Optiputer High Performance Transport Protocols Implementation Plans by johanpetro

VIEWS: 13 PAGES: 15

									Optiputer High Performance Transport
 Protocols & Implementation Plans




Ryan X. Wu, Andrew A. Chien, Eric Weigle, Nut Taesombut
                  CSAG, CSE, UCSD
                      01/25/2005

                                                    1
       High Performance Transport Problem
•   OptIPuter is Bridging the Gap Between High Speed Link Technologies and
    Growing Demands of Advanced Applications
•   Transport Protocols Are the Weak Link
    – TCP Has Well-Documented Problems That Militate Against its Achieving High
      Speeds
    – OptIPuter Pursuing Range of Transport Protocols
       – Shared, Routed Infrastructure:
            – XCP [USC/ISI],
            – UDT [Gu&Grossman, UIC]
        – Provisioned Lambda:
            – RBUDP/LambdaStream [He&Xiong&Leigh, EVL, UIC]
            – GTP [Wu&Chien, UCSD]
        – Cluster-based:
            – CEP [Weigle&Chien, UCSD]
    – How to make applications utilize these protocols?
•   OptIPuter Transport Protocol Implementation/Integration
    – Goal: Flexibility of Protocol Choices under Single Application Implementation
    – Solution: Globus/XIO

                                                                       2
Optiputer High Performance
    Transport Protocols




                             3
                    RBUDP/LambdaStream
•   Suitable for continuous data intensive applications.
•   Quickly grasp the underlying bandwidth and maintain very low jitter.
•   More in Eric He‟s talk




                                [He&Xiong&Leigh, EVL/UIC]
                                                       4
     XCP for High BDP Networks – Shared, Routed
                    Environment
•   Systematic Implementation and Evaluation
    – Build/Test FreeBSD Kernel implementation, Performance Evaluation
    – Full protocol specification and mature the protocol
    – Work with the community (researchers, applications developers, users, vendors,
      operators, IETF), To Develop Deployment plan
    – Initial Results are Promising: Match Simulations


       TCP Measured                                XCP Measured




                                [Bannister&Falk, USC-ISI]            5
UDT: UDP-based Data Transport Protocol

• UDP-based Data Transfer Protocol
 -- Successor of SABUL
• Congestion/Flow Control
  -- For shared networks
 -- Configurable congestion control
• Implementation
 -- User space: no need for root privilege or kernel recompilation
 -- Socket API (will also provide XIO API soon)
 -- Open source: udt.sf.net




                             [GU&Grossman, UIC]               6
                GTP: Group Transport Protocol:
            Receiver-based Congestion Management
• Request-response for Reliable Data Transfer
• Single Flow Adaptation and Capacity Estimation
• Receiver-based Flow Scheduling for Fairness and Low Loss Rate
   – Balance Concurrent Data Fetching from Multiple Sources
   – Fair across Varied Sender RTTs
   – Efficient Transitions under Rapid Changes
                                                                Applications
                                                  GTP


                                                  Single Flow
                                                  Control and            …...
                                                  Monitoring



                                                        Centralized Rate Allocation


                                                    UDP (data flow) / TCP (control flow)
       R1                                  R2
                                                                    IP
    Multipoint-to-point contention at receivers    GTP Receiver Architecture

                                                                               7
                                                                          [Wu & Chien, UCSD]
               CEP: Composite Endpoint Protocol
•   We need a High-performance, scalable, robust,
    flexible, adaptive, easy-to-use mechanism to
    terminate disproportionately large data
    transfers.
•   User controls transfer through simple sockets-
    like interface, and transfer constraint
    specification.
      – Can emulate GridFTP, MPI-IO, and parallel
         filesystem interfaces easily.
•   CEP handles actual transfers subject to user
    constraints via scheduling mechanism
      – Uses multiple cooperating nodes.
      – Heterogeneous hosts may take part
         unequally in a connection.
      – Scheduler handles failures, fairness,
         changing resources over time.
      – Does the “Heavy Lifting”
      – TCP, UDP, FAST, HSTCP, ... used as
         underlying protocol.
                                                     [Weigle & Chien, UCSD]
                                                          8
OptIPuter Transport Protocol
 Implementation/Integration




                               9
     Transport Protocol Deployment Problem

•   Issues:
     – How to make applications utilize OptIPuter transport protocols?
     – What is the relation between transport protocols and backbone management and
       configuration?
•   Each transport protocol has its own API and semantics
     – Different targeted network environments
         – Shared, routed network
         – Private, or QoS-provisioned network
     – Different semantics
         – Blocking/non-blocking
         – Point-to-point, multipoint-to-point
•   Applications need to be modified to use different protocols.
     – Changes may be complex
     – Need to maintain multiple versions with different transport protocols
•   Provide the same API for all OptIPuter transport protocols
     – Allow convenient integration with applications
     – Support further testing/evaluating different transport protocols
     – Solution: Globus XIO: eXtensive Input Output library




                                                                               10
                               Globus/XIO


• Globus XIO user API                                      User API
   – Single API/Single Set of semantics.                            Driver Stack
   – Simple open/close/read/write API




                                                    Framework
• Driver Abstraction                                                 Transform
   – Hides protocol details
   – Allows for extensibility                                        Transform
   – Drivers can be selected at runtime
• XIO interface for OptIPuter transport                              Transport
  protocols
   – A simple XIO wrapper (~1000 lines of code)                 XIO Framework
     on top of each transport protocol to provide
     XIO API
   – Almost no changes needed to the original
     protocol implementation



                                                                            11
                              XIO Example
•   Init the protocol stack
     – globus_xio_stack_init(&stack, NULL);

•   Load drivers
     – globus_xio_driver_load(“TCP", &txdriver);
         – or „UDP‟, “UDT‟‟, “GTP”…
     – globus_xio_stack_push_driver(stack, txdriver);

•   Init and set protocol attributes
     – globus_xio_attr_init(&attr);
     – globus_xio_attr_cntl(attr, txdriver, GLOBUS_XIO_SET_LISTENING_PORT,
       hosts, numhosts);

•   Create handle
     – globus_xio_handle_create(&handle, stack);

•   Send/receive data
     – globus_xio_open(&handle, NULL, target);
     – globus_xio_write(handle, "hello world\n", 12, 1, &nbytes, NULL);
     – globus_xio_close(handle, NULL);

                                                                   12
    OptIPuter Communication Software Architecture

                 Distributed Applications/ Web Services/Viz. Apps


                High-level IPC Library (e.g., MPI, PVM)         DVC. APIs


                  Socket Wrapper (C and Java)


                          XIO Interface

                              Lambda-                               DVC Config.
    GTP   CEP   XCP   UDT                     DVC Descriptor
                               Stream

•   Standard and Custom Communication Protocols are Accessible through a
    Unique API
•   DVC Descriptor Provides Information about Resource Configuration,
    Naming, and Other Properties
                                                                     13
                  Implementation Plan

• Current Implementation Status
   – The TCP, UDP and simple UDT XIO drivers are coming with the
     Globus/XIO
   – GTP XIO driver is available
• Implementation and Integration Plan
   – XIO driver for UDT 2.0 will be available early next month
   – XIO drivers for LambdaStream and CEP are in progress
   – Demonstrate applications running with different protocols via XIO
     interface
   – Further testing and large-scale experiments on different transport
     protocols




                                                                  14
                      For More Information
•   A. Falk, T. Faber, J. Bannister, A. Chien, R. Grossman, J. Leigh, Transport
    protocols for high performance, Communications of the ACM, Volume 46,
    Number 11, November 2003, pp. 42-49.
•   X. Wu and A. Chien, GTP: Group Transport Protocol for Lambda Grids, IEEE
    Symposium on Cluster Computing and the Grid (CCGrid), April 2004.
•   Y. Gu, X. Hong, and R. Grossman, Experiences in Design and Implementation of
    a High Performance Transport Protocol, (submitted for publication).
•   Eric Weigle and Andrew A. Chien, The Composite Endpoint Protocol (CEP):
    Scalable Endpoints for Terabit Flows Submitted to CCGrid 2005.
•   LambdaStream
•   Globus XIO: http://www.globustookit.org




                                                                 15

								
To top