and solitons over 13,500 km error-free in the same dispersion-managedsystem,” ZEEE Photon. Technol. Lett., vol. 9, pp. 1412-1414 (1997). [47] D. Marcuse and C . R. Menyuk, “Simulation of single-channeloptical systems at 100Gb/s,”J. Lightwave Technol., vol. 17, pp. 564-569 (1999). [43] T.Yu, E. A. Golovchenko, A. N. Pilipetskii, and C. R. Menyuk, “Dispersionmanaged soliton interaction in optical fibers,” Opt. Lett., vol. 22, pp. 793-795 (1997). [44] N. J. Smith, N. J. Doran, W. Forysiak, and E M. Knox, “Soliton transmission using periodic dispersion compensation,”J. Lightwave Technol., vol. 15, pp. 18081822 (1997). [45] E. A. Golovchenko, A. N. Pilipetskii, and C . R. Menyuk, “Collision-induced timing jitter reduction by periodic dispersion management in soliton WDM transmission,” Electron. Lett., vol. 33, 735-737 (1997). [46] E. A. Golovchenko, A. N. Pilipetskii, and C. R. Menyuk, “Periodic dispersion management in soliton wavelength-division multiplexing transmission with sliding filters,” Opt. Lett., vol. 22, 1156-1 158 (1997). [47] J. H. B. Nijhof, W. Forysiak, and N. J. Doran, “The averaging method for finding exactly periodic dispersion-managed solitons,” IEEE J. Select. Topics Quantum Electron., vol. 6, pp. 330-336 (2000). [48] N. J. Zabusky and M. D. Kruskal, “Interaction of solitonsin a collisionlessplasma and the recurrence of initial states,” Phys. Rev Lett., vol. 15, pp. 240-243 (1965). [49] V S. Grigoryan, G. M. Carter, and C. R. Menyuk, “Tolerance of dispersionmanaged soliton transmission to the shape of the input pulses,” IEEE Photon. Technol. Lett., vol. 12, pp. 1165-1167 (2000). [50] L. E Mollenauer, J. P Gordon, and S. G. Evangelides, “The sliding-frequency . guiding filter: An improved form of soliton jitter control,” Opt. Lett., vol. 17, pp. 1575-1577 (1992). E511 M. Nakazawa, E. Yamada, H. Kubota, and K. Suzuki, “lOGbit/s soliton data transmission over one million kilometres,” Electron Lett., vol. 27, pp. 1270-1272 (199i). [52] E. Desurvire, 0. Leclerc, and 0. Audouin, “Synchronous in-line regeneration of wavelength-division multiplexed solitons signals in optical fibers,” Opt. Lett., vol. 21, pp. 1026-1028 (1996). [53] Y Sun, J. W Sulhoff, A. K. Srivastava,J. L. Zyskind, T. A. Strasser, J. R. Pedraz. zini, C. Wolf, J. Zhou, J. B. Judkins, R. P. Espindola, and A. M. Vengsarkar, “80 nm ultra-widebanderbium-doped silica fibre amplifer,” Electron Lett., ~01.33, pp. 1965-1967 (1997). 328 Curtis R Menyuk et al. [54] L. Ding, E. A. Golovchenko, A. N. Pilipetskii, C. R. Menyuk, and P. K. A. Wai, “Modulated NRZ signal transmission in dispersion maps,” OS4 Trends in Optics and Photonics, vol. 12, System Technologies, A. E. Willner and C. R. Menyuk, eds. Optical Society of America: Washington, DC, pp. 204-206 (1997). [55] M. I. Hayee and A. E. Willner, “Pre- and post-compensation of dispersion and nonlinearities in 10-Gb/s WDM systems,” IEEE Photon. Technol. Lett., vol. 9, pp. 1271-1273 (1997). [56] D. Marcuse, “Derivation of analytical expressions for the bit-error probability in lightwave systemswith optical amplifiers,”J.Lightwave Technol., vol. 8, pp. 18161822 (1990). [57] P. A. Humblet and M. Azizoglu, “On the bit error rate of lightwave systems with optical amplifers,” J. Lightwave Technol., vol. 9, pp. 1576-1582 (1991). [58] R.-M. Mu, V. S. Grigoryan, C . R. Menyuk, G. M. Carter, and J. M. Jacob, “Comparison of theory and experiment for dispersion-managed solitons in a recirculating fiber loop,” Select. Topics Quantum Electron., vol. 6, pp. 248-251 (2000). [59] T. Yu, W. M. Reimer, V. S. Grigoryan, and C. R. Menyuk, “Amean field approach for simulating wavelength-divisionmultiplexed systems,” IEEE Photon. Technol. Lett., vol. 12, pp. 4 4 3 4 5 (2000). [60] W Shakespeare, Romeo and Juliet, Act 2, Scene 2. . Chapter 8 Metropolitan Optical Networks Nasir Ghani, Jin-Yi Pan, and Xin Cheng Sorrento Networh Inc., San Diego, California 1. Introduction The modem telecommunications age has created an unprecedented growth in networking infrastructures. With the recent explosion of Internet data communications, Internet Protocol (IP) data traffic volumes have surpassed voice and are projected to be over 75% of total network traflic within the next 2 years [60,65]. It is often stated that Internet traffic is experiencing “exponential” growth rates, and whilst the exact values are very case specific, it is safe to say that data growth is much larger than circuit voice growth [42]. Moreover, it is also well known that data traffic characteristics are very different from voice, exhibiting highly bursty and unpredictable behaviors, see [60, 1591. As the user base grows and more content-rich applications emerge, undoubtedly capacity demands will increase further. These fundamental trends are already having a profound impact on the overall networking space. Telecommunications networks are usually segmented in a three-tier hierarchy: access, metropolitan, and long haul (further delineations also are possible [65]). Long-haulbackbone networks span interregional/global distances (1000 km or more) and provide large tributary connectivity between regional and metro domains. Backbone networks are optimized for transmission, and related costs have been dominated by expensive line equipment, for example, regeneration gear [64, 731. On the other end of the hierarchy are access networks, providing connectivity to a plethora of customerswithin close proximity. Access networks use a very broad range of technologies/protocoIs and represent a continual flux. Straddled in the middle are metropolitan (metro) networks, averaging regions between 10-100h [22, 641 and interconnecting access and long-haul networks. Metro networks today are based upon synchronous optical network (SONET)/synchronous digital hierarchy (SDH) ring architectures. Namely, smaller tributary rings, for example, OC3/STM-1(155 Mb/s) or OC-12/STM-4(622 Mb/s), aggregatetrafficonto larger core interoffice (IOF) [64] rings that interconnect central office (CO) locations at higher bit rates, for example, OC-48/STM-16 (2.5 Gb/s). Overall, SONET/SDH has been very successful in delivering the first wave of end-user connectivity, namely voice. 329 OPTICAL FIBER TELECOMMUNICATIONS, VOLUME N B Copyright 0 2002, Elscvier Science (USA). All rights of repduction in any form m d . ISBN:&12-395173-9 330 Nasir Ghani et a. l Internet traffic growth has significantly altered the networking domains bordering the metro. Long-haul networks felt the Internet crunch first and have undergone large-scaleexpansionsusing optical dense wavelength-division multiplexing (DWDM) technology [64] (note that herein, the terms DWDM and WDM are used interchangeably). DWDM yields the best backbone costkapacity tradeoff [41,71], and many backbone networks now boast terabit capacities. Meanwhile, access networks have also seen their share of progress. Residential cable and digital subscriber loop (DSL) modems have increased user access rates from kilobits to megabits, and other advanced technologiespromise to further this trend. Also, large corporatecustomers are now deploying advanced switchinglrouting gear capable of direct line-rate inputs to the metro core, for example, concatenated OC-48~4192~~ lO-Gb/s Ethernet interfaces. Collectively, these increased access rates are blurring traditional access boundaries and beginning to stifle legacy “voicecentric” metro architectures. Many SONETEDH rings are experiencingcapacity exhaust at even OC-48/192 rates [62, 65, 1421, and costly ring (fiber) expansion is becoming overly slow and expensive. Furthermore, as market expansion and deregulation take form, increased competition is forcing operators to support a highly dynamic, increasingly diverse mixtures of client protocols, both legacy and asynchronous data [61-65, 871. Examples include IP, ATM (Asynchronous Transfer Mode), SONETBDH, Ethernet (10/100 Mb/s, 1.0/10Gb/s), multiplexed TDM voice (DS-n), and other more specialized data protocols such as FDDI (Fiber Distributed Data Interface), ESCON (Enterprise System Connectivity), FICON (Fiber Connectivity Channel), Fibre Channel, cable video, etc. (see Fig. 8.1, adapted from [158]). Again, legacy SONET/SDH networks are proving extremelyrestrictive here, exhibitinghigh bandwidth inefficiencies and overly complex, cumbersome provisioning procedures [141, 1421. Clearly new metro solutions are required that offer superior price/ performance alternatives to legacy SONET/SDH expansion, and from the above, a host of necessary features can be derived. For example, new platforms must offer high bandwidth scalabilityand carry multiple protocols over a common infrastructureto reduce costs [46]. Rapid, intelligent servicesprovisioning and survivabilityare also crucial, as operators look to compete via service differentiation and not just pricing [65], Le., advanced service level agreements (SLA). Moreover, given current infrastructure investments, new schemes must provide backwards compatibility, i.e., legacy support [22,44], to enable more cost-effective, staged migrations. Undoubtedly, DWDM technology meets many of these requirements, and given its success in the long-haul space, is being increasingly touted in the metro domain [22, 35-57, 61-65]. Declining costs are making this technology very cost competitive with SONET/SDH, and even though increased bit rate TDM solutions are emerging, for example, 40 Gb/s OC-768/STM-256, DWDM is much more scalable from a pure capacity perspective, Le., terabitdfiber. Moreover, DWDM provides genuine bit rate/protocol transparency, a very compelling advantage in the 8. Metropolitan Optical Networks 331 I I rP I Physical layer ( i e , f b rDWDM CWDM) Fig. 8.1 Sample protocol mappings at the metro edge. protocol-diverse metro space. Furthermore, advanced DWDM component technologies (such as amplifiers, switches, filters, lasers, fibers) are now enabling network-level wavelength routing and protection [ 1, 71 over more complex multi-hop fiber topologies, e.g., rings or meshes. These capabilities, coupled with intelligent control architectures [148, 1491, will allow operators to provision large amounts of capacity with an advanced degree of service definitioddifferentiation. Overall, the last several years have seen substantial point-to-point metro DWDM deployments along congested inter-CO spans [61,621, and many operators are already looking to evolve to much more capable dynamic rindmesh paradigms. A key issue herein is network migration, i l as cost sensitivities w l mandate infrastructure reuse and modular expansion. Over time, however, advanced DWDM technology will dominate the metro core, usurping slower legacy TDM overprovisioningloverengineering paradigms (e.g., excess capacity buildouts) with more expedient, proactive wavelength services frameworks [36]. Nevertheless, even though optical DWDM technology offers many benefits, its induction within the metro arena is more complicated than in the Iong-haul networks. In essence, DWDM is best suited for larger metro core networks, where scalable large granularity “lambda” tributary provisioning is required in the gigabits range. At the metro edge, where operators must interface with increased protocol heterogeneitieshnterface bit rates, the need to costeffectively handle finer “sub-wavelength” capacity increments is paramount. 332 Nasir Ghani et al. This contrasts sharply with long-haul networks where input signals comprise a few protocols (SONETlSDH or digital wrappers [12]) and interface bit rates (2.5, 10, possibly 4OGbls). As a result, the metro edge requires integrated, intelligent opto-electronic solutions to perform multiprotocol aggregation/groomingonto larger DWDM tributaries, with a particular focus on data protocol efficiency [44, 621. Here, various “data-centric” solutions are emerging, driven by advances in high-density electronic integrated circuit (IC) technology, such as application-specificintegrated circuits (ASIC). These include DWDM edge rings, “next-generation”SONETlSDH and multiservice provisioning architectures, and IP packet rings. By all accounts, the metro space is set for large growth, with a predicted market size well into the multiple billions of dollars within the next several years [61-651. This chapter presents a broad overview of current developments in this exciting arena. However, it is important to note that this is a continually evolving field [64], and further evolutions are bound to occur. A high-level review of existing legacy architectures is first presented in Section 2, and various pertinent trends and developments are identified in Section 3. Section 4 briefly reviews developments in DWDM technology, and actual metro DWDM solutions are presented in Section 5, detailing architectures and migration strategies. A wide range of complementary opto-electronic edge solutions are covered in Section 6, and the role of network standards is discussed in Section 7. Finally, a brief look at future evolutions is given in Section 8, and conclusions are presented in Section 9. 2. Traditional Architectures It is instructive to take a brief look at existing SONETlSDH metro architectures in order to better position subsequent discussions. Overall, the SONETlSDH framework has evolved from the need to standardize multivendor interconnectivity at the fiber level (i.e., mid-span meet), a capability lacking in precursor plesiochronous digital hierarchy (PDH) systems. Specifically, on a physical level, SONETlSDH defines a time-division multiplexing (TDM) frame format (125 ps duration) and an associated multiplexing hierarchy (i.e., OC-n [4]). Meanwhile, on a higher level, related networking fimctionalities are also provided, such as transport, multiplexing, add/drop, cross-connectionlswitching,regeneration, and protection [4]. These functions require control information, and this is carried via designated overhead bytes in the frame format, roughly 4% of the OC-1 signal. SONETlSDH systems provide excellent resiliency and can support various configurations, such as hub, linear chain, ring, and even mesh (see [4] for a detailed review). Of these, ring architectures are by far the most ubiquitous in metrolregional settings [12, 35, 37, 1071 and are commonly based upon a hierarchy mirroring the associated channel rates. Specifically, slower-speed OC-3lSTM-1 (155 Mbls) 8. Metropolitan Optical Networks RegronaVlong-haul 333 e m (IOF) SONET/SDH B U R ring (oc-19YSTM-64) Mem (IOF) SONET/SDH B U R nng (OC-48/STM-16) \ Edge UPSR nng Enterprise routed swtches (leased line TDM interfaces) Fig. 8.2 Traditional SONETISDH ring hierarchy in metropolitan region. or OC-l2/STM-4 (622 Mb/s) metro edge rings (or loops) are used to aggregate customer traffic onto larger OC-48/STM-16 (2.5 Gb/s) or OC-192/STM-64 (10 Gb/s) metro core rings [3,35, 37,641. A sample overview of this hierarchy is shown in Fig. 8.2 and details follow. As will be seen, these current metro architectures will have a large impact on future evolutions, with similar edge and core delineations being carried forward (Sections 5 and 6). 2.1 METRO EDGE RINGS The metro edge is commonly defined as the space between high-speed (metro core) CO locations and customer access facilities [35, 651. Most metro edge rings span 20-65 km (average span lengths of 1 0 4 0 km) [64], and collect traffic between several customer sites [35] (under lo), running at either OC3/STM-1 or OC-12/STM-4 rates. (Note that these rings are equivalent to the junction layer defined in [4]). The key network element in metro edge rings is the add/drop multiplexer (ADM) node, which aggregates traffic from low/moderate-speed interfacedfacilities in the access layer (i.e., rates such as DS1, DS3, up to lOO’sMb/s). Metro edge ADM devices are usually linked to customer premise (CP) networks that groom client traffic onto TDM tributaries [4]. Examples can include digital loop carrier (DLC) setups, enterprise networks (Le., T1 leased lines), telephony public branch exchanges (PBX), etc. As the boundary between the metro edge and customer access continues 334 Nasir Ghani et al. to blur, many ADM nodes are also providing aggregation capabilities for smaller DSO rates. Such “edge” aggregation significantlyreduces required port counts versus more centralized “back-hauling” configurations [1341(detailed subsequently), i.e., “multiplexer mountain.” Because most edge traffic is usually outbound from the immediate localized region, metro edge rings exhibit strongly hubbed tratXc-demand patterns [35, 541, and hence are well-suited (bandwidth efficient) for SONET/SDH unidirectional path-switched ring (UPSR) [8] designs. UPSR schemes use dual counterrotating rings and perform 1 1 receiver bridging [4, 51 to reduce the need for more complex (expensive) protection signaling. Metro edge ring tratXc is routed onto larger metro core rings at appropriate IOF CO hub locations [64]. Ring interconnection can be done in several ways [4]. A simple means is via direct connection of the associated ADM nodes (i.e., metro edge and core rings). Alternatively, more advanced (expensive) synchronous switching digital cross-connect (DCS) functionalities (Fig. 8.2) can also be used to perform fine-granularity time-slot interchange (TSI) and spatial switching [5]. In other words, DCS nodes basically perform as bandwidth managers, providing tributary termination, segregation, and grooming (i.e., multiplexing/demultiplexing)capabilities at various levels, namely VT1.5 (1.544Mb/s) or STS-1 (51.84Mb/s). Namely, incoming tributaries are fully demultiplexed (e.g., into STM-1 streams [3]), even if they are “pass-through,” and these are either dropped locally or cross-connectedhegroomedwith other tributaries onto larger outbound tributaries. Such centralized DCS-based processing is commonly called “back-ha~ling~’ Also note that DCS[4]. based ring interconnection can be done either via directly subtending ADM nodes on to a separate DCS (nonintegrated) node or by a more advanced integrated DCS node [2, 1341 with added ADM functionality. Note that the latter approach saves the cost of two ADM nodes [134] and reduces equipment cost, footprint space, and management complexity. However, integrated DCS nodes increase network vulnerability, as a failure can affect both rings. There are two types of DCS nodes, namely wideband DCS (W-DCS) and broadband DCS (B-DCS) [4, 5, 1341, largely discerned by their grooming granularities. W-DCS nodes handle tributary rates down to DSlNT1.5 levels and therefore enable fine granularity bandwidth control [4, 51. As such, they are well-suited for metro edge ring bandwidth management, where tributary rates are smaller due to closer customer proximity. Meanwhile, B-DCS nodes can cross-connect signals from DS-3 (Mb/s) up to OC-48 (2.5 Gb/s) and have granularities of STS-1 [91]. As such, these nodes are much better suited for metro core networks handling larger increments andlor interfaces over OC-1. In general, given the complexity of SONETISDH standards and the differentiating augmentations added in many vendor solutions, multivendor network element interoperability is rather difficult. This makes it more likely that a single vendor’s solution will be deployed in a ring, and DCS nodes used to + 8. Metropolitan Optical Networks 335 independently groom traffic between rings [134]. For more details on DCS designs, refer to [4, 51. Traditionalmetro networks have employed a “multilayering”[38]approach for asynchronous data transport, where enterprise customers use leasedline services (Tl, T3, OC-n) to interconnect their private networks. In essence, this approach has evolved from a gradual layering of data networking protocols onto voice-based infrastructures. Specifically, data protocols are mapped onto SONET frames via intermediate “adaptationyy protocols (equipment) such as ATM or frame relay (switches). As a result, different “layersyy used to implement a complete “service” definition, for are example, SONETEDH for protectiordtransport, ATM for t&c/bandwidth engineering, IP for connectivityh-outing. More recently, however, direct packet over SONET (POS) interfaces have also been defined, where IP packets are mapped to the high-level data link control (HDLC) protocol and then framed into SONET/SDH [ll, 121. In fact, today, many upperend gigabit routers/switchesprovide high-speed concatenated POS interfaces, namely 2.5 Gb/s (OC-48c/STM-l6c) and 10 Gb/s (OC-192c/STM-64c),capable of directly interfacing to larger metro core rings [7, 651. Furthermore, as edge boundaries continue to blur, newer ADM designs are even providing direct “non-TDM” native data interfaces with SONET/SDH mappings [4] (see also Section 6.2). Nevertheless, multilayering increases overall port costkomplexity significantly since more protocol layers are involved (i.e., implementation, maintenance, etc.). Although this was acceptable in past “voice-centric”networks, its viability in increasingly “data-centric” scenarios is highly questionable. 2.2 METRO CORE RZNGS The core fiber plant of many establishedmetro operatorsis very dense, and this infrastructure is typically provisioned as a series of metro core rings running through major CO hub locations, i.e., the core layer [4]. These rings are then usually interconnected as meshes using DCS nodes. Metro core rings (also termed as IOF or collector rings [3, 851) naturally represent a higher level of aggregation, and their boundaries are marked by larger tributary rates. OC48/STM-16(2.5 Gb/s) has been widely used, although recently, the demand for OC-192/STM-64 systems is growing rapidly, with deployment set to double over the coming years [142]. These rings interconnect large CO (ADM) locations across a metropolis, and feed into larger regional and long-haulnetworks for cross-regiordglobaltransport, for example, between serviceprovider points of presence (POP) or inter-exchangeCOS.Ring sizes can vary anywhere from 50-250 km, with average spans of 40-80 km [64]. Note that SONET/SDH performs per-hop regeneration, and hence transmission impairment concerns for most metro distances do not arise [47]. 336 Nasir Ghani et al. Due to larger geographic spread and customer base, traflic demands in metro core rings are much more meshed (versus edge rings), i.e., “any-toany” between arbitrary CO ADM locations. In such scenarios, bidirectional line-switched ring (BLSR) designs [9] are well-suited, allowing for improved bandwidth efficiency via timeslot reuse [75]. Two- and four-fiber BLSR ring designs are available, performing loopback span switching for logical (two-fiber) and physical (four-fiber) rings, respectively [5]. In a typical large metropolisthere can be well over 10 IOF rings, and hence ring interconnection is also required. Due to the increased scale of traffic, larger IOF rings are usually connected via reliable dual ring interworking @RI)/drop-and-continue [4,9,74]schemes(also termed dual homing [5]). These schemesprevent against major outages from all single-hubhnterfacefailures and a number of dual failures [74]. However, these schemes come at a price, namely more equipment complexity and fiber requirements. Simpler variants are also possible, such as a single DCS node between two subtending ADM rings, etc. [74, 1341. Note that BLSR-UPSR interconnectionis also specsed [9], applicablefor edge/core ring interconnection. In some cases, DCS (W-DCS) nodes are also used to interconnect metro core rings in larger “mesh” network topologies across a large metropolis. Mesh networks are typically deployed when internode fiber connectivity is significantly higher than that in rings (e.g., beyond degree two). DCS mesh networks provide path-level mesh restoratiodprotection capabilities(logicallayer recovery [5]), and generally, these schemes are usually much more efficient in terms of spare capacity utilization due to higher oversubscription [74]. Both distributed and centralized DCS recovery schemes are possible [SI, but most solutions are proprietary vendor-based offerings (see also [10,741). However, DCS mesh restoration schemes require more complex software (versus selfhealing rings) and very careful capacity preplanning. Hardware complexity also increases significantly, and intermediate demultiplexing is very inefficient for handling larger-tributary pass-through operations. Moreover, mesh recovery timescales are usually much longer (seconds to minutes) [5], and therefore most stringent services must still rely upon ring protection. Overall this gives a three-tiered survivabilityhierarchy [5] for large metrohegional networks (i.e., ring protection switching, DCS mesh restoration, and manual operator control). 3. Emerging Trends As stated previously, there are various technological and commercial factors driving today’s metro arena, such as increasing traffic demands, improving access technologies, and industry deregulation. A very brief review is presented here to highlight the growing disconnect with legacy SONET/SDH metro architectures and to better qualify further discussion on solutions and 8. Metropolitan Optical Networks 337 migration strategies. Interested readers are referred to the related references for complete details [22,42, 61-65]. 3.1 DEMAND GROWTH Undoubtedly, rapid growth in network traffic volumes is one of the key drivers for optical networking solutions. In particular, there are various aspects of this growth that are of significance in the metro space, and some details are now presented. 3.1.1 Internet Applications Residential Internet has produced sustained data traffic growth, with close to half of the households in North America now having Internet connectivity [42]. Meanwhile, penetration rates are also growing significantly in Europe and Asia [65]. Applications such as e-mail, FTP, Telnet, and Web are commonplace and other advanced “content-rich” (bandwidth-intensive) applications are also emerging (e.g.. “real-time” applications such as Internet telephony, video-on-demandvideo streaming, Internet gaming, etc.). More futuristic offerings such as multidimensional virtual reality are also envisioned. On the corporate side, many businesses are heavily utilizing existing Internet applications and busy developing new possibilities. For example, Internet teleconferencingkelecommutingis commonplace and Web hostinglmirroring and e-commerce are growing rapidly (incorporating more graphics, sounds, and video content). Other, more distant possibilities such as telemedicine and remote sensing are also being studied. Overall, unlike traditional “delay-insensitive”applications, many newer applications have much more stringent (guaranteed) bandwidth needs. Apart from applicationhandwidth growth, the number of simultaneous peered sessions is also increasing rapidly, further accelerating volumes [60]. In all, the combination of increased user populations, applications, and cconliney’ times is boosting capacity demands tremendously, with compound annual growth rates (CAGR) of over 35% [43]. Also, many studies indicate that Internet traffic exhibits highly bursty (some argue nonstationary), asymmetric behaviors [159], and overall customer demands can be more difficult to predict as compared to legacy voice [22, 601. This contrasts with long-haul networks, where the degree of traffic aggregation is much higher @e.,between large population centers), and consequently,traffic fluctuationsare more gradual. Furthermore, as communication distances expand, larger proportions of end-user traffic, roughly 80%, are destined for wider communities of interest (unlike legacy voice), averaging over 400 km [22, 521, @e., 80-20 rule). Overall, these traffic characteristics are very challenging, and metro networks must scale to handle them properly (see [60] for details). 338 Nasir Ghani et al. 3.1.2 VirtuaVTransparent LANs As corporate clients expand to diverse geographic locations, the concept of transparenthirtual LAN (TLANNLAN) servicesis gaining strong favor, providing seamless LAN interconnectionbetween selected enterprise and remote locations [38] (via port grouping). This will allow enterprise clients to maintain the “LAN-like” feel of their networks, yet scale across larger geographic areas. Note that this is not a novel concept, and some existing protocols such as ATM or frame relay already support LAN emulation capabilities, albeit these are costly and provide limited bandwidth [64, 1331. A key difference today is the use of lower-cost native Ethernet interfaces.Native VLANs represent a very large, emerging metro market given the predominance of Ethernet interfaces [133] (Le., over 80% of enterprise tr&c originates in Ethernet form [46] and low-speed Ethernet (10/100Mb/s) dominates the desktop). Overall, VLAN servicesmay become subsets of more generalized optical virtual private network (0-VPN) services [65]. More recently, there have also been significant efforts to extend Ethernet protocol standards to gigabit rates, namely 10 Gb/s Ethernet (10 GbE) [159, 1601. Specifically, new standards are being defined for both for both LAN (short reach) and WAN (long reach) applications (IEEE 802.3 working group), and these are expected to dominate the high-speed router interface market as components arrive on the market [47]. Note that the WAN interface type will reuse SONET/SDH OC-192c framing and require synchronization [158, 1601, making if significantly different from the short-reach interface. Overall, gigabit Ethernet interfaces will significantlyincrease enterprisetra& volumes in the metro. As an aside, such bit rates will also help reduce port counts on high-throughput Ethernet switches. 3.1.3 Storage Area Networks As information volumes expand, corporations are looking at storage area networks (SAN) for reliable back-up/recovery capabilities. This market has emerged recently, driven by falling storage and bandwidth costs, and is poised for significant growth [64]. SANs utilize various storage protocols (Fibre Channel, ESCON, and recently Gigabit Ethernet), to provide reliable, highspeed, native connectivity between large storage sites in a metro area [65]. Namely, these sites interconnect servers and backup storage devices using SAN hublswitch nodes. Since many large corporations have on the order of petabytes worth of stored data [59], clearly, very large transport capacities are required. Additionally, SANs can also play a key role in disaster recovery applications (see [59] for a detailed deployment study). 3.1.4 Virtual Line Services Given the broad range of client protocols, many operators are looking to provide bit-rate-independent services, also termed “lambda servi~es.’~ These 8. Metropolitan Optical Networks 339 offerings are particularly attractive for enterprise clients wanting to build optical VPNs [155] to interconnect multiple locations via a full variety of interfacedprotocols (e.g., Gigabit Ethernet, SONET, frame relay, ATM, etc.). Also, in the future more advanced services such as bandwidth leasindtrading can also be leveraged hereof Unlike legacy leased-line services, virtual-line services will provide genuine transparency (e.g., even for SONET/SDH overheads), enabling customers to manage their own networks [68]. See [44, 64, 65, 1411 for more details. 3.1.5 Legacy Voice Despite data traffic growth, the legacy voice business is not in decline. On the contrary, private-line telephony service continues to exhibit steady linear growth, with some studies indicating rates up to 7-25% per year [42,65]. This represents a strong, sizable revenue stream for most incumbent operators. Since many operators have yet to deploy “packet-voice” services (e.g., technological, standardization, operational issues), this growth will likely remain for the next several years. Moreover, many users may opt to stay with legacy voice services, unless operators offer compelling, competitive value-adds to their “packet-voice” offerings [65]. 3.2 IMPROMNG ACCESS TECHNOLOGIES With growing application bandwidth demands, there is strong push to bring broadband capacities closer to end-users. Many new (high-speed) access technologies have emerged, such as cable and digital subscriber loop (DSL) modems, high-speed wireless, air fiber, and more advanced optical offerings. The economics of these solutions are key factors here, because end-users are very cost-sensitiveand access infrastructures are difficult to amortize over multiple subscribers (as with core infrastructures). Nevertheless, research shows that average access rates will increase in the coming years, with the exact values contingent upon technology type (see [65]).As these new access technologies undergo deployment, the commensurateneed for high-capacitymetro network solutions will strengthen. Some of these solutions are briefly reviewed. 3.2.1 Cable and Digital Subscriber Loop Modems Residential Internet access is still largely based upon copper-plant modem technology, with bit rates limited to under 56 kb/s. However, DSL technology is now available, utilizing signal processing techniques to boost copper line rates to over 2.0 Mb/s. Many residentidsmall business DSL solutions use an ATM link-layer, and DSL access multiplexers (DSLAMs) help aggregate customers onto larger TDM-based access tributaries (e.g., ATM DS-3 and OC-3 rates). ATM offers good latency control, and hence, some have also proposed c‘voice-over-DSLyy packaging data and voice services [65]. for 340 Nasir Ghani et a. l Meanwhile, cable is another widely deployed “last-mile” solution that offers relatively high access bandwidth. Cable modem technology allows for bidirectional data transmission using radio frequency (RF) separation [19], and typical access rates are in the megabits. Most cable operators are deploying two-way services using hybrid fiber-coax (HFC) networks, and standards are emerging for data over cable [64]. Overall, cable operatorshave a strong, growing customer base and many have successfully rolled out bundled datahoice services. Market research predicts that both the DSL and cable markets will grow by nearly an order of magnitude over the next few years [65]. 3.2.2 High-speed Wireless Technologies High-speed wireless technologies present a good alternative for many new entrants who lack ownership of fiber/copper/cableplants. For example, many wireless operators are moving to deploy “next-generation” wireless solutions (e.g., 3G wireless), with proposed bit rates of up to 2.0Mb/s. Additionally, ultra-high-bandwidth local multipoint distribution service (LMDS) technology is also available. LMDS occupies a higher spectrum range (24-39 Ghz) and yields about 1 Ghz bandwidth for a range of between 1-3 miles (as limited by precipitation and other effects) [MI. is usually best suited for business This customers in dense metropolitan districtslindustrialparks, providing a mix of voice and data services (Tl/OC-3). Other wireless solutions use various “lineof-sight” communication techniques between end-users and hub nodes. For example, multichannel multipoint distribution system (MMDS) offers about 200Mhz bandwidth and ranges between 5-35 miles. This is well-suited for residentiahmall business customers and provides an alternativeto cable/DSL access. Meanwhile, in dense, fiber-constrainedbusiness settings, “air fiber” solutions also offer high-capacity,localized transmissions for distances under 10km. Specifically, these schemes use carefully aligned high-launch-power semiconductor lasers. However, air-fibex systems must contend with many transmission limitations, such as atmospheric loss, scatteringhefraction (via rain, fog, mist, even snow), etc. Although back-up microwave transmission techniques can be utilized, transmission rates will be much lower. Overall, these techniques are less ubiquitous due to the line-of-sight requirement, even though availability levels of over 99% have been stated [23]. 3.2.3 Optical Access Technologies As bandwidth growth continues, there is a strong push to deploy fiber closer to end-users [20,22]in order to overcomecapacity and distance limitations of aging copper plants. Much research has been done on “optical” access architectures such as fiber-to-the-home (FTTH) and fiber-to-the-curb(FTTC) [19]. In particular, many solutions based upon passive optical networks (PONS) [2, 3, 191 have been investigated. Earlier PON schemes used power-splitting setups with time-slotted multi-access (TDMA) protocols, but these designs 8. Metropolitan Optical Networks 341 suffer from excessive power loss and ranging limitations [58]. With the emergence of cost-effective DWDM mudde-mux devices and amplifiers, higher-capacityDWDM PON architectureshave now emerged [58,168]directing wavelengths to end-user terminals (e.g., via wavelength gratings [19]). DWDM PON solutions provide enhanced security since optical channels are passively routed to receivers and not broadcasted. A lot of research has been done on different PON architectures (e.g., combining TDM and DWDM PONS) [5, 19, 21, 1681, and commercial offerings are even available. With improvingeconomicsthese architectureswill dramaticallyimprove accessrates (gigabits range). Code-division multiplexing (CDM) is an asynchronous spread-spectrum technique that has been widely applied in satellite, wireless, cable, and more recently, optical networks. Optical CDM permits bandwidth sharing by distinguishing signals via distinct spectral or temporal codes and can be applied to shared medium access networks (e.g., passive star coupler configurations [3, 171). Code switching can be done using low-cost passive (linear) optical devices (spectralfilters) and only requires a single optical carrier frequency (see [181). Additionally, the added signal processing capabilities may also enable consistent operation over a whole range of fiber types (e.g., single mode, dispersion shifted, even older polarized mode) [44], a capability difficult to achieve in high-speed TDM systems (10 Gb/s or beyond). There are many emerging advances in coding devices and code design [ l q , and this technology may soon come on the market. 3.3 INDUSTRY DEREGULATION Competitive pressures and new regulations have resulted in a global push to privatize and deregulate telecommunications markets. Deregulation has progressed rapidly in North America and Europe, and other markets (Asia, Latin America) will likely follow suit. Within the metro arena, the effects of deregulation are very acute, as entry costs tend to be lower than those in long-haul markets. This has created an abundance of operators, ranging from incumbent local exchange carriers (LECs), newer competitive local exchange carriers (CLECs), cable operators, and even (traditionally long haul) interexchange carriers (IXC) and utilities corporations. Given this large operator pool, undoubtedly competition is stiff [133]. For example, many IXC companies are actively expanding into the metro/access markets to feed expanded capacities on their long-haul DWDM buildouts. Meanwhile, CLECs are aggressivelybuying/leasing/layingdark fiber to build out their infrastructures [65]. Also, a growing number of utility companies are entering the telecom market in order to diversify and increase their revenue base (usually as “carriers’ carriers” [64]). These outfits own “right-of-way”into most metro areas, a crucial advantage that allows them to lay new fiber routes. Most utilities with telecom operations are leasing out dark fiber, and this is also expediting 342 Nasir Ghani et al. market entry for many other nonincumbent operators. Moreover, utilities can also leverage their large regional customer bases and extensive service and construction capabilities.In all, deregulationhas driven down bandwidth costs significantly [62], and operators are increasingly competingvia advanced service differentiation capabilities.For further details, refer to [62,64,65]. 3.4 CHALLENGESAND REQUIREMENTS In light of the above trends/developments, legacy metro networks are now facing many limitations With increasing demands and improving access technologies, capacity exhaust on metro rings is a major issue [5, 41, 62, 681 for OC-48/STM-16 and in some cases even OC-19USTM-64 bit rates. In fact, market research indicates that OC-48L3TM-16 is fast becoming the new unit of cunrency among carriers, replacing lower DS3 rates [63, 641. Now SONET/SDH technology really only offers two lengthy solutions to capacity exhaust, either increase single-channelTDM rates (e.g., OC-192,OC-768) or deploy new rings (i.e., “stacked rings”). Neither of these are particularly scalable, and hence are aptly termed “fork-lift” upgrades [64, 1411. The former is expensive, requiring equipment upgrades on all ring nodes, and usually gives only fourfold capacity increases. Meanwhile, the latter is even more expensive, requiring complex forecasting [68] and new fiber routes and TDM gear at ring add/drop points (although fewer nodes can be deployed initially to reduce costs [SI). Fiber expansion costs are dominated by right-of-way and construction costs, and albeit high, can vary significantly (see [45, 62, 711). Although it is relatively cheaper to pull fiber through existing conduits [22], in many cases conduit exhaust is also occurring. Additionally, electronic crossconnects must still demultiplex large incoming tributaries (e.g., OC-48) down to the STM-1 levels and switch/recombine on the outbound side [3]. This increases electronic switching fabric complexity and power consumption and is not scalable in the least for multiple ring hops. Overall, these unscalabilities present a huge capacity bottleneck between increasingly higher-speed endusers and abundant long-haul DWDM bandwidth, often termed the “metro gap” [36], see Fig. 8.3. Apart from capacity limitations, the SONET/SDH “multilayering’y paradigm is overly outdated for dynamic, data-centric settings For example, IP volumes are surpassing legacy voice levels and Ethernet ports easily source most Internet data traflic. Yet at the same time, data interface rates are highly incongruent with SONET/SDH TDM hierarchies, and resultant mappings yield large bandwidth inefficienciesand consume more ports for a given data throughput (see Section 6.2). This has accelerated capacity exhaust and proven very challenging for operators, especially as bandwidth costs decline [66]. For example, from a revenue standpoint, (lower-volume) voice still earnsmore “per-bit” transmitted than data, yet data growth requiresmuch more expenditures. Moving forward, metro operators must also contend with 8. Metropolitan Optical Networks Access-to-lonp-haul“metro-pad’ .Mainly SONET/SDH-based infrastructures *Low provisioning flexibility, scalability *Abundant core capacity left stranded 343 ...................... .10-100 Mb/s, 1 Gbls Ethernet ,” ,’ Metro core TributaryRates .OC-48,OC-192,oC-728 .I,IOGbls Ethernet Solution Technologies .DWDM transmission ,DWDM oADM rings .DWDM OXC meshes Gbls Fiber Channel ,200 Mbls ESCON *Cable video (DI) .xDSL traffic e1 / / : ‘ \ ‘\\\ ‘-.. Solution Technologies .SONET/SDH (0(3-12/481192) .DWDM edge rings -“Next-generation” SONETISDH -MSPP platforms *Resilientpacket rings , s ‘ ‘%\ , ,, ,“ ‘-..\ ----_______.__----- Fig. 8.3 Metro coreledge network taxonomy. rapid traffic variations (e.g., volumes, geographical locations, client protocols). Again, legacy TDM architectures are very restrictive here, as (re)provisioning across multiple SONET/SDH rings requires careful capacity planning and usually takes weeks [ 1421. Moreover, supporting “specialized” high-speed SAN data protocols, such as ESCON (200 Mb/s), FICON (1.06 Gb/s), or Fibre Channel (1.O Gb/s), is even more difficult as no standardized mappings exist [141] and related bit rates do not fit well with SONET/SDH tributaries [64,141] (see also Section 6.2). Traditionally, dedicated fibers are provisioned here and this is very inefficient [64]. Finally, SONET/SDH provides very little latitude in service definitions (e.g., fixed bit rates, full protection overheads). Overall, it is becoming increasingly evident that SONET/SDH “multilayering” is not an ideal solution for the metro market and suffers from the “lowest common denominator” effect, i.e., where one layer (particularly SONET/SDH) limits the scalability of the entire network [154]. Undoubtedly, the requirements for new metro solutions/platforms are plentiful (e.g., see [44, 1411). Perhaps paramount is the need for abundant bandwidth to “future proof” networks for unexpected demand growth. However, as cost is usually an overriding concern in the metro, newer solutions must reuse existing fiber infrastructures and allow for more cost-effective, modular capacity expansions (Le., “pay-as-you-grow” [22]). Here, a related crucial requirement is bandwidth transparency, Le., the ability to support multiple protocols/signals over a single platform [44], unlike SONET/SDH. Bandwidthtransparent metro platforms can significantly reduce operations and management costs [22]. More importantly, transparency enables continued support for 344 Nasir Ghani et al. legacy equipment (e.g., voice, private line), and does not require operators to forgo their investments in SONET/SDH gear (i.e., “backwards compatibility” for cost-effective migration). Still, pure capacity expansion will hardly suffice, as operators need intelligent “network-level” provisioning capabilities to support a full range of client protocols and applications (i.e., “bandwidth-ondemand” [64]). Specifically, metro platforms must efficiently allocate capacity resources and at the same time provide very selective handling in order to enable competitive services (SLA) differentiation. Service differentiation can be achieved in many facets, such as turn-up speed, channel quality, priority, protection levelshpeed, etc. [156]. In particular, service protection is a key discerning factor, and SONETlSDH timescales must be matched. Overall, it has even been argued that transparency and rapid, intelligent service creation capabilities are even more important than raw capacity and equipment consolidation [64,65]. Additionally, with increased fiber leasing at dense multicarrier colocation points, new solutions have to address floorspace/footprint and power consumption concerns [45, 1331. Finally, given the plethora of competing vendors, standards-based interoperablenetwork control and management will become more important as operators gradually induct differing gears. Ideally, the highest degree of interoperability would be multivendor interconnection at the mid-span meet level (hardware and software levels) [22]. Optical DWDM technology can meet many of the above requirements for the metro space. On a very high level, the salient features of DWDM include scalable capacity, transparent carriage/legacy support, efficient largetributary switching, rapid “payload-agnostic” survivability, modular expansion, reduced footprints, etc. On a more detailed level, however, the metro environment also poses challenges for DWDM applicability, ranging from market-related cost/migration concerns to more detailed technical groominglinterworking issues. The overall induction of DWDM technology into the metro, therefore, is significantly more complicated than in the long-haul market. In general, metro DWDM costs will tend to lie more in advanced network elements and not expensive line transmission systems (as in the long-haul space) [40, 621. Metro DWDM solutions are now discussed in detail. 4. Enabling Component Technologies Continuing advances in optical component technologies and declining costs are making metro DWDM increasingly viable. These “enabling” technologies include lasers, amplifiers, filters, fibers, and switching devices. From a market standpoint, it is projected that the collective effects of increased production capacities, improved fabrication techniques, and competitive pressures will yield annual price erosions of 8-20% [61]. As a result, key networking functions, traditionally performed in the electronic (SONET/SDH) domain, are now possible optically, such as dynamic channel add/drop, 8. Metropolitan Optical Networks 345 cross-connection/switching/protection, amplification, and even performance monitoring. Moreover, many of these developments represent relatively early steps in the components field, as future advances in integration (microphotonics) will condense multiple devices into single integrated optical circuits and further curtail size and cost [29]. The main DWDM component technologies are now briefly reviewed. 4.1 BROADBAND AMPLIFIERS Multichannel optical amplification was the primary enabler for long-haul DWDM, as increased amplifier spacing eliminated more costly electronic regenerators (per-channel, shorter distances) [1071. In particular, erbiumdoped fiber amplifier (EDFA) [1-31 devices have seen strong commercial acceptance, and related costs continue to decline. These devices yield good performance levels (noise levels, gain flatness)for C-band (15361565 nm) and L-band (1570-1620 nm) amplification, boosting span distances to hundreds of kilometers. Note that L-band amplification requires increased pumping due to higher attenuation. Lately, more compact erbium-doped waveguide amplifier (EDWA) devices are also coming to market, in addition to variable optical attenuator (VOA) and liquid crystal technologies for amplifier power equalization (see [81]). However, erbium-based devices do not cover the currently unused S-band region (1480-1 5 10nm). Consequently, broadband Raman amplification schemes can be used to extend amplification beyond the C- and L-bands (see references in [%I). The related low noise figures here make ultra-long-haul transmission possible. Overall, even as costs decline, amplification still represents a significant component of overall network cost. Now strictly considering metro distances, the need for amplification may vary (e.g., dense metro ring drops can be under 10km, whereas larger outer-city spans can exceed 200 km).Nevertheless, as operators expand their networks and utilize more complex optical gear, optical amplifierswill be needed to compensate for transmission and nodal losses (Section 5.2.2). 4.2 FILTERLNG TECHNOLOGIES Filtering is key to wavelength channel (band) management (Le., multiplexing/de-multiplexing and pass-through). Today, improving technologies are yielding increasingly dense channel spacings and higher channel counts, for example, commercial 100Ghz (0.8 nm) and 50 Ghz (0.4 nm) filters give 40 and 80 channels. respectively (both C- and L-band). Overall, three filtering schemes are common in the metro, namely thin film, planar waveguides, and fiberbased grating. Thin-film filters are extensivelyused for wider channel spacings (e.g., 200, 400 Ghz) and show good temperature stability and passband isolation. However, these filters have difficulty in achieving 100Ghz spacing [30]. Meanwhile, planar waveguides, such as bulk arrayed waveguide gratings 346 Nasir Ghani et a. l (AWG), can give 100-Ghz spacing and lend well to high integratiordchannel counts (althoughtemperaturestability and insertionlossesconcerns can arise). Finally, fiber-based gratings are good for narrow channel spacings (25 Ghz, 0.4 nm demonstrated [30]), but usually give lower channel drop counts Newer free-space dense gratings are also detailed in [58]. Additionally, tunable filters are also emerging, and will represent the next evolution in this space (e.g., see tunable thin-film filters [Sl]). 4.3 LASER SOURCES Ten years ago, lower laser transmitter powers rarely allowed individual span lengthsto exceed 40 km (i.e., about 1mW power) [SS]. Nevertheless, continuing advances and new integration techniques have yielded lower-cost, narrowlinewidth ITU-T grid laser sources with good thermal/wavelengthstabilities. In particular, directly-modulated designs, such as distributed feedback laser (DFB) [3] sources, have found good favor in the metro space for bit rates up to 2.5 G/bs. As bit rates increase to 10 Gb/s (and beyond), more powerfullexpensive externally modulated designs will be required to overcome dispersion issues (see [3] for details). More importantly, tunable lasers [32] are also commercially availableand related pricepointsmay soon become competitive with fixed laser costs. Tunability will allow for an added level of flexibility in wavelength provisioning (see Section 5.2.3) and also help reduce sparing costs versus complete laser bank setups. Additional advances are also noted for multifrequency lasers (MFL) [28]. 4.4 FIBER TECHNOLOGY Over the last decade single-mode fiber (SMF) technology has largely displaced earlier multimode fiber (MMF) types in most metro domains [25,71]. SMF is well-tuned to single channel 1310-nm (SONETKDH) transmission and also has relatively low attenuation in the 1550nm region (0.2-0.3 db/km [47]). However, at higher bit rates of 10 Gb/s or more, C- and L-band chromatic dispersion in SMF is a major problem, as opposed to attenuation, averaging 17ps/km/nm [25]. High dispersion is very problematic for dense OC-192/STM-64systems, and dispersion compensation will likely be required [106, 1071. However, many new dispersion-optimized fibers, termed non-zero dispersion shifted fiber (NZDSF) or negative dispersion fiber (NDF), have been proposed, yielding longer uncompensated IO-Gb/s transmissions up to 200 km (versus under 60 km for standard SMF) [37,53,81]. Many new metro deploymentsare utilizing these optimizedfiber types, and moreover, dispersion compensation on existing SMF spans can be done using negative-dispersion fiber coils. Other “metro-optimized”fibers have also been proposed to expand the 1350-1450 nm spectrum (to about 100 channels) by reducing hydroxil ion contents (i.e., remove “water-peak”[25,42]). 8. Metropolitan Optical Networks 347 4.5 ADVANCED SWITCHING TECHNOLOGIES Optical switchingenables a host of network-levelroutinglprotection functions and is becoming increasingly important as line rates increase. A variety of switch technologies are available, such as waveguides (lithium niobate, semiconductor optical amplifier (SOA) gates, thermo-optic (silica- and polymerbased), beam-steering (micro- and macro-mechanical, free-space prisms), liquid crystal, and recently even bubble jet (see [33, 91, 1681). In particular, micro electro-mechanical system (MEMS) [33] switch designs utilizing tiny switching mirrors are in strong favor today. MEMS devices give submillisecond switching times (700 ps) [33,91]and have favorablecharacterizations (ie., low crosstalk levels, high extinction ratios, and negligible polarization losses, see [33]). Moreover, emergent three-dimensional MEMS designs are promising much larger port counts of hundreds or more. Thermo-optic switches, meanwhile, have higher power consumption and exhibit medium extinction ratios [79]. These switches are usually integrated onto planar lightwavecircuits (PLC) but may not be as scalable as MEMS switches. Port count scalability is also a concern for liquid crystal technology. Overall, many advances are expected in this field over the coming years (see also [29, 1681). 5. Metro DWDM Solutions Maturing optical technology will inevitably push electronics towards the edge of the metro domain. Namely, DWDM technology has heavily permeated long-haul backbones, and as operators move to provide genuine “all-distance” wavelength services, its induction in the metro arena is very timely. In particular, the terms “transparent” and “all-optical’’ are commonly used to refer to entities (nodes, networks) where client signals travel entirely in the optical domain, without any form of opto-electronic processinglconversion [121. By and large, the emergingmetro optical taxonomies will continue to mirror existing edgekore delineations (Section 2) [64, 651, as shown in Fig. 8.4. Namely, DWDM technology will permeate the metro core as it evolves to support intelligent, rapid provisioning of large, interconnect capacities. Over time, optics will also migrate into the metro edge, blending in with advanced IC technologies to form an intelligent, opto-electronicmetro edge (Section 6). The metro edge will play a vital role in grooming a diverse array of traffic protocols onto wavelengths. DWDM is a very good fit for the metro core, providing scalable capacity (terabitdfiber), reducing signal regeneration needs, and yielding format transparency. There are various possible metro DWDM solutions, ranging from high-density, point-to-point transmission and static wavelength routing systems to more flexible wavelength routing networks (e.g., ring, mesh) capable 348 Nasir Ghani et al. Metro core nngs. T h u d nng scheme$ (BLSFUBPSR decigns meshed demands) Metro-regional nng (100- “Next-generation SONET’ nng (oC-12/48/192) I rnbda loo Edge-multiplexing, multiple formathit-rate clients (TDM. IP, Ethernet, ATM. FR,Escon, Fiber Channel etc) Layer twolthree edge graommglswitchmg. increased TDM multiplexing gams Fig. 8.4 Overall architecture for metropolitan optical networks. of advanced SLA provisioning. However, the choice of a particular solution is very operator-specific and depends upon customer profiles, demands, economics, and current migration strategies. Some of these topics are now discussed. Note that DWDM transmission can encounter many challenges, especially for larger channel counts, higher bit rates, or older fiber types [106108]. Transmission concerns have been well studied [l-31, but wherever pertinent, their implications herein will be discussed. These include fiber-related concerns, such as chromatic and polarization-mode dispersion (CMD/PMD) and nonlinear effects. Additionally, there are componentrelated impairments such as interchannel crosstalk, loss, amplifier noise, and nonlinearities. 5.1 POINT-TO-POINT TRANSMISSION SYSTEMS Point-to-point DWDM systems represent “first-generation” (metro) optical networking systems [6, 22, 51, 731 and are usually used for fiber relief on congested spans (i.e., “virtual dark fibers” [1313). Commercially available components (filters, lasers) can now yield over a hundred channeldfiber with up to 10 Gb/s per channel, and this is usually more economical than traditional means of capacity expansion (i.e., deploying new fiberhystems or increasing TDM system rates) [6] (Section 3.4). These systems basically comprise 8. Metropolitan Optical Networks 349 client interface cards, wavelength transponderheceiver pairs, filter modules, and wavelength mudde-mux devices, as shown in Fig. 8.5. The transponders perform wavelength modulation (from short-reach client interfaces, typically 1310nm) for client signals. The corresponding wideband receivers translate received wavelengths, outputting at either 1310 or 850 nm. Alternatively, newer SONET/SDH devices and even IP routers can be equipped with ITU-T compliant lasers (1550-nm band) and wideband receivers, allowing for “direct” DWDM interfacing. Overall, these elements will allow different protocols to traverse the fiber and can eliminate separate dark fiber requirements for “nonTDM mapped” payloads (e.g., ESCON, Fibre Channel, cable video). Note that Ethernet interfaces (1.O, 10 Gb/s) may require added receiver-side buffering to handle flow-control features over larger distances, depending upon level of buffering at client gears. Various setups can be used for the filtering modulehtage in Fig. 8.5. For example, operators who want to modularly expand their channel counts can use channel interleaving and banding filtering techniques. Specifically, band filters add/drop sets of wavelengths, and subsequently, interleavers (with appropriate offsets) and filter cascades achieve closer channel spacings [313. The latter is a very appealing solution since wider channel spacing technology is more mature and less costly [30]. Thin-film filters are in wide use today [64], and well-suited for band filtering (Le., dropping channel sets). A sample setup is shown in Fig. 8.6, where two “coarse” (thin-film) 200-Ghz filters are used to generate 100-Ghz spacing. Alternatively, serial add-drop combinations are also possible, albeit these are usually more disruptive (see [89]). For larger drop counts (i.e., more channel density), meanwhile, single high-channel count filtering devices are more attractive [89] (e.g., wavelength gratings [80] or AWG Wavelength laser transponders Wideband receivers ,_____._..........._~~~~~~~.... Fiber span (OMS) protection ( l + l , l:l,or 1:Nvariants) . I Standard short-reach \ Passive splitter ( l + l protection) or 1x2 switch (I:l protection) Optional filter modules (banding, interleaving) \ interfaces (I330 nm) Fig. 8.5 Point-to-point DWDM transmission. 350 Nasir Ghani et al. C-Band L-hand 200 Ghz spacing 100 Ghz spacing 200 Ghz spacing I530 nm 1565 nm I530 nrn 1565 nm L-hand 1565 nm 1620 nm Fig. 8.6 Interleavindbanding techniques for finer filter spacing. filters [90]). This setup requires less components and also yields lower loss than cascaded setups [SO], but has higher initial per-channel costs. Note also that client bit rates may affect the chosen channel spacingdfilter setups. For example, 10-Gb/s OC- 192/STM-64 transmission at 50-GHz spacing (versus 100 Ghz) will mandate more stringent channel passbands to limit crosstalk. Depending upon the distances involved and nodal insertion losses, optical amplification may be needed. Nodal insertion losses are usually dominated by the overall filtering setup (number of stages, components), and typical values are below 2.5 dB for band-drop [47]. For shorter SMF spans (under 50 km) and bit rates under 2.5 Gb/s, optical amplification is usually not required (using DFB lasers). For example, in [80] a nonamplified 40-channel C-band (100-Ghz spacing) point-to-point transmission system is demonstrated, with insertion loss across a grating filter measured at under 4.0 dB for 80-km transmission. Meanwhile, field trials using amplifiers are also presented in [49]. At faster 10 Gb/s channel rates, PMD becomes the most limiting factor, reducing SMF span lengths to about 60 km [27]. Hence for larger metro spans, amplification and dispersion compensation will be required (see also Section 5.2.2). Note that forward error correction (FEC) [79,88] can also be done at receiver interfaces to boost gains, but this limits transparency and increases interface costs considerably at higher line rates (e.g., SONETISDH or OCh digital wrappers overhead checksum processing, Section 6.1). 8. Metropolitan Optical Networks 351 Given the large increase in per-fiber traffic, various protection schemes have been proposed for point-to-point transmission systems (see [40,96, loll). Specifically, these include optical multiplex section (i.e., fiber span) or optical channel (OCh) level protection setups. Fiber protection can be done in a dedicated 1 1 manner [2, 40, 96, 100, 1011 and requires passive splitters to bridge all wavelengths onto two fibers. This is a simple setup and precludes any (head-to-tail) protection signaling [75], Fig. 8.5. Meanwhile, more complex “nondedicated” 1 : 1 or 1 : N protection schemes are also possible [40], and these are usually more resource efficient (since protection resources can be shared by multiple working fibers or lower priority traffic [96]). Still, 1 : 1/1: N protection requires active switchingdevices at the sender and fast protection signaling and additional electronicsto exploit fiber reuse [103]. Overall, fiber-level protection can reduce the need for electronic (client) protection equipment significantly (e.g., optical fiber protection with 1 :N SONET/SDH (electronic) protection is found to be very efficient [22, 40, loll). This finding will likely hold for more advanced optical ring and mesh networks also. Meanwhile, for optical channel protection in point-to-point systems, bridginglswitchingis most commonly used to provide selective 1 1 protection (not shown in Fig. 8.5). Although this is a simple approach, it requires more hardware (splitters, switches) and wavelengthsthan correspondingfiber protection, especially if all demands are to be protected. + + 5.2 O P T I C .RING NETWORKS Given the success of initial point-to-point DWDM deployments, multihop (wavelength) lightpath [1, 21 provisioning is gaining strong favor. Specifically, the goal is to maintain large tributary signals within the optical domain as much as possible, thereby exploiting the transparency and costper-bit advantages of DWDM. Since ring topologies are ubiquitous within the metro space, multihop DWDM rings will clearly represent a logical progression towards multihop wavelength networks (from simpler point-to-point setups). DWDM rings draw strongly from SONET/SDH concepts, essentially replacing timeslots with wavelengths and performing “optically equivalent” channel operations (i.e., add-drop, pass-through, protection). However, unlike SONET/SDH, these rings offer high-capacity scalabilityand transparency and permit multiple data rates. In particular, optical bypass eliminates the need for detailed electronic knowledge of, or even access to, client signals [6] and yields significant cost savings over traditional ADM/DCS nodes [3] (e.g., 40% or more in ADM costs [68, 1361). For example, in high bit rate TDM rings, traffic add/drop ratios are typically 25% [89, 1321, yet full electronic multiplexingldemultiplexingof all (concatenated) tributaries is still required and is very unscalable [92]. Many different DWDM ring architectureshave been proposed, varying from simpler static setups (i.e., fixed nodes) to more advanced 352 Nasir Ghani et al. sharing schemes @e., dynamic nodes) [92]. Details are now presented along with considerations for the metro space. 5.2.1 Static Rings In many cases, metro demands are relatively long standing, with holding times of 3 months or more [56]. Moreover, edge ring networks wiU continue to show very “hubbed” demands, since data traffic exhibits larger communities of interest (Section 3.1 .l) (i.e., communication remains “fixed” between access locations and central outbound gatewayshubs). Carefully note that there is a linear relation between node count and required wavelengths for supporting hubbed traffic patterns in optical rings [3, 541. Given that metro edge rings typically have lower node counts (between two and six [35]) collectively, the above implies that lower-cost static/passive WDM rings [5] with moderate channel counts (e.g., 16-32 wavelengths) are very amenable solutions [62] (i.e., (‘second-generation”optical networks). In particular, optical UPSR schemes, i.e., dedicated protection rings (OCh-DPRINGs) [73,75],are very efficient for hubbed traffic patterns and lend well to static realizations. OCh-DPRINGs use two unidirectional fibers, working and protection, with counterpropagating directions. Channel protection is done using dedicated 1 1 setups, requiring head-end bridging (via an optical splitter) and receiver-based switching (e.g., based upon received power levels) [132]. Specifically, in case of a link or node failure, receivers switch to bridge traffic along the protection route (see Fig. 8.10). This is a relatively simple, robust architecture and does not require any complex protection-signaling protocols. Moreover, path switching also precludes any (optical) performance monitoring inside the ring, unlike more complex bidirectional designs (Section 5.2.2) (see [73,75] for full details). Now since user routes are essentially predetermined for hubbed demands, channel routing and wavelength assignment (RWA) [l,21 can be greatly simplified [35]. Specifically, by allocating a unique, fixed set of wavelengths (band) to each node, as per projected demand, wavelength blocking can be avoided (see [56]). Note also that more robust dual-hub rings are also possible [83, 1371. Nevertheless, UPSR rings are very inefficient for meshed demands, largely limiting metro core applicability, see Section 5.2.2. The basic building block of an optical ring is the optical ADM (OADM) node [67]. Generically, a static-ring OADM node is basically a “back-toback” configuration of point-to-point (DWDM) transmission systems [35, 851 (note that “back-to-back” interconnection has been routinely done in earlier SONETISDH and PDH networks [4]). Static OADM nodes use passive optical filters to implement their k e d wavelengthrelations. A generic overview of a static two-fiber UPSR ring OADM is given in Fig. 8.7, comprising band and mudde-mux filters and optional amplifiers. Here, various passive filtering technologies can be used, such as thin-film filters, fiber Bragg gratings, Mach-Zehnder filters, AWG devices, and diffraction gratings (see [SS, 90,921). + 8. Metropolitan Optical Networks 353 Incoming east (working) I Incoming west (protection) Outgoing east (protection) \I I Wideband receivers 999 k .9.3.;$. transmitters Laser I Fig. 8.7 Fixed-wavelength OADM node. Akin to the point-to-point case (Section 5. l), varying filteringhterleaving setups can also be used, depending upon the application desired. This is commonly termed “right-sizing’’ [35]. For example, given large add/drop ratios, as required in UPSR hub nodes (Fig. 8.8), it is more economical to fully muxldemux all channels via a single device [89] (e.g., AWG). Meanwhile, for remote “nonhub” nodes, thin-film band filters (Fig. 8.8) can be much more economical for isolating a required wavelength band, usually under eight wavelengths [35]. Band filtering lowers insertion losses considerably (about 1-2 dB [47]), and thereby significantly improves link budgets. Moreover, this technique also prevents channel passband-narrowing effects [811 associated with fullspectrum filters, i.e., since multihop filter concatenations (spectral function multiplication) induce time-domain signal distortion [88] (e.g., 25% reduction in 3-dB bandwidth for four filters [Sl]). Note also that modular interleaving techniques can also be used to selectively increase channel densities (e.g., Fig. 8.6). In general, static OADM devices do not use advanced optical subsystems, and thus exhibit lower through-loss and design complexities than corresponding dynamic OADM or OXC nodes [56]. However, optical transmission impairments can still reduce the shorter distance advantages of metro domains. For example, even though amplifiers may not be required for distances under 80 km, signal attenuation concerns can still arise if the number of ring nodes is sufficiently large. Here, “built-in” amplification uses preamplifiers to compensate for transmission losses and postamplifiers to compensate for node 354 Nasir Ghani et al. ,J\ Working 1x2 switch Band mud de-mux at access locations Fig. 8.8 Static hubbed ring using fixed OADM nodes (1+1 UPSRIOCh-DPRING). subsystem losses (Fig. 8.7). However, such “full” amplification can be overly expensive for smaller edge rings, and moreover, excessive amplification can result in noise amplification and even lead to reduced signal-to-noise ratios (SNR). A more cost-effective solution is to judiciously place amplifiers along specific ring segments to ensure link budgets for all paths (working, protection). Nevertheless, this approach requires tedious, manual provisioning to compute and then populate amplifier locations, and is therefore best suited for more static demand scenarios. A sample amplifier placement study is presented in [47] for a static C- and L-band ring of approximately 50 km. It is also found that chromatic dispersion is less limiting for data rates below OC-48/STM-16 (2.5 Gb/s) [47], as will be common in smaller edge rings. In smaller, highly cost-sensitivesettings with less explosive demand growth, coarse WDM (CWDM) [159, 1601technology can also be used for static rings. CWDM is intended for unamplified settings, and can use very wide channel spacings (outside EDFA spectrum, usually about 20nm), all the way from 1310 to 1560nm. Depending upon the fiber type, this can yield about 16-32 channelshber. Note that narrow WDM (NWDM) has also been proposed, using 10-nm spacings in the 1530-1560nm band [30]. CWDM cost savings arise from its use of low-powedwider-linewidth (uncooled) laser sources and coarser filters and the lack of amplification. Wider laser linewidths mitigate 8. MetropolitanOptical Networks 355 (center-wavelength) temperature drift concerns and relax filter stringencies. However, unamplified transmission distances are usually limited to 40 km, and per-channel bit rates are strictly below 2.5 Gb/s. Additionally, it has also been proposed to combine CWDM with (lower) rate-adaptive transceivers to derive some gain from older, higher-loss fiber infrastructures that would otherwise be incapable of supporting high-speed DWDM transmission [159]. This would be particularly appealing in very low-cost settings, giving acceptable capacity improvements without expensive fiber upgrades. Nevertheless, this lower “first-cost” advantage is likely to disappear as EDFA-based DWDM system pricepoints decline [ 1591. Note that lower launch powers and broader wavelength bands require full opto-electronic processing, which is expensive, to map CWDM channels across larger metro core rings (i.e., signal regeneration, relaunch on ITU-T lasers). Hence, CWDM is best suited for smaller domains requiring simple point-to-point or static ring configurations, for example, low-cost interconnection of gigabit routers. Note that from an architectural standpoint, static OADM rings primarily expand capacity and provide large tributary bypass functions. As such, they may not be viable as stand-alone metro edge solutions, especially for smaller client tributaries (see Section 5.4). Instead, these static OADM platforms are best integrated with various “subwavelength-rate” opto-electronic aggregation solutions to improve wavelength efficiency (e.g., DWDM edge rings, Section 6.1). 5.2.2 Dynamic Rings Tr&c demands between metro core CO locations are much heavier and more “meshed” as compared to metro edge rings [38, 54, 871, and therefore the required wavelength counts will be much higher, easily exceeding 50 channels [88]. With increasing customer dynamics, static unidirectional rings (Section 5.2.1) will become very limiting, requiring complex manual wavelength planning and yielding reduced wavelength efficiencies [54]. Here, reconfigurability is a key issue (in addition to scalability), and dynamic OADM rings are very amenable solutions (i.e., “third-generation” optical networks). These architectures use various filtering and switching techniques [56, 921 to implement wavelength programmability (Le., selective channel adddrop). Optical protection is central to OADM rings given the large degree of fibedwavelength multiplexing. Dynamic optical rings are commonly based upon wavelength sharing [73] paradigms, yielding improved wavelength efficiencies and more advanced protection definitions (versus static rings). The former is particularly important in metro cores, as the wavelengthhode-count relation grows quadratically for meshed demands [54]. Overall, the market for such gear is expected to grow considerably [67l, as operators look to deploy more cost-effective architectureswith broader service offerings. Optical 356 Nasir Ghani et al. rings can either be opaque (wavelength convertible) or transparent (wavelength inconvertible) [ 151, although various results indicate only modest gains with wavelength conversion (see Section 5.2.3). The latter offers a common “payload-agnostic’’transport/protection framework, and related standardization efforts have also emerged [ 15,751.This reduces overlapping functionalities in electronic layers (IP, SONETBDH) essentially “delayering” existing “multilayered” setups and reducing equipment/maintenance costs [40]. In particular, service survivability can be offered for native “non-TDM mapped” data protocols, such as Ethernet and Fibre Channel, a crucial advantage [47]. Details are now discussed. Dynamic OADM nodes [67, 851 are the building blocks of a flexible metro core and logical counterparts of dynamic TDM ADM nodes. These elements provide dynamic “on-demand” wavelength add/drop, along with advanced signaling (setup, protection). An overview of a sample dynamic OADM node is presented in Fig. 8.9, comprising incoming/outgoing transport (i.e., mux/demux, filters, amplifiers), an add/drop stage, intelligent signaling/control, and optional subrate DCS grooming. Further generalizations are also possible in which a subset of demultiplexed wavelengths are manually provisioned (very cost effective for significant numbers of static connections). There are several ways to implement OADM add-drop stages (Fig. 8.9). In transparent nodes, Channel setup, maintenance. and protection (APS) signaling Optionalpnd bypass osc control 7I I 1 1 T-----+----------------------------------------l I I I r--lL------------___---------------------$’ Channel bypassladdMux I I 1 Optical @Direct ITU wavelength inputs (e g , IO GbE) Direct wavelength outputs (to wideband receivers) Recewerltransponder bank (addldrop channels) Sub-rate TDM Sub-rate m M Optional DCS stage (integrated or sub-tended) Fig. 8.9 Sample integrated OADMlDCS node (two-fiber design). 8. Metropolitan Optical Networks 357 static filtering can be coupled with active switching devices (i.e., switch arrays [67, 77, 901 or complete fabrics [85]). For example, MEMS or thermo-optic technologies (Section 4.5) can yield under 5 ms switching times [77]. Because OADM nodes have smaller switch requirements (i.e., two or four fiber ports), MEMS technology provides a very good fit as pricepoints continue to decline. Additionally, Mach-Zender and liquid crystal switches have also been tested [88]. Another means of flexible optical adddrop is via tunable fiber Bragg gratings [56, 57, 851, although this requires a filter per dropped channel [56]. Meanwhile, in opaque OADM designs, dynamic add-drop is done electronically via electronic cross-point switches (EXC) or large DCS fabrics [45, 1321. EXC devices use IC technology, such as gallium arsenide (GaAs) [38j, and commercial offerings with high-port counts can process native bit rate channels up to 3.5 Ghz [26]. Although opaque designs can offer varying levels of regeneration (two, three) and possibly even efficient subrate switching (DCS only), scalability to 10Gb/s and beyond is difficult. As a cost-effective tradeoff, optical and electronic technologies can be coalesced into “hybrid’ OADM designs (with intermediate receivers/transponders) (Fig. 8.9), to exploit the benefits of both domains. These issues are discussed more completely in Section 5.2.3. More generally, both transparent and opaque nodes can fully leverage bandingfinterleavingtechniques [89] (Section 5. l), for example, band filters [5q, AWG devices [go], and diffraction gratings [88]. Note that larger metro core channel counts will mandate L-band transmission. As an aside, by appropriatelyexpanded switch-based adddrop stage, local tributary interconnection can also be done (e.g., add-tributary/drop-tributarycross-connection, “hair-pinning” [56]). For amplified all-optical (OADM) nodes, gain equalization is an integral requirement as amplifier behaviors can perturb nonparticipating wavelengths during channel setup/takedown/protectionswitching [26,93]. With increasing metro customer dynamics (site changes, demand variations), gain flatnessmust be maintained over wide input dynamic ranges [79]. This is usually done using output amplifier feedback control (Fig. 8.9), and the related settling times will inevitably determine switchingfprotection timescales. Various hardwarebased equalization schemes have been demonstrated with good results [55, 79, 88, 931. For example, in a severe condition test (i.e., 32 to 1 channel drop, fibercut) millisecond settling times are achieved, with surge levels under 2 dB. Moreover, subsecond network stabilization times have been claimed [88], and as these techniques improve, they will find increasing favor in dynamic metro cores. As important, dynamic gain equalization will also prove invaluable in simplifying installation procedures, precluding operations staff from complex manual tuningkesting procedures. Using the above dynamic OADM designs, various dynamic DWDM selfhealing ring (SHR) architectures are possible. These solutions draw heavily from SONET/SDH concepts and include both dedicated &e., unidirectional) and more complex shared (i.e., bidirectional) rings [l-3,73,96]. Although the 358 Nasir Ghani et al. UPSR concept (OCh-DPRING)was detailed earlier for static “hubbed))rings, it can also be applied to dynamic two-fiber rings. However, this solution suffers from some serious limitations. For example, a protected bidirectionaldemand consumes wavelengths on all fibers [73], preventing any protection resource sharing between demands. Moreover, UPSR rings prevent spatialhavelength reuse [75] (i.e., different connections over different parts of a ring still cannot use the same wavelength). As a result, these schemes w l require more wavei l lengths to provision “meshed))demands, as compared to shared rings [75] (see [54, 1321). Although shared path protection is also possible (e.g., 1 : 1 UPSR) [96], wavelength reuse is still an issue (see [73, 751). Additionally, fiber-cut prevention overheads are excessive with 1 1 path protection, requiring perchannel splitters/switches,and this can be problematic in the metro core, where such occurrences are common (averaging one failure per 10-20 kndyear). Advanced signaling-basedshared protection ring (SPRING) schemes with wavelength reuse capabilities are much better suited for the metro core. Both fiber (OMS-SPRING)and channel (OCh-SPRING)variants are possible and utilize a bidirectional wavelength plan, namely bidirectional channels are routed between the same set of ring nodes and workinglprotection traffic can travel in both directions (wavelengthreuse). Two- and four-fiber bidirectional line switched ring (BLSR) [75, 96, 1321 schemes have been defined for fiber protection, drawing upon earlier SONETISDH BLSR concepts. Generally, fiber protection is deemed more efficient (albeit less selective) for recovering large traffic volumes [82]. Two-fiber BLSR schemes partition intrafiber wavelengths between working and protection groups [1321and perform “loopback” switching during failure events (node, fiber). Specifically, all failed wavelengths are rerouted along protection fibers between failure-adjacent nodes in the opposite direction (i.e., loop-back) (Fig. 8.10). Note that transparent rings require on-overlapping wavelength plans to ensure switchovers [96, 1321. Meanwhile, four-fiber BLSR [82] utilizes two separate fibers each for counterpropagatingworkinglprotection traiEc (Fig. 8. lo), and both loopback and span switching are possible. However, loop-back protection is more disruptive [82] and increaseschannel distances significantly (worst case, nearly double [75]). As a result, related analog impairments can be problematic for larger all-optical rings. Conversely, span switching (four-fiber ring) simply routes all failed wavelengths to a protection fiber between the same nodes, but cannot protect against node failures [75]. Finally, channelized protection is also possible for shared bidirectional rings (i.e., bi-directional path-switched ring (BPSR) or OCh-SPRING [73,75]). Here, full (edge-to-edge) ring switching is performed for failed channels, and this only required fault detection at the “edge” (e.g., at receiver interfaces). BPSR allows for more selective, per-channel protection, and has no counterpart in SONETISDH due to prohibitive signaling overheads for small tributaries. Overall, shared protection rings yield much better wavelength efficiency for ccmeshed” demands due to spatial reuse features, and four-fiber BLSR rings in particular provide the + 8. Metropolitan Optical Networks Two-Fiber BLSR 359 Faher WorXing wavelengths Four-Fiber BLSR Two-Fiber BPSR Fihcr ...... ...... pTotefeCll0” span . . . FTotcet,on . .. Protection fibers (Inner) Working wavelengths Fig. 8.10 Dynamic optical ring protection. most overall capacity (see [3, 54, 78, 1191for analytical details and [73, 751 for complete architectural details). Note that optical ring interconnection is also required, and this is detailed in Section 5.3. Shared ring schemes require real-time distributed protection signaling protocols, termed “optical APS” [75], to implement the above-mentioned protection schemes (fiber, channel). Most likely, operators will demand “SONET-like” recovery times (50 ms). Signaled recovery enables much more flexible, efficient resource utilization, such as sharing of protection wavelengths between multiple working connections (Le., backup multiplexing [96]) and/or by lower-priority traffic [82]. This contrasts with SONET/SDH protection, which requires full overhead for protected demands. Operators can use these features to offer multiple, diverse “wavelength service” levels (SLA categories), such as dedicated, shared, preemptable, and even nonprotected, as proposed in [1561 (see also proposed gold/silver/bronze/coppercategorizations in [65]). For example, carrier voice traffic can be fully protected (1 : l), whereas ISP data traffic can be left as unprotected or even preemptable. Nevertheless, there are no standardized optical APS schemes or protocols today, and much work remains to define architectures that can guarantee “SONET-like” recovery, discussed further in Section 7.2. Fast, signaled recovery will have a strong impact on optical signaling architectures [7]. More generally, signaling architectures will transport many other advanced control protocols used 360 Nasir Ghani et al. in dynamic ring provisioning (e.g., resource/topology autodiscovery, channel setup/maintenance, management, see Section 7). Here, two related architeo tures are possible, namely inband and outband control [3, 7, 161. The former is used in opaque OADM rings, where control information is carried on data wavelengths via designated overhead bytes in “opaque” framing formats, for example, SONET/SDH [9] or OCh (digital wrappers) framing [12, 13, 1451. Both of these framing formats reserve specialized bytes for APS switching, ensuring protection signaling bandwidth. Meanwhile, transparent networks more clearly decouple control/data planes using outband [2, 71 control. The most common approach here is via an optical supervisory channel (OSC) [2,7] on the 1510-nm wavelength, and this requires per-fiber filters, receivers, and lasers (Fig. 8.9). Note that typical OSC filters can add roughly 1.5 dB loss and this needs to be incorporated in the network design. Clearly, outband OSC control standards are strong prerequisites for interoperable all-optical rings (see early discussions in [16, 1451). Dynamic optical rings pose many practical engineering concerns, as highlighted by various studies [38, 49, 50, 51, 54-57, 77-79, 82, 85-89]. In [56], two dynamic OADM designs with differing add-drop stages are compared, one using tunable fiber Bragg grating cascades and the other using an AWG demux and 2 x 2 switch arrays. The latter technique requires more switches and internal fibers and yields higher crosstalk. However, both adddrop stages yield roughly equal add/drop attenuation (2-3.5 dB). In [77], a four-fiber “bridgeand-switch” (1 : 1 shared) path-protection OADM design using 32-channel Gband AWG filters and 2 x 2 mechanical switch arrays is presented. Measurements for varying “intra-OADM” working/protectionpaths indicate insertion loss fluctuations between 2-6 dB, stressing the need for dynamic gain equalization. In an all-optical node cascade test (transmission span losses of 10 dB, about 40 km), results through six nodes show average losses of 11.5dB and good BER performance for up to 5-Gb/s transmissions (positive simulation results for large OADM cascades are also reported in [3]). Meanwhile, lO-Gb/s transmission is found to suffer from PMD effects and temperature control issues with the AWG filters therein. Because C-band dispersion will limit such transmission to under about 60 km on SMF [27] (and intraband dispersion can vary), electronic regeneration (of TDM payloads only) andor DCF modules will be required. Here, detailed analyses are required for optimizing distributed DCF placement in optical rings (e.g., demand, channel routes, etc.), and increased laser powers may also be of concern (i.e., nonlinearities [1061).Note that advanced MEMS-based dispersion compensationtechniques are also being studied [26]. Additionally, [79] demonstrates a 32-channel L-band UPSR (1 1protection) OADM node with dynamic gain control using (PLC-integrated) thermo-optic 2 x 2 switch arrays. System performance was verified for lO-Gb/s transmission over 240 km of DCF through two nodes (distances with SMF would likely be less). In-band FEC is also performed at the receiver interface (via modified SDH frame format), providing about 2-3 dB + 8. Metropolitan Optical Networks 361 improvement. Finally, a four-fiber BLSR ring with protection fiber reuse via low priority traffic is also demonstrated in [82], yielding under 50 ms recovery. Overall, larger metro networks will require careful planning and analyses to ensure user performance levels (see other studies in [49, 57,87,92]). Because many SONET/SDH tributariedrings will be carried over wavelengths (e.g., stacked SONET rings, “path-in-lambda” [39]), protection interworking is a big concern as timescalescan easily overlap. Particularly, signaled BLSR schemescan exhibit cascaded APS behaviors [98] between the two layers, prolonging recovery intervals well beyond 50 ms. For example, 8 1.2-ms DS3 recovery times are measured in [lo41for a stacked TDMDWDM BLSR ring, where the SONET/SDH layer attempts both span and ring switches. These values violate the SONET/SDH 50-ms bound, but may still be acceptable for data services. Meanwhile, nonsignaled 1 1 UPSR (OCh-DPRING) recovery is usually much faster than BLSR (5 ms or less), and may limit APS cascading behaviors. Overall, coordinating interlayer recovery falls under the broader topic of multilayer (protocol) escalation strategies [loo] (see Section 7.2). As far as SONET/SDH is concerned, the simplest practical solution may be an “all or nothing” approach [SI, where protection on one layer is disabled, for example, [ 1041 proposes SONET/SDH carriage on unprotected wavelengths (or bands). + 5.2.3 Transparency Considerations There has been much debate on the pros and cons of optical transparency. Opaque networks perform per-hop opto-electronic conversions and map all client signals to standardized frame formats, for example, SONET/SDH or OCh digital wrappers [12]. Although this precludes full transparency, frame mappings allow for per-hop signal regeneration and detailed performance monitoringlfault isolation [ 12, 1051. This largely mitigates analog impairment concerns (e.g., loss, dispersion, crosstalk) for most metro networks with spans under 100km.Additionally, opaque networks allow for wavelength conversion [l-3, 1151to reduce call-blocking rates. Although many findings indicate that wavelength conversion yields good blocking probability improvements in mesh topologies (10-30%), it is not as effective in rings, as connectivity is limited (yielding high load correlation on subsequent links) [113, 114, 116, 117, 122-1261. More interestingly, judicious placement of ring wavelength conversion shows very good blocking performances (i.e., relatively close to full wavelength-conversion). For example, results in [116] indicate a significant reduction in the ratio of averaged blocking rates for longer-path versus shorter-path connections for larger multiring networks (13 nodes) where only 10-20% of the nodes can perform full wavelength conversion (see also [117]). Blocking reduction in smaller rings is less dramatic, and moreover, wavelength conversion at ring boundaries (Le., via OXC/DCS nodes, Section 5.3) yields almost the full benefits of wavelength conversion [116, 117, 1241. In other 362 Nasir Ghani et al. studies, increased wavelengths are found to be more effectivethan wavelength converters [122, 1261,particularly for increased traffic peakedness (peawmean ratios) [122]. As an aside, note that all-optical wavelength conversion (and regeneration) techniques may become available soon, for example, four-wave mixing and cross-modulation [120, 121, 124, 1271. Most likely, these offerings will not perform full range conversion (i.e., wavelength dependence between inputloutput). Regardless, architectural studies are still very promising, indicating that a relatively small degree of dependent wavelength conversion (25% of full range) can yield ring blocking close to that of full range conversion [121, 1241(similar to partial wavelength conversion). It is important to note that the findings are affected by ring channel RWA algorithms, discussed subsequently. Apart from limited wavelength conversion gains, opaque OADM rings present many other limitations. Clearly, footprints will be much larger and power consumption higher, as transpondersh-eceivers(and related electronic control hardware) are required for all fiber wavelengths [53, 541. Likely, transponder costs alone will be excessive for metro systems with large channel counts and bit rates over 10 Gb/s [88]. Furthermore, laser transponders present added reliability concerns and require sparing (further cost and complexity). Another huge limitation is that multirate wavelengths are difficult to provision, as payloads must be mapped onto a common TDM framing formatdrates, usually OC-48/STM-16.As in legacy SONETEDH networks, such mappings will inhibit transparency and complicate support for various data protocols (see Section 3.4). With increasing channel rateddemands, opaque OADM nodes will require complete laser bank upgrades, which are very restrictive to future growth. The actual wavelength tributary processing, meanwhile, will require complex DCS switching fabrics that must scale to handle large wavelength channel counts, for example, thousands of DS3 or STM-1 ports [3], further increasingcosts. Overall, suchfuZZ opto-electronicregeneration is more crucial in long-haul networks, where attenuation concerns are much more severe. In transparent networks, meanwhile, performance monitoring is a crucial concern, as standards and (to an extent) advanced features are lacking. Optical monitoring solutions use nonintrusive monitoring techniques to analyze such parameters as fibedwavelength power and optical signal-to-noise ratios (0SNR), and other factors such as Q-factors are being studied [94, 1051. Power monitoring, however, can only detect hard failures (node faults, fiber cuts) and not degenerative system faultskonditions, such as increased BER [91]. Moreover, costlpackaging restrictions still prohibit dedicated power monitoring for larger wavelength counts. As such, optical monitoring will suffice for BLSR-type schemes, and optical path-level protection (BPSR, UPSR) can use end-point (optical or electronic) detection [75]. Nevertheless, emerging advances in packaging techniques [29] and integrated monitoring capabilities will likely mitigate these concerns in the future (e.g., MEMS photodetectors [33], power detector arrays, 0-SNR [SO]). Moving forward, it is expected that there will be a reduced need for “complete” bit-level information at all nodal 8. Metropolitan Optical Networks 363 locations, akin to the trend of optical amplifiers usurping digital regenerators. Overall, optical monitoring will find good applicability in metro cores, where there is a premium on transparency. All-optical rings use intelligent RWA algorithms [l-3,7] to determine channel routedwavelength assignments. Ideally, if demands are known a priori, static (offline) RWA can optimize routes, and some algorithms are shown to largely mitigate the need for any wavelength conversion (i.e., 0.25% wavelength gain [114]). Note also that a priori demand knowledge can be used for virtual Nevertheless, realtopology [l] design before the actual RWA phase (see [A). istically, metro connection arrivalddepartures will occur in succession over many timescales, requiring dynamic (online) [2] RWA algorithms. Here, existing connections most likely cannot be disrupted (i.e., rerouted, preempted) to improve wavelengthutilizations(even though various virtual topology “migrationhuning” algorithms have been studied for slowly/periodicallyreadjusting circuit assignments, see [l, 1361for details). Dynamic RWA is usually done in two steps, namely route resolution and wavelength selection, and various algorithms have been proposed for each, for example, shortestllongestneast-loaded path selection, and randodfirst-availablefleast-usedwavelength selection, etc. (see [113, 122, 1281for details). Note that wavelength selection is not an issue in opaque rings, and simpler SONET/SDH ring algorithms can be used [114, 1221. Overall results for dynamic all-optical ring RWA are good, for example, [114] shows 90% of utilization of static “a priori” RWA and only 5% below full wavelength conversion RWA (see also [113]). Theoretical details of optical ring RWA are also considered in [119, 1231. Dynamic RWA algorithms will require ring resource information (e.g., wavelength usages/values, connection maps, etc.), and emerging standards are addressing these concerns (Section 7.1). Note that transponder sources must be “matched” to assigned wavelengths, and new tunable laserdater designs can be very advantageous here, precluding manual line-card changes. However, transparency poses further routing implications in larger metro networks, as impairment concerns can arise [106,108]. A simple solution here may be to enforce geographic bounds on channels, such as distances or hopcounts (e.g., enforced homogeneity [1061). However, as operators progressively expand their networks or provision higher bit rates (10 Gb/s or more), this is not feasible. Now, most RWA algorithms strictly focus on “logical” resource control (e.g., wavelengths, converters, channel delays, etc.) and assume perfect analog conditioning (e.g., flat amplifier gains, no losskrosstalkldispersion). Consequently, such “idealized” algorithms accept more calls (i.e., yield lower blocking probabilities) than may be practically possible for a given channel BER threshold [108]. For example, consider amplifier saturation, where simulation results for nine nodes indicate that heavier meshed demands reduce span lengths by 30% over lighter hubbed traffic [54], even if wavelengths are available. Consequently, physical layer concerns may have to be incorporated into the RWA phase [106, 108-1101. In a sample proposal, [110] presents a 364 Nasir Ghani et al. two-phase logical/physical RWA scheme that incorporates transmission loss, crosstalk, and amplifier saturation (see additional work in [108, 1091 also). In another solution, translucent routing (Le., partial wavelength conversion) is considered, where signals are transmitted optically “as far as possible” before signal degradation mandates opto-electronic regeneration [1181. This is especially appealing in the metro space, as transparency is very important. Specifically, a translucent routing algorithm (modeling componentlfiber loss and crosstalk)is presented and results for a metro-sized multiring indicategood gains for larger crosstalk values (see [l 181). Overall, the computed complexities of such advanced “optical-layer”routing algorithms are sigdcant, making them more suited for centralized control architectures[lo61(Section 7.1). Note that translucent routing requires subtending opto-electronic stages, and DCS fabrics can be used here (i.e., “hybrid” OADM, Fig. 8.9). Carefully note that using DCS switching fabrics for wavelength conversion/translucent routing also offers another key benefit, namely edge subrate TDM grooming [48, 651 between metro core and access rings. Namely, bundled subrateTDM tributaries (DS3,0C-3/12)emanatingfrom different access rings can be unbundled and rearranged onto larger core wavelengths to improve efficiency. Such subrate grooming has also been studied more broadly under the SONETLDWDM circuit-wavelength grooming (e.g., “stacked” SONET rings) [68, 1361. Specifically, studies have focused on demand grouping and lightpath routing to minimizing network-wide electronic ADM/DCS costs for a given OADM design (i.e., exploit optical bypassing). For example, [ 1391derives bounds and presents results for various intelligent grooming (circle construction) algorithms for uniform and nonuniform traflic. Simulations indicate significantADM cost savings in both uni- and bidirectional dynamic rings (from 30 to 60%). Detailed discussions and analytical bounds for arbitrary traffic patterns are also presented in [140] (Le., primitive ring grooming). Additionally, [137, 1381 also study various OADM-DCS ring combinations for static OADM designs, for example, point-to-point, single/dual-hub configurations. Again, cross-connection is found to lower overall network-wide ADM costs for dynamic demand conditions. This is an important area, and related findings can be exploited to help appropriately size subtending DCS stages on dynamic OADM rings, such as dropped wavelength counts, DCS port counts/types, etc. 53 OPTICALHY3MDMESHNETWORg;S . Metro topologies are rarely homogenous and usually comprise a meshed interconnection of many IOF rings, up to 20 in larger areas [143] (commonly termed hybrid topologies [75,76]).Now, due to larger communitiesof interest, user demandsmust be routed/protectedbetween multiple rings, requiring ring gateway interconnection [43, 541. Alternatively, in the future, operators may choose to move to mesh routing paradigms (either “greenfield” builds or via 8. Metropolitan Optical Networks 365 selective ring fiberplant expansions) in order to improve resource efficiencies further. In these hybrid/mesh networks, optical cross-connect switch (OXC) nodes will be the central elements, handling increased spatial (fiber) connectivities and serving in an advanced generalization of the role performed by SONET/SDH DCS nodes. Overall, “mesh” optical switching has emerged from the long-haul space [6, 91-93], and first-generation switch designs used electronic DCS switching fabrics [58,91, 1321 (Le., electronic wavelength crossconnects [2]). Specifically, a DCS fabric functioned as a bandwidth manager [48], performing subwavelength grooming/aggregation and full signal regeneration for long spans [93]. However, akin to OADM evolutions (Section 5.2.3), associated opto-electronic fabric costs proved prohibitive for large-tributary switching, and hence optical switching technologies (Section 4.5) gained favor. Nevertheless, since multiple, diverse add-drop systems can converge at large metro switching interchange locations, extensive bandwidth management capabilities will still be required, ranging from subrate TDM streams to large-granularity optical wavelengths. Consequently, integrated OXC/DCS nodes [46, 64, 91, 1681 are most germane here (Fig. 8.11), comprising both optical and electronic switching cores. In these “hybrid” designs, optical switching of larger tributaries will significantly reduce required DCS fabric filtenng (adddrop channels) Sub-rateTDM Sub-rate TDM Electmnic DCS grwmmg. channel adddrop stage Fig. 8.11 Integrated OXC/DCS node. 366 Nasir Ghani et a. l sizes, yet DCS fabrics will still permit subrate terminatiodgrooming andlor interdomain regeneratiordmanagementfunctions. OXC cores can use a variety of switching technologies (Section 4.5), although MEMS is most common today [33]. Currently, 16 x 16 twodimensional MEMS modules have been demonstrated [33] and are commercially available, whereas a practical limit of 32 x 32 is stated in [26]. Nevertheless, future metro switch fabric requirements will clearly increase further, driven by denser channel counts and larger multiring interconnections. For example, port counts are proportional to the square of wavelength channels for multiring interconnection [58]. Therefore, switch fabric expansiodscalability is a major issue [91]. One solution here is to use multistage interconnection [34] of multiple smaller switches, for example, Clos architectures [3, 91, 1681. Here, nonblocking configurations are highly desirable, as transparent networks already suffer from wavelength blocking. However, complex multistage connections increase insertion loss/crosstalk between stages, and ideally a large single stage is most desirable [3,91]. Although dense threedimensional MEMS switches are being studied [33] (256 x 256 and beyond), cost concerns will likely preclude metro application for the next several years. Consequently, a more immediate and scalable solution is to use band filters (Section5.2.1) to perform coarserwavelength band switching[46]. This reduces port count requirements significantly, as a single MEMS port can switch a whole band, even fiber. Note, however, that band switching has various higher-layer provisioning implications, and therefore is best applied to traflic with similar characteristics, for example, routes, application types, protection levels, etc. [93]. Apart from switching, note that many considerationsfor the transport stage (filtering, amplification) are also similar to those for dynamic OADM nodes. Ring tributary interconnection will likely be the fist application of integrated OXC/DCS designs [1131. Now in many cases, multivendor rings using differingoptical gear will interfacevia standardizedinterfaces (e.g., OC-48/192 at 1310nm) at administrative boundaries. Most likely, bit-level monitoring will be necessary at these interchange points in order to monitor crossconnecting signals (e.g., security concerns, SLA BER monitoring, etc.) [105]. Moreover, operators will want cleaner interdomain hand-offs, fully regeneratinghormalizingincoming wavelength tributaries before transmission across their own networks. For example, long-haul networks use standardized payloads (SONET/SDH, OCh digital wrappers) and require relaunching via more powerful lasers. Hence for the foreseeable future, ring interconneo tion will be done opto-electronicallyvia DCS devicedfabrics (Fig. 8.1 l), even for larger tributaries. Moving forward, however, optical ring interconnection via OXC switches will also become important, especially in single-operator domains [56]. Such “clear-channel” interconnection [58] can extend the reach of “non-SONET/SDHmapped” services (not possible with DCS interconnection). Although no standards currently exist for optical ring interconnection, 8. Metropolitan Optical Networks 367 evolutions will likely be similar to those for SONET/SDH ring interconnection, for example, manual "back-to-back'' patch-panels to intermediate OXC/DCS nodes, and later possibly integrated OADM/OXCnodes [73]. Overall, many of the pertinent SONET/SDH ring interconnection concepts [8,9] will also apply to optical interconnection. For example, nonsignaled single and dual gateway (homing) configurations are described in [54,74], see Fig. 8.12.In the former, two gateways (one from each ring) are connected to protect against failures in each ring, but these represent single points of failure. Consequently, in the latter, four nodes (two from each ring) are connected in a drop-and-continue [74] configuration to provide added gateway failure recovery (note that control mechanisms of two rings can be different [9]). Since this is a nonsignaled setup, added hardware considerations are necessary (optical splitters, switches, Fig. 8.12). Detailed availability analysis of dual gateway interconnection is presented in [74] and more advanced path-based (dedicated) and line-based (shared) strategieshave also been trialed [86]. Overall results indicate that more traffic can be carried over interconnections with more fibers (or nodes) between the homing points. Nevertheless, optical interconnection increases transmission domains/distances, and this will inevitably complicate network design. Additionally, DWDM mesh networks have also been studied [l-3, 1491, using OXC nodes to route/protect wavelength circuits over arbitrary fiber topologies. By and large, OXC devices cost significantly more than OADM nodes, relegating such solutions to long-haul backbone applications [1131. Single-Hub Ring Interconnection Dual-Hub Ring Interconnection HvbridMesh Touologv Fig. 8.12 OXC applications in metro networks. 368 Nasir Ghani et al. Nevertheless, as costs decline and metro operators expand their fiber plants, mesh or hybrid (ring-mesh) topologies (Fig. 8.12) will gain importance. Many theoretical results indicate that mesh topologies are usually more wavelength efficient than rings (assuming larger connectivities) (i.e., rings need more wavelengths to route a given demand set, see [113, 1321). A whole range of mesh RWA algorithms have been developed, including online/offline and/or distributed/centralized computation schemes, see [128] for a survey. Many mesh survivability mechanisms have also been tabled, broadly classified as restoration and protection schemes [7, 961. The former refers to more active, postfault signaled recovery (usually path level), whereas the latter refers to predesigned protection (link or path level) [7, 1131. Restoration schemes are very wavelengthkber efficient, but protection schemes are by far the most studied, and many variations are possible, for example, nonsignaled and signaled dedicated/shared schemes, etc. [7, 95-97, 155, 1561. Although further details are beyond the scope herein, results show that shared protection over increased mesh connectivity can yield significant capacity/efficiency improvements (see referencesin [96,97]). A more ominous concern with DWDM mesh protection (and restoration) is recovery timescales, as topologies can be arbitrary. Typical values are larger than those for ring protection (i.e., hundreds of milliseconds [48, 96, 97,991). Nevertheless, these values are very adequate for many Internet data services [159], and overall, mesh networks can offer very rich service definition capabilities (similar to those mentioned for advanced bidirectional rings, Section 5.2.2). As an aside, given these larger recovery timescales, slower optical sampling and/or spectrum scanning techniques [93, 1051 can be considered for monitoring/fault localization. Meanwhile, standardized control architectures are also emerging for optical mesh networks, and detailed protection protocols have been tabled (Section 7). Overall, transparent DWDM mesh routing will face many of the same challenges as dynamic OADM nodes (i.e., attenuation/dispersion, performance monitoring). As per previous discussions, rings present ubiquitous, fast protection mechanisms to which operators are very well-accustomed [76]. Consequently, many operators may still want to provision rings over expanded hybrid (mesh ring) topologies (see also migration issues, Section 5.5). In particular “virtual ring” emulation may become a key requirement, allowing operators to define arbitrary ring topologies on top of mesh plants, traversing a mix of OADM and OXC nodes. Over time, these requirements will likely blur the lines between OXC and OADM designs, as integrated OADM/OXC/DCS nodes (Fig. 8.12) collapse mesh routing and UPSR/BLSR ring protocol functions onto a common platform [85]. Various techniques have been studied to resolve ring overlays onto mesh topologies, termed ring covers [3], and [73,97,113] present surveys. For example, rings can be grouped via “communities of interest” to minimize interring traffic. Alternatively, rings can also be defined to help “localizey~ recovery domains, ensuring that network faults do not disrupt t r a c in other parts of a larger hybrid network [97]. In [84], the ring-covers problem 8. Metropolitan Optical Networks 369 is solved using an optimization formulation to resolve “optimal” ring sets and ensuing connection routes (hierarchical, nonhierarchical rings). Meanwhile, [97l presents a simulated annealing solution. Some early standardization discussions on ring-mesh interworking have also appeared, for example, ring identifier extensions to mesh routing protocols [76]. 5.4 ECONOMIC CONSIDERATIONS: TDM VS. D WDM Several years ago, DWDM technology was largely considered as a longhaul solution, as optical amplifiers provided better economics over new fiber builds (and opto-electronic regenerators). At the time, the viability of metro DWDM was questionable, since related subsystem costs could not justify applications over smaller unamplified distances. For example, in [71], for very modest demand growth (12 STS-1 circuits/year), OC- 192TDM yielded better economics than DWDM (at OC-48 rates). Similar results were presented in [41, 511. Nevertheless, recent studies are indicating a reversal of this trend, especially for bit rates over OC-IUSTM-4. Most current studies compare DWDM node (and possibly amplifier) costs against higher-rate TDM systems with increased fiber counts, commonly termed as space division multiplexing (SDM) [49]. For example, in [54], metro core dimensioning shows good cost reduction using DWDM for larger STM-16/64 tributaries (3549%). Smaller access networks, however, give a much closer tradeoff, with DWDM approaching cost parity for more aggressive demand scenarios (see studies in [51] also). Meanwhile, results for point-to-point spans [49] find DWDM more economical than OC-192 for capacity expansion up to eight OC-48 channels (assuming no amplification). Most likely, it will also be more cost effective to transport OC-192 signals via DWDM than SONETISDH (e.g., no per-hop complexity) [3]. Additionally, in a comprehensive study [40], three different metro (two-fiber ring) network scenarios show over 20% cost savings in capacity exhaust scenariosfor both moderate and aggressivedemands. Furthermore, findings for direct OC-12 tributary transport indicate near parity in cost between SONET OC-48, OC-192, and 24-wavelength systems [40]. The authors even hint at direct OC-3 cost parity in the future as costs decline. In general, market research also confirms overall DWDM cost effectiveness for bit rates over OC-lUSTM-4 [62,65]. Other discussions are presented in [3,511. Overall, many of these techno-economic analyses are very encouraging, especially as DWDM component prices continue to decline. Moreover, in the studies, cost modeling was largely done for equipment/infrastructure. Many crucial operator-relatedbenefits of DWDM systems are more difficult to incorporate, whose inclusion, nonetheless, would further enhance the DWDM value proposition. These include rapid provisioning, advanced service definitions, transparent routing/protection of “non-TDM mapped” payloads, and likely reduced operations costs (e.g., power consumption, floorspace). For example, cost studies for rapid provisioning in regionalllong-haul networks 370 Nasir Ghani et a . l show significant revenue potential over slower legacy provisioning [72]. As the TDM-DWDM metro debate abates, -the two technologies are now seen as more complementary, with focus shifting on when to deploy and achieving the right mix between the two for transition purposes (e.g., TDM edge multiplexinglswitching, see Section 6) [22, 641. Generally speaking, transport-related functionalities will migrate to the optical layer, and SONET/SDH will serve as an aggregation/framing (sub)layer [6] closer to the edge. 5.5 MIGRATIONSTRATEGIES Despite the many advantages of DWDM technology in the metro core, it is unrealistic to expect immediate, full-scale deployments. Instead, a more gradual, cost-effective migration is likely, as operators pair added expenditures with expanded revenue growth. Moreover,many metro operators have invested (and continue to invest) heavily in SONETISDH technology [46,63,142] and may be unwilling to undertake further expenditures without realizing significant returns thereof. Additionally, DWDM represents a new technology, and most operators will first want to gain valuable field/market experience with smaller-scale deployments. Most operators will look for strategies that offer low “firstcost” and economical lifecycle costs [22]. In particular, a migration path that reuses existing fiber infrastructures/equipment will help improve initial costs significantly. Overall, it is expected that metro DWDM will evolve from less costly point-to-poindring deployments to more advanced dynamic rindmesh (hybrid) optical networking solutions offering wavelength services [22,49, 50,731, as shown in Fig. 8.13. For most incumbent metro operators, fiber-reliefkapacityexpansion is the first likely application of DWDM technology. In cases where there is still available conduit space, it is usually more economical to simply pull more fiber through [22]. However, in many largeddenser cities, conduit space may be lacking and laying new fiber can be expensiveand very time consuming(involving lengthy right-of-wayconcerns). Moreover, such spatial expansion may not meet projected demand increases, and DWDM offers the best price/bandwidth return. In fact, initial linear point-to-point fiber-relief deployments are well underway in the metro, see [22, 37, 49, 501 for deployment details. DWDMenabled spans offer a large number of “clear-channels” (is., 32 wavelengths and beyond), with which operators can easily address immediate capacity concerns and further expand “non-TDM” client services (e.g., Fiber Channel, Escon, etc.). Note here that band-filteringhterleaving techniques (Section 5.1) will also permit modularly wavelengthexpansion, helping to lower initial costs. In order to minimize disruptions for existing clients, metro DWDM can be phased in with existing SONET/SDH infrastructures by using inexpensive 1310/1550-nm broadband (mux/de-mux) filters to create two “virtual fibers” [70] (Fig. 8.14). Specifically, legacy SONET/SDH interfaces can continue to use the lower 1310-nm region, whereas the upper 1550-nm region 8. Metropolitan Optical Networks SONET/SDH Ring Hierarchy Inter-nng DCS hub 371 SONET/SDH w. Partial WDM Increased WDM penetration Multiple Programmable Optical Rings High capacity, fast rotection witchi J Hybrid (Ring-Mesh)Networks Hybrid ringlmeshes. virtual overlay nngs \ \ w lambda3 Network expansion/ Fig. 8.13 High-level overview of metropolitan (core) network migration phases. Edge sub-rate TDM DWDM mudde-mux Fig. 8.14 Wavelength-band filtering for SONETBDH-DWDM coexistence. 312 Nasir Ghani et al. (C-, L-bands) can be used by DWDM transmission (note that filters are required at all nodes under 1 dB loss). This filtering technique is very cost effective and expedient since it reuses existing fiberplants. Moreover, it also allows a more staged DWDM induction on a hop-basis, for example, DWDM transmission only added on two “capacity-exhaust” spans (Fig. 8.14). As the traffickonnectivity profiles develop, more advanced OADM devices can be subtended, allowing for TDM/DWDM ring coexistance.A detailed economic analysis of this technique is done in [70], and results show improved cost effectivenessover counterpart SONET-ringupgrades for larger-ratetributaries (e.g., from OC-12 to OC-48 and from OC-48 to OC-192). Larger ring sizes are also found to favor the DWDM-based solution [70]. Alternatively, many newer SONETKDH systems can be equipped with 1510-nm band interfaces, allowing for more direct integration with optical transmission gear. As metro point-to-point DWDM penetration increases and larger portions of ring topologies become wavelength capable, the next logical step is a move to genuine optical ring architectures (Fig. 8.13). In fact, many operators are already at this evolution point, effectively migrating SONET/SDH out of the metro core. Meanwhile, many upstart operators building new “greenfield” networks are also utilizing ring topologies for various reasons, for example, most transport gear (TDM, DWDM) is specialized for rings. Optical rings will extend “clear channel” services across multiple CO hops, allowing for “network-level”wavelength services provisioning. For many carriers’ carriers, this is becoming much more profitable than directly leasing dark fiber [64]. Overall, many factors will determine the types and sizes of optical rings deployed, for example, cost, trafEc types, service requirements, projected growth, etc. [78]. In highly cost-sensitiveareas with moderate growth and/or hubbed traffic patterns, operators will likely choose simpler static UPSR designs. For example, in [35] it is stated that eight wavelengths will suffice for most small (metro edge) rings. Conversely, for larger-growth regions with dynamic demand patterns, advanced BLSR/BPSR rings will be more scalable, for example, advanced service definitions, better capacity utilization/reuse, etc. Over time, integrated OADM/OXC/DCS nodes will also replace SONETBDH DCS nodes at ring interconnection points. However, a precursor requirement for these advanced ring architectures is the need for detailed architecturalkignaling standards. Although still evolving, these are expected to mature over the next several years (Section 7). Finally, migration to hybrid/mesh metro core topologies may also occur in larger markets, where marketldemand growth can justify the associated OXC-based costs. For example, it is conceivable that after several years, some operators may experience congestion in their optical ring architectures, for example, at interconnection gateways(interring traffic) or at specificring nodes (intraring traffic). As a result, they may be forced to deploy new fiber routes to essentially “break” existing ring topologies to even demand distribution [45,76], for example, fiber routes between nonadjacent ring nodes, additional 8. Metropolitan Optical Networks 373 intergateway fiber routehodes (see findings in [ 6 ) Alternatively, opera8l. tors may experience unpredictable demand growth in altogether new areas, requiring added fiber builds from existing ring plants. In these scenarios, operators will deploy a mix of OADM and advanced OXC gear, dependingupon the fiber build and economic conditions. Nevertheless, most locations terminating higher fiber counts will utilize more costly, integrated OADM/OXC/DCS gear. In these topologies, mesh-ring interworking (Section 5.3) will be crucial from a services point of view, in order to minimize service disruption. Overall, much more defining studies are required for planning migration from ring to hybrid topologies. 6. Metro Edge Solutions Similar to traditional taxonomies, the metro edge will continue to represent a merging between the core interoffice and the client-access spaces. Here it is becoming increasinglyevident that SONETKDH is not the best unifying layer [42]. Conversely, since DWDM is bandwidth-inefficient for subgigabit linerates, advanced electronic multiplexing technologies are needed to aggregate diverse end-user protocols onto large-granularityoptical (DWDM) tributaries [46]. Dense IC technologies are finding particular favor here, helping collapse legacy multiplexing hierarchies (i.e., “system-on-a-shelfkard” [141]) and further blurring traditional access boundaries. Many new metro edge solutions, broadly termed as optical edge devices (OED) [63], have been proposed, including DWDM edge rings, next-generation SONET/multiservice provisioning, and IP routindpacket rings. Some details are now presented. 6.1 D WDM EDGE RINGS In Section 5.2.1 it was stated that low-cost passive optical rings are generally well-suited for hubbed traffic patterns and can serve as metro edge solutions, Le., metro DWDM access [64,65]. However, a key issue is mapping client protocols onto the underlying wavelengths, and several solutions are possible (see Fig. 8.15). A straightforward, transparent approach is to assign a complete wavelength to each client signal, regardless of bit rate (termed “protoc01per-lambda”). This works well for low demandnode counts and s a c i e n t wavelength channels. For example, [87] proposes to back-haul individual subrate client tributaries (DSI, DS3,0C-3) across a dual-homed DWDM (access) ring to a CO for aggregationhermination. Nevertheless, given the propensity of metro-edge DSl/DS3 subrate tributaries, this approach cannot scale in wavelengths and is very costkapacity inefficient for rates below the “breakeven” OC-IUSTM-4value, see Section 5.4. Clearly, edge aggregation must be coupled with static DWDM rings in order to improve efficiency and reduce wavelength port requirements, as shown in Fig. 8.4. 374 Nasir Ghani et al. Sub-rate TDM multiplexing (line cards) Fixed wavelength/ Fiber Channel Ethernet switch Fig. 8.15 Generic overview of metro edge DWDM ring node (single direction). A variety of subrate “circuit” multiplexing schemes can be implemented. For legacy TDM support, dense IC chipsets can reduce many multiplexing hierarchies onto line cards, e.g., 4 : 1 OC-3 to OC-12, and even OC-12 to OC-48 (Fig. 8.15). These “integrated SONET/DWDM” interfaces [40], also termed “thin mux” [51], can combine the benefits of both technologies and further improve cost effectiveness. For example, in [40], integration of SONET/SDH line termination functionality with DWDM transport is found to yield significant savings in electronic protection overheads. The availability of newer software-programmable SONET/SDH transceivers (OC-3/12/48) further improves flexibility, as line rates can be adjusted per demand, eliminating the need for constant line card upgrades. In some cases, a complete subrate DCS switching unit can also be added to aggregate multiple flows. However, SONET/SDH multiplexing is only amenable for legacy voice/leased-line traffic and not native packet interfaces (e.g., 10/100Mb/s, 1 Gb/s Ethernet). The latter require expensive “telecom adapter” (mapping) interfaces, more than quadrupling overall interface costs (electronics, labor) and yielding high bandwidth inefficiencies [46, 1331, see also Section 6.2. More flexible TDM multiplexing techniques can also be used for metro edge-aggregation. For example, proprietary “asynchronous” multiplexing can combine multiple subrate circuits onto a higher-bit-rate time-division carrier. The resultant protocol concurrencies can be much more efficient and can include a mixture of SONET/SDH and other alternate data tributaries (e.g., 155 Mb/s OC-3, 200 Mb/s ESCON, 1 Gb/s Ethernet). Recent 8. Metropolitan Optical Networks 375 developments in digital wrapper standards [12,13]will also facilitatemore flexible TDM multiplexing schemes. Digital wrappers define client-independent overheads for transport and management of payload bit streams across optical domains, for example, bytes for management, monitoring, protection signaling, even FEC (about 6% FEC overhead [145]). These overheads are processed at “electronic” (opaque) monitoring points, such as boundaries between access/core rings/domains. Currently, several “wavelength” rates are defined, namely 2.5, 10, and 40 Gb/s, albeit subrate multiplexing hierarchies are not defined [12, 1451. Conceivably, a full variety of protocols can be multiplexed into the payload section, unlike rigid SONET hierarchies, although there can i l be FEC implications (see [1451). Nevertheless, digital wrappers w l inevitably entail similar overhead processing complexity as SONET/SDH and related chipset costs will likely relegate this technology to long-haulhegional transport and/or for larger (metro) interdomain interfacing functionality for the nearlmedium term. Optical frequency division multiplexing (0-FDM) has also been proposed for edge multiplexing, using a single laser to modulate “subrateyy carriers (ie., client channels) within a spectral band. Specifically, all signals undergo quadrature amplitude modulation (QAM) and are subsequently frequencymultiplexed onto a fibedwavelength(i.e., 0-FDM/DWDM ring combination). 0-FDM is genuinely bit rate transparent and can improve spectral efficiency over on-off keying (OOK)modulation schemes used in SONET/SDH or Gigabit Ethernet encoding (between 20-50%, 20 Gb/s per wavelength possible). 0-FDM transmission is also more dispersion tolerant, and can work well on older fibers, for example, high-PMD types: unlike 10- or 4O-Gb/s TDM [44]. Additionally, related electronic costs are lower, since speeds need only match slower subcarrier channels. Studies for moderate demand scenarios show 0-FDM to be more fiber efficientthan OC-48 andmore capacity efficient than OC-192 [44]. Nevertheless, DWDM-induced transmission impairments for 0-FDM transmission may be problematic, and this needs proper characterization. Note that only DWDM technology can transport 0-FDM signals in their native formats. 6.2 “NEXT-GENERATION99 SONETMULTISER WCE PRO VISIONINGPARADIGMS Even though legacy TDM technology has many shortcomings (Section 3.4), it will continue to play a significant role in the convergence of data and optical networks at the metro edge. Demand for short-haul SONETBDH gear is still high and may continue to grow for the next several years [63,65]. A large part of this market comprises larger OC-48/STM-16and OC-192/STM-64systems, although smaller OC-3/STM-1 and OC-12/STM-4 systems Will also see increased deployments (see [63]). Moreover, many existing routershwitches have SONET/SDH interfaces (e.g., POS, AALS), and recent efforts to define 376 Nasir Ghani et al. a broader generic framing protocol (GFP) [141 (for mapping “nonstandard” data protocols) may further propagate the ubiquity of such framing [158]. In light of this, many proposals have sought to “enhance” SONETLSDH paradigms to better suit data traffic needs [130, 131, 133-135, 141-1441. Although these proposals have appeared under different names (e.g., “super SONET” [63], “data-awareSONET” [130]), herein the term “next-generation SONET” (NGS) is chosen (Fig. 8.4). Overall, all these solutions share two main features, namely efficient data tributary mappings and integrated higher layer (twokhree) protocol functionalities, as shown in Fig. 8.16. Concurrently, these solutions also leverage ubiquitous SONET/SDH performance monitoring, protection switching, and network management capabilities. Some details are presented. SONET/SDH mapping of smaller packet interfaces (10, 100Mb/s Ethernet) is usually done in “coarse” STM-1 increments and the resultant bit-rate incongruencies usually yield large amounts of stranded bandwidth [129, 1441 (e.g., lO-Mb/s Ethernet allocated a full STS-1, 80% unused capacity). Bursty data profiles can further exacerbate bandwidth inefficiencies. Nevertheless, advanced IC technologies are permitting high-density switching fabrics with much finer TDM granularities,particularly at smaller/fractionalVTl.5 levels [141, 1581. By combining finer tributaries, for example, virtual concatenation [158] (wideband packet-over-SONET[130]), native packet interface rates “NextGeneration” SONETBDH Node Tributary switching, addldmp, protection _______....... I ______________I Input buffcrs Scndcr side Output buffcrs MultiService Provisioning Platform (MSpPr Rmiver ride Fig. 8.16 “Next-generation”SONETKDH and MSPP. 8. Metropolitan Optical Networks 377 can now be matched much more closely (e.g., lO-Mb/s Ethernet via seven VT1.55). Multiple “matched” tributaries can then be more efficiently packed into existing standardized tributaries, and this will help collapse multiplexing (equipment) hierarchies. Switching designs can also use multilevel DCS fabrics to assign capacities in both VTl.5 and larger STM-1 increments to better scale electronic complexities [129] (Fig. 8.16). Overall, advanced DCS designs will extend ubiquitous TDM tributary add-drop/switching/protectionfunctions to cover a full range of combined streams, in addition to intenvorking with legacy streams (i.e., from ADM, W-DCS, B-DCS gears). Furthermore, more advanced renditions are possible that dynamically adjust allocations to “match” bursty loads on incoming interfaces (albeit layer two/three buffering/processing and end-to-end signaling are required here). Along these lines, there have been notable developments in the link capacity adjustment scheme (LCAS) [172] mechanism. LCAS defines a control protocol that allows for “hitlessly” increasing/decreasing the number of “trails” (e.g., STS-1 circuits) assigned to a connection. Moreover, each circuit trail can be diversely routed to improve resiliency and failed trails can be removed altogether. Additionally, connection asymmetry also can be achieved by assigning a different number of trail counts to a given connection direction. Overall, LCAS defines a very powerful new capability for exploiting virtual concatenation techniques and improving capacity utilization (see [172] for more details). To further improve data efficiency/scalability,NGS designs intend to provide a full range of higher-layer “non-SONET/SDH” protocol functionalities (i.e., “data-aware” TDM interfaces). Examples include IP routing, ATM switching, LAN switching, and even frame-relay aggregation, see [4, 130, 133, 135, 142, 1441. Namely, intelligent layer two/three cell/packet processing (e.g., bfiering, scheduling, switching, routing, Fig. 8.16) capabilities are used to increase capacity oversubscription ratios (e.g., statistical multiplexing gains) between multiple customer ports [143], a step beyond “circuit” aggregation. Many designs also ,providedirect data (Ethernet) packet interfaces, eliminating the need for more expensive “telecom adapter” private-line interfaces at client switches/routers.Additionally, line-terminationcapabilities can also be added to extract and process payloads from existing privateline data interfaces [1301. This essentially “decouples” link interfaces from their associated data payloadprotocols, an important step in extending the benefits of oversubscription to private-line traffic (Fig. 8.16). Traffic multiplexing coupled with tributary concatenation achieves aggregation closer to the edge, leaving more free capacity inside the ring and yielding very good bandwidth efficiencies. For example, three 10-Mb/s Ethernet streams averaging 3 Mb/s can be edge-buffered and packed into six VT1.5 circuits versus three OC-1 legacy interfaces, a bandwidth savings of 94%. Note that edge-multiplexing of multiple packet interfaces also reduces port counts and the need for complex, costly centralized back-hauling setups [1311. 378 Nasir Ghani et al. Overall, integrating formerly distinct packetkell and SONET/SDH protocols onto a common platform removes multiple subtending aggregation gears (routers, switches, ADMs) and their associated management systems. This reduction can yield significant operational cost/provisioning complexity improvements and reduce footprint space/complex cabling considerably. Moreover, the emerging generalized MPLS framework (Section 7.1) presents a comprehensive control setup for NGS systems, Le., edge equivalence mappings between circuit and packet labels and more recently, even provisioning of concatenated SONETlSDH tributaries (see [153]). As an aside, note that some earlier schemes proposed using ATM as the primary multiservice SONETEDH aggregation layer [4, 1441. However, these designs suffered from high bandwidth inefficiency (about 20% [l58]) and hardware scalabilitykost concerns, and have been largely usurped by improving IP paradigms [65]. Despite its capacity improvements, NGS still reuses rigid, synchronized electronic payload framinglencapsulation formats. As a result, this solution is not truly capacity scalable (i.e., electronic bit rate and cost limitations) and is much better suited to improving time-slot packing on existing rings (OC-3, OC-12) and/or for areas with limited demand growth or high fiber count [65l. SONETBDH framing again precludes transparency, making it difficult to support data protocols such as Fibre Channel, ESCON, or FICON without proprietary handling [51, 1411 (until the formalization of GFP at least [14]). Moreover, the associated functionalitiesof such protocols may be too specialized, stillmandating the use of subtendinggear. A more ominous concern with NGS is that its associated packet/cell functionalities/featureslikely may not match those of “best-in-class” solutions offered by specialized routerhwitch vendors [65]. Here, many operators already have (or plan to deploy) separate “best-in-class” geadmanagement systems and will be unwilling to accept single-vendor solutions. Consequently, more generalized multiservice provisioning platforms (MSPP) attempt to address these limitations by further integratingDWDM functionalityto boost transparency/scalability. essence, In MSPP solutions combine NGS with DWDM (ring) technology (Section 6.1), and have also been more aptly termed as integratedmetro DWDM [64,65].An overview of an MSPP node is given in Fig. 8.16, where added passive DWDM transport/ring functions are shown. To lower costs and increase flexibility, many MSPP designs intend to add “optical” functionalities (multiplexing, transport, filtering) in a modular fashion via line-card additions. Again, “allin-one” MSPP solutions may be overly expensive and impractical, especially if the existing base of legacy TDM and/or “best-in-class” routing/switching infrastructures is large (see also Section 6.4). On the subject of SONETlSDH enhancements, recent advances have proposed increasing SONETBDH line rates to 40 Gb/s (OC-768/STM-256), further propagating existing TDM-paradigms. Currently, 40-Gbh transmission is usually done by optically interleaving [2] four 10-Gb/s streams 8. Metropolitan Optical Networks 379 (channelized), as commercial availabilityof electronic SONET/SDK overhead processing/clockrecovery circuitry at direct 4O-Gb/s rates is still a ways off (i.e., OC-768cconcatenated interfaces). Regardless of the interface type, dispersion (slope) effects at these bit rates will restrict transmission distances significantly (e.g., chromatic dispersion at 40 Gb/s is 16 times larger than at 10 Gb/s, yielding a dispersion limit of about 25 kni [106, 1121). This will hinder applicability beyond small metro domains, and usually extensive dispersion compensation and fiber characterization considerations will be necessary (as used in most studies, see [I 111). Moreover, bandwidth scalability at these increased bit rates still falls well short of those yielded by DWDM. Furthermore, mapping OC768/STM-256 tributaries onto wavelengths (e.g., for transmission across core metro rings) will likely require larger 100-Ghz spacings (and not 50 Ghz) due to interchannel crosstalk limitations. To an extent, this mitigates the gains of increasing the channel bit rate. Overall, 4O-Gb/s OC-768/STM-256 solutions have yet to be deployed, and related technical and cost concerns will adversely affect or delay their applicability in the highly cost-sensitive metro edge [88]. When they do emerge, such large TDM interfaces will likely interface with larger metro core wavelength routing gears. Moreover, it is also conceivable that cheaper multiplexed 4O-Gb/s Ethernet router interfaces will emerge first. Namely, these interfaces will simply multiplex four lO-Gb/s Ethernet streams together, thereby avoidingmany complexitiesassociated with genuine 4O-Gb/s TDM clock recovery and/or header processing. 6.3 PACKET-BASED SOLUTIONS Although DWDM rings provide significant improvements over TDM rings, as discussed previously, they still embody a circuit-switching paradigm. It is well known that circuit switching is generally less bandwidth efficient than packet switching[4], and bandwidth utilization oncircuit-multiplexedDWDM rings can be very low for bursty data profiles [60, 1661. Nevertheless, the packet-switching devices (Ethernet switches, emergence of ‘cnext-generation’y IP routers) is helping resolve many of these data inefficiencies. Specifically, advanced hardware-based packet filtering [1641 and switching technologies [168] can now support line-rate inputloutput switching at full “wavelength” tributary rates (0C-48c/192cylO-Gb/s Ethernet) (e.g., via custom high-speed ASIC solutions or even generalized network-processor chips). Hence these nodes can serve as direct (POP) aggregation boxes for metro edge, even core, DWDM rings, completely collapsing inefficient, legacy “leased-lineyy data hierarchies (e.g., “evolutionary delayering,” see Fig. 8.18). There have also been significant improvements in the overall IP routing framework to support “TDM-style” guarantees (bandwidth, delay, loss, etc.), namely the multiprotocol label switching (MPLS) framework and its associated resource reservation protocol (RSVP). Overall, this emergent framework can support very fine quality of service (QOS) levels via the integrated services model (Intserv) 380 Nasir Ghani et ai. or more coarse (scalable) class of service (COS) levels via the differentiated services (Diffserv) model (see [154, 1561 and related references). These capabilities provide “soft circuit” setups that achieve high statistical multiplexing gains and can vary bandwidth allocations per any given criterion (e.g., perport, client group, application, etc.). However, various provisioning concerns still need to be addressed before “carrier-class’’ services can be offered (e.g., hardware/control scalability, service survivability). In light of these, more specialized packet-switching schemes are being developed. Recently, the concept of “packet rings” has been proposed, aiming to combine the salient features of “TDM-origin” ring topologies (i.e., simple connectivity, high resiliency) with the advantages of packet switching (statistical multiplexing, finer QOS), namely resilient packet rings (RPR, IEEE 802.17) [162, 1631. Specifically, a new Ethernet-layer media access control (MAC) protocol is defined to statistically multiplex multiple IP packets onto Ethernet packets (Le., layer-two). The MAC protocol itself is “media-independent,’’ and will be capable of running over various underlying networking infrastructures, including SONETEDH, DWDM, or dark fiber. RPR designs can provide separate “layer-two’’ packet bypass capabilities at coarser granularities, and this will relieve packet loads at the IP (layer-three) routing level and improve QoS provisioning [ 1.581. For example, sample RPR node design in Fig. 8.17 shows Packet ring \ 10GhEWAN interface? 4 Metro core M nng Enterpnqe router (100Mb/sEth) Integrated packet ring node Fig. 8.17 Sample packet ringlnode architectures. 8. Metropolitan Optical Networks 381 two data priorities, or COS categories. Additionally, packet rings will also provide a rapid “layer-twoyy protection signaling protocol, designed to match the 50-ms timescales yielded by SONET/SDH [158]. The current RPR framework focuses on two (i.e., dual) counterpropagating “ringsyy that can both carry working traffic (i.e., no reserved protection bandwidth for bandwidth efficiency). All control messages are carried “in-stream,” making the control strictly in-band. Additionally, (layer-two) destination stripping is performed for unicast flows, unlike earlier source-stripping FDDI rings, permitting spatial reuse of bandwidth (note that multicast and broadcast still require source stripping, however). Collectively, the above features significantly improve ring capacity utilizatiodthroughput (i.e., bandwidth multiplication, see [1581 for details). Currently, a spatial reuse protocol (SRP) framework has been tabled for standardization and is commerciallyavailable (amongst others), aiming to provide all RPR features (e.g., protection switching, topology discovery,bandwidth fairness, etc.). In particular, the related protection switching protocol, termed the intelligent protection switching (IPS) protocol, is an architectural counterpart to the SONET/SDH K 1-K2 byte protocol. Nevertheless, since packet rings have emerged from enterprise LAN requirements, they clearly cannot support legacy TDM t r a f h (without proprietary mappings). Moreover, since RPR nodes must perform “electronic” packet processing operations, realistically, their scalability to high speeds (10 Gb/s and beyond) and large node counts needs to be proven [ 1581. As such, they are most suitablefor new IP-based carriers [65], at least initially, providing very low-cost metro solutions. Most likely3initial deployments will run over smaller-scale metro edge rings, and here, packet ring/optical ring interworking will become an important issue (see [75]for early discussions on this topic). Nevertheless, it is likely that future advances in optical packet switching (Section 8) will be leveraged to design substantially faster terabit packet rings. Overall, this is an evolving area, and more work will emerge (i.e., standardization, design, and performance evaluation). 6.4 MIGRATION STRATEGIES Metro edge evolution will likely exhibit high variability due to the diversity of subrate client protocols and available solutions. Ultimately, any chosen solution will depend very much upon existing infrastructures, economic considerations, and clientloperational needs. Many metro edge networks still run at lower TDM bit rates (OC-3/STM-l, OC-12/STM-4), and therefore, relatively large capacity expansions can be cost effectively achieved by simply upgrading to higher bit rate TDM systems [54]. This is especially true for moderate demand growth and smaller tributary rates (DS3, OC-3/STM-1) and/or fiber-rich scenarios. Meanwhile, newer operators with littlelno existing gear and more constrained fiber counts may prefer compact “data-efficient” NGSiMSPP platforms to rapidly provision a full range of services (TDM and 382 Nasir Ghani et al. layer two/three). Furthermore, those incumbents with costly “best-in-class” routinglswitching gear may adopt a more cautious strategy towards NGS, choosing to deploy full solutions only for highly compellingprice/performance alternatives. Still other incumbents may prefer NGS, given its strong origins from existing paradigms and finer-capacity allocation capabilities. Meanwhile, for native Ethernet traffic, clearly packet-ring technology will be much more cost-effective than solutions using “telecom adapter” interfaces (SONETEDH, NGS). Moving forward, this will likely be the solution of choice for newer data-centric operators without legacy clients. For example, packet rings will offer very low-cost aggregationbetween residential (Internet) cable and DSL hubs. Although the above alternatives may prevent immediate deployment of DWDM technology in the metro edge, in the longer term it remains the most scalable and complementary solution [63, 641. Namely, DWDM rings can agnostically support all other solutions (e.g., by reserving different wave length sets for SONET/SDH, NGS, and IP routing solutions) and will clearly decouple operators from continuing fluctuations in technology directions. More importantly, DWDM rings will allow operators to easily expand service offerings (e.g., legacy TDM voice/private line to data or vice versa). Most likely, many larger operators with existing legacy gear and diverse, specialized “higher-layer” systems (IP routers, ATM switches, Fibre Channel hubs, telephony switches) will deploy passive DWDM rings with flexible edge aggregation interfaces to consolidate their architectures [35]. This will ensure abundant capacities for any future demand “spikes,” and also address the growing “wavelength services” market (gigabits to customer edge [62, 641). Other operators who choose NGS gear may also move to modularly add DWDM capabilities in the future (e.g., flexible MSPP solutions). A key planninglcosting activity will be choosing when to cross over to optical ring architectures (see also Section 5.4). For example, some operators may move from OC-48/STM-4 rings to DWDM rings rather than upgrade to lessscalable OC- 192/STM-16systems. Clearly, metro edge evolutionrequires more defining studies, see [63-651 for market-related considerations 7. Network Standards Interoperable standards are a key factor in ensuring the success and adoption of next-generation metro optical networking solutions. Standards help to properly formalize both features and functionality, and will also help insulate operators from single-vendor solutions. The key components of optical interoperabilityare now beginning to emerge. At the physical and link layers, many interface standards are well-defined, for example, SONET/SDH concatenated formats, ITU-T wavelength grids, IEEE Ethernet interfaces, OIF interfaces, etc. Increasingly, higher-layer control and architectural issues are 8. Metropolitan Optical Networks Data Plane ATM AALS mapping to SONETISDH frames Packet-overGigabit Ethernet framing, LAN and WAN standards 383 IP packet routing IP, ATM traffic engineennglresource management, SONETlSDH protechon switching I . I , 7 IP/MPLS routing and traffic SONETISDH switching Packet, wavelength, fiber rouhng. Setup signaling, resource/traffic engineenng, rotechodrecovery Trends ......... ........ ........ P \ ~ Physical opticslfiber layer Fig. 8.18 IP over optics: data and control plane “delayering.” now being considered, and the ITU-T optical transport network (OTN) architecture defines three layers of transport (channel, multiplex, transport) [ 12, 1451. Meanwhile the IETF and OIF are beginning to tackle more detailed network signaling/protocols issues [157] and the ANSI T l X l is studying optical ring frameworks. Perhaps the most notable development is the multiprotocol lambda switching (MPAS) [148, 149, 1571framework, superseded recently by the more generalized (emerging) multiprotocol label switching (GMPLS) framework [150, 1541. GMPLS represents a strong push to increase horizontal control plane integration (data and optical) by extendingheusing existing data networking concepts/protocols. The overall aim is to replace the features of multiple protocol layers in traditional multilayered models (e.g., separate addressing schemes, SONET/SDH protection, ATM traffic engineering) with a more unified solution, as shown in Fig. 8.18. A brief summary is presented here (refer to related references for details). 7.1 CHANNEL PROVISIONING There are several major required components for dynamic channel provisioning and advanced SLA management in metro optical networks, namely setup signaling, resource discovery, and constraint-based routing [7]. GMPLS implements all of these requirements by extending MPLS signaling and 384 Nasir Ghani et al. resource discovery protocols and defining multiple link-specific abstractions of the original MPLS label-swappingparadigm (ie., “implicit labels” for timeslots, wavelengths, and fibers), see [148,149,153,154]. Thesedefinitionscanbe further coupled with hierarchical label-stacking schemes to exploit scalability (e.g., packet labels into TDM circuit labels into lambda labels). In particular, this ubiquity make GMPLS an ideal control framework for multiservicemetro edge platforms (Sections 6.1 and 6.2). First, optical channel setup signaling is accomplished by extensions to MPLS signaling protocols, namely RSVP-TE (RSVP traffic engineering) and CR-LDP (constraint-routing label distribution protocol) (see [148, 1541 and references therein). Here, the explicit-routing (ER) capability [ 1491 is used to indicate the channel route and reserve resources. Meanwhile, actual route computation (Le., RWA, Section 5.2.3) is done via constrained routing/path computation (Le., constraint-basedrouting (CBR) [1481). Moreover, CBR can also incorporate advanced trafiic/resourceengineering algorithms for dynamic ring/mesh networks. Finally, route computation requires network topological/resourceinformation (i.e., self-inventorycapability), and this is propagated via extensions to pertinent routing protocols, namely open-shortest path fist (OSPF) and intermediate-systemto intermediate-system (IS-IS) (see [154] for full details). Examples include fiber-types,wavelength counts, wavelength conversion resources, and possibly even analog metrics. More recently, resource diversity information has also been proposed to explicitly capture risk associations (physical, logical) [106, 1551, and this can help channel-routing algorithms improve the “disjointedness” between working/protection paths. A key concern is provisioning architectures, namely centralized or distributed architectures [7]. Data routing traditionally uses distributed control (signaling, routing), whereas optical ring/mesh routing is much more amenable to centralized implementations [7, 106, 1551. For example, many shared protection schemes (Sections 5.2.2 and 5.3) require advanced optical ring/mesh RWA algorithms with global per-connection state. Distributinghlooding such information to all nodes is clearly unscalable. In other cases, if transmission impairments are incorporated, the resulting computations themselves are less amenable to distributed renditions. Nevertheless, many distributed shortestpath heuristic RWA algorithms are still possible, and possible future advances (components, algorithms) may permit more feasible distributed renditions (see [l, 7,1281).Regardless, the GMPLS framework can be applied for either model (e.g., appropriate LSA extensions (distributed) and/or policylroute servers (centralized) 11611). 7.2 PROTECTION SIGNALING Dynamic optical rings, and likely even hybrid/mesh architectures (Sections 5.2.2 and 5.3), must provide fast optical protection signaling protocols in order to match the capabilities of SONET/SDH APS (i.e., 50-ms 8. Metropolitan Optical Networks 385 recovery). Moreover, these protocols are necessary to implement advanced service definitions (e.g., multilevel resource sharing, Section 5.2.2). Various standardization efforts for optical protection are underway [15,75,95], but no signaling standards currently exist. However, early proposals for fast optical protection signaling in GMPLS have appeared [75, 951. For example, [75, 941 discusses extending existing MPLS LSP protection signaling or defining an altogether new optical A P S protocol. Initially, APS protocol(s) can be defined for rings (ie., leveraging on SONETEDH concepts), but subsequent generalizations to hybrid ring-mesh networks can also be considered. Meanwhile, [95] presents a new “lightweight” restoration signaling protocol in lieu of RSVP/CR-LDP signaling. In general, until such standards are dehed, metro operators will continue to rely upon SONET/SDH protection, ultimately delaying the introduction of dynamic optical services provisioning. Assuming that fast optical protection signaling schemes will emerge, interlayer protection coordination becomes an issue. Many metro-area protocols have their own recovery mechanisms, operating across multiple domains (e.g., optical channel protection, SONET APS, MPLS LSP protection switching, IP flow rerouting), and the simultaneous interference of such functionalities can be very detrimental. Specifically, problems can include reduced resource utilization, increased recovery times, or routing instabilities [96, 100,102, 1031 (e.g., prolonged SONETISDH recovery times, Section 5.2.2). DWDM can also compromise higher-layer survivability, as the high degree of multiplexing can lower higher-layer connectedness without proper preplanning [1021. Additionally, replicated (excessive) protection functionality across layers can be very inefficient [98]. To date, no standards exist for multilayer protection interworking, and this is largely done via careful preplanning, see [1021 for a detailed study. Moving forward, more formalized mechanisms are needed for coordinating interlayer recovery actions between the packet/wavelength/fiber levels, termed escalation strategies [94, 100, 1031. Various escalation strategies are possible, such as bottom-uphop-down [7, 1001 or serial/parallel[103], and these will require complex interlayer signaling and hold-off timer mechanisms. In particular, metro edge (NGS, MSPP) platforms handling many protocols and their associated control/monitoring functions present some very unique protection coordination possibilities. For example, routing diversity information can be used to ensure higher-layer workinglprotection resource separation. Overall, this is a complex area that requires much more work [98]. 7.3 DOMAIN INTERFACING Edge clients will need intelligent interfaces in order to automatically requesthelease “optical” bandwidth (i.e.,“bandwidth-on-demand” applications. Here, several interworking models have been defined to propagate routing/connectivity information between the data (IP) and optical routing domains, namely overlay, peer, and integrated models [154, 1551. The overlay 386 Nasir Ghani et aL model achieves maximum separation using separate routinglsignaling protocols in each domain and defining an intermediate optical user network interface (0-UNI) [146, 147, 1511. Conversely, the peer model achieves maximum integration running the same (extended) routinglsignalingprotocols in both domains. However, this proves overly complex/cumbersome, requiring routers (optical nodes) to maintainlprocess optical (data) routing information. The integrated model strikes a balance between the above two schemes, running different instances of the same protocols (e.g., signaling, routing with extensions) and using gateway protocols for end-point exchange (see [155] for full details). In the near term, however, the overlay model will see most favor since related UNI standards are available and proprietary optical control protocols can be accommodated. Moreover, this model provides better multiservice support, not just IP, and thus is well-suited for the metro space. At the core of "optical" channel provisioning is the concept of a service definition, as extended via an 0-UNI or elementhetwork management system (EMS/NMS) interface. The 0-UNI is intended improve vertical integration between layers by allowing automated service discovery along with bandwidth signaling functions (e.g., request/release/modify operations). In addition, a set of generic signaled attributes are defined that can be mapped to subsequent channel requests (e.g., RSVP/CR-LDP, Section 7.1; framing type; bit rate; protection type; priority; etc.) [147, 151, 157. These mappings can cover a broad range of underlying capabilities (e.g., ring protection, mesh restoration, etc.) [94]. Signaled attributes will help facilitate multiple service levels for differing customer requirements, a necessary requirement in metro networks. Overall, 0-UNIs are very germane to metro edge platforms, and even metro core nodes with direct wavelength interfaces. Several 0-UNI definitions have been tabled for standardization of which both the ODSI interface [147] and OIF standard [146] have been completed. Meanwhile, interdomain channel routing and protection coordination between operator networks will require (optical) network-to-network interface (0-NNI) definitions and early considerations are also underway here [ 152, 1551. 7.4 NETWORKMANAGEMENT As metro network elements continue to integrate many more diverse inter- faceskapabilities, especially at the metro edge, integrated network management is obviously a major requirement [22]. Network management is a very large focus area in its own right, and here only a brief discussion is provided due to scope limitations. In general, the well-accepted telecommunication network management (TMN) framework defines a hierarchical management model comprising vendor element management systems (EMS) entities interfacing with multivendor network management systems (NMS). Although most early metro (DWDM) systems only provided proprietary EMS support for point-to-point transport nodes [156], more advanced solutions are now 8. Metropolitan Optical Networks 387 being offered. Optical channel (and link) visibility is of particular concern here, and is complicated by the fact that there are still no standards for related parameters. As an interim, detailed SONET/SDH B 1/JO overhead byte monitoring (or digital wrappers equivalent) can be used at ‘ ‘ ~ p a q u e ~ ~ to points measure bit-level performance (e.g., errored/severely-enored seconds, etc.). These “opaque” points can either be inside opaque nodes and/or at “edge” interfaces in transparent networks. Note that some vendors are also beginning to offer various (proprietary) sets of optical monitoring parameters, such as laser powers/current/temperature, amplifier power, etc. [68]. Meanwhile, with metro networks supporting many more protocols, advanced NMS solutions will be the key enablers for “end-to-end” services management operating across multiple vendors’ equipment [22]. Ideally, NMS solutions should provide operators with a full range of functionalities that are derived across multiple protocol domains, such as remote configuration, performance monitoring, rapid fault detection/alarm processing, failure isolation, diagnostics testing, and comprehensivelogginglreporting, well-defined graphical interfaces, etc. [43]. However, genuine multivendor services management requires widescale adoption of standardized management frameworks, and overall this area is still in its infancy. Going forward, the common approach here will likely be to adapt TMN concepts and develop appropriate management information models between EMS and NMS systems [156]. 8. Future Directions As data t r a c volumes continue to increase, packet-switching paradigms will gain increasing favor, owing to their inherent statistical efficiencies [7, 1651681. Particularly, optical packet switching (OPS) designs are under study, utilizing optical techniques to perform as much of the packet-routing operations as possible (e.g., switching, buffering, even processing, i.e., “fourth generation” optical networks). On the transport level, data packets are sent directly over wavelength channels (i.e., “packet-over-lightwave” (POL), Fig. 8.1). OPS nodes intend to achieve ultra-high packet throughputs, in the multiterabits range, largely surpassing current gigabit router designs. Even though all packet-switchinglrouting functions are difficult to perform optically (and may remain so for the foreseeable future), various multifaceted opto-electronic designs are being studied, and inevitably this work will lead to significant improvements in packet-routing performance. In particular, the three main functionalities pertaining to OPS are buffering, switching, and header processing (i.e., label lookup) [2, 1681. Some brief details are reviewed, and readers are referred to related references for more complete treatments. By and large, OPS nodes have the same architecture as electronic packet switches (i.e., input buffering, space switching, output buffering [168,169], see Fig. 8.19). In packet switching, contention can occur if multiple packets are 388 Nasir Ghani et al. Metropolitan hybrid packetkircuit routing network Header extraction (electmnic. possihly optical logic) Integrated opticallelectronic packetkircuit switching node : Coupleror space - Fig. 8.19 Hybrid metro optical packet/circuit switching networkhode. routed to the same output port [165], and resolution is commonly achieved via buffering. Most OPS designs use fiber delay line buffers, and recently, the use of fast tunable laserskonverters has also been proposed to exploit the wavelength dimension to store multiple (wavelength) packets in a given delay line [167]. However, fiber delay line buffers constrain packets to multiples of a fixed length, as there is no means to retrieve packets before minimum buffer delays. Furthermore, large buffer sizes become costly/bulky (requiring complex sharing setups), and additionally, fiber attenuation concerns will limit the number of “circulations,” (usually under 100 [165]). Hence, as a tradeoff, a mixture of electronic memory and delay line buffering can be utilized [166]. Note that there are also very interesting, early developments in all-optical memories (e.g., molecular transistors, see references in [ 1671). OPS switching fabrics, meanwhile, can also exploit the wavelength dimension to reduce contention and boost throughputs by orders of magnitude. However, nanosecond timings are needed for packet transfers between switch ports, and this can be problematic for MEMS technology. SOA gate technology has been considered here, although careful design is necessary to control crosstalk [I 661. Another switching setup couples ultra-fast tunable lasers and wavelength converters with passive wavelength routing devices (e.g., AWG) [ 1691. As component tunability performances improve and integration technologies mature, this approach 8. Metropolitan Optical Networks 389 may become very feasible. Finally, carefully note that unlike circuit switching, OPS is still linked to the bit rate of the client signal, at least the header. Specifically, even though payload flow is optically transparent, electronic header processing/synchronization (and subsequent control of switchinglbuffering resources) must be done electronically, as “all-optical” processing is not currently feasible (see [1681).Clearly, electronic costhcalability concerns can arise for high-wavelengtldfiber count systems and/or very small packet sizes, and this may limit the complexity/range of label processing/packet filtering operations performed. Consequently, various bit-serial packet-coding techniques and/or guard-band schemes have been proposed to reduce electronic processing bottlenecks [169]. Note also that transparent payload sections will suffer from multi-hop optical degradations (loss, crosstalk), and this may require all-optical regeneration to maintain transmission distances. Overall, OPS will provide a good match for limited metro distances and can reuse much of the existing packet-switching (MPLS, DiffServ) protocol suites [156, 166, 1671. Moreover, metro OPS will prove highly complementary to emerging packet-based PON access solutions. Most likely, future metro packet switching solutions will evolve towards hybrid opto-electronic switching architectures (Fig. 8.19). Specifically, electronic switchinglbuffering will be utilized for finer-granularity/more complex label-processing “edge” operations, whereas OPS will implement less complex label-swapping functionalities, achieving higher throughputs for more “aggregated” packet flows (e.g., 10Tb/s stated in [166]). This delineation potentially lends well to “optical” packet rings, where more “coarse” stages (layer two COS)can be implemented using OPS. Even more germane to the metro arena, optical packet- and circuit-switching paradigms can be integrated onto a common platform, as the packet switches are still optically transparent to data payloads [167]. Namely, lightpaths can be switched over the same OPS switching fabric by simply “decoupling”the switching state from header-processingcontrol logic @e.,bypass control logic and apply zero delay lines, Fig. 8.19). This optical circuit bypassing will permit transparent legacy support (in exactly the same manner as current DWDM systems, Section 5): and when applied to the packet domain, will further improve scalability (Le., eliminate per-node processing of large transithabeled packet flows). Overall, OPS is an exciting new frontier, and future advances in optical processing/bufferingwill undoubtedly yield fundamental conceptual evolutions in the metro domain. Note also that slotted DWDM rings have also been proposed for accesdmetro data transport, utilizing fast tunable transmitters and/or receiver devices and fixed slot timings [170, 1711. Here, no optical buffering is performed inside the ring, and instead, advanced multiaccess (MAC) protocols are used to arbitrate DWDM channel slots in a fair manner between multiple users/traffic classes. These rings require edge packet buffering but can yield improved efficiencies versus circuit-provisioned rings, as the wavelengths are shared between multiple source/destinationpoints. A sample, advanced design using subcarrier sensing 390 Nasir Ghani et a. l techniques is studiedin [171]. However,fixed slot timings arevery inefficientfor variable-length IP packets, and proposed protocol amendments (e.g., multiple slot sizes, backoff schemes [171]) entail further desigrdprotocolcomplexity. Moreover, access-controlprotocol timingsmay adverselyaffect scalabilityover larger metro core distances. 9. Conclusions Metropolitan networks occupy a strategic place in the overall network hierarchy, bridging end-users with abundant long-haul capacities. Traditionally, hierarchicalSONET/SDH architectureshave dominated the metro landscape, with slower speed access rings interconnectingto larger, faster metro core rings. However, as metro operators look to the future, many foresee surging bandwidth demands, primarily driven by data traffic, and a plethora of diverse clients with differing protocols and service requirements. As competition intensifies, legacy multilayered architectures are proving overly sluggish and unsalable in meeting complex, stringent service requirements. Clearly, metro operators are in urgent need of scalable,flexible, multiservice bandwidthprovisioning solutionsthat allow for achieving a high level of servicedifferentiation. DWDM technology provides many benefits in the metro arena, including scalable capacity, transparency, and survivability. Moreover, many technoeconomic studies have confirmed the cost-effectivenessof DWDM for bit rates beyond OC-lUSTM-4, bolstered further by falling componentpricepoints As a result, DWDM technology has gained strong favor as a metro core solution, and various architectures are possible (ranging from simpler point-to-point transmissionsystemsto dynamicwavelength-routingring and mesh networks). Nevertheless, given the large existing base of (SONETEDH) fiber rings in the metro area, network migration is a key issue. Very likely, the first step in this migration will be a move to point-to-pointDWDM “fiber-relief”applications, and then onwards to more advanced optical ringlhybrid architectures Meanwhile, metro edge networks are evolving to represent a merging of the optical and electronic domains, aggregating many user protocols onto large metro core wavelength tributaries Many metro edge solutions have been proposed, ranging from DWDM edge rings, next-generation SONETEDH, and multiservice provisioningplatforms, to “IP-based”packet rings. The choice of edge solution will clearly depend upon an individual operator’sneeds,but over time, those incorporating DWDM technologywill be most beneficial and hence will likely gain prominence. Acknowledgments The authors are indebted to the editors, Dr. Tingye Li and Dr. Ivan Kaminow, for their support, guidance, and overall patience in the preparation of 8. Metropolitan Optical Networks 391 this manuscript. The authors would also like to thank Mrs. Amber Alam, Mr. Mohamed Haiba, Mr. Paul Bonenfant, Dr. Hongxing Dai, Dr. De Y u Zang, Dr. James Fu, Dr. Zhensheng Zhang, Dr. Xinyi Liu, Mr. Ajay Sarkar, Dr. Xinhong Wang, Dr. Zhijian Wang, Mr. Bill Berry, and Mr. Don Buell for their invaluable feedback and discussions. Additionally, the authors are extremely thankful to Mrs. Rozsa Punkosti and Dr. Ti-Shiang Wang for their assistance with related references. Abbreviations ADM ASIC ATM AWG B-DCS BLSR BPSR CAGR CDM CLEC CMD Add-drop multiplexer Application-specific integrated circuit Asynchronous transfer mode Arrayed waveguide grating Broadband digital cross-connect Bidirectional line-switched ring Bidirectional path-switched ring Compound annual g o t rate rwh Code-division multiplexing Competitive local exchange carrier Chromatic mode dispersion Central office Class of service Customer premise Coarse WDM Digital cross-connect Distributed feedback laser Differentiated services Digital loop carrier Dedicated protection ring Dual-ring interconnection Digital subscriber loop DSL access multiplexer Dense WDM Erbium-doped fiber amplifier Erbium-doped waveguide amplifier Element management system Enterprise system connectivity Electronic cross-point switch Fiber distributed data interface Forward error correction Fiber connectivity channel Fiber to the curb co cos CP CWDM DCS DFB DifTServ DLC DPRING DRI DSL DSLAM DWDM EDFA EDWA EMS ESCON EXC FDDI FEC FICON FTTC 392 Nasir Ghani et a. l FTTH GaAs GFP GMPLS HDLC HFC IC IntServ IOF IP IS-IS IXC LCAS LEC LMDS MAC MEMS MFL MMDS MMF MPLS MSPP NDF NGS NMS NWDM NZDSF OADM OCh OED 0-FDM OMS 0-NNI OOK OPS osc 0-SNR OSPF OTN 0-UNI 0-VPN oxc PBX PDH Fiber to the home Gallium arsenide Generic framing protocol Generalized MPLS High-level data link control Hybrid fiber coax Integrated circuit Integrated services Interoffice fiber Internet protocol Intermediate-system to intermediate-system Interexchange carrier Link capacity adjustment scheme Local exchange carrier Local multipoint distribution service Media access control Micro electro-mechanicalsystem Multifrequency laser Multichannelmulti-point distribution system Multimode fiber Multiprotocol label switching Multiservice provisioning platform Negative dispersion fiber Next-generation SONET Network management system Narrow WDM Non-zero dispersion shifted fiber Optical add-drop multiplexer Optical channel Optical edge device Optical frequency division multiplexing Optical multiplex section Optical network-to-network interface On-off keying Optical packet switching Optical supervisory channel Optical signal-to-noiseratio Open-shortest path first Optical transport network Optical user network interface Optical virtual private network Optical cross-connect Public branch exchange Plesiochronousdigital hierarchy 8. Metropolitan Optical Networks 393 PLC PMD POL PON POP POS QAM QOS RF RPR RSVP RWA SAN SDH SDM SHR SLA SMF SNR SOA SONET SPRING TDM TDMA TLAN TMN TSI UPSR VLAN VPN W-DCS WDM Planar lightwave circuit Polarization-mode dispersion Packet over lightwave Passive optical network Provider points of presence Packet over SONET Quadrature amplitude modulation Quality of service Radio frequency Resilient packet ring Resource reservation protocol Routing and wavelength assignment Storage area network Synchronous digital hierarchy Space division multiplexing Self-healing ring Service level agreement Single mode fiber Signal-to-noise ratio Semiconductor optical amplifier Synchronous optical network Shared protection ring Time-division multiplexing Time-slotted multiaccess Transparent local area network Telecommunication Management Network Time-slot interchange Unidirectional path-switched ring Virtual local area network Virtual private network Wideband digital cross-connect Wavelength division multiplexing References [I] B. Mukherjee, Optical CommunicationsNetworks, McGraw-Hill, New York, NY, 1997. [2] R.Ramaswami, K. Sivarajan, Optical Networks: A Practical Perspective, Morgan Kaufmann Publishers, San Francisco, CA, 1998. [3] T. E. Stern, K. Bala, Multiwavelength Optical Networks: A Layered Approach, Addison-Wesley, Reading, MA, 1999. 394 Nasir Ghani et a. l [4] M. Chow, Understanding SONET/SDH: Standards and Applications, Andan Publisher, New Jersey, 1995. [5] T. H. Wu, Fiber Network Service Survivability, Artech House, Boston, 1992. [6] R.Alferness, et al., “A Practical Vision for Optical Transport Networking,” Bell Labs Technical Journal, Vol. 4, No. 1, January-March 1999, pp. 3-18. [7] N. Ghani, et al., “On IP Over WDM Integration,” IEEE Communications Magazine, Vol. 38, No. 3, March 2000, pp. 72-84. [8] GR-1400-CORE, SONET Dual-Fed UnidirectionalPath Switched Ring (UPSR) Equipment Generic Criteria, Bellcore, Issue 1,Revision I , October 1995. [9] GR-1230-CORE, SONET Bi-directional Line-Switched Ring Equipment Generic Criteria, Issue 4, December 1998. [101 J. Sosnosky, “Service Applications for SONET DCS Distributed Restoration,” IEEE Journal on Selected Areas in Communications, Vol. 12, January 1994, pp. 59-68. [ll] A. Mali% “PPP over SONETEDH,” Internet Engineering Task Force (IETF) Requestfor Comments (RFC) 2615, June 1999. [121 P. Bonenfant, A. Moral, “Optical Data Networking,” IEEE Communications Magazine, Vol. 38, No. 3, pp. 63-70, March 2000. [13] “Draft ITU-T Rec. G. 709,” ANSI TlX1.5/2000-246, October 2000. [141 “GFP Draft Specification-Revision2,” ANSI TlX1.5/2001-024, January 2001. [15] M. Souillere, “Proposed ITU-T Contribution on Transparent OCh SPRings,” ANSI TlX1.5/20OI-027, January 2001. [161 L. Szerenyi, “Approach to OSC Standardization,” ANSI T1X1.5/2001-002, January 200 1. [17] A. Stok, E. Sargent, “Lighting the Local Area: Optical Code-Division Multiple Access and Quality of Service Provisioning,” IEEE Network, Vol. 14, No. 6, NovemberDecember 2000, pp. 4246. [18] T. Mossberg, M. Raymer, “Optical Code-Division Multiplexing,” Optics & Photonics News, Vol. 12, No. 3, March 2001, pp. 50-54. [19] N. J. Frigo, “A Survey of Fiber Optics in Local Access Architectures,” in Optical Fiber Telecommunications IIZA, Academic Press, San Diego, CA, 1997, pp. 461-522. [20] I. P. Kaminow, “Advanced Multiaccess Lightwave Networks,” in Optical Fiber Telecommunications IIIA, Academic Press, San Diego, CA, 1997, pp. 560-593. . [21] J Senior, M. Handley, M. Leeson, “Developmentsin Wavelength Division Multiple Access Networking,” IEEE Communications Magazine, Vol. 36, No. 12, December 1998, pp. 28-36. [22] P. V. Hatton, F. Cheston, “WDM Deployment in the Local Exchange Network,” IEEE CommunicationsMagazine, Vol. 36, No. 2, February 1998, pp. 5661. [23] D. J. T. Healty, et al., “Optical Wireless: The Story So Far,” IEEE Communications Magazine, Vol. 36, No. 12, December 1998, pp. 72-82. [XI H. Salloum, “How Does Local Multipoint Distribution Service (LMDS) Compare to Fiber to the Home (FTTH),” National Fiber Optic Engineers Conference (NFOEC) 1998, Orlando, FL, September 1998. 8. Metropolitan Optical Networks 395 [25] J. Refi, “Optical Fibers for Optical Networking,” Bell Labs Technical Journal, Vol. 4, No. 1, January-March 1999, pp. 24G261. [26] E. Goldstein, L. Lin, J. Walker, “Lightwave Micromachines for Optical Networks,” Optics & Photonics News, Vol. 12, No. 3, March 2001, pp. 60-65. [27] Y. Danziger, D. Askegard, “High-Order-Mode Fiber,” Optical Networks Magazine, Vol. 2, No. 1, Januarymebruary 2001, pp. 40-50. [28] M. Zirngibl, “Multi-Frequency Laser and Applications in WDM Networks,” IEEE Communications Magazine, Vol. 37, No. 2, February 1999, pp. 39-41. . [29] P Fairley, “The Microphotonics Revolution,” Technology Review, July/August 2000, pp. 38-44. [30] J. Bautista, “Multiplexers Bring DWDM to Metro/Access Markets,“ WDM Solutions, February 2000, pp. 11-14. [311 C. Huang, et al., “Ultra-Low Loss, Temperature-Insensitive16-channel100-Ghz Dense Wavelength Division Multiplexers Based on Cascaded AU-Fiber Unbalanced Mach-Zehnder Structure,” Optical Fiber Communications Conference (OFC) 1999, San Diego, CA, February 1999. [32] B. Broberg, et al., “Widely Tunable Semiconductor Lasers,” Optical Fiber CommunicationsConference (OFC) 1999, San Diego, CA, February 1999. [33] L. Lin, “Free-Space Micromachined Optical-SwitchingTechnologies and Architectures,” OpticalFiber CommunicationsConference (OFC) 1999, San Diego, CA, February 1999. [34] Y.Pan,et al., “Optical Multi-Stage Interconnection Networks,” IEEE Communications Magazine, Vol. 37, No. 2, February 1999, pp. 50-56. [35] Y. Chen,et al., “Metro Optical Networking,” Bell Labs TechnicalJoirrnal, Vol. 4, No. 1, January-March 1999, pp. 163-186. [36] X. Cheng, “Beyond Transport: Creating the Metro Network,” Telephony, May 22,2000. [37] T. Thakur, “Deploying Metro DWDM Systems,” National Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [38] E. Almstrom, et al., “Requirements and Solutions For Reconfigurable Metro WDM Networks,” National Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [39] T. E. Adams, et al., “Metropolitan Area Optical Networking,” National Fiber Optic Engineers Conference (NFOEC) 1998, Orlando, FL, September 1998. [40] R. H. Cardwell. 0. J. Wasem, H. Kobrinski, “WDM Architectures and Economics in Metropolitan Areas.” Optical Networks Magazine, Vol. 1, No. 3, July 2000, pp. 41-50. [41] R. Cardwell, et al., “WDM Network Economics,” National Fiber OpticEngineers Conference (NFOEC) 1997, San Diego, CA, September 1997. [42] R. Boncek, et al., “Fiber and Systems for Metropolitan Optical Networks,” National Fiber Optic Engineers Conference (NFOEC), Orlando, FL, September 1998. 396 Nasir Ghani et sl. [43] C. D. Webb, T. D. Krause, “Drivers and Enabling Technologies for WavelengthBased Services in The All Optical Network,” National Fiber Optic Engineers Conference (NFOEC) 1998,Orlando, FL, September 1998. [44] S. Wilkinson, “Metropolitan Area Transport Requirements,” National Fiber Optic Engineers Conference (NFOECJ 2000, Denver, CO, August 2000. [ 5 l? Wong, “Network Architectures in the Metro DWDM Environment,” National 4] Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [46] M. Simcoe, “A Layered Architecture for Metro Optical Networks,” National Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [47] M. Daoust, B. Lavallee, “Designing and Planning Metropolitan DWDM Solutions,” National Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [48] M. Wendel, R. Ward, M. Jary, “Transparent Optical Networking: Future Trends in Design and Deployment,”National Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [49] S. Greenholtz, et al., “Deployment of Dense Wave Division Multiplexing (DWDM) Technology in the Bell Atlantic Network,” National Fiber Optic Engineers Conference (NFOEC) 1998,Orlando, FL, September 1998. [50] M. Williams, D. Klepp, “Introduction of DWDM Technology in BellSouth,” National Fiber Optic Engineers Conference (NFOEC) 1998,Orlando, FL, September 1998. [51] G. Ocakoglu, N. Wauters, P Falcao, “The Technical Issues in the Selection . and Deployment of WDM Systems in a Pan-European Carriers’ Carrier Network,” National Fiber Optic Engineers Conference (NFOEC) 1998,Orlando, FL, September 1998. [52] S. Nathan, “Overcoming the Challenges of Designing, Planning, and Building an Optical Global Network That Will Support the Smooth Migration Towards the Intelligent Optical Network,” EuroForum: Intelligent Optical Networks, May 2000. [53] M. Estep, “Deploying Advanced Optical Fiber Solutions for Metropolitan Networks,” EuroForum: Intelligent Optical Networks, May 2000. [54] S. Johansson, et al., “A Cost Effective Approach to Introduce an Optical WDM Network in the Metropolitan Environment,” IEEE Journal on Selected Areas in Communications,Vol. 16, No. 7, September 1998, pp. 1109-1122. [55] L. D. Garrett, et al., “The MONET New Jersey Network Demonstrator,” IEEE Journal on Selected Areas in Communications, Vol. 16, No. 7, September 1998, pp. 1199-1219. [56] D. Stoll, et al., “Metropolitan DWDM: A Dynamically Configurable Ring for the KomNet Field Trial in Berlin,” IEEE CommunicationsMagazine, Vol. 39, No. 2, February 2001, pp. 106-113. [57] A. Richter, et al., “Germany-Wide DWDM Field Trial: Transparent Connection of a Long-Haul Link and a Multichannel Metro Network,” Optical Fiber Communications Conference (OFC) 2001, Anaheim, CA, March 2001. 8. Metropolitan Optical Networks 397 [58] A. Stavdas, “Architectural Solutions Towards a 1,000 Channel Ultra-Wideband WDM Network,” Optical Networks Magazine, Vol. 2, No. 1, JanuarylFebruary 2001, pp. 5140. [59] C. Decusatis, “Dense Wavelength Division Multiplexing for Parallel Sysplex and Metropolitan/Storage Area Networks,” Optical Networks Magazine, Vol. 2, No. 1, JanuaryRebruary 2001, pp. 6!%80. [60] B. St. Amaud, “Current Optical Network Designs May Be Flawed,” Optical Networks Magazine, Vol. 3,No. 2, March/April2000, pp. 21-28. [61] “WDM and Optical Networks-Metro WDM: Market Forecast and Analysis,” Ryan Hankin Kent W K ) Report, February 2001. [62] D. Cooperson, “Short-Haul WDMlOpticalNetworking: Is It (Finally)for Real?” National Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [63] “SONET Transport and DCS-Metro WDM: Market Forecast and Analysis,“ Ryan Hankin Kent (RHK) Report, March 2001. [64] “Metro Optical Networks: Metro DWDM and the New Public Network,” Pioneer ConsultingReport, 1999. [65] “Optical Edge Networks: Market Opportunities for Integrated Optical Network Solutions in Metro Networks,” Pioneer Consulting Report, July 2000. [66] J. P. Ryan, T h e Golden Age of Telecommunications?” Ryan Hankin Kent ( M K ) STARTRAX2OOO Conference, Palm Springs, CA, November 2000. [67] J. Montgomery, “Optical AddlDrop Multiplexers Match Network Growth,” WDMSolutions, February 2000, pp. 4-8. [68] K. Struyve, et al., “Application, Design, and Evolution of WDM in GTS’s PanEuropean Transport Network,” IEEE ComrniinicationsMagazine, Vol. 38, No. 3, March 2000, pp. 114-121. [69] G. Ocakoglu, K. Struyve, P. Falcao, “The Business Case for DWDM Metro Systems in a Pan-European Carrier Environment,’’ National Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [70] 0. Gerstel, R. Ramaswami, “Upgrading SONET Rings with WDM Instead of TDM: An Economic Analysis,” Optical Fiber ComrniinicationsConference (OFC) 1999, San Diego, CA, February 1999. [71] G. Redifer, “DWDM vs. TDM In Metro and Long Haul Applications,”National Fiber Optic Engineers Conference (NFOEC) 1997, San Diego, CA, September 1997. [72] L. Nederlof, et al., “Value Proposition for Configurable Optical Network Elements,” Optical Fiber Communications Conference (OFC) 2001, Anaheim, CA, March 2001. [73] P. Arijs, et aZ., “Design of Ring and Mesh Based WDM Transport Networks,” Optical Networks Magazine, Vol. 1, No. 3, July 2000, pp. 2 5 4 0 . [74] W. Grover, “High Availability Path Design in Ring-Based Optical Networks,” IEEE/ACM Transactions on Networking, Vol. 7, No. 4, August 1999, pp. 558-574. 398 Nasir Ghani et al. [75] N. Ghani, et al., “ArchitecturalFramework for Automatic Protection Provisioning In Dynamic Optical Rings,” Internet Draj?, draft-ghani-optical-rings-01. txt, March 2001. [76] D. Guo, et d., “Hybrid Mesh-Ring Optical Networks and Their Routing Information Distribution Using Opaque LSA,” Internet Draj?, draj?-guo-opticalmesh-ring-OO.txt, December 2000. [77] T. Miyazaki, T. Kato, S. Yamamoto, “A Demonstration of an Optical Switch Circuit w t ‘Bridge-and-Switch’Function in WDM Four-Fiber Ring Networks,” ih ZEICE Transactions on Communications, Vol. E82-B, No. 2, February 1999, pp. 326-334. [78] N. Nagatsu, et al., “Flexible OADM Architecture and Its Impact on WDM Ring Evolutionfor Robust and Large-Scale Optical Transport Networks,”IEZCE Transactions on Communications,Vol. E82-B, No. 8, August 1999, pp. 1105-1 114. [79] S. Kuwano, H. Uematsu, “Two-Fiber Uni-Directional OADM Ring System for L-Band,” National Fiber OpticEngineers Conference (NFOEC)2000, Denver, CO, August 2000. [80] W Emkey, “Transparent DWDM Network with Optical Management,” National . Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [81 N. Antoniades, “Engineering the Performance of DWDM Metro Networks,” 1 National Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [82] R. Iraschko, et al., “An Optical 4-Fiber Bi-Directional Line-Switched Ring,” OpticalFiber CommunicationsConference (OFC) 1999, San Diego, CA, February 1999. [83] W. Zhong, “New Bi-Directional WDM Ring Networks with Dual Hub Nodes,” IEEE Global CommunicationsConference (GLOBECOM’97),Phoenix, AZ, 1997. OBECOM’97. [84] A. Zolfghari, “A Self-Healing Ring Network Topology Optimization Method,” National Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [85] C. R. Giles, M. Spector, “The Wavelength Add/Drop Multiplexer for Lightwave Communication Networks,” Bell Labs Technical Journal, Vol. 4, No. 1, JanuaryMarch 1999, pp. 207-229. [86] G. Ellinas, et al., “Architecture Considerations in Merging Multi-Vendor WDM Rings for the MONET Washington D.C. Network,” Uptical Fiber CommunicationsConference (OFC) 1999, San Diego, February 1999. [871 A. DiMichele, “Building High Capacity Broadband Metro Networks Using Optical Add-Drop Multiplex Systems,” National Fiber Optic Engineers Conference (NFOEC) 1999, Chicago, IL, 1999. [88] S. Ferguson, “Meeting OADM Performance Goals,” National Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [89] W J. Tomlinson, “Comparison of Approaches and Technologies for Wave. length AddIDrop Network Elements,”National Fiber OpticEngineers Conference (NFOEC) 1998, Orlando, FL, September 1998. 8. Metropolitan Optical Networks 399 [90] G. E. Bodeep, et al., “16 Channel ReconflgurableOptical AddlDrop Multiplexer: An Engineering Model,” National Fiber Optic Engineers Conference (NFOEC) 1998, Orlando, FL, September 1998. [91] N. A. Jackman, et al., “Optical Cross Connects for Optical Networking,” Bell Labs TechnicalJournal, Vol. 4, No. 1, January-March 1999, pp. 262-281. [92] P. Jaggi, H. Onaka, S. Kuroyanagi, “Optical ADM and Cross Connects: Recent Technical Advancements and Future Optimal Architectures,” National Fiber Optic Engineers Conference (NFOEC) 1998, Orlando, FL, September 1998. [93] J. Gruber, P. Roorda, F. LaLonde, “The Photonic SwitcWCross-Connect (PSX)-Its Role in Evolving Optical Networks,” National Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [94] N. Ghani, “SurvivabilityProvisioningin Optical MPLS Networks,” 5th European Conference on Networks and Optical Communications(NOC ZOOO), June 2000. [95] B. Rajagopalan, et al., “Signaling for Fast Restoration in Optical Mesh Networks,” Internet Draj?, draj?-bala-restoration-signcrling-OO.~t, February 2001. [96] D. Zhou, S. Subramaniam, “Survivabilityin Optical Networks,” IEEE Network Magazine, Vol. 14, No. 6, November/December 2000, pp. 16-23. [97] A. Fumagalli, L. Valcarenghi, “IP Restoration vs WDM Protection: Is There an Optimal Choice?” ZEEE NetworkMagazine, Vol. 14, No. 6, NovembedDecember 2000, pp. 34-41. 1981 S. Ayandeh, P. Veitch, “Dynamic Protection and Restoration in Multi-Layer Networks,” Optical InternetworkingForum, OIF2001.166, April 200 1. [99] R. Doverspike, et al., “Fast Restoration in a Mesh Network of Optical CrossConnects,” Optical Fiber Communications Conference (OFC) 1999, San Diego, February 1999. [loo] P. Demeester, et al., “Resilience in Multilayer Networks,” IEEE Communications Magazine, Vol. 37, No. 8, August 1999, pp. 44-51. [loll R. Batchellor, “Optical Layer Protection: Benefits and Implementation,” National Fiber Optic Engineers Conference (NFOEC)1998, Orlando, FL, September 1998. [1021 J. Sosnosky, Z. Lin, “Planning for Broadband Multilayer Survivability,” NationaZ Fiber Optic Engineers Conference (NFOEC) 1998, Orlando, FL, September 1998. [1031 J. Manchester, P. Bonenfant, “Fiber OpticNetwork Survivability: SONET/Optical Protection Layer Intenvorking,” National Fiber Optic Engineers Conference (NFOEC) 1996, Denver, CO, 1996. [lo41 R. Iraschko, et al., “Interoperability of SONET and Optical Layer Protection,” National Fiber Optic Engineers Conference (NFOEC) 1999, Chicago, IL, 1999. [lo51 C. Larsen, I? Anderson, “Signal Quality Monitoring in Optical Networks,” Optical Networks Magazine, Vol. 1, No. 4, October 2000, pp. 17-23. [lo61 J. Strand, A.Chiu, “Issues for Routing in the Optical Layer,’’ IEEE Commuizications Magazine, Vol. 39, No. 2, February 2001, pp. 81-87. [lo71 C. Fan, J. P. Kunz, “Terrestrial Amplified Lightwave System Design,” in Opfical Fiber 22iecornrnunications IIU, Academic Press, San Diego, CA, 1997, pp. 265-30 1. 400 Nasir Ghani et al. [lo81 B. Ramamurthy, et al., “Impact of Transmission Impairments on the Teletraffic Performance of Wavelength-Routed Optical Networks,” IEEE/OSA Journal o f Lightwave Technology, Vol. 17, No. 10, October 1999, pp. 1713-1723. [lo91 D. Datta, et al., “BER-based Call Admission in Wavelength-Routed Optical Networks,” OpticaI Fiber Communications Conference (OFC’98), San Jose, CA, February 1998. [110] M. Ali, B. Ramamurthy, J. Deogun, “A Genetic Algorithm for Routing in WDM Optical Networks with Power Considerations, Part I: Unicast Case,” 8” IEEE International Conference on Computer Communicationsand Network (ICCCN’99), Boston, MA, October 1999. [l 111 L. Nelson, “Challenges of 40 Gb/s WDM Transmission,” Optical Fiber Communications Conference (OFC) 2001, Anaheim, CA, March 2001. [112] S. Wilkinson, S. Sasaki, Y Nakano, “Technological Advances Towards OC768 Transmission,” National Fiber Optic Engineers Conference (NFOEC) 1998, Orlando, FL, September 1998. [113] D. Marcenac, “Benefits of Wavelength Conversion in Optical Ring-Based Networks,” Optical Networks Magazine, Vol. 1 No. 2, April 2000, pp. 23-29. [114] D. Marcenac, C. Felicite, “Practical Benefits of 4 Fiber WDM Rings With Realistic Ti-aEc,” Optical Fiber Communications Conference (OFC) 1999, San Diego, CA, February 1999. [115] B. Ramamurthy, B. Mukherjee, “Wavelength Conversion in WDM Networks,” IEEE Journal on Selected Areas in Communications, Vol. 1, No. 7, September 1998, pp. 1061-1073. [116] K. Bala, “The Benefits of Minimal Wavelength Interchange in WDM Rings,” Optical Fiber Communications (OFC) Conference 1997, Dallas, TX, February 1997. [117] E. Bouillet, K. Bala, “The Benefits of Wavelength Interchange in WDM Rings,” IEEE International Conference on Communications (ICC), June 1997, pp. 41 1 4 5 . [1181 B. Ramamurthy, “Transparent vs. Opaque vs. Translucent Wavelength-Routed Optical Networks,” Optical Fiber Communications Conference (OFC) 1999, San Diego, CA, February 1999. [119] 0. Gerstel, R. Ramaswami, G. H. Sasaki, “Fault-Tolerant Multi-Wavelength Optical Rings with Limited Wavelength Conversion,” IEEE Journal on Selected Areas in Communications,Vol. 16, No. 7, September 1998, pp. 1166-1 178. [1201 J. Elmirghani, H. Mouftah, “All-Optical Wavelength Conversion: Technologies and Applications in DWDM Networks,” IEEE Communications Magazine, Vol. 38, No. 3, March 2000, pp. 8692. [121] J. Yates, et al., “Limited-Range Wavelength Translation in All-Optical Networks,” IEEEINFOCOM, Vol. 3, March 1996, pp. 954-961. [122] G. Maier, et al., “Performance of WDM Rings With Wavelength Conversion Under Non-Poisson Tralk,” IEEE International Conference on Communications (ICC’99), Vancouver, B.C., June 1999. 8. Metropolitan Optical Networks 401 [123] 0.Gerstel, S. Kutten, “DynamicWavelength Allocation in All-OpticalRing Networks,” IEEE International Conference on Communications (ICC’97), Montreal, Canada, June 1997. [ l a ] J. Yates, M. Rumsewicz, J. Lacey, “Wavelength Converters in DynamicallyReconfigurable WDM Networks,” IEEE CommunicationsSurveys, Vol. 2, No. 2, 2ndQuarter, 1999 (available at http://www.comsoc.org/pubs/surveys). [125] M. Kovacevic, A. Acampora, “Benefitsof Wavelength Translationin All-Optical Clear-Channel Networks,” IEEE Journal on Selected Areas in Communications, Vol. 14, No. 5, June 1996, pp. 868-880. [126] S. Subramaniam, M. Azizoglu, A. Somani, “AU-Optical Networks with Sparse Wavelength Conversion,”IEEE Journal on SelectedAreasin Communications,Vol. 4, No. 4, August 1996, pp. 544556. [127] M. Stephens, et al., “All-Optical Regeneration and Wavelength Conversion in an Integrated Semiconductor Optical Amplifier/Distributed Feedback Laser’’ OpticalFiber CommunicationsConference (OFC) 1999, San Diego, CA, February 1999. [128] H. Zang, J. Jue, B. Mukherjee, “Review of Routing and Wavelength Assignment Approachesfor Wavelength-RoutedOpticalWDM Networks,” OpticalNetworks Magazine, Vol. 1, No. 1, January 2000, pp. 47-60. [139] G. A. Young, “Planning SONET Architectures for Improved Capacity Utilization,” National Fiber @tic Engineers Confemce (NFOEC) 1998, Orlando, FL, September 1998. [130] A. Repech, “Architecting Data-Aware SONET Networks,” National Fiber Optic Engineers Conference (NFOEC) 1999, Chicago, IL, 1999. [131] D. O’Connor, “Internet Optimized SONET ADM,” National Fiber Optic Engineers Confermce (NFOEC) 1998, Orlando, FL, September 1998. [132] K. Bala, “WDM Optical Network Architectures for a Data-Centric Environment,“ National Fiber Optic Engineers Conference (NFOEC) 1998, Orlando, FL, September 1998. [133] E Moore, ”Extending the LAN to the Transport Network,” National Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [134] D. K. Gordon, “Broadband Digital Cross Connect Application Guidelines,” National Fiber Optic Engineers Conference (NFOEC)1998, Orlando, FL, September 1998. [135] R. Goudreault, et al., “Impact of Integrated Architecture for Bandwidth Management in a Broadband Network Infrastructure,” Bel1 Labs Technical Journal, Vol. 4, No. 1, January-March 1999, pp. 19-41. 11361 E. Modiano, A. Nola-Tam, “Mechanisms for Providing Optical Bypass in WDM-Based Networks,” Optical Networks Magazine, Vol. 1, No. 1, January 2000, pp. 11-18. [137] 0. Gerstel, R. Ramaswami, G. Sasaki, “Cost-Effective Traffic Grooming in WDM Rings,” IEEWACM Transactions on Networking, Vol. 8, No. 5, October 2000, pp. 618630. 402 Nasir Ghani et al. [138] G. Sasaki, 0.Gerstel, “Minimal Cost WDM SONET Rings That Guarantee No Blocking,” Optical Networks Magazine,Vol. 1, No. 4, October 2000, pp. 51-57. [139] X.Zhang, “An Effective and Comprehensive Approach for Traffic Grooming and Wavelength Assignment in SONETlWDM Rings,” IEEEIACM Transactions on Networking, Vol. 8, No. 5, October 2000, pp. 608-617. [140] P. Wan, et al., “Grooming of Arbitrary TraEic in SONETlWDM BLSRs,” IEEE Journal on Selected Arem in Communications,Vol. 18, No. 10, October 2000, pp. 1995-2003. [141] L. Castellon, T. Rado, “The Future of SONET,” National Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [142] A. Shivji, “SONET Gear for the Optical Generation,” National Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [143] N. Sayer, S. Donnelly, G. Winslow, “Positioning for Automated Provisioning of Broadband Services Over SONET,” National Fiber Optic Engineers Conference (NFOEC) 1998, Orlando, FL, September 1998. [144] R. Newman, A. Dobrushin, S. White, “Effects of Data and New Access Technologieson Optical Access Networks,” NationalFiber Optic Engineers Conference (NFOEC) 1998, Orlando, FL, September 2000. [145] L. Steinhorst, “Enablingthe Optical Transport Network Through Digital Wrappers,” National Fiber Optic Engineers Conference (NFOEC) 2000, Denver, CO, August 2000. [146] L. McAdams, J. Yates, K. Bala, “User to Network Interface (UNI) Service. Defjnition and Lightpath Attributes,” OIF Forum, OIF2000.061, September 2000. [147] K. Arvind, et al., “Optical Domain Service Interconnect (ODSI) Signaling Control Specification,”Version 1A.5, ODSI Coalition,November 2000. [148] D. Awduche, Y Rekhter, “Multiprotocol Lambda Switching: CombiningMPLS Trallic Engineering Control with Optical Crossconnects,”ZEEE Communications Magazine, Vol. 39, No. 3, March 2001, pp. 111-116. [149] N. Ghani, “Lambda-Labeling:A Framework for IP over WDM Using MPLS,” Optical Networks Magazine, April 2000. [150] Y Xu, et al., “Generalized MPLS Control Plane Architecture for Automatic Switched Transport Network,” Internet Draft, drafr-xu-mpls-ipo-gmpls-archOOM, November 2000. [15 1 Y Xue, et al., “Carrier Optical Services Framework and AssociatedUNI Require1 . ments,” Internet Draft, draft-many-cam‘er-framavork-uni-00. November 2000. at, [152] D. Papadimitriou, et al., “Optical Network-to-Network Interface Framework and Signaling Requirements,” Internet Draft, draft-papadimihiou-onni-ffameOl.brt,November 2000. [153] G. Bernstein, E. Mannie, V. Sharma, “Framework for MPLS-based Control of Optical SDWSONET Networks,” IETF Draft, draft-bms-optical-sdhsonet-mplscontml-Jimwrk-00. November 2000. at, 8. MetropolitanOptical Networks 403 [154] A. Banerjee, et al., “Generalized Multiprotocol Label Switching: An Overview of Routing and Management Enhancements,” IEEE CommunicationsMagazine, Vol. 39, No. 1, January 2001, pp. 144-150. I [155] B. Rajagopalan, et al., “ POver Optical Networks: Architectural Aspects,” IEEE CommunicationsMagazine, Vol. 39, No. 9, September 2000, pp. 94-102. [156] N. Golmie, T. Ndousse, D. Su, “A Differentiated Optical Services Model for WDM Networks,” IEEE CommunicationsMagazine, Vol. 38, No. 2, February 2000, pp. 68-73. [157] D. H. Su, D. W. Griffith, “Standards: Standards Activities for MPLS Over WDM Networks,” Optical Networks Magazine, Vol. 1, No. 3, July 2000, pp. 6b-69. . [158] P Bonenfant, A. Moral, “Framing Techniques for IP Over Fiber,” IEEE Network Mugmine: Vol. 15, No. 4, JulylAugust 2001, pp. 12-18. [159] B. St. Amaud, “Overview of the Latest Developments in Optical Internets,” Optical Networks Magazine, Vol. 1, No. 3, July 2000, pp. 51-54. [160] D. H. Su, “Standards: The IEEE P802.3ae Project for 10GbIsEthernet,“ Optical Networks Magazine, Vol. 1, No. 4, October 2000, pp. 58-59. [161] N. Ghani, et al., “COPS Usage for ODSI,” Optical Domain Service Interconnect (ODSI) Coalition, Policy/Accounting Specification, January 200 1. [162] A. Herrera, et al., “A Framework for IP Over Resilient Packet Rings,” Internet February 2001. Draj?, draft-ietf-iporpr-framework-OO.Lxt, [163] IEEE 802.17 Resilient Packet Ring Working Group, http://www.ieee802.org/17/. 11641 P. Gupta, N. McKeown, “Algorithms for Packet Classification,” IEEE Network Magazine, Vol. 15, No. 2, MarcldApril2001,pp. 24-32. [165] B. Li, et al., “Photonic Packet Switching: Architectures and Performance,” Optical Networks Magazine, Vol. 2, No. 1, January/February 2001, pp. 27-39. n [166] A. Jourdan, etal., “The Perspective of Optical Packet Switchingi IP-Dominant Backbone and Metropolitan Networks,” IEEE CommunicationsMagazine, Vol. 39, No. 3, March 2001, pp. 136141. [167] S. Yao, B. Mukhejee, “All-Optical Packet Switching for Metropolitan Area Networks: Opportunities and Challenges,”IEEE CommunicationsMagazine, Vol. 39, NO.3, March 2001, pp. 142-148. [168] A. Pattavina, et al., “Techniques and TechnologiesTowards All-Optical Switching,” Optical Networks Magazine, Vol. 1, No. 2, April 2000, pp. 75-92. [169] L. Xu, “Techniques for Optical Packet Switching and Optical Burst Switching,” IEEE CommunicationsMagazine, Vol. 39, No. 1, January 2001, pp. 136143. WDM Multi-Rings with Differentiated QOS,” [170] M. Marsan, et al., c‘All-Optical IEEE CommunicationsMagazine, Vol. 36, No. 12, February 1999, pp. 58-66. [171] K. Shrikhande, et al., “HORNET A Packet-Over-WDM Multiple Access Metropolitan Area Ring Network,” IEEE Journal on SelectedAreas in Communications, Vol. 18, No. 10, October 2000, pp. 2004-2016. [172] N. Jones, H. Helvoort, T. Wilson, “Proposed Text on Link Capacity Adjustment Scheme (LCAS) for SDH Virtually Concatenated VCs to be contributed to ITUT SG-15,” ANSI TlX1.5/2001429, January, 2001. Chapter 9 The Evolution of Cable TV Networks Xiaolin Lu Morning Forest, Highlands Ranch, Colorado Oleh Sniezko Oleh-Lightcom,Highlands Ranch, Colorado 1. Introduction The telecommunications industry is facing tremendous challenges and new opportunities brought about by deregulation, competition, and advanced technologies and services The opportunities in broadband subscriber access and its promise of eliminatingthe bottleneck that has limited progress towards a global information infrastructure have been stimulating large-scale business efforts and technology evolution. Network operators and service providers must upgrade their “embedded infrastructure” in order to protect their current revenue while searching for new market potentials. This becomes more critical as the existing services (telephone, broadcast analog TV)are reaching the saturationpoint, while the emergenceof new serviceopportunities requires different network characteristics (asynchronousvs synchronous, switched vs broadcast, broadband vs narrowband, etc.). Depending on the economic situation and projected revenue potential, different network operators and service providers may choose different network upgrade paths and business strategies involving different technologies [ 141. Historically, communication networks have been established to support specific services and their related business models. Therefore, each type of network has its own characteristics that serve special needs. With entertainment and communication being the two major and distinguished services, at least historically, the access networks can be categorized as broadcast-based (point-to-multipoint) and switch-based (point-to-point). Examples include tree-and-branch-based network topology used by the cable industry and twisted-pair-based star architecture used by the telephony industry. The cable industry’s primary business is broadcasting multichannel video information to customers and to subscribers’homes. The notion of “pushingyy allowing them to choose naturally leads to the use of the tree-and-branch architecture and coaxial cable (broadband capacity) for signal distribution utilizing radio frequency (RF) subcarrier multiplexing technique. Although the pros and cons of this architecture versus that of the star architecture continue to be debated, the use of tree-and-branch topology with tapped coaxial 404 OPTICAL FIBER TELECOMMUNICATIONS VOLUME IVB Copyright 0 2002, Elsevier Science (USA). All rights of reproduetion i any form reserved. n ISBN 0-12-395173-9 9. The Evolution of Cable TV Networks 405 bus provides the cable industry the most cost-effective way of building and operating the network for two reasons. First, in the United States, residential houses are constructed based on a typical grid topology, which resembles the bus architecture. The tapped coaxial bus allows cable operators to easily connect and disconnect customers and provision potential growth in certain areas without overbuilding or underbuilding the network. For example, along a typical coaxial bus, one can easily provision many taps (within certain limitations defined by the signal quality requirements). Each tape can have many different numbers of ports for connecting customers (customer activation), and the number of ports can be larger than the current number of customers in that area, therefore leaving room for growth. Second, multiple system operators (MSOs) can extend the network reach by extending the coaxial bus @e., adding coaxial amplifiers), as long as the signal level requirements are met. The use of Frequency Division Multiplexing (FDM) or Subcarrier Multiplexing (SCM) schemes throughout the network also makes the network transparent to signal formats. Any change (e.g., bandwidth requirement) can be made by modifying the headend equipment and customer premise equipment (CPE), therefore simplifying the operation. Armed with the advent of linear lightwave technology, the advanced R F modem, and DSP technology, cable network operators have embarked on extensive upgrades and on the transition to hybrid fiber coax (HFC) architectures in the past 2 decades. The main goal has been to evolve the infrastructure from a broadcast-type trunk-and-branch plant to a high-capacity two-way network with superior quality and reliability that is ready to deliver advanced telecommunications services. This chapter will discuss cable network evolution and related technology innovations, and their impacts on business and operations. Instead of discussing all the physical and engineering details, we will try to provide an intuitive view of this evolution. Section 2 provides a general overview of the conventional coaxial cable network and its migration to the HFC architecture. Section 3 discusses the commonly used modern broadband cable infrastructure, including the metro networks, secondary hub networks, and the Fiber-to-the ServingArea (FSA) distribution networks. Section 4discusses the further evolution of the HFC network based on deep fiber penetration and digital technology. Section 5 provides background on some of the technology advances that continually drive the network evolution. 2. Historical Overview In 1998, the cable television industry celebrated its fiftieth anniversary. The first cable television system in the United States was built with twin-lead 406 Xiaolin Lu and Oleh Sniezko symmetricalantenna wires and without ampliliers [5-6]. At approximatelythe same time, similar systems were also built in London, England and Ontario, Canada. Soon after these first, amplificationlesssystems, tube ampwers and then later silicon-basedamplilierswere added to extend the reach on the coaxial distribution systems. A drive towards increased bandwidth in the cable television distribution systems led to the deployment of 400+-MHz systems with 54+ analog video channels by the end of the 1970s. This expansion was made possible by significant strides in semiconductor technology that created push-pull amplifiers, feed-forward amplifiers, and power-doubling amplifiers. Further improvement in these basic technologies resulted in forward bandwidth being expandedto 550 MHz by the end of the 1990sand to 750-4360MHz by the end of the century. This also allowed for 25 and more amplifiers being cascaded while still meeting the minimal regulatory end-of-line performance requirements. At the end of the 1970s and the beginning of the 1 9 8 0 ~ ~ two-way cable TV systems became operational, although the wide deployment of two-way broadband systems did not start until early in the 1990s when the technology innovation broke the tech-economic barrier and enabled cable operators to support public demand for broadband telecommunication services. In a modern two-way cable system, 5-42MHz is typically used for upstream transmission, and 50 MHz and above is used for downstream transmission, in which 50 -500 MHz is used for carrying broadcast analog video and the upper-frequency band (550 MHz) is used for emerging digital services. In the late 1980%linear fiber-optic technology achieved the level of performance suitable for analog cable TV applications. The deployment of this technology allowed for improved quality and reliability of cable television systems by shorteningthe cascadesof activedevices. In the 1990% affordablelinear reverse lasers became available. Both these contributed to an unprecedented architecturalprogress, which eventually convertedtraditionaltree-and-branch coaxial architecture into a node-based HFC architecture. 2.1 TRADITIONAL COAXIAL SYSTEM Traditionally, the majority of television sets were built to receive AM-VSB (analog modulation-vestigial sideband) signals. The most cost-effective way to deliver those signals to customers was to maintain the same format and avoid the cost of format conversion. Therefore, the AM-VSB signals were frequency multiplexed and transmitted over coaxial cable with multiple coaxial amplifiers in cascade [5-71. At the end of these cascades, the signals had to maintain a minimal level of performance (noise and distortions) to meet regulatory requirements and customer satisfaction. The minimal performance requirements changed over time with increased subjective perceptibility of the video impairments brought about by customer education, larger TV screens, and high-quality alternative means of video content reproduction (e.g., DVD ,,,, ,,,, 9. The Evolution of Cable TV Networks 407 Line Extender - Trunk Coax Cable _____ TaDwd Coax Cable Fig. 9.1 Traditional tree-and-branch (point-to-multipoint) coaxial network. players). Today, it is commonly accepted that the minimal performance levels at the TV receiver are approximately 49 dB for carrier-to-noise ratio (CNR) and better than -53 dBc for nonlinear distortions, includingcomposite second order (CSO) and composite triple beat (CTB). In the past, broadcasting many channels of entertainment video to all customers was cable TV operators’ main business. Given metro markets were historically fragmented and belonged to different operators, each operator in its respective franchise areas used broadcast networks that had the following characteristics: 0 0 Single collection points (signal importation, direct feeds, and local origination), Large broadcast coverage in the same area. Under these circumstances, the traditional tree-and-branch (point-tomultipoint) broadband coaxial network served the cable operators well. This architecture contains three major components, as shown in Fig. 9.1: 0 0 0 Headend, Trunk system, Distribution system. 2.1.1 Headend The headend of a traditional cable TV system served as a collection point for all signals from external sources, whether national, regional, or local. The 408 Xiaolin Lu and Oleh Sniezko signalswere received with off-air antennas, satellite antennas, and/or via direct feeds from remote locations (local studios, local broadcasters, remote satellite stations, remote off-air antenna farms, etc.). Optical links were fist deployed for those direct feeds in the early 1980s and were in common use by the end of the decade. In this system, video was coded with proprietary codecs, applied to directly modulated lasers, and transmitted over those optical links in TDM fashion. After collecting all the signals at the headend, certain processing equipment and RF-combining networks were used to assemble the signals into a channel line-up (subcarrier multiplexing), and to launch those signals to the trunk and distribution networks. A basic configuration of the processing equipment and RF-combining network are shown in Fig. 9.2. 1 4 Addltlonal Modulators for Channeb 19.20,21,22. 7 , 8 , 9, io, I t 1 2 AddMonal Modulators for Channels 3,4.5.6. Ax. AY, 37.0 dBmV 1 Modulator W, Channel 13 56'o dBmV DC20 t amplifier Fig. 9.2 Con6guration of signal processing and combining equipment in headend. 9. The Evolution of Cable TV Networks 409 It should also be noted that, until recently, most headends were not interconnected and were feeding independent cable television systems. This was due to first, the lack of high capacity and affordable transport technology to support consolidation of the signal origination and processing equipment, and second, the fragmentation of the market ownership. 2.1.2 Trunk System The trunk system was built for long reach, while maintaining good signal quality for distribution systems further delivering those signals to customers’ homes. Therefore, the trunk amplifier cascade was optimized for the lowest possible impairment contribution to the signals being distributed at the longest possible reach. This was a balancing act between the theoretical optimal gain (equal to 1Np or 8.69 dB) [8] and practical network uncertainties related to frequency response of the amplifiers, their thermal stability, and many other factors causing a shift from stable conditions. Because the passive devices were not used extensively in the trunk (express) lines, trunk amplifiers had to mostly overcome cable loss that was highly frequency dependent. All those then determined the internal trunk amplifier configuration and the selection of input and output levels. An example of the trunk amplifier is shown in Fig. 9.3a. A typical practice is to set the trunk amplifier gain between 15 and 25 dB in the presence of these uncertainties. In the past, trunk cascades of 25 amplifiers were common. Some cascades, feeding remote pockets of subscribers, consisted of more than 50 amplifiers carrying 50 to 80 analog channels (400-550 MHz bandwidth) [9-121. Although the required performance level can be achieved through progress in semiconductor technology, linearization techniques, and internal architecture of the trunk amplifiers [13], the long cascades resulted in low reliability and high maintenance cost. 2.1.3 Distribution System Branching out from the trunk systems, the distribution system delivers adequate signal level to the outlets at each residential dwelling. Two- or three-high output power amplifiers, called either distribution amplifiers or line extenders, were cascaded to serve the area covered by each branch of the distribution system. These amplifiers were optimized for high output capability to compensate for high passive loss induced by RF power splitting. An example of a typical distribution amplifier is shown in Fig. 9.3b. Beginning late in the 1980s, a concept of superdistribution was developed. In this architecture, tapped coaxial buses that delivered signals to the customers were separated from bridging lines between distribution amplifiers. This is shown in Fig. 9.4. The so-called bridger amplifier, similar to the trunk amplifier, is used for this purpose. This architecture allows for increased reaGh while improving reliability of the coaxial network. Xiaolin Lu and Oleh Sniezko I PORT 4 EPF IDC INTERSTAGE PIN E9 PAD Diode m AC Power toniwm *rI 4 b AC Power Powa suppl: REV REVLPF IGC Power Supply tolfmm E9 PAD STATION WRT 4 . . REVERSE TP * Fig. 9.3 (a) Configuration of a trunk coax amplifier. (b) Contiguration of a line extender. 2.1.4 Upstream System Both amplifierspresented in Figs 9.3a and 9.3b are equipped with the reverse (upstream) path components to allow for two-way communication over an integrated R F coaxial system with diplexers separating the upstream and downstream bands. The initial use of the reverse path was limited to the transmission of status-monitoringinformation and reverse signals for impulse pay per view. 9. The Evolution of Cable TV Networks 411 Trunk and Super-D Coax Cable Fig. 9.4 Superdistribution architecture. 2.2 MOVING TO HYBRID FIBEWCOAX NETWORKS Meeting the requirements for end-of-lineperformance has been a major challenge for traditional coaxial networks due to the accumulation of noise and distortion over long, cascaded coax amplifier chains. The advent of linear lightwave technology allows operators to replace trunk parts of the cable networks with optical fiber and only keep the last-mile coax plants for distribution. Pioneered by ATC (part of today's AOWTime Warner) and followed by many other cable operators, the fiber-backbonearchitecture was developed in the late 1980s [14-171. By shortening the coax cascade, cable operators can increase network reliability and improve signal quality, while expanding the serving areas covered by a single headend (Fig. 9.5). The ability of transmitting multichannel analog video over optical fiber has had a profound impact on the cable industry. The commonly used systemswere intensity modulated/direct detected systems [ 181. Enabled by the innovation in semiconductor technology, continuing improvement in laser structure, especially the advent of the single-frequency-distributed-feedback(DFB) lasers, led to increased output power, lower relative-intensity noise (RIN), and better linearity. The linear lightwave family quickly grew to include 1.3 and 1.5 km DFB lasers, external modulators, and erbium-doped fiber amplifiers (EDFAs) to further extend the reach. The system performance was also enhanced by sophisticated predistortion and noise reduction techniques. All of these enabled linear lightwave systems capable of delivering more than 80 channels of analog video plus a broad spectrum of digital-RF signals over distances in excess of 60 km, with performance approaching the theoretical optimum. 412 Xiaolin Lu and Oleh Sniezko ___ Trunk Coax Cable - Optical Fiber Cable Fig. 9.5 1000 900 800 N I 700 3 600 .5 500 U s U 400 S m m 300 w200 100 01 1962 1967 Fiber backbone architecture. - I 1 I I I I 1972 1977 1982 1987 1992 1997 2002 Years Fig. 9.6 Bandwidth expansion of cable network. 3. An End-to-End HFC Network Linear lightwave technology not only enables the cable operator to replace coaxial trunk with optical fiber to achieve high capacity, quality, and reliability, but also allows for the interconnection of many distribution networks to increase network flexibility and to further reduce operating costs. The total bandwidth in a cable network has expanded exponentially over the past 35 years, initially enabled by technology innovations in R F devices and further accelerated by the deployment of fiber-optics(Fig. 9.6). Most recently, 9. The Evolution of Cable TV Networks 413 Metro Network Secondary Ring Access Network Fig. 9.7 A modern hybrid fiberlcoax cable network. emerging interactive services, always-on applications, and therefore substantially increased network usage increase the demand for further bandwidth expansion per user in both downstream and upstream directions. Competitive pressures, on the other hand, motivated network evolution toward cost reduction and operating savings. All of these, as well as increased customer expectation for quality and reliability, led to technology innovations in transforming the traditional broadcast cable network into a broadband twoway infrastructure, and further to an end-to-end digital platform with wide implementation of the advances in lightwave and digital technologies. In North America, a modern cable network typically consists of three major sections, as shown in Fig. 9.7: 1. Metro network, 2. Secondary ring, 3. Access network. 3.1 METRO MARKET ARCHITECTURE In the late 1980s and early 1990s, metro market consolidation (or clustering) among cable TV operators accelerated significantly.In parallel with this trend, new services and changes in the competitive landscape increased pressure on the cable TV industry to develop the capacity to deliver a multitude of services at a much larger scale and serve those clustered areas in a more cost-effective way. Pioneered by engineers of Rogers Cablesystems in Canada, a metro fiber ring architecture to interconnect clustered HFC distribution networks was established and was later adopted by CableLabs as a recommended metro market architecture for the North America cable industry [19, 201. Today, most of the metro networks in markets exceeding 100,000 homes resemble ring configurations (single or dual rings). Numerous so-called mater headends, sometimes operated by different multiple system operators (MSOs), 414 Xiaolin Lu and Oleh Sniezko are connected by the metro rings with full redundancy and survivability.These mater headends typically serve as the primary signdcontent sources, and potentially also serve as interconnection points with other service providers (ISPs, ILECs, and CLECs) for high-speed data and telecommunication services. Traditional separated (stand-alone) headends are consolidated into a so-called primary hub, which are interconnectedby the market rings with the mater headends and serve as furtber signal distribution points. These primary hubs typically serve 60,000 to 100,000 homes. In the past, the choice of transmission technology over the metro ring was based mostly on the requirements for signal quality and on technology availability (including its cost). Also, the transport systems in the metro markets were historically built to meet the requirements of certain services at that time. Since the headend consolidation began when broadcast video services dominated, the transport systems were therefore optimized for these signals. An analysis of the video transport systems used in metro markets is shown in Table 9.1, which reflects the results of several studies performed by major MSOs. It should be noted that the deployment of the SONET-based systems in the cable industry was slow in the past because video distribution over the metropolitan rings had been monopolized by the cost-effective, proprietary, linear PCM systems. However, these systems provided only limited support for transport of voice and data services. This situation changed in the mid-1990s when high-speed data, digital video, and digital telephony services were introduced. To support delaysensitive voice service, as well as high-speed data services, operators started deploying SONET rings in parallel to the embedded proprietary systems. Recently, the increasing need for a common IP infrastructure to support integrated serviceswith reduced OAM&P(operation, administration, and provision) cost further stimulated the integration of the layered protocol stack. AU these resulted in a mixed transport systembeing deployed in the metro markets, therefore imposinga big challengeto metro network migration. Any new metro network infrastructure has to support not only legacy services, but also emerging services, both of which employ a wide variety of formats and protocols. Theserequirements demand the capability of multiplexingand demultiplexing to DSl level, while being able to perform optical aggregation at OC-48, OC- 192, and even higher speeds. The metro network needs to support standard interfaces for MPEG-2 video service, as well as native Ethernet interface for data service. It also needs to accommodate many different protocols for networking and management purposes, such as 802.1 p/Q Layer 2 protocol, Border Gateway Protocol (BGP), Open Shortest Path First (OSPF) Layer 3 protocols, as well as emerging multiprotocol label switching (MPLS)-types of protocols. In addition, efficient trafEc protection and recovery mechanisms continue to be a critical feature in the new transport platforms. Preeminent among 9. The Evolution of Cable TV Networks Table 9.1 Transport TechnologiesUsed in Metro Networks Comments Technology Analog FM 415 Description Baseband video signals frequency modulated and subcarrier multiplexed pros Good overall signal quality Cons Difficult to encounter different analog scrambling systems Proprietary systems Limited cascading High cost of interfacing to RF Analog AM 1550-nm window Baseband video signals amplitude modulated and subcarrier multiplexed Baseband or IF video signals digitized and transmitted for long distances Proprietary systems SONET-like features Numerous interfaces available (DS3 and subsets) Potentially interface with any digital service Standard systems Survivability Perfect for system interconnects Dropladd capability interfacing to RF Linear PCM SONET Digital TDM optical hierarchy system High cost of interfacing to RF these is the need for sub-50-ms Automatic Protection Switching (APS)for ring recovery. Two emerging platforms are becoming more popular and are competing to address these requirements. The first is the SONET Multiservice Provisioning Platform (MSPP). This platform is based on existing SONET standards but 416 Xiaolin Lu and Oleh Sniezko can also support data processing capabilities without the need for third-party hardware. It incorporates a high-density circuit switching matrix for TDM services and separate packet processing engines implementing Layer 2 and Layer 3 protocols for routing of IP or other packet-based traffic. Allocating transport bandwidth among mixed TDM and packet-based services may also be supported in smaller increments than that of traditional SONET platforms through VT1.5 virtual concatenation as per ITU-T G.707 standard or through vendor-specific means. This permits a much more efficient approach to bandwidth usage around metropolitan fiber transport networks. Examples include: 1. MSPPs that implement full SONET TDM transport and traffic protection features with added statistical multiplexing capabilities to support packet-based data traffic. In this case, delay-sensitive voice and video traffic is transported over standard TDM channels. Packet-based IP traffic is encapsulated and framed, and then statistically multiplexed into clear OC-3c/OC-12c/OC-48~ channels. 2. MSPPs that implement a streamlined version of SONET (SONET-lite) only to provide basic framing and to support functions such as failure notification. In these implementations, all trafEc, TDM and packet-based, is encapsulated in frames prior to transport over an optical channel at SONET rates. Another platform competing against SONET in this arena is based on pure packet transport platforms that use Resilient Packet Ring (RPR) architectures and protocols currently being standardized under the IEEE 802.17 working group. Theseplatforms promisemore efficientbandwidth allocationoptimized for high-burst, variable-rate packet services. They are not constrained by the SONET limit of bandwidth allocation in fixed increments and support sub50-ms ring failure recovery. In addition, some of the implementations support transport of legacyvoice and other delay-sensitivet r a c via TDM circuit emulation. Packet-processingengines implementingLayer 2 and Layer 3 protocols are also incorporated in these platforms. RPR architecturessupport service transport over two counterrotating optical fiber rings interconnecting a number of RPR nodes. The bandwidth around the ring is shared across all RPR nodes. Unlike SONET, however, each RPR node is topology-aware. This enables implementation of a logical meshed network as opposed to point-to-point nailed-up connections. The RPR protocols arbitrate bandwidth access among all nodes in the rings and implement congestion avoidance mechanisms to support near-full usage of the available bandwidth on both rings. Protection mechanisms enable service restoration in sub-50-ms timeframes while preserving all service connections for high-priority t r a c . 9. The Evolution of Cable TV Networks 417 In addition to these two platforms, the advanced DWDM technology makes it feasible and attractive to gradually incorporate more intelligence into the optical layer. With the t r a c being classified at the edge of the network using electronic switches and routers, an intelligent and transparent metro optical core could be established for networking purposes (switching/routing, adddrop, etc.), therefore accommodatingwhatever serviceformat and protocols. 3.2 SECONDARY HUB ARCHITECTURE The secondary hub, or distribution hub, architecture was introduced to facilitate the headend consolidation. In this configuration, secondary hubs (SH) serve as signal concentration and distribution points to limit the number of fibers between the primary hub ring and secondary or distribution hubs. This configuration leads to additional cost reduction (HE consolidation) and improved mean-time-to-repair (MTTR) (shorter fiber cable restoration times) and allows for cost-effective backup switching (redundancy) [19-26]. Two topologies commonly used in the secondary hub architecture are: 1. Star architecture in which the secondary hub performs physical aggregation between the primary hub and many fiber nodes. This then realizes an end-to-end ring-star-star-bus architecture (primary ring-secondary star-distribution start and bus). 2. Ring architecture in which secondary hubs are interconnected with the primary hub in a ring, with star architecture from secondary hubs to the nodes (an end-to-end ring-ring-star-bus architecture). The star architectureis very flexible in selectingthe transmissiontechnology. For example, the subcarrier multiplexing (SCM) technology can be deployed throughout the network between the primary hub ring and the customer. The SCM scheme can support the combination of digital and analog transmission technology and multiple accessprotocols compatiblewith the terminal devices. This therefore results in a simple architecture that is largely format transparent between the headend or primary hub and the optical node. Unfortunately, this architecture employs high fiber count cables that increases capital costs and incurs single point of failure with long MTTR. The ring topology, on the other hand, enables a cost-effective and highly reliable network with a limited number of fibers between the primary and secondary hubs, and is preferred by all major MSOs. Several methods of sharing the fibers (multiplexing) were implemented. Commonly used multiplexing technologies range from standard time-division multiplexing (TDM) of data streams for target service delivery through frequency-division multiplexing (FDM) or frequency conversion (especially 418 Xiaolin Lu and Oleh Sniezko practical in the reverse direction) to wavelength-division multiplexing, including dense wavelength-division multiplexing (DWDM). Figure 9.8 shows the use of FDM technology for fiber sharing. In this arrangement, broadcast analog and digital videos are carried by one fiber and distributed to all the secondary hubs in the ring. The target services (telephony, data, etc.) addressing each fiber node (FN) are block upconverted and (a) PRIMARY HUB Broadcast (in ring configuration) SECONDARY HUB OPTICAL NODES 200 MHz per Block - - 3* 600 HP Digital Feeds per one Block Up Converter 3' 600 HP Digital 200 MHz per Block - - 3* 600 HP Digital Feeds per one Block Up Converter 3* 600 HP Digital Feeds per one Block (b) PRIMARY HUB SECONDARY HUB OPTICAL NODES E* 600 HP Digita 'eed, per one Bloc Fig. 9.8 FDM-based transport over primary-secondary-node link. (a) Downstream transport. (b) Upstream transport. 9. The Evolution of Cable TV Networks 419 combined at the primary hub and transmitted to the secondary hub over a separate fiber. They are then downconverted to the original frequency at the secondary hub, and are directed to each fiber node accordingly. The upstream transmission uses the same principle. DWDM technology has been considered a very desirable solution, but has matured to support SCM signals in the cable TV environment only in recent years. As shown in Fig. 9.9, instead of multiplexing target services (addressing each FN) over RF frequency (as in the FDM case), these signals are multiplexed over optical wavelength [26]. With the fast-paced DWDM (a) PRIMARY HUB SECONDARY HUB (b) I I I I I I I I I I PRIMARY HUB I I SECONDARY HUB ( _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ OPTICAL NODES I I 1 --------; I 1~~~~~~~~~~~~~~~~~~~~ I I I I I I I I I I I 1 I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I Fig. 9.9 DWDM-based transport over primary-secondary-node link. (a) Downstream transport. (b) Upstream transport. 420 Xiaolin Lu and Oleh Sniezko technology, this approach is proven to be the most cost-effective and flexible solution for the secondary hub architecture. 3.3 ACCESS NETWORK: FIBER TO THE SERVING AREA (FSA) As we discussed previously, the HFC network was originally designed for broadcast services with point-to-multipoint architecture utilizing the coaxial bus. Enabled by lightwave technology with different levels of fiber deployment, upgrade alternatives to support target services can be realized with a certain degree of virtual point-to-cell configuration. Fiber to the Serving Area (FSA) with fiber node segmentation has been proven to be a very cost-effective solution and has been widely deployed in the cable industry [16, 27-29]. The differences between implementations are mostly related to the node sizes, with particular emphasis on the design optimization for individual needs (power consumption, time-to-market, or end-of-line performance and bandwidth) and the level of redundancy. Figure 9.10 shows an end-to-end ring-ring-star-bus cable network with FSA as the distribution architecture. In this situation, a fiber-based star topology is used to connect many fiber nodes to the secondary hub. Below the FN, coaxial bus is used for final connection to customers. Each FN is the center of the “cell”-the serving areas covered by the FN. Each FN typically serves 500 -2000 households, with the associated coax plant being segmented into several coaxial buses. With this segmentation, the bandwidth capacity can be adjusted based on the service needs. For example, the entire FN serving area originally could share the 5-42 MHz upstream band. When the service take rate or usage rate increases, a predefined four-way segmentation can break the FN serving area into four small cells with each segment (bus) occupying the entire 5-42-MHz band. Those four 5-42-MHz bands are then multiplexed at 4 DWDM Transport End-to-end Transparency b 4 Segmentation 4~ capacity Fig. 9.10 A FSA-based HFC network with four-way segmentation at fiber node. 9. The Evolution of Cable TV Networks 421 FN using DWDM technology. This effectively increases the upstream bandwidth by a factor of four without adding new fibers and still maintaining the end-to-end transparency. 4. Photonic Moore’s Law and Deep Fiber Penetration The innovati