Get documents about "International Journal of Computer Science Volume 8 April 2010
Description
International Journal of Computer Science and Information Security (IJCSIS) provides a major venue for rapid publication of high quality computer science research, including multimedia, information science, security, mobile & wireless network, data mining, software engineering and emerging technologies etc. IJCSIS has continued to make progress and has attracted the attention of researchers worldwide, as indicated by the increasing number of both submissions and published papers, and also from the web statistics.. It is included in major Indexing and Abstracting services. We thank all those authors who contributed papers to the April 2010 issue and the reviewers, all of whom responded to a short and challenging timetable. We are committed to placing this journal at the forefront for the dissemination of novel and exciting research. We should like to remind all prospective authors that IJCSIS does not have a page restriction. We look forward to receiving your submissions and to receiving feedback. IJCSIS April 2010 Issue (Vol. 8, No. 1) has an acceptance rate of 35%.
Shared by: ijcsiseditor
Categories
Tags
sensor network, Agile development, pervasive computing, Paper Submission, Information Technology, Computer Vision, Ubiquitous Networks, International Journal, Computer Science, Information Security, Networks Security, signal processing, Best Practices, pervasive systems, IJCSIS, Journal of Computer Science, Research
-
Stats
- views:
- 6464
- posted:
- 5/9/2010
- language:
- English
- pages:
- 360
Document Sample
IJCSIS Vol. 8 No. 1, April 2010
ISSN 1947-5500
International Journal of
Computer Science
& Information Security
© IJCSIS PUBLICATION 2010
Editorial
Message from Managing Editor
International Journal of Computer Science and Information Security (IJCSIS)
provides a major venue for rapid publication of high quality computer science research,
including multimedia, information science, security, mobile & wireless network, data
mining, software engineering and emerging technologies etc. IJCSIS has continued to
make progress and has attracted the attention of researchers worldwide, as indicated by
the increasing number of both submissions and published papers, and also from the
web statistics.. It is included in major Indexing and Abstracting services.
We thank all those authors who contributed papers to the April 2010 issue and the
reviewers, all of whom responded to a short and challenging timetable. We are
committed to placing this journal at the forefront for the dissemination of novel and
exciting research. We should like to remind all prospective authors that IJCSIS does
not have a page restriction. We look forward to receiving your submissions and to
receiving feedback.
IJCSIS April 2010 Issue (Vol. 8, No. 1) has an acceptance rate of 35%.
Special thanks to our technical sponsors for their valuable service.
Available at http://sites.google.com/site/ijcsis/
IJCSIS Vol. 8, No. 1, April 2010 Edition
ISSN 1947-5500 © IJCSIS 2010, USA.
Indexed by (among others):
IJCSIS EDITORIAL BOARD
Dr. Gregorio Martinez Perez
Associate Professor - Professor Titular de Universidad, University of Murcia
(UMU), Spain
Dr. M. Emre Celebi,
Assistant Professor, Department of Computer Science, Louisiana State University
in Shreveport, USA
Dr. Yong Li
School of Electronic and Information Engineering, Beijing Jiaotong University,
P. R. China
Prof. Hamid Reza Naji
Department of Computer Enigneering, Shahid Beheshti University, Tehran, Iran
Dr. Sanjay Jasola
Professor and Dean, School of Information and Communication Technology,
Gautam Buddha University
Dr Riktesh Srivastava
Assistant Professor, Information Systems, Skyline University College, University
City of Sharjah, Sharjah, PO 1797, UAE
Dr. Siddhivinayak Kulkarni
University of Ballarat, Ballarat, Victoria, Australia
Professor (Dr) Mokhtar Beldjehem
Sainte-Anne University, Halifax, NS, Canada
Dr. Alex Pappachen James, (Research Fellow)
Queensland Micro-nanotechnology center, Griffith University, Australia
TABLE OF CONTENTS
1. Paper 29031048: Buffer Management Algorithm Design and Implementation Based on Network
Processors (pp. 1-8)
Yechang Fang, Kang Yen, Dept. of Electrical and Computer Engineering, Florida International University,
Miami, USA
Deng Pan, Zhuo Sun, School of Computing and Information Sciences, Florida International University,
Miami, USA
2. Paper 08031001: Multistage Hybrid Arabic/Indian Numeral OCR System (pp. 9-18)
Yasser M. Alginaih, Ph.D., P.Eng. IEEE Member, Dept. of Computer Science, Taibah University, Madinah,
Kingdom of Saudi Arabia
Abdul Ahad Siddiqi, Ph.D., Member IEEE & PEC, Dept. of Computer Science, Taibah University,
Madinah, Kingdom of Saudi Arabia
3. Paper 30031056: Attribute Weighting with Adaptive NBTree for Reducing False Positives in
Intrusion Detection (pp. 19-26)
Dewan Md. Farid, and Jerome Darmont, ERIC Laboratory, University Lumière Lyon 2, Bat L - 5 av.
Pierre Mendes, France, 69676 BRON Cedex, France
Mohammad Zahidur Rahman, Department of Computer Science and Engineering, Jahangirnagar
University, Dhaka – 1342, Bangladesh
4. Paper 30031053: Improving Overhead Computation and pre-processing Time for Grid Scheduling
System (pp. 27-34)
Asgarali Bouyer, Mohammad javad hoseyni, Department of Computer Science, Islamic Azad University-
Miyandoab branch, Miyandoab, Iran
Abdul Hanan Abdullah, Faculty Of Computer Science And Information Systems, Universiti Teknologi
Malaysia, Johor, Malaysia
5. Paper 20031026: The New Embedded System Design Methodology For Improving Design Process
Performance (pp. 35-43)
Maman Abdurohman, Informatics Faculty, Telecom Institute of Technology, Bandung, Indonesia
Kuspriyanto, STEI Faculty, Bandung Institute of Technology, Bandung, Indonesia
Sarwono Sutikno, STEI Faculty, Bandung Institute of Technology, Bandung, Indonesia
Arif Sasongko, STEI Faculty, Bandung Institute of Technology, Bandung, Indonesia
6. Paper 30031060: Semi-Trusted Mixer Based Privacy Preserving Distributed Data Mining for
Resource Constrained Devices (pp. 44-51)
Md. Golam Kaosar, School of Engineering and Science, Victoria University, Melbourne, Australia
Xun Yi, Associate Preofessor, School of Engineering and Science, Victoria University, Melbourne,
Australia
7. Paper 12031005: Adaptive Slot Allocation And Bandwidth Sharing For Prioritized Handoff Calls
In Mobile Netwoks (pp. 52-57)
S. Malathy, Research Scholar, Anna University, Coimbatore
G. Sudha Sadhasivam, Professor, CSE Department, PSG College of Technology, Coimbatore.
K. Murugan, Lecturer, IT Department, Hindusthan Institute of Technology, Coimbatore
S. Lokesh, Lecturer, CSE Department, Hindusthan Institute of Technology, Coimbatore
8. Paper 12031009: An Efficient Vein Pattern-based Recognition System (pp. 58-63)
Mohit Soni, DFS, New Delhi- 110003, INDIA.
Sandesh Gupta, UIET, CSJMU, Kanpur-208014, INDIA.
M.S. Rao, DFS, New Delhi-110003, INDIA
Phalguni Gupta, Professor, IIT Kanpur, Kanpur-208016, INDIA.
9. Paper 15031013: Extending Logical Networking Concepts in Overlay Network-on-Chip
Architectures (pp. 64-67)
Omar Tayan
College of Computer Science and Engineering, Department of Computer Science, Taibah University, Saudi
Arabia, P.O. Box 30002
10. Paper 15031015: Effective Bandwidth Utilization in IEEE802.11 for VOIP (pp. 68-75)
S. Vijay Bhanu, Research Scholar, Anna University, Coimbatore, Tamilnadu, India, Pincode-641013.
Dr.RM.Chandrasekaran, Registrar, Anna University, Trichy, Tamilnadu, India, Pincode: 620024.
Dr. V. Balakrishnan, Research Co-Supervisor, Anna University, Coimbatore.
11. Paper 16021024: ECG Feature Extraction Techniques - A Survey Approach (pp. 76-80)
S. Karpagachelvi, Mother Teresa Women's University, Kodaikanal, Tamilnadu, India.
Dr. M.Arthanari, Tejaa Shakthi Institute of Technology for Women, Coimbatore- 641 659, Tamilnadu,
India.
M. Sivakumar, Anna University – Coimbatore, Tamilnadu, India
12. Paper 18031017: Implementation of the Six Channel Redundancy to achieve fault tolerance in
testing of satellites (pp. 81-85)
H S Aravinda *, Dr H D Maheshappa**, Dr Ranjan Moodithaya ***
* Department of Electronics and Communication, REVA ITM, Bangalore-64, Karnataka, India.
** Director & Principal, East Point College of Engg, Bidarahalli, Bangalore-40, Karnataka, India.
*** Head, KTMD Division, National Aerospace Laboratories, Bangalore-17, Karnataka, India.
13. Paper 18031018: Performance Oriented Query Processing In GEO Based Location Search
Engines (pp. 86-94)
Dr. M. Umamaheswari, Bharath University, Chennai-73, Tamil Nadu,India,
S. Sivasubramanian, Bharath University, Chennai-73,Tamil Nadu,India,
14. Paper 20031027: Tunable Multifunction Filter Using Current Conveyor (pp. 95-98)
Manish Kumar, Electronics and Communication, Engineering Department, Jaypee Institute of Information
Technology, Noida, India
M.C. Srivastava, Electronics and Communication, Engineering Department, Jaypee Institute of
Information Technology, Noida, India
Umesh Kumar, Electrical Engineering Department, Indian Institute of Technology, Delhi, India
15. Paper 17031042: Artificial Neural Network based Diagnostic Model For Causes of Success and
Failures (pp. 95-105)
Bikrampal Kaur, Chandigarh Engineering College, Mohali, India
Dr. Himanshu Aggarwal, Punjabi University, Patiala-147002, India
16. Paper 28031045: Detecting Security threats in the Router using Computational Intelligence (pp.
106-111)
J. Visumathi, Research Scholar, Sathyabama University, Chennai-600 119
Dr. K. L. Shunmuganathan, Professor & Head, Department of CSE, R.M.K. Engineering College, Chennai-
601 206
17. Paper 31031091: A Novel Algorithm for Informative Meta Similarity Clusters Using Minimum
Spanning Tree (pp. 112-120)
S. John Peter, Department of Computer Science and Research Center, St. Xavier’s College, Palayamkottai,
Tamil Nadu, India
S. P. Victor, Department of Computer Science and Research Center, St. Xavier’s College, Palayamkottai,
Tamil Nadu, India
18. Paper 23031032: Adaptive Tuning Algorithm for Performance tuning of Database Management
System (pp. 121-124)
S. F. Rodd, Department of Information Science and Engineering, KLS’s Gogte Institute of Technology,
Belgaum, INDIA
Dr. U. P. Kulkarni, Department of Computer Science and Engineering, SDM College of Engineering and
Technology, Dharwad, INDIA
19. Paper 26031038: A Survey of Mobile WiMAX IEEE 802.16m Standard (pp. 125-131)
Mr. Jha Rakesh, Deptt. Of E & T.C., SVNIT, Surat, India
Mr. Wankhede Vishal A., Deptt. Of E & T.C., SVNIT, Surat, India
Prof. Dr. Upena Dalal, Deptt. Of E & T.C., SVNIT, Surat, India
20. Paper 27031040: An Analysis for Mining Imbalanced Datasets (pp. 132-137)
T. Deepa, Faculty of Computer Science Department, Sri Ramakrishna College of Arts and Science for
Women, Coimbatore, Tamilnadu, India.
Dr. M. Punithavalli, Director & Head, Sri Ramakrishna College of Arts & Science for Women, Coimbatore,
Tamil Nadu, India
21. Paper 27031039: QoS Routing For Mobile Adhoc Networks And Performance Analysis Using
OLSR Protocol (pp. 138-150)
K.Oudidi, Si2M Laboratory, National School of Computer Science and Systems Analysis, Rabat, Morocco
A. Hajami, Si2M Laboratory, National School of Computer Science and Systems Analysis, Rabat, Morocco
M. Elkoutbi, Si2M Laboratory, National School of Computer Science and Systems Analysis, Rabat,
Morocco
22. Paper 28031047: Design of Simple and Efficient Revocation List Distribution in Urban Areas for
VANET’s (pp. 151-155)
Ghassan Samara , National Advanced IPv6 Center, Universiti Sains Malaysia, Penang, Malaysia
Sureswaran Ramadas, National Advanced IPv6 Center, Universiti Sains Malaysia, Penang, Malaysia
Wafaa A.H. Al-Salihy, School of Computer Science, Universiti Sains Malaysia, Penang, Malaysia
23. Paper 28031044: Software Process Improvization Framework For Indian Small Scale Software
Organizations Using Fuzzy Logic (pp. 156-162)
A. M. Kalpana, Research Scholar, Anna University Coimbatore, Tamilnadu, India
Dr. A. Ebenezer Jeyakumar, Director/Academics, SREC, Coimbatore, Tamilnadu, India
24. Paper 30031052: Urbanizing the Rural Agriculture - Knowledge Dissemination using Natural
Language Processing (pp. 163-169)
Priyanka Vij (Author) Student, Computer Science Engg. Lingaya‟s Institute of Mgt. & Tech, Faridabad,
Haryana, India
Harsh Chaudhary (Author) Student, Computer Science Engg. Lingaya‟s Institute of Mgt. & Tech,
Faridabad, Haryana, India
Priyatosh Kashyap (Author) Student, Computer Science Engg. Lingaya‟s Institute of Mgt. & Tech,
Faridabad, Haryana, India
25. Paper 31031073: A New Joint Lossless Compression And Encryption Scheme Combining A
Binary Arithmetic Coding With A Pseudo Random Bit Generator (pp. 170-175)
A. Masmoudi * , W. Puech **, And M. S. Bouhlel *
* Research Unit: Sciences and Technologies of Image and Telecommunications, Higher Institute of
Biotechnology, Sfax TUNISIA
** Laboratory LIRMM, UMR 5506 CNRS University of Montpellier II, 161, rue Ada, 34392
MONTPELLIER CEDEX 05, FRANCE
26. Paper 15031012: A Collaborative Model for Data Privacy and its Legal Enforcement (pp. 176-182)
Manasdeep, MSCLIS, IIIT Allahabad
Damneet Singh Jolly, MSCLIS, IIIT Allahabad
Amit Kumar Singh, MSCLIS, IIIT Allahabad
Kamleshwar Singh, MSCLIS, IIIT Allahabad
Mr Ashish Srivastava, Faculty, MSCLIS, IIIT Allahabad
27. Paper 12031010: A New Exam Management System Based on Semi-Automated Answer Checking
System (pp. 183-189)
Arash Habibi Lashkari, Faculty of ICT, LIMKOKWING University of Creative Technology,
CYBERJAYA, Selangor,
Dr. Edmund Ng Giap Weng, Faculty of Cognitive Sciences and Human Development, University Malaysia
Sarawak (UNIMAS)
Behrang Parhizkar, Faculty of Information, Communication And Technology, LIMKOKWING University
of Creative Technology, CYBERJAYA, Selangor, Malaysia
Siti Fazilah Shamsudin, Faculty of ICT, LIMKOKWING University of Creative Technology, CYBERJAYA,
Selangor, Malaysia
Jawad Tayyub, Software Engineering With Multimedia, LIMKOKWING University of Creative Technology,
CYBERJAYA, Selangor, Malaysia
28. Paper 30031064: Development of Multi-Agent System for Fire Accident Detection Using Gaia
Methodology (pp. 190-194)
Gowri. R, Kailas. A, Jeyaprakash.R, Carani Anirudh
Department of Information Technology, Sri Manakula Vinayagar Engineering College, Puducherry – 605
107.
29. Paper 19031022: Computational Fault Diagnosis Technique for Analog Electronic Circuits using
Markov Parameters (pp. 195-202)
V. Prasannamoorthy and N.Devarajan
Department of Electrical Engineering, Government College of Technology, Coimbatore, India
30. Paper 24031037: Applicability of Data Mining Techniques for Climate Prediction – A Survey
Approach (pp. 203-206)
Dr. S. Santhosh Baboo, Reader, PG and Research department of Computer Science, Dwaraka Doss
Goverdhan Doss Vaishnav College, Chennai
I. Kadar Shereef, Head, Department of Computer Applications, Sree Saraswathi Thyagaraja College,
Pollachi
31. Paper 17021025: Appliance Mobile Positioning System (AMPS) (An Advanced mobile
Application) (pp. 207-215)
Arash Habibi Lashkari, Faculty of ICT, LIMKOKWING University of Creative Technology,
CYBERJAYA, Selangor, Malaysia
Edmund Ng Giap Weng, Faculty of Cognitive Sciences and Human Development, University Malaysia
Sarawak (UNIMAS)
Behrang Parhizkar, Faculty of ICT, LIMKOKWING University of Creative Technology, CYBERJAYA,
Selangor, Malaysia
Hameedur Rahman, Software Engineering with Multimedia, LIMKOKWING University of Creative
Technology, CYBERJAYA, Selangor, Malaysia
32. Paper 24031036: A Survey on Data Mining Techniques for Gene Selection and Cancer
Classification (pp. 216-221)
Dr. S. Santhosh Baboo, Reader, PG and Research department of Computer Science, Dwaraka Doss
Goverdhan Doss Vaishnav College, Chennai
S. Sasikala, Head, Department of Computer Science, Sree Saraswathi Thyagaraja College, Pollachi
33. Paper 23031033: Non-Blind Image Watermarking Scheme using DWT-SVD Domain (pp. 222-228)
M. Devapriya, Asst.Professor, Dept of Computer Science, Government Arts College, Udumalpet.
Dr. K. Ramar, Professor & HOD, Dept of CSE, National Engineering College, Kovilpatti -628 502.
34. Paper 31031074: Speech Segmentation Algorithm Based On Fuzzy Memberships (pp. 229-233)
Luis D. Huerta, Jose Antonio Huesca and Julio C. Contreras
Departamento de Informática, Universidad del Istmo Campus Ixtepéc, Ixtepéc Oaxaca, México
35. Paper 30031058: How not to share a set of secrets (pp. 234-237)
K. R. Sahasranand , Nithin Nagaraj, Department of Electronics and Communication Engineering, Amrita
Vishwa Vidyapeetham, Amritapuri Campus, Kollam-690525, Kerala, India.
Rajan S., Department of Mathematics, Amrita Vishwa Vidyapeetham, Amritapuri Campus, Kollam-690525,
Kerala, India.
36. Paper 30031057: Secure Framework for Mobile Devices to Access Grid Infrastructure (pp. 238-
243)
Kashif Munir, Computer Science and Engineering Technology Unit King Fahd University of Petroleum
and Minerals HBCC Campus, King Faisal Street, Hafr Al Batin 31991
Lawan Ahmad Mohammad, Computer Science and Engineering Technology Unit King Fahd University of
Petroleum and Minerals HBCC Campus, King Faisal Street, Hafr Al Batin 31991
37. Paper 31031076: DSP Specific Optimized Implementation of Viterbi Decoder (pp. 244-249)
Yame Asfia and Dr Muhamamd Younis Javed, Department of Computer Engg, College of Electrical and
Mechanical Engg, NUST, Rawalpindi, Pakistan
Dr Muid-ur-Rahman Mufti, Department of Computer Engg, UET Taxila, Taxila, Pakistan
38. Paper 31031089: Approach towards analyzing motion of mobile nodes- A survey and graphical
representation (pp. 250-253)
A. Kumar, Sir Padampat Singhania University, Udaipur , Rajasthan , India
P.Chakrabarti, Sir Padampat Singhania University, Udaipur , Rajasthan , India
P. Saini, Sir Padampat Singhania University, Udaipur , Rajasthan , India
39. Paper 31031092: Recognition of Printed Bangla Document from Textual Image Using Multi-
Layer Perceptron (MLP) Neural Network (pp. 254-259)
Md. Musfique Anwar, Nasrin Sultana Shume, P. K. M. Moniruzzaman and Md. Al-Amin Bhuiyan
Dept. of Computer Science & Engineering, Jahangirnagar University, Bangladesh
40. Paper 31031081: Application Of Fuzzy System In Segmentation Of MRI Brain Tumor (pp. 261-
270)
Mrigank Rajya, Sonal Rewri, Swati Sheoran
CSE, Lingaya’s University, Limat, Faridabad India, New Delhi, India
41. Paper 30031059: E-Speed Governors For Public Transport Vehicles (pp. 270-274)
C. S. Sridhar, Dr. R. ShashiKumar, Dr. S. Madhava Kumar, Manjula Sridhar, Varun. D
ECE dept, SJCIT, Chikkaballapur.
42. Paper 31031087: Inaccuracy Minimization by Partioning Fuzzy Data Sets - Validation of
Analystical Methodology (pp. 275-280)
Arutchelvan. G, Department of Computer Science and Applications Adhiparasakthi College of Arts and
Science G. B. Nagar, Kalavai , India
Dr. Srivatsa S. K., Dept. of Electronics Engineering, Madras Institute of Technology, Anna University,
Chennai, India
Dr. Jagannathan. R, Vinayaka Mission University, Chennai, India
43. Paper 30031065: Selection of Architecture Styles using Analytic Network Process for the
Optimization of Software Architecture (pp. 281-288)
K. Delhi Babu, S.V. University, Tirupati
Dr. P. Govinda Rajulu, S.V. University, Tirupati
Dr. A. Ramamohana Reddy, S.V. University, Tirupati
Ms. A.N. Aruna Kumari, Sree Vidyanikethan Engg. College, Tirupati
44. Paper 27031041: Clustering Time Series Data Stream – A Literature Survey (pp. 289-294)
V.Kavitha, Computer Science Department, Sri Ramakrishna College of Arts and Science for Women,
Coimbatore, Tamilnadu, India.
M. Punithavalli, Sri Ramakrishna College of Arts & Science for Women, Coimbatore ,Tamil Nadu, India.
45. Paper 31031086: An Adaptive Power Efficient Packet Scheduling Algorithm for Wimax
Networks (pp. 295-300)
R Murali Prasad, Department of Electronics and Communications, MLR Institute of technology,
Hyderabad
P. Satish Kumar, professor, Department of Electronics and Communications, CVR college of engineering,
Hyderabad
46. Paper 30041037: Content Base Image Retrieval Using Phong Shading (pp. 301-306)
Uday Pratap Singh, LNCT, Bhopal (M.P) INDIA
Sanjeev Jain, LNCT, Bhopal (M.P) INDIA
Gulfishan Firdose Ahmed, LNCT, Bhopal (M.P) INDIA
47. Paper 31031090: The Algorithm Analysis of E-Commerce Security Issues for Online Payment
Transaction System in Banking Technology (pp. 307-312)
Raju Barskar, MANIT Bhopal (M.P)
Anjana Jayant Deen,CSE Department, UIT_RGPV, Bhopal (M.P)
Jyoti Bharti, IT Department, MANIT, Bhopal (M.P)
Gulfishan Firdose Ahmed, LNCT, Bhopal (M.P)
48. Paper 28031046: Reduction in iron losses In Indirect Vector-Controlled IM Drive Using FLC (pp.
313-317)
Mr. C. Srisailam , Electrical Engineering Department, Jabalpur Engineering College, Jabalpur, Madhya
Pradesh,
Mr. Mukesh Tiwari, Electrical Engineering Department, Jabalpur Engineering College, Jabalpur, Madhya
Pradesh,
Dr. Anurag Trivedi, Electrical Engineering Department, Jabalpur Engineering College, Jabalpur, Madhya
Pradesh
49. Paper 31031071: Bio-Authentication based Secure Transmission System using Steganography (pp.
318-324)
Najme Zehra, Assistant Professor, Computer Science Department, Indira Gandhi Institute of Technology,
GGSIPU, Delhi.
Mansi Sharma, Scholar, Indira Gandhi Institute of Technology, GGSIPU, Delhi.
Somya Ahuja, Scholar, Indira Gandhi Institute of Technology, GGSIPU, Delhi.
Shubha Bansal, Scholar, Indira Gandhi Institute of Technology, GGSIPU, Delhi.
50. Paper 31031068: Facial Recognition Technology: An analysis with scope in India (pp. 325-330)
Dr.S.B.Thorat, Director, Institute of Technology and Mgmt, Nanded, Dist. - Nanded. (MS), India
S. K. Nayak, Head, Dept. of Computer Science, Bahirji Smarak Mahavidyalaya, Basmathnagar, Dist. -
Hingoli. (MS), India
Miss. Jyoti P Dandale, Lecturer, Institute of Technology and Mgmt, Nanded, Dist. - Nanded. (MS), India
51. Paper 31031069: Classification and Performance of AQM-Based Schemes for Congestion
Avoidance (pp. 331-340)
K.Chitra Lecturer, Dept. of Computer Science D.J.Academy for Managerial Excellence Coimbatore, Tamil
Nadu, India – 641 032
Dr. G. Padamavathi Professor & Head, Dept. of Computer Science Avinashilingam University for Women,
Coimbatore, Tamil Nadu, India – 641 043
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
Buffer Management Algorithm Design and
Implementation Based on Network Processors
Yechang Fang, Kang Yen Deng Pan, Zhuo Sun
Dept. of Electrical and Computer Engineering School of Computing and Information Sciences
Florida International University Florida International University
Miami, USA Miami, USA
{yfang003, yenk}@fiu.edu {pand, zsun003}@fiu.edu
Abstract—To solve the parameter sensitive issue of the network QoS, and also the key method to solve the network
traditional RED (random early detection) algorithm, an congestion problem. Queue management consists of buffer
adaptive buffer management algorithm called PAFD (packet management and packet scheduling. Generally the buffer
adaptive fair dropping) is proposed. This algorithm supports management is applied at the front of a queue and
DiffServ (differentiated services) model of QoS (quality of cooperates with the packet scheduling to complete the queue
service). In this algorithm, both of fairness and throughput are operation [2, 3]. When a packet arrives at the front of a
considered. The smooth buffer occupancy rate function is queue, the buffer management decides whether to allow the
adopted to adjust the parameters. By implementing buffer packet coming into the buffer queue. From another point of
management and packet scheduling on Intel IXP2400, the view, the buffer management determines whether to drop the
viability of QoS mechanisms on NPs (network processors) is packet or not, so it is also known as dropping control.
verified. The simulation shows that the PAFD smoothes the
The control schemes of the buffer management can be
flow curve, and achieves better balance between fairness and
analyzed from two levels, data flow and data packet. In the
network throughput. It also demonstrates that this algorithm
data stream level and viewed form the aspect of system
meets the requirements of fast data packet processing, and the
resource management, the buffer management needs to
hardware resource utilization of NPs is higher.
adopt certain resource management schemes to make a fair
Keywords-buffer management; packet dropping; queue and effective allocation of queue buffer resources among
management; network processor flows through the network nodes. In the data packet level
and viewed from the aspect of packet dropping control, the
I. INTRODUCTION buffer management needs to adopt certain drop control
Network information is transmitted in the form of data schemes to decide that under what kind of circumstances a
flow, which constitutes of data packets. Therefore, different packet should be dropped, and which packet will be dropped.
QoS means different treatment of data flow. This treatment Considering congestion control response in an end-to-end
involves assignment of different priority to data packets. system, the transient effects for dropping different packets
Queue is actually a storage area to store IP packets with may vary greatly. However, statistics of the long-term
priority level inside routers or switches. Queue management operation results indicates that the transient effect gap is
algorithm is a particular calculation method to determine the minimal, and this gap can be negligible in majority of cases.
order of sending data packets stored in the queue. Then the In some specific circumstances, the completely shared
fundamental requirement is to provide better and timely resource management scheme can cooperate with drop
services for high priority packets [1]. The NP is a dedicated schemes such as tail-drop and head-drop to reach effective
processing chip to run on high speed networks, and to control. However, in most cases, interaction between these
achieve rapid processing of packets. two schemes is very large. So the design of buffer
management algorithms should consider both of the two
Queue management plays a significant role in the control
schemes to obtain better control effects [4, 5].
of network transmission. It is the core mechanism to control
1 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
II. EXISTING BUFFER MANAGEMENT ALGORITHMS that QoS of service flows with poor transmission conditions
Reference [6] proposed the RED algorithm for active cannot be guaranteed. Packet scheduling algorithms usually
queue management (AQM) mechanism [7] and then use generalized processor sharing (GPS) as a comparative
standardized as a recommendation from IETF [8]. It model of fairness. During the process of realization of
introduces congestion control to the router's queue packet scheduling algorithms based on GPS, each service
operations. RED uses early random drop scheme to smooth flow has been assigned a static weight to show their QoS.
packet dropping in time. This algorithm can effectively The weight φi actually express the percentage of the service
reduce and even avoid the congestion in network, and also flow i in the entire bandwidth B. φi will not change with
solve the TCP protocol global synchronization problem. packet scheduling algorithms, and meet
N
However, one concern of the RED algorithm is the ∑ φi = 1 (1)
i =1
stability problem, i.e., the performance of the algorithm is
very sensitive to the control parameters and changes in where N expresses the number of service flows in the link.
network traffic load. During heavy flow circumstances, the And the service volume is described by
performance of RED will drop drastically. Since RED
φi
algorithm is based on best-effort service model, which does giinc = B (2)
∑ φj
not consider different levels of services and different user j∈B
flows, it cannot provide fairness. In order to improve the where i, j denotes two different service flows. In GPS based
fairness and stability, several improved algorithms have algorithms, the bandwidth allocation of different service
been developed, including WRED, SRED, Adaptive-RED, flows meets the requirement Bi/φi = Bj/φj, where Bi is the
FRED, RED with In/Out (RIO) [9, 10] etc. But these allocated bandwidth of the service flow i. By assigning a
algorithms still have a lot of problems. For example, a large smaller weight φ to an unimportant background service flow,
number of studies have shown that it is difficult to find a the weight of service flow with high priority φhigh will be
RIO parameter setting suitable for various and changing much larger than φlow, so that the majority of the bandwidth
network conditions. is accessed by high-priority service flows.
III. THE PAFD ALGORITHM A. Algorithm Description
In this paper, we propose a new buffer management In buffer management algorithms, how to control the
algorithm called PAFD (Packet Adaptive Fair Dropping). buffer space occupation is very key [11]. Here we define
This algorithm will adaptively gain balance between
congestion and fairness according to cache congestion Ci C j (3)
=
Wi W j
situation. When there is minor congestion, the algorithm will
tend to fairly drop packets in order to ensure all users access
where Ci is the buffer space occupation, and Wi expresses
the system resources to their scale. For moderate congestion,
the synthetic weight of the service flow i. When the cache is
the algorithm will incline to drop the packet of low quality
full, the service flow with the largest value of Ci /Wi will be
service flows by reducing its sending rate using scheduling
dropped in order to guarantee fairness. Here the fairness is
algorithm to alleviate congestion. In severe congestion, the
reflected in packets with different queue length [12, 13].
algorithm will tend to fairly drop packets, through the upper
Assume that ui is the weight, and vi is the current queue
flow control mechanism to meet the QoS requirements, and
length of the service flow i. The synthetic weight Wi can be
reduces sending rate of most service flows, in order to speed
calculated as described by
up the process of easing the congestion.
In buffer management or packet scheduling algorithms, Wi = α × ui + (1 − α ) × vi (4)
it will improve the system performance to have service
where α is the adjust parameter of the two weighting
flows with better transmission conditions reserved in
coefficients ui and vi . α can be pre-assigned, or determined
advance. But this operation will make system resources such
in accordance with usage of the cache. ui is related to the
as buffer space and bandwidth be unfairly distributed, so
2 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
service flow itself, and different service flows are assigned cycling times is related to the ratio between the longest and
with different weight values. As long as the service flow is the shortest packets. At this moment, the time complexity
active, this factor will remain unchanged. vi is time varying, overhead is still small based on practices.
which reflects dropping situation of the current service flow.
In Step 2, α, a function of shared buffer, is a parameter
Suppose a new packet T arrives, then the PAFD for adjusting proportion of the two weighting coefficients u
algorithm process is described as follows: and v. For a large value of α, the PAFD algorithm will tend
to fairly select and drop packets according to the synthetic
• Step 1: Check whether the remaining cache space
weight W. Otherwise, the algorithm tends to select and drop
can accommodate the packet T, if the remaining
the service flow with large queue length. A reasonable value
space is more than or equal to the length of T, add T
for α can be used to balance between fairness and
into the cache queue. Otherwise, drop some packets
performance. Here we introduce an adaptive method to
from the cache to free enough storage space. The
determine the value of α. This adaptive method will
decision on which packet will be dropped is given in
determine α value based on the congestion situation of the
the following steps.
cache, and this process does not require manual intervention.
• Step 2: Calculate the weighting coefficients u and v
When there is a minor congestion, the congestion can be
for each service flow, and the parameter α. Then get
relieved by reducing the sending rate of a small number of
the values of new synthetic weights W for each flow
service flows. The number of service flows in wireless
according to (4).
network nodes is not as many as that in the wired network.
• Step 3: Choose the service flow with the largest So the minor congestion can be relieved by reducing the
weighted buffer space occupation (Ci/Wi), if the sending rate of any one of service flows. We hope this
service flow associated to the packet T has the same choice is fair, to ensure that all user access to the system
value as it, then drop T at the probability P and resources according to their weights.
returns. Otherwise, drop the head packet of the
When there is a moderate congestion, the congestion can
service flow with the largest weighted buffer space
not be relieved by reducing the sending rate of any one of
occupation at probability 1−P, and add T into the
service flows. Reducing the rate of different service flows
cache queue. Here Probability P is a random number
will produce different results. We hope to reduce the rate of
generated by the system to ensure the smoothness
service flows which are most effective to the relief of
and stability of the process.
congestion. That is, the service flow which current queue
• Step 4: Check whether the remaining space can length is the longest (The time that these service flow
accommodate another new packet, if the answer is occupied the cache is also the longest). This not only
yes, the packet will be transmitted into the cache. improves system throughput, but also made to speeds up the
Otherwise, return to Step 3 to continuously choose congestion relief.
and drop packets until there is sufficient space.
When there is a severe congestion, it is obvious that
If all packet lengths are the same, the algorithm only reducing the sending rate of a small portion of the service
needs one cycle to compare and select the service flow with flows cannot achieve the congestion relief. We may need to
the largest weighted buffer space occupation. Therefore, the reduce the rate of a lot of service flows. Since the TCP has a
time complexity of the algorithm is O(N). In this case, we characteristic of additive increase multiplicative decrease
also need additional 4N storage space to store the weights. (AIMD), continuous drop packets from one service flow to
Taking into account the limited capacity of wireless network, reduce the sending rate would adversely affect the
N is usually less than 100. So in general the algorithm's performance of the TCP flow. While the effect on relieving
overhead on time and space complexity are not large. On the system congestion will become smaller, we gradually
other hand, if packet lengths are different, then it is increase the values of parameters, and the algorithm will
necessary to cycle Step 3 and Step 4 until the cache has choose service flows to drop packet fairly. On one hand, at
enough space to accommodate the new packet. The largest this point the "fairness" can bring the same benefits as in the
3 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
minor congestion system; on the other hand this is to avoid In the DiffServ model, we retain the implement process
continuously dropping the longer queue service flow. of PAFD, and only modify (4) into
Congestion is measured by the system buffer space
Wi = (α × u i + (1 − α ) × vi ) × β (6)
occupation rate. α is a parameter relevant to system
congestion status and its value is between 0 to 1. Assume where β is a new parameter used to adjust the fairness
that the current buffer space occupation rate is denoted by among service flows of different service levels. As
Buffercur, and Buffermedium, Buffermin, and Buffermax represent mentioned above, we can set the value of parameter α
threshold value of the buffer space occupation rate for different from that shown in Figure 1 to satisfy different
moderate, minor, and severe congestion, respectively. requirements. α is the parameter which balances fairness and
When Buffercur is close to Buffermin, the system enters a transmission conditions. For high-priority services, the curve
state of minor congestion. When Buffercur reaches Buffermax, in Figure 1 is reasonable. The fairness is able to guarantee
the system is in a state of severe congestion. Buffermedium the QoS for different service flows, and also is required to
means moderate congestion. If we value α by using linear relief congestion quickly. For high-priority services which
approach, the system will have a dramatic oscillation. have no delay constraints and high fairness requirements, a
Instead we use high order nonlinear or index reduction to get higher throughput is more practical. Therefore, we can get
smooth curve of α as shown in Figure 1. the value of the parameter α for low-priority services, which
is slightly less than that for high-priority services as shown
in Figure 2.
Fig.1. An adaptive curve of α
The value of α can also be calculated as below
⎧0, if Buffercur < Buffermin
2 2
⎪ 2 2
⎪ Buffercur − Buffermin
α = ⎨1− 2 2
2 2 2
, if Buffermin ≤ Buffercur ≤ Buffermax (5)
⎪ Buffermax − Buffermin
⎪1, if Buffer 2 < Buffer 2
⎩ cur max
B. DiffServ Model Support
In the PAFD algorithm, we can adopt the DiffServ model
Fig.2. Values of α for different priority services
to simplify the service flows by dividing them into
high-priority services such as assurance services and Now we check the effects of the parameter β. For
low-priority services such as best-effort services. We use the high-priority services, β is a constant with value 1. For
queuing method for the shared cache to set and manage the low-priority services, the value of β is less than 1, and
cache. When a new packet arrives at the cache, first the influenced by the network load. When network load is low,
service flow is checked to see whether it matches the service β equals to 1. In this case, different level service flows have
level agreement (SLA). If it does, then this new packet the same priority to share the network resources. As network
enters the corresponding queue. Otherwise, the packet is load increases, in order to guarantee the QoS of high-priority
assigned to low-priority services, and then enters the services, low-priority services gradually give up some
low-priority queue. transmission opportunities, so the value of β decreases. The
4 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
higher network load is, the smaller the values of β and W are. channel transmission condition will give higher priority and
Therefore, the probability of a low-priority packet being result effective throughput.
dropped is higher. Values of β are shown below.
Fig.3. Values of β for different priority services
IV. SIMULATION RESULTS
A. Simulation for Commen Services Fig.4. Throughputs of RED and PAFD
We compare the PAFD algorithm with two commonly
used buffer management algorithms RED and tail drop (TD).
We choose two common packet scheduling algorithms Best
Channel First (BCF) and Longest Queue First (LQF) to
work with PAFD, RED and TD. Here the LQF uses the
weighted queue length for packet scheduling. So there are 6
queue management algorithm combinations, which are
PAFD-BCF, PAFD-LQF, RED-BCF, RED-LQF, TD-BCF,
and TD-LQF. The performance comparisons of these
algorithms are carried out with respect to throughput
effectiveness, average queuing delay, and fairness.
We use K1297-G20 signaling analyzer to simulate Fig.5. Average Queuing Delay for TD, RED and PAFD
packet sending, and the operation system for K1297-G20 is
Windows NT 4.0. ADLINK 6240 is used as the NP blade. From Figure 5, we find that RED has better performance
Based on the simulation configuration, there are 8 different on the average queuing delay due to the capability of early
packet length configurations for the data source. They are detection of congestion and its drop mechanism. BCF has
fixed length of 64 bytes, fixed length of 65 bytes, fixed better performance on queuing delay than that of LQF. As
length of 128 byte, fixed length of 129 bytes, fixed length of the load increases, the average queuing delay of PAFD first
256 bytes, random length of 64-128 bytes, random length of increases, then decreases. This is because RAFD does not
64-256 bytes, and random length of 64-1500 bytes. use tail drop, and instead searches a service flow with the
largest weighted buffer space occupation to drop the head
Figure 4 shows that all the algorithms have similar
packet to reduce the average queuing time.
throughputs for low network load. When the load increases,
the throughput effectiveness of BCF is higher than that of Both TD and RED use shared cache instead of flow
other scheduling algorithms. This figure shows that queuing so that they fail to consider the fairness. Here the
PAFD-BCF provides significant higher throughput than the fairness index F is given by
other algorithms. PAFD does not randomly drop or simply
Gi 2
N
tail drop packets, but fully considers fairness and (∑)
Wi 1 (7)
transmission conditions. In this way, service flows under F= N
G
N ∑ ( i )2
poor transmission condition receive high probability of 1 Wi
packet dropping, thus a relatively short virtual queue. When where Gi is the effective throughput of service flow i, and N
BCF is working with PAFD, the service flow under better
5 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
is the total number of service flows. It is not difficult to
prove that F∈(0, 1). When F has a lager value, the fairness
of the system is better. If the value of F equals to 1, the
system resource is completely fair. We can use (7) to
calculate the fairness index and compare the fairness of
different algorithms. In ON-OFF model with the assumption
that there are 16 service flows, the ON average rate of flows
1-8 is twice of that of 9-16. That is, Wi : Wj = 2 : 1, where
i∈[1, 8] and j∈[9, 16]. Using round robin algorithms
without considering W, we can calculate the reference value
of fairness index F = 0.9. Table I gives the fairness index of
TD, RED and PAFD which are combined with packet Fig.6. Throughputs of RED and DS-PAFD
scheduling algorithms.
TABLE I. FAIRNESS INDEX
Algorithms Fairness
TD-BCF 0.8216
TD-LQF 0.9162
RED-BCF 0.8855
RED-LQF 0.9982
PAFD-LQF 0.9988
PAFD-BCF 0.8902
The table indicates that the fairness index of BCF is
lower when combined with TD and RED. Since PAFD takes Fig.7. Average Queuing Delay of RED and DS-PAFD
the fairness into consideration, the fairness index of PAFD is
Table II gives the comparison of fairness index of theses
higher than that of TD when there are congestions. The
algorithms. Comparing these numbers with those shown in
combination of PAFD and LQF has higher throughput and
Table I, we can draw a similar conclusion. However, the
more fair distribution of cache and bandwidth resources. By
difference in values is that the fairness index of low-priority
changing the value of parameter α, we can conveniently
services is slightly lower than that of high-priority services
balance the system performance and fairness based on the
as a result of different values of parameter α selected.
requirements.
TABLE II. COMPARISON OF FAIRNESS INDEX
B. Simulation for DiffServ Model
In this section we adopt the same environment as TD-BCF TD-LQF
described in the previous section to test the PAFD Flow 0.8346 0.9266
performance based on the DiffServ model. The only DSPAFD-BCF DSPAFD-LQF
difference is that half of the services are set to high-priority, High-priority Service Flow 0.8800 0.9922
and another half to low-priority. DSPAFD-BCF DSPAFD-LQF
Low-priority Service Flow 0.8332 0.9488
Figures 6 and 7 show the throughput and average
queuing delay of those algorithms. The only difference in
these two tests is that the value of parameter α for half of the As shown in Figures 2-3, 6 and 7, when network load is
service flows used in the second simulation is slightly lower light, the throughputs are similar for different priority
than the one in the first simulation. So the curves in Figures services. This means different priority services have the
7 and 8 are very similar to those shown in Figures 4 and 5. same priority to share network resources. As network load
6 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
increases, the throughput gradually decreases. However, are 1024 queues in total. As we adopt the SRAM structure, it
even in the case of heavy load, the PAFD algorithm still is very easy to enqueue.
allocates small portion of resources to low-priority services
The dequeuing operation is similar to the enqueuing
to meet the fairness requirement. And this operation will
operation. In order to maintain the performance of the
prevent high-priority services from fully occupying the
system, micro engine threads of NPs must operate in strict
network resources.
accordance with the predetermined sequence. This is
V. IMPLEMENTATION BASED ON NETWORK PROCESSORS controlled by internal thread semaphore. When a queue
changes from empty to non-empty in an enqueuing
Here we adopt NP Intel IXP2400 to implement the
operation, or from non-empty to empty in a dequeuing
PAFD algorithm. Intel IXP2400 provides us with eight
operation, the buffer manager of PAFD will send a message
micro-engines, and each micro-engine can support up to
to packet scheduling module through the adjacent loop.
eight hardware threads. When the system is running, each
micro-engine deals with one task. During the thread VI. CONCLUSIONS
switching, there is no need for protection, each hardware
Buffer management algorithm is the core mechanism to
thread has its own register, so the switching speed is very
achieve network QoS control. It also plays an important role
fast. Also Intel IXP2400 is appropriate for DiffServ model.
in network resource management. In this paper, a novel
The PAFD Algorithm executes enqueuing and dequeuing buffer management algorithm called PAFD is proposed
operations in the transmission, which are implemented using based on NPs. The PAFD algorithm takes into account the
chained list of the SRAM of IXP2400. The buffer manager impact of transmission environment on packets. It can
of PAFD receives enqueuing request from the functional adaptively balance between queue congestion and fairness
pipeline, and accepts dequeuing request through the micro according to cache congestion. PAFD also supports the
engines of NPs. In the PAFD algorithm, Q-Array in the DiffServ model to improve network QoS based on NPs. The
SRAM controller is used to the chained list, and a queue simulation results show that the throughput and fairness are
descriptor is stored in the SRAM. The buffer manager uses better balanced after this algorithm is applied. Finally, the
content associative memory (CAM) to maintain queue PAFD algorithm is implemented based on IXP2400, which
buffer of the descriptor. When enqueuing request arrives, the means that the hardware resource utilization of NPs is
buffer manager will check CAM to see if the queue higher.
descriptor is in the local buffer. If so, PAFD will be run to
The future network has two development requirements:
decide whether the new packets should enter the queue. If
high-speed bandwidth and service diversification. Research
not, the descriptor is excluded from the Q-Array, and then
on buffer management algorithms is able to suit for these
stored in the SRAM. Therefore, another specified queue
requirements. In the future, buffer management will become
descriptor is read into the Q-Array, and then PAFD is run to
more complex. Therefore, the requirements for NPs and
decide whether to drop the new packets. When a queue
other hardware will be more stringent. It is very important to
enters a queue, Q-Array logic moves the first four bits to the
consider the comprehensive performance of the algorithms
SRAM controller. Q-Array can buffer 64 queue descriptors
while pursuing simplicity and easy implementation.
in each SRAM channel. The PAFD algorithm only reserves
16 entrances for the buffer manager, and the rest are for free ACKNOWLEDGEMENTS
idle chained list and SRAM loops. The current count of This work was supported by Presidential Fellowship
packets is stored in the local memory. This operation needs 2007-2009 and Dissertation Year Fellowship 2009-2010,
16 bits, and each bit represents the number of packets Florida International University.
through the 16 entrances. The packet counter is initialed REFERENCES
when entrances are read into the Q-Array, and then it [1] Intel Corporation, “Intel internet exchange architecture software
building blocks developer’s manual [M/ CD],” Document Number:
executes the operation of plus one or minus one base on the 278664 - 010: 279-289, 73-86, 2006.
response. The implemented system we designed supports 64 [2] F. Buccafurri et. al., “Analysis of QoS in cooperative services for real
time applications,” Data & Knowledge Engineering, Vol.67, No.3,
virtual ports, and each port supports 16 queues. Thus, there 2008.
7 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
[3] Yoshihiro Ito, Shuji Tasaka, “Feasibility of QoS control based on QoS Deng Pan received his Ph.D. and M.S.
mapping over IP networks,” Computer Communications, Vol.31,
No.10, 2008. degree in Computer Science from State
[4] Anunay Tiwaria and Anirudha Sahoo, “Providing QoS in OSPF based University of New York at Stony Brook in
best effort network using load sensitive routing,” Simulation 2007 and 2004. He received M.S. and B.S.
Modelling Practice and Theory, Vol.15, No.4, 2007. in Computer Science from Xi'an Jiaotong
[5] Daniel A. Menascéa, Honglei Ruana, and Hassan Gomaa, “QoS University, China, in 2002 and 1999,
management in service-oriented architectures,” Performance
Evaluation, Vol.64, No.7, 2007. respectively. He is currently an Assistant
[6] S. Floyd, V. Jacobson, “Random Early Detection Gateways for Professor in the School of Computing and
Congestion Avoidance”, IEEE/ACM Transactions on Networking Information Sciences, FIU. He was an
(TON), August, 1993.
Assistant Professor in School of Computing and Information Sciences, FIU
[7] Nabeshima, Masayoshi, “Improving the performance of active buffer
from 2007 to 2008. His research interests include high performance routers
management with per-flow information,” IEEE Communications
Letters, Vol.6, No.7, July, 2002. and switches, high speed networking, quality of service, network processors
[8] RFC: Recommendations on Queue Management and Congestion and network security.
Avoidance in the Internet.
[9] W. Feng, Kang G. Shin, D.D. Kandlur, and D. Saha, “The Blue active Zhuo Sun received her BS degree in
queue management algorithms,” IEEE/ACM Transactions on
Networking Vol.10, No.4, pp.513–528, 2002. computer science from Guangxi
[10] C.V. Hollot, V. Misra, D. Towsley, and W. Gong, “Analysis and University, Nanning, China, in 2002, and
design of controllers for AQM routers supporting TCP flows,” IEEE the MS degree in software engineering
Transactions on Automatic Control, Vol.47, No.6, pp.945−959, 2002. from Sun Yat-sen University, Guangzhou,
[11] M. Ghaderi and R. Boutaba, Call admission control for voice/data China, in 2005. Then she worked at Nortel
integration in broadband wireless networks, IEEE Transactions on
Mobile Computing, Vol.5, No.3, 2006. Guangzhou R&D, Guangzhou, China. She
[12] G. Ascia, V. Catania, D. Panno, “An Efficient Buffer Management is currently a second year Ph.D student in
Policy Based On an Integrated Fuzzy-Ga Approach,” Proc. of IEEE Florida International University. Her
INFOCOM 2002, New York, USA, 23-27 Jun. 2002.
research interests are in the areas of
[13] Ellen L. Hahne, Abhijit K. Choudhury, “Dynamic queue length
high-speed network.
thresholds for multiple loss priorities,” IEEE/ACM Transactions on
Networking, Vol.10, No.3, June 2002.
AUTHORS PROFILE
Yechang Fang received his M.S. in
Electrical Engineering from Florida
International University (FIU), Miami,
USA in 2007. From 2006 to 2007, he
served as an IT specialist at IBM China
to work with Nokia, Motorola and
Ericsson. He is currently a Ph.D.
candidate with a Dissertation Year
Fellowship in the Department of
Electrical and Computer Engineering,
FIU. His area of research is
telecommunication. Besides, his research interests also include computer
networking, network processors, fuzzy Logic, rough sets and classification.
Kang K. Yen received the M.S. degree
from University of Virginia in 1979 and
Ph.D. degree from Vanderbilt University
in 1985. He is currently a Professor and
Chair of the Electrical Engineering
Department, FIU. He is also a registered
professional engineer in the State of
Florida. He has been involved in
theoretical works on control theory and on
parallel simulation algorithms development for real-time applications in the
past several years. In the same periods, he has also participated in several
industry supported projects on real-time data processing and
microprocessor-based control system designs. Currently, his research
interests are in the security related issues and performance improvement of
computer networks.
8 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
Multistage Hybrid Arabic/Indian
Numeral OCR System
Yasser M. Alginaih, Ph.D., P.Eng. IEEE Member Abdul Ahad Siddiqi, Ph.D., Member IEEE & PEC
Dept. of Computer Science Dept. of Computer Science
Taibah University Taibah University
Madinah, Kingdom of Saudi Arabia Madinah, Kingdom of Saudi Arabia
yginahi@taibahu.edu.sa asiddiqi@taibah.edu.sa
Abstract— The use of OCR in postal services is not yet numeral OCR systems for Postal services have been used
universal and there are still many countries that process in some countries, but still there are problems in such
mail sorting manually. Automated Arabic/Indian numeral systems, stemming from the fact that machines are unable
Optical Character Recognition (OCR) systems for Postal to read the crucial information needed to distribute the
services are being used in some countries, but still there are mail efficiently. Historically, most civilizations have
errors during the mail sorting process, thus causing a different symbols that convey numerical values, but the
reduction in efficiency. The need to investigate fast and Arabic version is the simplest and most widely
efficient recognition algorithms/systems is important so as to
correctly read the postal codes from mail addresses and to
acceptable. In most Middle Eastern countries both the
eliminate any errors during the mail sorting stage. The Arabic (0,1,2,3,4,5,6,7,8,9) and Indian
objective of this study is to recognize printed numerical (۰,۱,۲,۳,٤,٥,٦,۷,۸,۹) numerals are used. The objective of
postal codes from mail addresses. The proposed system is a this work is to develop a numeral Arabic/Indian OCR
multistage hybrid system which consists of three different system to recognize postal codes from mail letters
feature extraction methods, i.e., binary, zoning, and fuzzy processed in the Middle Eastern countries. A brief history
features, and three different classifiers, i.e., Hamming Nets, on the development of postal services is qouted from [1].
Euclidean Distance, and Fuzzy Neural Network Classifiers. “The broad development of mechanization in postal
The proposed system, systematically compares the operations was not applied until the mid-1950s. The
performance of each of these methods, and ensures that the
numerals are recognized correctly. Comprehensive results
translation from mechanization to automation of the U.S.
provide a very high recognition rate, outperforming the Postal Services (USPS) started in 1982, when the first
other known developed methods in literature. optical character reader was installed in Los Angeles.
The introduction of computers revolutionized the postal
industry, and since then, the pace of change has
Keywords-component; Hamming Net; Euclidean Distance; accelerated dramatically [1].”
Fuzzy Neural Network; Feature Extration; Arabic/Indian
Numerals In the 1980s, the first OCRs were confined to reading
the Zip Code. In the 1990s they expanded their
I. INTRODUCTION capabilities to reading the entire address, and in 1996, the
Optical Character Recognition (OCR) is the electronic Remote Computer Reader (RCR) for the USPS could
translation of images of printed or handwritten text into recognize about 35% of machine printed and 2% of
machine-editable text format; such images are captured handwritten letter mail pieces. Today, modern systems
through a scanner or a digital camera. The research work can recognize 93% of machine-printed and about 88% of
in OCR encompasses many different areas, such as handwritten letter mail. Due to this progress in
pattern recognition, machine vision, artificial intelligence, recognition technology the most important factor in the
and digital image processing. OCR has been used in efficiency of mail sorting equipment is the reduction of
many areas, e.g., postal services, banks, libraries, cost in mail processing. Therefore, a decade intensive
museums to convert historical scripts into digital formats, investment in automated sorting technology, resulted in
automatic text entry, information retrieval, etc. high recognition rates of machine-printed and handwritten
addresses delivered by state-of- the-art systems [1 – 2]
The objective of this work is to develop a numerical OCR
system for postal codes. Automatic Arabic/Indian
9 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
According to the postal addressing standards [3], a been implemented. Many OCR systems are available in
standardized mail address is one that is fully spelled out the market, which are multi font and multilingual.
and abbreviated by using the postal services standard Moreover, most of these systems provide high recognition
abbreviations. The standard requires that the mail rate for printed characters. The recognition rate is
addressed to countries outside of the USA must have the between 95% - 100%, depending on the quality of the
address typed or printed in Roman capital letters and scanned images, fed into the systems, and the application
Arabic numerals. The complete address must include the it is used for [9]. The Kingdom of Saudi Arabia has also
name of addressee, house number with street address or initiated its efforts in deploying the latest technology of
box number/zip code, city, province, and country. automatic mail sorting. It is reported in [10], that Saudi
Examples of postal addresses used in the Middle East are Post has installed an advanced Postal Automation System,
given in table 1. working with a new GEO-data based postal code system,
an Automatic Letter Sorting Machine, and an OCR for
TABLE 1: Examples of postal addresses used in the Middle East simultaneous reading of Arabic and English addresses. It
comprises components for automatic forwarding,
Address with Arabic numerals Address with Indian Numerals
Mr. Ibrahim Mohammad ﺍﻟﺴﻴﺪ ﻣﺤﻤﺪ ﻋﻠﻲ sequencing, and coding
P.O. Box 56577 ٥۲۱۰٦ :ﺹ. ﺏ
RIYADH 11564 ۱۲۳٤٥ :ﺍﻟﺮﻳﺎﺽ In his in-depth research study, Fujisawa, in [11]
SAUDI ARABIA ﺍﻟﻤﻤﻠﻜﺔ ﺍﻟﻌﺮﺑﻴﺔ ﺍﻟﺴﻌﻮﺩﻳﺔ reports on the key technical developments for Kanji
(Chinese character) recognition in Japan. Palumbo and
Standards are being developed to make it easy to Srihari [12] described a Hand Written Address
integrate newer technologies into available components Interpretation (HWAI) system, and reported a throughput
instead of replacing such components, which is very rate of 12 letters per second. An Indian postal automation
costly; such standards are the OCR/Video Coding based on recognition of pin-code and city name, proposed
Systems (VCS) developed by the European Committee by Roy et al in [13] uses Artificial Neural Networks for
for standardization. The OCR/VCS enables postal the classification of English and Bangla postal zip codes.
operators to work with different suppliers on needed In their system they used three classifiers for the
replacements or extensions of sub-systems without recognition. The first classifier deals with 16-class
incurring significant engineering cost [1] [4]. problem (because of shape similarity the number is
reduced from 20) for simultaneous recognition of Bangla
Many research articles are available in the field of and English numerals. The other two classifiers are for
automation of postal systems. Several systems have been recognition of Bangla and English numerals, individually.
developed for address reading, such as in USA [5], UK Ming Su et al. [14], developed an OCR system, where the
[6], Japan [7], Canada [8], etc. But very few countries in goal was to accomplish the automatic mail sorting of
the Middle East use automated mail-processing systems. Chinese postal system by the integration of a mechanized
This is due to the absence of organized mailing address sorting machine, computer vision, and the development of
systems, thus current processing is done in post offices OCR. El-Emami and Usher [15] tried to recognize postal
which are limited and use only P.O. boxes. Canada Post address words, after segmenting these into letters. A
is processing 2.8 billion letter mail pieces annually structural analysis method was used for selecting features
through 61 Multi-line Optical Character Readers of Arabic characters. On the other hand, U.Pal et.al.,
(MLOCRs) in 17 letter sorting Centers. The MLOCR – [16], argues that under three-language formula, the
Year 2000 has an error rate of 1.5% for machine print destination address block of postal document of an Indian
reading only, and the MLOCR/RCR – Year 2003 has an state is generally written in three languages: English,
error rate of 1.7% which is for print/script reading. Most Hindi and the State official language. Because of inter-
of these low read errors are on handwritten addresses and mixing of these scripts in postal address writings, it is
on outgoing foreign mail [9]. very difficult to identify the script by which a pin-code is
written. In their work, they proposed a tri-lingual
The postal automation systems, developed so far, are (English, Hindi and Bangla) 6-digit full pin-code string
capable of distinguishing the city/country names, post and recognition, and obtained 99.01% reliability from their
zip codes on handwritten machine-printed standard style proposed system whereas error and rejection rates were
envelopes. In these systems, the identification of the 0.83% and 15.27%, respectively. In regards to
postal addresses is achieved by implementing an address recognizing the Arabic numerals, Sameh et.al. [17],
recognition strategy that consists of a number of stages, described a technique for the recognition of optical off-
including pre-processing, address block location, address line handwritten Arabic (Indian) numerals using Hidden
segmentation, character recognition, and contextual post Markov Models (HMM). Features that measure the image
processing. The academic research in this area has characteristics at local, intermediate, and large scales
provided many algorithms and techniques, which have were applied. Gradient, structural, and concavity features
10 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
at the sub-regions level are extracted and used as the (0, 1, 2, 3…) and Indian (۰, ۱, ۲, ۳, ٤….) numerals.
features for the Arabic (Indian) numeral. The achieved Therefore, this system can be considered a bi-numeral
average recognition rate reported was 99%. recognition system.
Postal services are going to remain an integral part of The proposed system includes more than one feature
the infrastructure for any economy. For example, recent extraction and classification methods. As a result, the
growth in e-commerce has caused a rise in international hybrid system will help reduce the misclassification of
and domestic postal parcel traffic. To sustain the role of numerals. The system can be used specifically in the
mail as one of most efficient means of business Middle East and countries which use Arabic and Indian
communication, postal services have to permanently numerals in their documents. The proposed design
improve their organizational and technological methodology includes a character recognition system,
infrastructure for mail processing and delivery [4]. which goes through different stages, starting from
Unfortunately, as explained above the character preprocessing, character segmentation, feature extraction
recognition process is not perfect, and errors often occur. and classification. The main building blocks of a general
OCR system are shown in Figure 2 and the design of the
A simplified illustration of how an OCR system is proposed hybrid system is shown in Figure 3.
incorporated into postal services is shown in Figure 1.
This figure, in no way reflects the current technology used
in available mail processing systems. Typically, an OCR
system is developed for the application of postal services
in order to improve the accuracy of mail sorting by
recognizing the scanned Arabic and Indian numerical
postal codes from addresses of mail letters.
Figure 2: A General OCR System
Figure 2, represents the stages a general OCR system
goes through. The process here ignores all the steps
before the OCR step and assumes the availability of the
Figure 1: OCR in Postal Services mail document as a grey-level bitmap graphic file. The
proposed OCR system in Figure 3 show the
The proposed method combines different feature preprocessing, feature extraction, and classification
extraction and classification algorithms to produce a high stages. It also shows stage for comparison to produce the
recognition rate in such application. The proposed hybrid output recognized numeral. After the preprocessing
system is explained in section II of this paper, which stage, features are extracted using the first two feature
explains, the different feature extraction, training and extraction methods, namely feature1 and feature2, then
classification techniques, where as section III of this these two feature vectors are passed through classifiers,
paper presents the results and observations and finally the namely classifier1 and classifier2 respectively. The
concluding remarks are stated in section IV. output from both classifiers is compared, if there is a
match then the output is accepted and no further
II. PROPOSED HYBRID OCR SYSTEM processing is required for this numeral, otherwise the
third feature is calculated, and then passed through
The significance of this research project is in classifier3. The output from classifier3 is then compared
recognizing and extracting the most essential information with both outputs of classifier1 and classifier2. If there is
from addresses of mail letters, i.e., postal zip codes. This a match with the output of classifier3 with either outputs
system will have a profound effect in sorting mail and of classifier1 and classifier2, then the output is accepted,
automating the postal services system, by reading the
postal codes from letter addresses, which contain Arabic
11 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
otherwise the output is rejected and the postal letter needs locating the postal code, the characters are segmented so
to go through either post-processing or manual sorting. that each can be processed individually for proper
In the next subsections of this paper, the preprocessing, recognition. At this point, all numerals were normalized
feature extraction, training and classification techniques to a size of 25 x 20, which was decided experimentally
used in this system are explained in details. according to a 12-font size numeral scanned at a
resolution of 300 dpi. The normalization step aims to
remove the variations of printed styles and obtain
standardized data.
B. Feature Extraction
The proposed hybrid OCR system, Figure 3, is based
on the feature extraction method of character recognition.
Feature extraction can be considered as finding a set of
vectors, which effectively represents the information
content of a character. The features were selected in such
a way to help in discriminating between characters. The
proposed system uses a combination of three different
methods of feature extraction, which are extracted from
each normalized numeral in the postal code, these features
are: the 2D array of the pixel values after the conversion
of the address image into binary, the array of black pixel
distribution values from square-windows after dividing
each normalized character into a 5x5 equal size windows
[19], and finally the maximized fuzzy descriptive
features, [20 – 21], are obtained using equation (1).
N1 N2
S ij = max(max( w[i − x, j − y ] f xy )) − − − − > (1)
x =1 y =1
for i = 1 to N1 , j = 1 to N 2
Figure 3: Proposed Hybrid OCR System S ij gives the maximum fuzzy membership pixel value
using the fuzzy function, w[m, n] , equation (2). Where
A. Preprocessing
f xy is the ( x, y ) binary pixel value of an input pattern
Postal mail images were assumed to be free of noise
with a skew angle not exceeding ± 2 o . The preprocessing (0 ≤ f xy ≤ 1) . N1 and N 2 are the height and width of
tasks performed are: localization of the address, the character window.
conversion from grey scale images to binary images,
localization of the postal code on the image, and character
w[m, n] = exp(− β 2 (m 2 + n 2 )) − − − − > (2)
segmentation. The first step in pre-processing locates the Through exhaustive search, β = 0.3 is found to be the
address to be processed, such as the incoming/outgoing most suitable value to achieve higher recognition rate.
addresses, as long as the address is in the proper standard This maximized membership fuzzy function, equation (2),
format there will not be a problem in specifying its was used in the second layer of the Fuzzy Neural Network
location. Following the localization of the postal code, presented in [20 – 21], which will be used as one of the
thresholding was used to convert the image into binary. If classifiers of the proposed system. S ij gives a 2D fuzzy
the pixel value was above a threshold value then it
becomes white (background) otherwise black feature vector whose values are between 0 and 1, and has
(foreground) [18]. Here, the average of the pixels in the the same size as the normalized image window of the
image was taken to be the threshold value. At this stage, numeral. It is known from the fuzzy feature vector
most of the noise was eliminated using thresholding and method, that the features which resemble the shape of the
only slight distortion to characters was observed, which character will be easily recognized due to this
suggests that pixels were either lost or added to the characteristic of the descriptive fuzzification function.
characters during the thresholding process. Isolated noise Therefore, the feature values closer to the boundary of the
was removed during the character segmentation process. character will result in higher fuzzy membership value.
Then the zip or postal code was located according to the Similarly, further from the boundary of the character will
location according to the postal services standards. After result in lower fuzzy membership value.
12 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
Arial, Times New Roman, Lucida Console, and New
C. Training Courier. Each typeset contained 20 numerals for both
A bitmap image file containing the sets of the numbers Arabic and Indian with 5 different font sizes (12, 14, 16,
from 0 – 9 Arabic and from ۰ – ۹ Indian was used in the 18, 20), providing us with a total of 100 numerals for each
training process to calculate the prototype features. The typeset and 400 numerals for the complete training set.
training sets from which the feature prototypes were Figure 4 shows a full set for one typeset with all the font
calculated contained four different typesets, these are: sizes, note the figure is not to scale.
Figure 4: The training sets used to extract the prototype features. (Figure not to scale)
The prototype numerals were then normalized to a size The prototype feature file for binary features contained 80
of 25 x 20. The three different features explained above feature vectors, each having a vector size of 25x20
were calculated from the normalized characters, and then features. Figure 6 shows an example of a 32-feature
stored in a separate file as prototypes to be compared with vector for a normalized numeral. As explained in the
the features extracted from the images under test. Figure feature extraction section, each of these features
5(a) shows a normalized image for the Arabic numeral (1) represents the black pixel distribution in a window size
and Figure 5(b) shows a normalized image for the Indian 5x5. The features from font size 12 for all sets of Arabic
numeral (٤). From the image above, 1 represents the and Indian numerals were used as the prototype features
foreground and 0 represents the background of the to be passed to the Euclidean Distance Classifier. The file
numeral. Here, only the features for the normalized contained 80 vectors, each containing 32 features. Figure
characters with font size 12 were used as the prototype 6 shows an example of a 32-feature vector for an Arabic
features to be passed to the Hamming Net classifier since numeral.
size 12 is considered as a standard size.
(a) (b)
Figure 5: Normalized Images showing Arabic Numeral 1 and Indian Numeral 4.
Figure 6: Zoning features for a Normalized numerals.
13 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
Figure 7 shows the fuzzy features for the Arabic numeral numeral, the higher the fuzzy feature value and, the
1. The highlighted area resembles the shape of the further from the boundary of the numeral, the lower the
numeral, which shows the fuzzy feature value equals to 1. fuzzy feature values.
It is from Figure 7 that, the closer to the boundary of the
Figure 7: Fuzzy features for the normalized numeral 1 (Arabic) – Size 25 x 20
The prototype features were calculated from the numerals, respectively, providing us with a total of 80
normalized characters. For each font, the prototypes of the prototype feature vectors each containing 25 x 20 features
five font sizes for each numeral in both Arabic and Indian as shown in Figure7. Many Arabic/Indian numeral sets
were averaged by adding them then dividing the sum by for the 4 typesets were scanned at different resolutions
5. This resulted in 20 prototype feature vectors for each and were used during the testing process. Figure 8 shows
typeset, 10 for Arabic numerals and 10 for Indian some examples of some numeral sets used for testing.
Figure 8: The results of testing some complete Arabic and Indian numeral sets
D. Classification d = ( p0 − q0 ) 2 + ( p1 − q1 ) 2 + + ( p N −1 − q N −1 ) 2
A multistage OCR system with three-feature extraction N
and three classification algorithms is employed to
maintain the accuracy in the recognition of the postal
= ∑( p − q )
i =1
i i
2
− − − − − − − − − − − − − − > (3)
codes. The first classifier used is the Euclidean distance
which provides the ordinary distance between two points. Where
To recognize a particular input numeral feature vector, the
system compares this feature vector with the feature p = [p0 p1 p N −1 ]T and
vectors of the database of feature vectors of normalized q = [q 0 q1 q N −1 ]T
numerals using the Euclidean distance nearest-neighbor
classifier [22]. If the feature vector of the input is q and
and N is the size of the vector containing the features.
that of a prototype is p, then the Euclidean distance
Here, the match between the two vectors is obtained by
between the two is defined as:
minimizing d.
The second classifier is the Hamming Net classifier,
[23 – 24] shown in Figure 9 below.
14 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
x is the input vector. 1 − 0.2 − 0.2
k
y1
O is the output of the Maxnet and it is
y is the input to the Maxnet − 0.2 − 0.2 k
y2
o = wM y = = wM net k − − > (8)
c is the encoded class prototype vector
M is the number of classes. − 0.2 − 0.2 1 k
yM
Where
k k
net1 f (net1 )
0 when net j < 0
net =
k
net k and O = f (net k ) f (net j ) =
j j
net j when net j ≥ 0
k
net M f (net M ) k
The third classifier used in this work is the Fuzzy
Neural Network, FNN, developed by Kwan and Cai, [20].
It uses the fuzzy descriptive features explained in the
feature extraction section. Figure 10 shows the structure
of the network which is a four-layer FNN. The first layer
is the input layer; it accepts patterns into the network
which consists of the 2D pixels of the input numeral. The
second layer of the network is a 2D layer of MAX fuzzy
Figure 9: Hamming net with Maxnet as the second layer
neurons whose purpose is to fuzzify the input patterns
through the weighted function w[m, n], equation (2). The
The algorithm designed for the minimum Hamming third layer produces the learned patterns. The fourth
distance classifier which was adopted from [23] is as layer is the output layer which performs defuzzification
follows: and provides non-fuzzy outputs; it chooses the maximum
similarity as the activation threshold to all the fuzzy
Step1: initialize the weight matrix wj and the biases: neurons in the fourth layer (Refer [20 - 21] for details on
the FNN). After passing through the different stages of
c ji
w ji − − − −− > (4) the classifier, the character is identified and the
2 corresponding class is assigned. In the post-processing
n step, recognized postal codes will be compared against
bj = − − − −− > (5)
2 valid postal codes stored in a database. In the
i = 1,2, . . ., n; j = 1, 2, … , M classification phase, feature vectors of an unknown
character are computed and matched with the stored
Step 2: For each input vector x, do steps 3 to 5. prototypes. Matching is done by calculating distance
Step 3: Computer the netj, j = 1, 2, … , M: (dissimilarity) measure between the character and stored
net j = b j + ∑xi w ji
i
− − − −− > (6) prototypes.
The proposed system, shown in Figure 3 suggests that
i = 1,2, . . ., n; j = 1, 2, … , M
if the Hamming Net Classifier and the Euclidean Distance
Classifier did not provide a match then the fuzzy features
are calculated and passed through the FNN classifier. The
Step 4: Initialize the activation yj for the Maxnet, the
FNN classifier is different from a traditional Neural
second
Network because the function of each fuzzy neuron is
layer of the network which represents the
identified and its semantics is defined. The function of
Hamming
such networks is the modeling of inference rules for
similarity.
classification. The outputs of a FNN provide a
y j = net j − − − −− > (7)
measurement of the realization of a rule, i.e. the
Step5: Maxnet compares the outputs of the netj and membership of an expected class. Typically, a FNN is
enforces represented as special four-layer feed-forward neural
the largest one as the best match prototype, while network, in which the first layer corresponds to the input
suppressing the rest to zero. variables, the second layer symbolizes the fuzzy rules, the
Step6: Recurrent processing of the Maxnet: third layer produces the learned patterns and the fourth
15 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
layer represents the output variables. It is trained by III. RESULTS AND OBSERVATIONS
means of a data-driven learning method derived from The authors presented the initial results of this research
neural network theory. Therefore, the result of the FNN study in [25], in which only one font was used and no
classifier is compared to both other classifiers and if there thorough testing of the system was conducted. The
is a match found between the FNN’s result and any of the proposed system can handle small amount of skew in the
previously calculated classifier results the numeral is range of –2 to +2 degrees. The system supports BMP
accepted, otherwise it is rejected, Figure 3. image formats; with image scan resolution of 100 – 300
dpi and above. The documents used were of multiple
fonts with multiple sizes. The fonts used in the system for
testing were: Arial, New Times Norman, Lucida Console
and New courier. Font sizes of 10 – 20, with font styles
normal, and bold were incorporated in the system.
Extensive testing of the proposed OCR system has
been done on approximately 200 mail address images of
different quality printed documents with different
resolutions, font styles and sizes. Figure 11 shows an
example of a processed mail address.
The proposed hybrid system produced successful
results in recognizing Arabic and Indian numerals from
postal letters. The proposed hybrid system provided a
100% recognition rate with no misclassification of
numerals and a rejection rate of less than 1%. When
combining the recognition rate for all images at different
resolutions, the average recognition rate considering the
rejected numerals as misclassified the recognition rate
was 99.41% for all the images which varied in
resolutions, typesets and brightness. This recognition rate
assumed that the rejected characters were misclassified.
This shows the effectiveness of the system in providing
high recognition rate using 4 different fonts and suggests
that more fonts could be applied, if desired.
Figure 10: Four-Layer Feed forward FNN.
Figure 11: A processed envelope containing a postal code written in Indian numerals.
16 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
Future work will use more fonts, and will incorporate TABLE 5: Recognition rate including rejected characters
for the proposed hybrid method
a post-processing step to check the availability of the
postal codes so as to ensure the character recognition of Proposed Hybrid Method
Middle Eastern countries' addresses for mail sorting and Resolution 100% 200% 300% 400%
proper distribution of mail according to postal zip codes Recognition Rate 99.27 99.31 99.33 100
and cities. Tables 2 – 5 show the recognition rates for the
proposed hybrid method and the three methods used IV. CONCLUSION
separately. The proposed hybrid method outperformed
the other three methods, if used separately, as shown in In this work, a hybrid numeral OCR system for
Table 2. The recognition rate calculated in Table 1 did Arabic/Indian postal zip codes was successfully developed
not include any of the rejected numerals. It can also be and thoroughly tested. The system used three different
observed that, the higher the resolution, the better the feature extraction methods and three different classifier
recognition rate. techniques in order to guarantee the accuracy of any
numeral processed through it. Over 200 letter images
TABLE 2: Recognition rate for all methods using images with different were used where the postal code was localized, and then
resolutions recognized through the proposed system. Four different
font styles with sizes ranging from 10 to 20 points were
Resolution 100% 200% 300% 400% used in testing the system and the recognition accuracy
No. of Characters 5460 4340 5690 2710
was 99.41%, when considering the rejected numerals as
Recognition Rate
Hamming 99.08% 99.39% 98.80% 98.89%
un-recognized numerals.
Euclidean Distance 99.36% 99.08% 98.76% 99.88%
Fuzzy Neural
Network 98.13% 99.31 99.43% 100% ACKNOWLEDGMENT
Proposed Hybrid
Method 100% 100% 100% 100% The authors would like to acknowledge the financial
support by the Deanship of Scientific Research at Taibah
Table 3 shows the total number of misclassified University, KSA, under research reference number
numerals at different resolutions. 429/230 academic year 2008/2009 to carry out the
research to design and develop the postal OCR system for
TABLE 3: Number of misclassified characters using images with the recognition of address postal codes in the Middle
different resolutions Eastern countries.
Resolution 100% 200% 300% 400%
REFERENCES
No. of Characters 5460 4340 5690 2710
[1] John Buck, “The postal service revolution- A look at the past and
Hamming 72 70 68 30
Misclassified
where are we headed,” Mailing Systems Technology magazine,
Characters
Euclidean Distance 50 40 70 3
Dec. 2007. <http://www.parascript.com/objects/1107MST.pdf>
Fuzzy Neural [2] Simon Jennings, “Speed Read,” OCR Systems, March 2008, pp.
Network 146 30 32 0 44-50. http://www.parascript.com/objects/0308OCRSystems.pdf>
Proposed Hybrid [3] United States Postal Services, "Postal Addressing Standards".
Method 0 0 0 0 Updated, Viewed on, May 28, 2009,
<http://pe.usps.gov/text/pub28/28c2_001.html>
Table 4 shows the number of rejected characters when [4] John Buck, “International mail-sorting automation in the low-
volume environment,” The Journal of Communication
using the proposed hybrid method. As shown, there were
Distribution, May/June 2009.
no misclassified or rejected numerals with the 400% <http://www.parascript.com/objects/MailMagazine.InternationalM
U
resolution - due to the fact that larger size numerals ailSorting.May2009.pdf> U
provide good quality numerals when normalized. [5] United States Postal Services, http://www.usps.com Viewed 23rd of
U U
June, 2009.
[6] A.C. Downton, R.W.S. Tregidgo and C.G. Leedham, and
Hendrawan, "recognition of handwritten british postal addresses,"
TABLE 4: Number of rejected characters using the proposed hybrid From Pixels to Features 111: Frontiers in Handwriting
method Recognition, pp. 129-143, 1992.
Proposed Hybrid Method [7] Y. Tokunaga, "History and current state of postal mechanization in
Japan", Pattern Recognition Letters, vol. 14, no. 4, pp. 277-280,
Resolutions 100% 200% 300% 400% April 1993
No. of Rejected [8] Canada Post, <www.canadapost.ca>, Viewed 13th of July, 2009.
U U
Characters 40 30 38 0 [9] Canada Post, In Quest of More Advanced Recognition
Technology, 28th of October 2004. Viewed March
Table 5 presents the recognition rate for the 2009,<http://www.hqrd.hitachi.co.jp/iwfhr9/AfterWS/Pict
U
hybrid method when considering the rejected numerals as ures/Panel-presen/Panel-Ulvr.pdf.> U
misclassified. [10] Udo Miletzki, Product Manager Reading Coding, Siemens AG and
Mohammed H. Al Darwish, "Significant technological advances in
17 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
Saudi Post, the first Arabic address reader for delivery point
sorting", Saudi Post Corporation World Mail Review, Nov. 2008
[11] Hiromichi Fujisawa, “A view on the past and future of character
and document recognition,” International Conference on
Document Analysis and Recognition, Brazil, Sept. 23-26, 2009.
[12] P.W. Palumbo and S. N. Srihari, "postal address reading in real
time" Inter. J. of Imaging Science and Technology, 1996
[13] K. Roy, S. Vajda, U. Pal, B. B. Chaudhuri, and A. Belaid, "A
system for Indian postal automation", Proc. 2005 Eight Intl. Conf.
on Document Analysis and Recognition (ICDAR'05) Yasser M. Alginahi, became a member of IEEE
[14] Yih-Ming Su, and Jhing-Fa Wang, "recognition of handwritten in 2000. He earned a Ph.D., in electrical engineering from the University
Chinese postal address using Neural Networks", Proc. of Windsor, Ontario, Canada, a Masters of Science in electrical
International Conference on Image Processing and Character engineering and a Bachelors of Science in biomedical engineering from
Recognition, Kaohsiung, Taiwan, Rep. of China, 1996, pp.213-219 Wright State University, Ohio, U.S.A. Currently, he is an Assistant
[15] El-Emami and M. Usher, "On-line recognition of handwritten Professor, Dept. of Computer Science, College of Computer Science and
Arabic characters," Pattern Analysis and Machine Intelligence, Engineering, Taibah University, Madinah, KSA. His current research
IEEE Transactions on, vol. 12, pp. 704-710, 1990 interests are Document Analysis, Pattern Recognition (OCR), crowd
[16] U. Pal, R. K. Roy, and F. Kimura, "Indian multi-script full pin- management, ergonomics and wireless sensor networks. He is a
code string recognition for postal automation", In the 10th licensed Professional Engineer and a member of Professional Engineers
International Conference on Document Analysis and Recognition, Ontario, Canada (PEO). He has over a dozen of research publications
Barcelona, Spain, 2009, pp.460-465 and technical reports to his credit.
[17] Sameh M. Awaidah & Sabri A. Mahmoud, "A multiple
feature/resolution scheme to Arabic (Indian) numerals recognition
using hidden Markov models", In Signal Processing, Volume 89
,No. 6, June 2009
[18] R. C. Gonzalez and R. E. Woods, Digital Image Processing, 3rd
Edition, Prentice Hall, Upper Saddle River, New Jersey, 2008
[19] Y. M. Alginahi, “Thresholding and character recognition for
security documents with watermarked background”, Conference
on Document Image Computing, Techniques and Applications,
Canberra, Australia, December, 2008
[20] H.K. Kwan and Y.Cai, “A Fuzzy Neural Network and its
Dr. Abdul Ahad Siddiqi received a PhD and a MSc
applications to pattern recognition,” IEEE Trans. On Fuzzy
in Artificial Intelligence in year 1997, and 1992 respectively from
Systems, Vol.2. No.3, August 1994, pp. 185-193.
University of Essex, U.K. He also holds a bachelor degree in Computer
[21] P.M. Patil, U.V. Kulkarni and T.R. Sontakke, “Performance
Systems Engineering from NED University of Engineering and
evaluation of Fuzzy Neural Network with various aggregation
Technology, Pakistan. He is a Member of IEEE, and Pakistan
operators,” Proceedings of the 9th International Conference on
Engineering Council. Presently he is an Associate Professor at College
Neural Information Processing, Vol. 4, Nov 2002, pp. 1744-1748
of Computer Science and Engineering at Taibah University, Madinah,
[22] Duda, R.O. and Hart, P.E. 1973. Pattern classification and scene
KSA. He has worked as Dean of Karachi Institute of Information
analysis. Wiley: New York, NY.
Technology, Pakistan (affiliated with University of Huddersfield, U.K.)
[23] Bow, S-T, Pattern Recognition and Image Processing, 2nd Edition,
between 2003 and 2005. He has over 18 research publications to his
Marcel Dekker, Inc. New York, Basel, 2002.
credit. He has received research grants from various funding agencies,
[24] Artificial Neural Networks Technology, Viewed Jan 2009,
notably from Pakistan Telecom, and Deanship of Research at Taibah
<https://www.dacs.dtic.mil/techs/neural/neural8.php>
University for research in are areas of Intelligent Information Systems,
[25] Yasser M. Alginahi, and Abdul Ahad Siddiqi, "A Proposed Hybrid
Information Technology, and applications of Genetic Algorithms.
OCR System for Arabic and Indian Numerical Postal Codes ", The
2009 International Conference on Computer Technology &
Development (ICCTD), Kota Kinabalu, Malaysia, November,
2009, pp-400-405.
18 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
Attribute Weighting with Adaptive NBTree for
Reducing False Positives in Intrusion Detection
Dewan Md. Farid, and Jerome Darmont Mohammad Zahidur Rahman
ERIC Laboratory, University Lumière Lyon 2 Department of Computer Science and Engineering
Bat L - 5 av. Pierre Mendes, France Jahangirnagar University
69676 BRON Cedex, France Dhaka – 1342, Bangladesh
dewanfarid@gmail.com, jerome.darmont@univ-lyon2.fr rmzahid@juniv.edu
Abstract—In this paper, we introduce new learning algorithms small network, several data mining algorithms, such as decision
for reducing false positives in intrusion detection. It is based on tree, naïve Bayesian classifier, neural network, Support Vector
decision tree-based attribute weighting with adaptive naïve Machines, and fuzzy classification, etc [10]-[20] have been
Bayesian tree, which not only reduce the false positives (FP) at widely used by the IDS community for detecting known and
acceptable level, but also scale up the detection rates (DR) for unknown intrusions. Data mining based intrusion detection
different types of network intrusions. Due to the tremendous algorithms aim to solve the problems of analyzing the huge
growth of network-based services, intrusion detection has volumes of audit data and realizing performance optimization
emerged as an important technique for network security. of detection rules [21]. But there are still some drawbacks in
Recently data mining algorithms are applied on network-based
currently available commercial IDS, such as low detection
traffic data and host-based program behaviors to detect
intrusions or misuse patterns, but there exist some issues in
accuracy, large number of false positives, unbalanced detection
current intrusion detection algorithms such as unbalanced rates for different types of intrusions, long response time, and
detection rates, large numbers of false positives, and redundant redundant input attributes.
attributes that will lead to the complexity of detection model and A conventional intrusion detection database is complex,
degradation of detection accuracy. The purpose of this study is to dynamic, and composed of many different attributes. The
identify important input attributes for building an intrusion problem is that not all attributes in intrusion detection database
detection system (IDS) that is computationally efficient and
may be needed to build efficient and effective IDS. In fact, the
effective. Experimental results performed using the KDD99
use of redundant attributes may interfere with the correct
benchmark network intrusion detection dataset indicate that the
proposed approach can significantly reduce the number and
completion of mining task, because the information they added
percentage of false positives and scale up the balance detection is contained in other attributes. The use of all attributes may
rates for different types of network intrusions. simply increase the overall complexity of detection model,
increase computational time, and decrease the detection
Keywords-attribute weighting; detection rates; false positives; accuracy of the intrusion detection algorithms. It has been
intrusion detection system; naïve Bayesian tree; tested that effective attributes selection improves the detection
rates for different types of network intrusions in intrusion
I. INTRODUCTION detection. In this paper, we present new learning algorithms for
network intrusion detection using decision tree-based attribute
With the popularization of network-based services, weighting with adaptive naïve Bayesian tree. In naïve Bayesian
intrusion detection systems (IDS) have become important tools tree (NBTree) nodes contain and split as regular decision-trees,
for ensuring network security that is the violation of but the leaves contain naïve Bayesian classifier. The proposed
information security policy. IDS collect information from a approach estimates the degree of attribute dependency by
variety of network sources using intrusion detection sensors, constructing decision tree, and considers the depth at which
and analyze the information for signs of intrusions that attempt attributes are tested in the tree. The experimental results show
to compromise the confidentiality and integrity of networks that the proposed approach not only improves the balance
[1]-[3]. Network-based intrusion detection systems (NIDS) detection for different types of network intrusions, but also
monitor and analyze network traffics in the network for significantly reduce the number and percentage of false
detecting intrusions from internal and external intruders [4]-[9]. positives in intrusion detection.
Internal intruders are the inside users in the network with some
authority, but try to gain extra ability to take action without The rest of this paper is organized as follows. In Section II,
legitimate authorization. External intruders are the outside we outline the intrusion detection models, architecture of data
users without any authorized access to the network that they mining based IDS, and related works. In Section III, the basic
attack. IDS notify network security administrator or automated concepts of feature selection and naïve Bayesian tree are
intrusion prevention systems (IPS) about the network attacks, introduced. In Section IV, we introduce the proposed
when an intruder try to break the network. Since the amount of algorithms. In Section V, we apply the proposed algorithms to
audit data that an IDS needs to examine is very large even for a the area of intrusion detection using KDD99 benchmark
19 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
network intrusion detection dataset, and compare the results to the IDS alert the network security administrator or automated
other related algorithms. Finally, Section VI contains the intrusion prevention system (IPS). The generic architectural
conclusions with future works. model of data mining based IDS is shown in Fig 1.
II. INTRUSION DETECTION SYSTEM: IDS
A. Misuse Vs. Anomaly Vs. Hybrid Detection Model
Intrusion detection techniques are broadly classified into
three categories: misuse, anomaly, and hybrid detection model.
Misuse or signature based IDS detect intrusions based on
known intrusions or attacks stored in database. It performs
pattern matching of incoming packets and/or command
sequences to the signatures of known attacks. Known attacks
can be detected reliably with a low false positive using misuse
detection techniques. Also it begins protecting the
computer/network immediately upon installation. But the major
drawback of misuse-based detection is that it requires
frequently signature updates to keep the signature database up-
to-date and cannot detect previously unknown attacks. Misuse
detection system use various techniques including rule-based
expert systems, model-based reasoning systems, state transition
analysis, genetic algorithms, fuzzy logic, and keystroke
monitoring [22]-[25].
Anomaly based IDS detect deviations from normal
behavior. It first creates a normal profile of system, network, or
program activity, and then any activity that deviated from the
normal profile is treated as a possible intrusion. Various data
mining algorithms have been using for anomaly detection
techniques including statistical analysis, sequence analysis,
neural networks, artificial intelligence, machine learning, and
artificial immune system [26]-[33]. Anomaly based IDS have
the ability to detect new or previously unknown attacks, and
insider attacks. But the major drawback of this system is large
Figure 1. Organization of a generalized data mining based IDS
number of false positives. A false positive occurs when an IDS
reports as an intrusion an event that is in fact legitimate
• Audit data collection: IDS collect audit data and
network/system activity.
analyzed them by the data mining algorithms to detect
A hybrid or compound detection system detect intrusions suspicious activities or intrusions. The source of the
by combining both misuse and anomaly detection techniques. data can be host/network activity logs, command-based
Hybrid IDS makes decision using a “hybrid model” that is logs, and application-based logs.
based on both the normal behavior of the system and the
• Audit data storage: IDS store the audit data for future
intrusive behavior of the intruders. Table I shows the
reference. The volume of audit data is extremely large.
comparisons of characteristics of misuse, anomaly, and hybrid
Currently adaptive intrusion detection aims to solve the
detection models.
problems of analyzing the huge volumes of audit data
TABLE I. COMPARISONS OF INTRUSION DETECTION MODELS and realizing performance optimization of detection
Characteristics Misuse Anomaly Hybrid rules.
Detection Accuracy High (for Low High
known attacks) • Processing component: The processing block is the
Detecting New Attacks No Yes Yes heart of IDS. It is the data mining algorithms that apply
False Positives Low Very high High for detecting suspicious activities. Algorithms for the
False Negatives High Low Low analysis and detection of intrusions have been
Timely Notifications Fast Slow Rather Fast traditionally classified into two categories: misuse (or
Update Usage Patterns Frequent Not Frequent Not Frequent signature) detection, and anomaly detection.
B. Architecture of Data Mining Based IDS • Reference data: The reference data stores information
An IDS monitors network traffic in a computer network about known attacks or profiles of normal behaviors.
like a network sniffer and collects network logs. Then the • Processing data: The processing element must
collected network logs are analyzed for rule violations by using frequently store intermediate results such as
data mining algorithms. When any rule violation is detected,
20 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
information about partially fulfilled intrusion into a normal instance, known attack, or new attack. In 2004,
signatures. Amor et al. [43] conducted an experimental study of the
performance comparison between NB classifier and DT on
• Alert: It is the output of IDS that notifies the network KDD99 dataset. This experimental analysis reported that DT
security officer or automated intrusion prevention outperforms in classifying normal, denial of service (DoS), and
system (IPS). R2L attacks, whereas NB classifier is superior in classifying
• System security officer or intrusion prevention system Probe and U2R attacks. With respect to running time, the
(IPS) carries out the prescriptions controlled by the authors pointed out that NB classifier is 7 times faster than DT.
IDS. Another naïve Bayes method for detecting signatures of
specific attacks is motivated by Panda and Patra in 2007 [44].
C. Related Work From the experimental results implemented on KDD99 dataset,
the authors give a conclusion that NB classifier performs back
The concept of intrusion detection began with Anderson’s propagation neural network classifier in terms of detection rates
seminal paper in 1980 [34] by introducing a threat and false positives. It is also reported that NB classifier
classification model that develops a security monitoring produces a relatively high false positive. In a later work, the
surveillance system based on detecting anomalies in user same authors Panda and Patra [45] in 2009, compares NB
behavior. In 1986, Dr. Denning proposed several models for classifier with 5 other similar classifiers, i.e., JRip, Ridor,
commercial IDS development based on statistics, Markov NNge, Decision Table, and Hybrid Decision Table, and
chains, time-series, etc [35], [36]. In 2001, Lindqvist et al. experimental results shows that the NB classifier is better than
proposed a rule-based expert system called eXpert-BSM for other classifiers.
detecting misuse of host machine by analyzing activities inside
the host in forms of audit trails [37], which generates detail
reports and recommendations to the system administrators, and III. FEATURE SELECTION AND ADAPTIVE NB TREE
produces low false positives. Rules are conditional statements
that derived by employing domain expert knowledge. In 2005, A. Feature Selection
Fan et al. proposed a method to generate artificial anomalies Feature selection becomes indispensable for high
into training dataset of IDS to handle both misuse and anomaly performance intrusion detection using data mining algorithms,
detection [38]. This method injects artificial anomaly data into because irrelevant and redundant features may lead to complex
the training data to help a baseline classifier distinguish intrusion detection model as well as poor detection accuracy.
between normal and anomalous data. In 2006, Bouzida et al. Feature selection is the process of finding a subset of features
[39] introduced a supplementary condition to the baseline from total original features. The purpose of feature selection is
decision tree (DT) for anomaly intrusion detection. The idea is to remove the irrelevant input features from the dataset for
that instead of assigning a default class (normally based on improving the classification accuracy. Feature selection in
probability distribution) to the test instance that is not covered particularly useful in the application domains that introduce a
by the tree, the instance is assigned to a new class. Then, large number of input dimensions like intrusion detection.
instances with the new class are examined for unknown attack Many data mining methods have been used for selecting
analysis. In 2009, Wu and Yen [21] applied DT and support important features from training dataset such as information
vector machine (SVM) algorithm to built two classifiers for gain based, gain ratio based, principal component analysis
comparison by employing a sampling method of several (PCA), genetic search, and classifier ensemble methods etc
different normal data ratios. More specifically, KDD99 dataset [46]-[53]. In 2009, Yang et al. [54] introduced a wrapper-based
is split into several different proportions based on the normal feature selection algorithm to find most important features from
class label for both training set and testing set. The overall the training dataset by using random mutation hill climbing
evaluation of a classifier is based on the average value of method, and then employs linear support vector machine
results. It is reported that in general DT is superior to SVM (SVM) to evaluate the selected subset-features. Chen et al. [55]
classifier. In the same way, Peddabachigari et al. [40] applied proposed a neural-tree based algorithm to identify important
DT and SVM for intrusion detection, and proven that decision input features for classification, based on an evolutionary
tree is better than SVM in terms of overall accuracy. algorithm that the feature contributes more to the objective
Particularly, DT much better in detecting user to root (U2R) function will consider as an important feature.
and remote to local (R2L) network attacks, compared to SVM.
In this paper, to select the important input attributes from
Naïve Bayesian (NB) classifier produces a surprising result training dataset, we construct a decision tree by applying ID3
of classification accuracy in comparison with other classifiers algorithm in training dataset. The ID3 algorithm constructs
on KDD99 benchmark intrusion detection dataset. In 2001, decision tree using information theory [56], which choose
Barbara et al. [41] proposed a method based on the technique splitting attributes from the training dataset with maximum
called Pseudo-Bayes estimators to enhance the ability of information gain. Information gain is the amount of
ADAM intrusion detection system [42] in detecting new information associated with an attribute value that is related to
attacks and reducing false positives, which estimates the prior the probability of occurrence. Entropy is the quantify
and posterior probabilities for new attacks by using information information that is used to measure the amount of randomness
derived from normal instances and known attacks without from a dataset. When all data in a set belong to a single class,
requiring prior knowledge about new attacks. This study there is no uncertainty then the entropy is zero. The objective
constructs a naïve Bayes Classifier to classify a given instance of ID3 algorithm is to iteratively partition the given dataset into
21 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
sub-datasets, where all the instances in each final subset belong Adaptive naïve Bayesian tree splits the dataset by applying
to the same class. The value for entropy is between 0 and 1 and entropy based algorithm and then used standard naïve Bayesian
reaches a maximum when the probabilities are all the same. classifiers at the leaf node to handle attributes. It applies
Given probabilities p1, p2,..,ps, where ∑i=1 pi=1; strategy to construct decision tree and replaces leaf node with
s
naïve Bayesian classifier.
Entropy: H(p1,p2,…ps) = ∑ (pi log(1/pi)) (1)
i =1 IV. PROPOSED LEARNING ALGORITHM
Given a dataset, D, H(D) finds the amount of sub-datasets of
original dataset. When that sub-dataset is split into s new sub- A. Proposed Attribute Weighting Algorithm
datasets S = {D1, D2,…,Ds}, we can again look at the entropy of In a given training data, D = {A1, A2,…,An} of attributes,
those sub-datasets. A subset is completely ordered if all where each attribute Ai = {Ai1, Ai2,…,Aik} contains attribute
instances in it are the same class. The ID3 algorithm calculates values and a set of classes C = {C1, C2,…,Cn}, where each
the gain by the equation “(2)”. class Cj = {Cj1, Cj2,…,Cjk} has some values. Each example in
the training data contains weight, w = {w1,w2…, wn}. Initially,
s
Gain (D,S) = H(D)- ∑ p(Di)H(Di) (2) all the weights of examples in training data have equal unit
i =1 value that set to wi = 1/n. Where n is the total number of
training examples. Estimates the prior probability P(Cj) for
After constructing the decision tree from training dataset, each class by summing the weights that how often each class
we weight the attributes of training dataset by the minimum occurs in the training data. For each attribute, Ai, the number
depth at which the attribute is tested in the decision tree. The of occurrences of each attribute value Aij can be counted by
depth of root node of the decision tree is 1. The weight for an summing the weights to determine P(Aij). Similarly, the
attribute is set to1 d , where d is the minimum depth at which conditional probability P(Aij |Cj) can be estimated by summing
the attribute is tested in the tree. The weights of attributes that the weights that how often each attribute value occurs in the
do not appear in the decision tree are assigned to zero. class Cj in the training data. The conditional probabilities P(Aij
|Cj) are estimated for all values of attributes. The algorithm
B. Naïve Bayesian Tree then uses the prior and conditional probabilities to update the
weights. This is done by multiplying the probabilities of the
Naïve Bayesian tree (NBTree) is a hybrid learning
different attribute values from the examples. Suppose the
approach of decision tree and naïve Bayesian classifier. In
training example ei has independent attribute values {Ai1,
NBTree nodes contain and split as regular decision-trees, but
Ai2,…,Aip}. We already know the prior probabilities P(Cj) and
the leaves are replaced by naïve Bayesian classifier, the
conditional probabilities P(Aik|Cj), for each class Cj and
advantage of both decision tree and naïve Bayes can be utilized
attribute Aik. We then estimate P(ei |Cj) by
simultaneously [57]. Depending on the precise nature of the
probability model, NB classifier can be trained very efficiently P(ei | Cj) = P(Cj) ∏ P(Aij | Cj) (5)
in a supervised learning. In many practical applications,
parameter estimation for naïve Bayesian models uses the To update the weight of training example ei, we can
method of maximum likelihood. Suppose the training dataset, estimate the likelihood of ei for each class. The probability that
D consists of predictive attributes {A1, A2,…,An}, where each ei is in a class is the product of the conditional probabilities for
attribute Ai = {Ai1, Ai2,…,Aik} contains attribute values and a set each attribute value. The posterior probability P(Cj | ei) is then
of classes C = {C1, C2,…,Cn}. The objective is to classify an found for each class. Then the weight of the example is
unseen example whose class value is unknown but values for updated with the highest posterior probability for that example
attributes A1 through Ak are known. The aim of decision tree and also the class value is updated according to the highest
learning is to construct a tree model: {A1, A2,…,An}→C. posterior probability. Now, the algorithm calculates the
Correspondingly the Bayes theorem, if attribute Ai is discrete information gain by using updated weights and builds a tree.
or continuous, we will have: After the tree construction, the algorithm initialized weights
for each attributes in training data D. If the attribute in the
P(Aij | C j )P(C j ) training data is not tested in the tree then the weight of the
P(Cj | Aij) = (3) attribute is initialized to 0, else calculates the minimum depth,
P(Aij ) d that the attribute is tested at and initialized the weight of
attribute to 1 d . Finally, the algorithm removes all the
Where P(Cj|Aij) denote the probability. The aim of
Bayesian classification is to decide and choose the class that attributes with zero weight from the training data D. The main
maximizes the posteriori probability. Since P(Aij) is a constant procedure of proposed algorithm is described as follows.
independent of C, then: Algorithm 1: Attribute Weighting
*
(
C = arg max P C j | Aij ) Input: Training Dataset, D
Output: Decision tree, T
cεC
Procedure:
(
= arg max P Aij | C j P C j )( ) (4) 1. Initialize all the weights for each example in D,
cεC
wi=1/n, where n is the total number of the examples.
22 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
2. Calculate the prior probabilities P(Cj) for each class create a NB classifier for the current node. Partition the
∑wi
Cj in D. P(Cj) = Ci
training data D according to the test on attribute Ai. If Ai is
continuous, a threshold split is used; if Ai is discrete, a multi-
n
way split is made for all possible values. For each child, call
∑w
i=1
i
the algorithm recursively on the portion of D that matches the
3. Calculate the conditional probabilities P(Aij | Cj) for test leading to the child. The main procedure of algorithm is
described as follows.
each attribute values in D. P(A | C ) = P(Aij )
ij j
∑w i Algorithm 2: Adaptive NBTree
Ci Input: Training dataset D of labeled examples.
4. Calculate the posterior probabilities for each example Output: A hybrid decision tree with naïve Bayesian
in D. classifier at the leaves.
P(ei | Cj) = P(Cj) ∏ P(Aij | Cj) Procedure:
5. Update the weights of examples in D with Maximum 1. Calculate the prior probabilities P(Cj) for each class
Likelihood (ML) of posterior probability P(Cj|ei);
wi= PML(Cj|ei) Cj in D. P(Cj) = Ci
∑wi
6. Change the class value of examples associated with n
maximum posterior probability, Cj = Ci→ PML(Cj|ei). ∑w
i=1
i
7. Find the splitting attribute with highest information
gain using the updated weights, wi in D. 2. Calculate the conditional probabilities P(Aij | Cj) for
Information Gain = each attribute values in D. P(A | C ) = P(Aij )
ij j
k ∑ wi
n ∑ wi
∑w
∑ wi
i
− i=Cij log w
Ci
− i =Ci log i =Ci
∑ n ∑ ∑ wi
∑ i 3. Classify each example in D with maximum posterior
n i =C m
j =1
∑
i =1
wi ∑ wi
i =1
i=Ci
i =1 ij
probability. P(ei | Cj) = P(Cj )∏P(Aij | Cj )Wi
i=1
8. T = Create the root node and label with splitting 4. If any example in D is misclassified, then for each
attribute. attribute Ai, evaluate the utility, u(Ai), of a spilt on
9. For each branch of the T, D = database created by attribute Ai.
applying splitting predicate to D, and continue steps 1 5. Let j = argmaxi(ui), i.e., the attribute with the highest
to 8 until each final subset belong to the same class or utility.
leaf node created. 6. If uj is not significantly better than the utility of the
10. When the decision tree construction is completed, for current node, create a naïve Bayesian classifier for
each attribute in the training data D: If the attribute is the current node and return.
not tested in the tree then weight of the attribute is 7. Partition the training data D according to the test on
initialized to 0. Else, let d be the minimum depth that attribute Ai. If Ai is continuous, a threshold split is
the attribute is tested in the tree, and weight of the used; if Ai is discrete, a multi-way split is made for all
attribute is initialized to1 d . possible values.
11. Remove all the attributes with zero weight from the 8. For each child, call the algorithm recursively on the
training data D. portion of D that matches the test leading to the child.
B. Proposed Adaptive NBTree Algorithm V. EXPERIMENTAL RESULTS AND ANALYSIS
Given training data, D where each attribute Ai and each
example ei have the weight value. Estimates the prior A. Dataset
probability P(Cj) and conditional probability P(Aij | Cj) from Experiments have been carried out on KDD99 cup
the given training dataset using weights of the examples. Then benchmark network intrusion detection dataset, a predictive
classify all the examples in the training dataset using these model capable of distinguishing between intrusions and normal
prior and conditional probabilities with incorporating attribute connections [58]. In 1998, DARPA intrusion detection
weights into the naïve Bayesian formula: evaluation program, a simulated environment was set up to
m acquire raw TCP/IP dump data for a local-area network (LAN)
P(ei | Cj) = P(Cj )∏P(Aij | Cj )Wi (6) by the MIT Lincoln Lab to compare the performance of various
i=1 intrusion detection methods. It was operated like a real
Where Wi is the weight of attribute Ai. If any example of environment, but being blasted with multiple intrusion attacks
training dataset is misclassified, then for each attribute Ai, and received much attention in the research community of
evaluate the utility, u(Ai), of a spilt on attribute Ai. Let j = adaptive intrusion detection. The KDD99 dataset contest uses a
argmaxi(ui), i.e., the attribute with the highest utility. If uj is version of DARPA98 dataset. In KDD99 dataset each example
represents attribute values of a class in the network data flow,
not significantly better than the utility of the current node,
23 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
and each class is labeled either normal or attack. Examples in C. Experiment and analysis on Proposed Algorithm
KDD99 dataset are represented with a 41 attributes and also Firstly, we use proposed algorithm 1 to perform attribute
labeled as belonging to one of five classes as follows: (1)
selection from training dataset of KDD99 dataset and then we
Normal traffic; (2) DoS (denial of service); (3) Probe,
surveillance and probing; (4) R2L, unauthorized access from a use our proposed algorithm 2 for classifier construction. The
remote machine; (5) U2R, unauthorized access to local super performance of our proposed algorithm on 12 attributes in
user privileges by a local unprivileged user. In KDD99 dataset KDD99 dataset is listed in Table IV.
these four attack classes are divided into 22 different attack TABLE IV. PERFORMANCE OF PROPOSED ALGORITHM ON KDD99 DATASET
classes that tabulated in Table II.
Classes Detection Rates (%) False Positives (%)
TABLE II. ATTACKS IN KDD99 DATASET Normal 100 0.04
4 Main Attack Classes 22 Attack Classes Probe 99.93 0.37
Denial of Service (DoS) back, land, neptune, pod, smurt, teardrop DoS 100 0.03
ftp_write, guess_passwd, imap, multihop, phf, U2R 99,38 0.11
Remote to User (R2L) R2L 99.53 6.75
spy, warezclient, warezmaster
User to Root (U2R) buffer_overflow, perl, loadmodule, rootkit
Table V and Table VI depict the performance of naïve
Probing ipsweep, nmap, portsweep, satan
Bayesian (NB) classifier and C4.5 algorithm using the original
The input attributes in KDD99 dataset are either discrete or 41 attributes of KDD99 dataset.
continuous values and divided into three groups. The first TABLE V. PERFORMANCE OF NB CLASSIFIER ON KDD99 DATASET
group of attributes is the basic features of network connection,
which include the duration, prototype, service, number of bytes Classes Detection Rates (%) False Positives (%)
from source IP addresses or from destination IP addresses, and Normal 99.27 0.08
Probe 99.11 0.45
some flags in TCP connections. The second group of attributes DoS 99.68 0.05
in KDD99 is composed of the content features of network U2R 64.00 0.14
connections and the third group is composed of the statistical R2L 99.11 8.12
features that are computed either by a time window or a
TABLE VI. PERFORMANCE OF C4.5 ALGORITHM USING KDD99 DATASET
window of certain kind of connections. Table III shows the
number of examples of 10% training data and 10% testing data Classes Detection Rates (%) False Positives (%)
in KDD99 dataset. There are some new attack examples in Normal 98.73 0.10
testing data, which is no present in the training data. Probe 97.85 0.55
DoS 97.51 0.07
TABLE III. NUMBER OF EXAMPLES IN TRAINING AND TESTING KDD99 U2R 49.21 0.14
DATA R2L 91.65 11.03
Attack Types Training Examples Testing Examples Table VII and Table VIII depict the performance of NB
Normal 97277 60592
Denial of Service 391458 237594
classifier and C4.5 using reduces 12 attributes.
Remote to User 1126 8606 TABLE VII. PERFORMANCE OF NB CLASSIFIER USING KDD99 DATASET
User to Root 52 70
Probing 4107 4166 Classes Detection Rates (%) False Positives (%)
Total Examples 494020 311028 Normal 99.65 0.06
Probe 99.35 0.49
B. Performance Measures DoS 99.71 0.04
U2R 64.84 0.12
In order to evaluate the performance of proposed learning R2L 99.15 7.85
algorithm, we performed 5-class classification using KDD99
network intrusion detection benchmark dataset and consider TABLE VIII. PERFORMANCE OF C4.5 ALGORITHM USING KDD99 DATASET
two major indicators of performance: detection rate (DR) and Classes Detection Rates (%) False Positives (%)
false positives (FP). DR is defined as the number of intrusion Normal 98.81 0.08
instances detected by the system divided by the total number of Probe 98.22 0.51
intrusion instances present in the dataset. DoS 97.63 0.05
U2R 56.11 0.12
DR = Total _ det ected _ attacks * 100 (7) R2L 91.79 8.34
Total _ attacks We also compare the intrusion detection performance
FP is defined as the total number of normal instances. among Support Vector Machines (SVM), Neural Network
(NN), Genetic Algorithm (GA), and proposed algorithm on
FP = Total _ misclassif ied _ process * 100 (8) KDD99 dataset that tabulated in Table IX [59], [60].
Total _ normal _ process TABLE IX. COMPARISON OF SEVERAL ALGORITHMS
All experiments were performed using an Intel Core 2 Duo SVM NN GA Proposed Algorithm
Processor 2.0 GHz processor (2 MB Cache, 800 MHz FSB) Normal 99.4 99.6 99.3 99.93
with 1 GB of RAM. Probe 89.2 92.7 98.46 99.84
DoS 94.7 97.5 99.57 99.91
U2R 71.4 48 99.22 99.47
R2L 87.2 98 98.54 99.63
24 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
VI. CONCLUSIONS AND FUTURE WORKS model,” Computer Physics Communications, Vol. 180, Issue 10,
October 2009, pp. 1795-1801.
This paper presents a hybrid approach to intrusion detection [13] Chih-Forn, and Chia-Ying Lin, “A triangle area based nearset neighbors
based on decision tree-based attribute weighting with naïve approach to intrusion detection,” Pattern Recognition, Vol. 43, Issuse 1,
Bayesian tree, which is suitable for analyzing large number of January 2010, pp. 222-229.
network logs. The main propose of this paper is to improve the [14] Kamran Shafi, and Hussein A. Abbass, “An adaptive genetic-based
performance of naïve Bayesian classifier for network intrusion signature learning system for intrusion detection,” Expert System with
detection systems (NIDS). The experimental results manifest Applications, Vol. 36, Issue 10, December 2009, pp. 12036-12043.
that proposed approach can achieve high accuracy in both [15] Zorana Bankovic, Dusan Stepanovic, Slobodan Bojanic, and Octavio
NietopTalasriz, “Improving network security using genetic algorithm
detection rates and false positives, as well as balanced detection approach,” Computers & Electrical Engineering, Vol. 33. Issues 5-6,
performance on all four types of network intrusions in KDD99 2007, pp. 438-541.
dataset. The future works focus on applying the domain [16] Yang Li, and Li guo, “An active learning based TCM-KNN algorithm
knowledge of security to improve the detection rates for current for supervised network intruison detection,” Computers & security, Vol.
attacks in real time computer network, and ensemble with other 26, Issues 7-8, December 2007, pp. 459-467.
mining algorithms for improving the detection rates in intrusion [17] Wun-Hwa Chen, Sheng-Hsun Hsu, and Hwang-Pin Shen, “Application
detection. of SVM and ANN for intrusion detection,” Computers & Operations
Research, Vol. 32, Issue 10, October 2005, pp. 2617-1634.
[18] Ming-Yang Su, Gwo-Jong Yu, and Chun-Yuen Lin, “A real-time
ACKNOWLEDGMENT network intrusion detection system for large-scale attacks based on an
incremental mining approach,” Computer & Security, Vol. 28, Issue 5,
Support for this research received from ERIC Laboratory, July 2009, pp. 301-309.
University Lumière Lyon 2 – France, and Department of
[19] Zeng Jinquan, Liu Xiaojie, Li Tao, Liu Caiming, Peng Lingxi, and Sun
Computer Science and Engineering, Jahangirnagar University, Feixian, “A self-adaptive negative selection algorithm used for anomaly
Bangladesh. detection,” Progress in Natural Science, Vol. 19, Issue 2, 10 February
2009, pp. 261-266.
REFERENCES [20] Zonghua Zhang, and Hong Shen, “Application of online-training SVMs
for real-time intrusion detection with different considerations,”
Computer Communications, Vol. 28, Issue 12, 18 July 2005, pp. 1428-
[1] Xuan Dau Hoang, Jiankun Hu, and Peter Bertok, “A program-based 1442.
anomaly intrusion detection scheme using multiple detection engines [21] Su-Yun Wu, and Ester Yen, “Data mining-based intrusion detectors,”
and fuzzy inference,” Journal of Network and Computer Applications, Expert Systems with Applications, Vol. 36, Issue 3, Part 1, April 2009,
Vol. 32, Issue 6, November 2009, pp. 1219-1228. pp. 5605-5612.
[2] P. Garcia-Teodoro, J. Diaz-Verdejo, G. Macia-Fernandez, and E. [22] S. R. Snapp, and S. E. Smaha, “Signature analysis model definition and
Vazquez, “Anomaly-based network intrusion detection: Techniques, formalism,” In Proc. of the 4th Workshop on Computer Security Incident
systems and challenges,” Computers & Security, Vol. 28, 2009, pp. 18- Handling, Denver, CO. 1992.
28. [23] P. A. Poras, and A. Valdes, “Live traffic analysis of TCP/IP gateways,”
[3] Animesh Patch, and Jung-Min Park, “An overview of anomaly detection In Proc. of the Network and Distributed System Security Symposium,
techniques: Existing solutions and latest technological trends,” San Diego, CA: Internet Society, 11-13 March, 1998.
Computer Netwroks, Vol. 51, Issue 12, 22 August 2007, pp. 3448-3470. [24] T. D. Garvey, and T. F. Lunt, “Model based intrusion detection,” In
[4] Lih-Chyau Wuu, Chi-Hsiang Hung, and Sout-Fong Chen, “Building Proc. of the 14th National Conference Security Conference, 1991, pp.
intrusion pattern miner for Snort network intrusion detection system,” 372-385.
Journal of Systems and Software, Vol. 80, Issue 10, October 2007, pp. [25] F. Carrettoni, S. Castano, G. Martella, and P. Samarati, “RETISS: A real
1699-1715. time security system for threat detection using fuzzy logic,” In Proc. of
[5] Chia-Mei Chen, Ya-Lin Chen, and Hsiao-Chung Lin, “An efficient the 25th IEEE International Carnahan Conference on Security
network intrusion detection,” Computer Communications, Vol. 33, Issue Technology, Taipei, Taiwai ROC, 1991.
4, 1 March 2010, pp. 477-484. [26] T. F. Lunt, A. Tamaru, F. Gilham, R. Jagannathan, P. G. Neumann, H. S.
[6] M. Ali Aydin, A. Halim Zaim, and K. Gokhan Ceylan, “A hybrid Javitz, A. Valdes, and T. D. Garvey, “A real-time intrusion detection
intrusion detection system for computer netwrok security,” Computer & expert system (IDES),” Technical Report, Computer Science
Electrical Engineering, Vol. 35, Issue 3, May 2009, pp. 517-526. Laboratory, Menlo Park, CA: SRI International.
[7] Franciszek Seredynski, and Pascal Bouvry, “Anomaly detection in [27] S. A. Hofmeyr, S. Forrest, A. Somayaji, “Intrusion detection using
TCP/IP networks using immune systems paradigm,” Computer sequences of system calls,” Journal of Computer Security, Vol. 6, 1998,
Communications, Vol. 30, Issue 4, 26 February 2007, pp. 740-749. pp. 151-180.
[8] Jr, James C. Foster, Matt Jonkman, Raffael Marty, and Eric Seagren, [28] S. A. Hofmeyr, and S. Forrest, “Immunity by design: An artificial
“Intrusion detection systems,” Snort Intrusion detection and Prevention immune system,” In Proc. of the Genetic and Evolutionary Computation
Toolkit, 2006, pp. 1-30. Conference (GECCO 99), Vol. 2, San Mateo, CA: Morgan Kaufmann,
[9] Ben Rexworthy, “Intrusion detections systems – an outmoded network 1999, pp. 1289-1296.
protection model,” Network Security, Vol. 2009, Issus 6, June 2009, pp. [29] J. M. Jr. Bonifacio, A. M. Cansian, A. C. P. L. F. Carvalho, and E. S.
17-19. Moreira, “Neural networks applied in intrusion detection systems,” In
[10] Wei Wang, Xiaohong Guan, and Xiangliang Zhang, “Processing of the Proc. of the International Conference on Computational Intelligence
massive audit data streams for real-time anomaly intrusion detection,” and Multimedia Application, Gold Coast, Australia, 1997, pp. 276-280.
Computer Communications, Vol. 31, Issue 1, 15 January 2008, pp. 58- [30] H. Debar, M. Becker, and D. Siboni, “A neural network component for
72. an intrusion detection system,” In Proc. of the IEEE Symposium on
[11] Han-Ching Wu, and Shou-Hsuan Stephen Huand, “Neural network- Research in Security and Privacy, Oakland, CA: IEEE Computer Society
based detection of stepping-stone intrusion,” Expert Systems with Press, 1992, pp. 240-250.
Applications, Vol. 37, Issuse 2, March 2010, pp. 1431-1437. [31] W. Lee, S. J. Stolfo, and P. K. Chan, “Learning patterns from Unix
[12] Xiaojun Tong, Zhu Wang, and Haining Yu, “A research using hybrid precess execution traces for intrusion detection,” AAAI Workshop: AI
RBF/Elman neural netwroks for intrusion detection system secure
25 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
Approaches to Fraud Detection and Risk Management, Menlo Park, CA: Recognition (ICPR 2002), Quebec: IEEE Computer Society, 2002, pp.
AAAI Press, 1999, pp. 50-56. 568-571.
[32] W. Lee, S. J. Stolfo, and K. W. Mok, “Mining audit data to built [52] S. Mukkamala, and A.H. Sung, “Identifying key features for intrusion
intrusion detection models,” In Proc. of the 4th International Conference detection using neural networks,” In Proc. of the ICCC International
on Knowledge Discovery and Data Mining (KDD-98), Menlo Park, CA: Conference on Computer Communications, 2002.
AAAI Press, 2000, pp. 66-72. [53] W. Lee, and S. J. Stolfo, “A framework for constructing features and
[33] S. Forrest, S. A. Hofmeyr, A. Somayaji, and T. A. Longstaff, “A sence models for intrusion detection systems,” ACM Transactions on
of self for Unix Precesses,” In Proc. of the 1996 IEEE Symposium on Information and System Security, 3(4), 2000, pp. 227-261.
Security and Privacy, Oakland, CA: IEEE Computer Society Press, [54] Y. Li, J.L. Wang, Z.H. Tian, T.B. Lu, and C. Young, “Building
1996, pp. 120-128. lightweight intrusion detection system using wrapper-based feature
[34] James P. Anderson, “Computer security threat monitoring and selection mechanisms,” Computer & Security, Vol. 28, Issue 6,
surveillance,” Technical Report 98-17, James P. Anderson Co., Fort September 2009, pp. 466-475.
Washington, Pennsylvania, USA, April 1980. [55] Y. Chen, A. Abraham, and B. Yang, “Hybrid flexible neural-tree-based
[35] Dorothy E. Denning, “An intrusion detection model,” IEEE Transaction intrusion detection systems,” International Journal of Intelligent
on Software Engineering, SE-13(2), 1987, pp. 222-232. Systems, 22, pp. 337-352.
[36] Dorothy E. Denning, and P.G. Neumann “Requirement and model for [56] J. R. Quinlan, “Induction of Decision Tree,” Machine Learning Vol. 1,
IDES- A real-time intrusion detection system,” Computer Science 1986, pp. 81-106.
Laboratory, SRI International, Menlo Park, CA 94025-3493, Technical [57] R. Kohavi, “Scaling up the accuracy of naïve Bayes classifiers: A
Report # 83F83-01-00, 1985. Decision Tree Hybrid,” In Proc. of the 2nd International Conference on
[37] U. Lindqvist, and P. A. Porras, “eXpert-BSM: A host based intrusion Knowledge Discovery and Data Mining, Menlo Park, CA:AAAI
detection solution for Sun Solaris,” In Proc. of the 17th Annual Computer Press/MIT Press, 1996, pp. 202-207.
Security Applications Conference, New Orleans, USA, 2001, pp. 240- [58] The KDD Archive. KDD99 cup dataset, 1999.
251. http://kdd.ics.uci.edu/databases/kddcup99/kddcup99.html
[38] W. Fan, W. Lee, M. Miller, S. J. Stolfo, and P. K. Chan, “Using artificial [59] Mukkamala S, Sung AH, and Abraham A, “Intrusion dection using an
anomalies to detect unknown and known netwrok intrusions,” ensemble of intelligent paradigms,” Proceedings of Journal of Network
Knowledge and Information Systems, 2005, pp. 507-527. and Computer Applications, 2005, 2(8): pp. 167-182.
[39] Y. Bouzida, and F. Cuppens, “Detecting known and novel network [60] Chebrolu S, Abraham A, and Thomas JP, “Feature deduction and
intrusions,” Security and Privacy in Dynamic Environments, 2006, pp. ensemble design of intrusion detection systems.” Computer & Security,
258-270. 2004, 24(4), pp. 295-307.
[40] S. Peddabachigari, A. Abraham, and J. Thomas, “Intrusion detection
systems using decision tress and support vector machines,” International
Journal of Applied Science and Computations, 2004. AUTHORS PROFILE
[41] D. Barbara, N. Wu, and Suchil Jajodia, “Detecting novel network
intrusions using Bayes estimators,” In Proc. of the 1st SIAM Conference Dewan Md. Farid was born in Dhaka, Bangladesh in 1979. He is currently a
on Data Mining, April 2001. research fellow at ERIC Laboratory, University Lumière Lyon 2 - France. He
obtained B.Sc. Engineering in Computer Science and Engineering from Asian
[42] D. Barbara, J. Couto, S. Jajodia, and N. Wu, “ADAM: A tested for
University of Bangladesh in 2003 and Master of Science in Computer Science
exploring the use of data mining in intrusion detection,” Special Interest
Group on Management of Data (SIGMOD), Vol. 30 (4), 2001. and Engineering from United International University, Bangladesh in 2004.
He is pursuing Ph.D. in the Department of Computer Science and
[43] N. B. Amor, S. Benferhat, and Z. Elouedi, “Naïve Bayes vs. decision Engineering, Jahangirnagar University, Bangladesh. He is a faculty member in
trees in intruison detection systems,” In Proc. of the 2004 ACM the Department of Computer Science and Engineering, United International
Symposium on Applied Computing, New York, 2004, pp. 420-424. University, Bangladesh. He is a member of IEEE and IEEE Computer
[44] M. Panda, and M. R. Patra, “Network intrusion deteciton using naïve Society. He has published 10 international research papers including two
Bayes,” International Journal of Computer Science and Network journals in the field of data mining, machine learning, and intrusion detection.
Security (IJCSNS), Vol. 7, No. 12, December 2007, pp. 258-263.
[45] M. Panda, and M. R. Patra, “Semi-naïve Bayesian method for network Jérôme Darmont received his Ph.D. in computer science from the University
intrusion detection system,” In Proc. of the 16th International Conference of Clermont-Ferrand II, France in 1999. He joined the University of Lyon 2,
on Neural Information Processing, December 2009. France in 1999 as an associate professor, and became full professor in 2008.
[46] P.V.W. Radtke, R. Sabourin, and T. Wong, “Intelligent feature He was head of the Decision Support Databases research group within the
extraction for ensemble of classifiers,” In Proc. of 8th International ERIC laboratory from 2000 to 2008, and has been director of the Computer
Conference on Document Analysis and Recognition (ICDAR 2005), Science and Statistics Department of the School of Economics and
Seoul: IEEE Computer Society, 2005, pp. 866-870. Management since 2003. His current research interests mainly relate to
[47] R. Rifkin, A. Klautau, “In defense of one-vs-all classification,” Journal handling so-called complex data in data warehouses (XML warehousing,
of Machine Learning Research, 5, 2004, pp. 143-151. performance optimization, auto-administration, benchmarking...), but also
include data quality and security as well as medical or health-related
[48] S. Chebrolu, A. Abraham, and J.P. Thomas, “Feature deduction and applications.
ensemble design of intrusion detection systems,” Computer & Security,
24(4), 2004, pp. 295-307.
[49] A. Tsymbal, S. Puuronen, and D.W. Patterson, “Ensemble feature Mohammad Zahidur Rahma is currently a Professor at Department of
selection with the simple Bayesian classification,” Information Fusion, Computer Science and Engineering, Jahangirnager University, Banglasesh. He
4(2), 2003, pp. 87-100. obtained his B.Sc. Engineering in Electrical and Electronics from Bangladesh
University of Engineering and Technology in 1986 and his M.Sc. Engineering
[50] A.H. Sung, and S. Mukkamala, “Identifying important features for
in Computer Science and Engineering from the same institute in 1989. He
intrusion detection using support vector machines and neural networks,”
obtained his Ph.D. degree in Computer Science and Information Technology
In Proc. of International Symposium on Applications and the Internet
from University of Malaya in 2001. He is a co-author of a book on E-
(SAINT 2003), 2003, pp. 209-217.
commerce published from Malaysia. His current research includes the
[51] L.S. Oliveira, R. Sabourin, R.F. Bortolozzi, and C.Y. Suen, “Feature development of a secure distributed computing environment and e-commerce.
selection using multi-objective genetic algorithms for handwritten digit
recognition,” In Proc. of 16th International Conference on Pattern
26 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Improving Overhead Computation and pre-processing
Times for Grid Scheduling System
1
Asgarali Bouyer, 2Mohammad javad Hoseyni, Abdul Hanan Abdullah
Department of Computer Science Faculty of Computer Science and Information Systems
1,2
Islamic Azad University-Miyandoab branch UNIVERSITI TEKNOLOGI MALAYSIA
Miyandoab, Iran Johor, Malaysia
1
basgarali2@live.utm.my, smj.hosseini@gmail.com hanan@utm.my
Abstract— Computational Grid is enormous environments with on the place of tasks or applications. For example, suitable
heterogeneous resources and stable infrastructures among other node selection can reduce overhead communication and cost
Internet-based computing systems. However, the managing of and makespan and even execution time. Resource discovery is
resources in such systems has its special problems. Scheduler important but not enough because of the dynamic variation in
systems need to get last information about participant nodes from the grid, such that resource prediction is necessary for grid
information centers for the purpose of firmly job scheduling. In system to predict coming status of nodes and their workloads.
this paper, we focus on online updating resource information Therefore, for prediction of node's status, schedulers need to
centers with processed and provided data based on the assumed get up-to date or last information about nodes. Another
hierarchical model. A hybrid knowledge extraction method has
problem is how to get up-to date information about nodes. In
been used to classifying grid nodes based on prediction of jobs’
features. An affirmative point of this research is that scheduler
most of the grid scheduling systems, there are some special
systems don’t waste extra time for getting up-to-date information centers that maintain last information about grid node's status
of grid nodes. The experimental result shows the advantages of that periodically updated by its management section such as
our approach compared to other conservative methods, especially Meta-computing Directory Services [1] in Globus toolkit. In
due to its ability to predict the behavior of nodes based on the Globus Toolkit, Resource and status information is
comprehensive data tables on each node. provided via a LDAP-based network directory called Meta-
computing Directory Services (MDS). It has a grid information
Keywords-component; job scheduling; hierarchical model; Grid service (GIS) that is responsible for collecting and predicting
nodes modul; Grid resource information center the resource status information, such as CPU capacities,
memory size, network bandwidth, software availabilities, and
I. INTRODUCTION load of a site in a particular period. GIS can answer queries for
In computational grid systems, a job or application can be resource information or push information subscribers [2]. n our
divided into tasks and distributed to grid nodes. These tasks can research, we have used GIS idea to maintain nodes’
be executed independently at the same time in parallel ways to information, but a little different from Globus’ GIS, for
minimize completion time of job execution. Therefore, grid predicting in a local fashion. For this aim, we used a special
nodes dynamically share their resources to use by another grid component on all participant Grid nodes that is called grid
application. In order to perform job scheduling and resource node’s module (GNM). In Globus, all processing information
management at Grid level, usually it is used a meta-scheduler. is done by MMDS, and it does not use local processing for this
A resource scheduler is fundamental in any large-scale Grid purpose. However, we have used a local information center
environment. The task of a Grid resource broker and scheduler each node to maintain a complete information or background
dynamically is to identify and characterize the available about its status n order to exactly exploration of knowledge for
resources, and to select and allocate the most appropriate valuation and scheduling.
resources for a given job. In a broker-based management The rest of this paper is ordered as follow. In section two, a
system, brokers are responsible for selecting best nods, problem formulation is described. Some, related works on
ensuring the trustworthiness of the service provider. Resource earlier research have been reviewed in section 3. Our proposed
selection is an important issue in a grid environment where a approach has been discussed in section4. In section 5, the
consumer and a service provider are distributed geographically experimental results and evaluations have been mentioned.
across multiple administrative domains. Choosing the suitable Finally, the paper is concluded in section 6.
resource for a user job to meet predefined constraints such as
deadline, speedup and cost of execution is an main problem in II. PROBLEM FORMULATION
grids. As you know, each task has some conditions that must be
considered by schedulers to select the destination nodes based One motivation of Grid computing is to aggregate the
power of widely distributed resources, and provide non-trivial
27 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
services to users. To achieve this goal, an efficient Grid AppLeS (Application Level Scheduling) [6] focuses on
scheduling system must be considered as an essential part of developing scheduling agents for individual Grid applications.
the Grid. Since the grid is a dynamic environment, the It applies agents for individual Grid applications. These agents
prediction and detection of available resources and then use an use application oriented scheduling, and select a set of
economic policy in resource scheduling for coming jobs with resources taking into consideration application and resource
consider some sensible criteria is important in scheduling information. AppLeS is more suitable for Grid environment
cycle. In a Grid environment, prediction of resource with its sophisticated NWS[7] mechanism for collecting system
availability, allocation of proper nodes to desired tasks, a fairly information [8]. However, it performs resource discovering and
price adapter for participant nodes is the prerequisite for a scheduling without considering resource owner policies.
reasonable scheduling guarantee. Many approaches for grid AppLeS do not have powerful resource managers that can
meta-scheduler are discussed from different points of view, negotiate with applications to balance the interests of different
such as static and dynamic policies, objective functions, applications [8]. EMPEROR [9] provides a framework for
application models, adaptation, QOS constraints, and strategies implementing scheduling algorithms based on performance
dealing with dynamic behavior of resources that have some criteria. The implementation is based on the Open Grid
weaknesses (e.g., complexity time, predicting problems, using Services Architecture (OGSA) and makes use of common
out of date data, unfair, unreliable, nonflexible, etc.). Based on Globus services for tasks such as monitoring, discovery, and
the current researches, a new approach has been proposed as a job execution. EMPEROR is focused on resource performance
helpful tool for meta-scheduler to do a dynamic and intelligent prediction and is not distributed nor does it support economic
resource scheduling for grid with considering some important allocation mechanisms.
criteria such as dynamism, fairness, response time, and
reliability. Singh et al. proposed an approach for solving the Grid
resource management problem by taking into
The job scheduling problem is defined as the process of consideration[10]. The paper proposed an approach aimed at
making decision for scheduling tasks of job based on grid obtaining guarantees on the allocation of resources to task
resources and services. Grid scheduling problem is formally graph structured applications. In mentioned research, resource
represented by a set of the given tasks and resources. A grid availabilities are advertised as priced time slots, and the authors
system is composed of a set on nodes as N = {N1 , N 2 ,..., N n } presented the design of a resource scheduler that generates and
and each node consists of several resources, that is, advertises the time slots. Moreover, Singh et al. demonstrated
that their proposed framework (incorporating resource
i
N i = {R 1 , R i2 ,…, R ir } and each resource is appeared often in all reservation) can deliver better performance for applications
nodes within different characteristics. By a set of the given jobs than the best effort approach.
in time period T, it consists of several jobs within different
{ }
characteristics, that is, J = J1 , J 2 ,..., J j that belong to c Another work has been done by Chao et al. that is a
coordination mechanism based on group selections of self-
consumers C = {C1 , C 2 ,..., Cc } . Each job necessarily is divided organizing agents operating in a computational Grid [18]. The
{ }
into several tasks, that is, J i = T1i , T2i , T3i ,..., Tti . The main authors argued that due to the scale and dynamicity of
computational Grids, the availability of resources and their
objective in most scheduling systems often is to design a varying characteristics, manual management of Grid resources
scheduling policy for scheduling submitted jobs with the goal is a complex task, and automated and adaptive resource
of maximizing throughput and efficiency and also minimizing management using self-organizing agents is a possible solution
job completion times. Job’s scheduling is generally broken to address this problem. Authors have pointed out that for Grid
down into three steps: resource management, examples in which performance
1- To define a comprehensive and versatile method and divide fairly enhancement can be achieved through agent-based
job between grid nodes. coordination include: decision making in resource allocation
2- The allocation of tasks to the computing nodes based on user and job scheduling, and policy coordination in virtual
requirement and grid facilities. organizations.
3- The monitoring of running grid tasks on the nodes over time and
reliability factors. Kertész and Kacsuk have argued that there are three
With a large number of users attempting to execute jobs possible levels of interaction to achieve interoperability in
concurrently on the grid computing, parallelism of the Grids: operating system level, Grid middleware level, and
applications and their respective computational and storage higher-level services level [11]. Authors described three
requirements are all issues that make the resource scheduling approaches to address the issue of Grid interoperability,
problem difficult in these systems. namely: 1) extending current Grid resource management
systems; 2) multi-broker utilization; and 3) meta-brokering. In
III. RELATED WORKS extending current Grid resource management systems, they
Condor’s Matchmaker [3-5] adopts a centralized developed a tool called GTBroker that interacts with Globus
mechanism to match the advertisement between resource resources and performs job submission. This proposed meta-
requesters and resource providers. However, these centralized brokering service is designed for determining which Grid
servers can become bottlenecks and points of failures. So the broker should be best selected and concealing the differences in
system would not scale well when the number of the nodes utilizing them. Extra to the meta-brokering service, they
increases. proposed a development of the Broker Property Description
Language (BPDL) that is designed for expressing metadata
28 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
about brokers. The authors have also implemented their ideas for job submission and resource allocation like other methods,
in the Grid meta-broker architecture that enables users to we only focus on Grid Node’s Module (GNM) as a significant
access resources of different Grids through their own broker(s). part of our research. Note that the model described here does
not prescribe any implementation details; the protocols,
Many other considerable approach such as hierarchical grid programming languages, operating systems, user interfaces and
resource management [12], a new prediction-based method for other components. Proposed architecture uses a hierarchical
dynamic resource provisioning and scaling of MMOGs in grid model with minimum communication cost and time.
[13], aggregated resource information for resource selection
methods by grid broker[14] has been offered with considerable In this research, the knowledge extraction module is
idea that is recommended for researches as hopeful methods. devolved to Provider Node (PN). In many approaches [24], the
needed information is gathered in special places in order to
IV. GRID NODE’S MODULE FOR OPTIMIZED SCHEDULING manage by Grid Resource Brokers or Meta-Schedulers that
Most of grid scheduling systems consist of two main surely take much time or have the problem of out-of-date
components: nodes, and schedulers. Scheduler can be information. Here, the proposed module for provider node
considered as local schedulers and meta-schedulers. In some saves all required information in the local database and it will
earlier methods [3, 15-18] meta-scheduler, as the main do knowledge extraction methods in a local fashion. Finally,
component, are responsible for job scheduling. However, there the summarized information about each grid node’s status is
saved in local scheduler’s data tables and dynamically is
updated by an online method [25]. A new illustration of GNM
Local DB
is depicted in Fig. 2 with more details.
Grid Node Module
GNM
Knowledge extraction (RS and CBR)
Knowledge Extraction
Status announcer and
Task Management
adjustment Local-DB Announcer Section
Urgent Change
Task-Management
Price adjusting
Task Submission
Sender/ Receiver Node Status
Pre-processing Announcing
Local -Scheduler module
LS-DB Preparing message Updating Data File
Job Manager
Fault Tolerance provider Info Collector
Interface (input/output)
Bidding services
Figure 2. The Grid Node’s Module (GNM) with more details.
Coordinator layer
Meta-Scheduler module
A. Knowledge Extraction
Auction Manager
The applied methods for knowledge extraction are Rough
Ms-DB Set theory and Case-based Reasoning. GNM uses Case-Based
Reasoning (CBR) technique and Rough Set Analysis. These
New Arrival Job Queues techniques work together to discover knowledge in order to
Fault Tolerance Management supply a satisfied recommendation to send to local scheduler.
The exploration is based on previous data or experiments on
the “node data table” that is helpful to make a better decision.
Figure 1. A hierarchical architecture for optimized scheduling.
Use of multiple knowledge extraction methods is beneficial
because the strengths of individual technique can be leveraged.
is other scheduling methods [19-23] in which local schedulers In previous research [18], we proposed a learning method for
perform most of job scheduling steps. These mentioned resource allocation based on the fuzzy decision tree. This
methods have not applied the impact of using grid nodes in method observed that it has a successful potential to increase
scheduling system and they only map jobs to nodes. In this accuracy and reliability if the job length is large. However, in
section we are going to devolve some steps of scheduling this section, we use a hybrid of CBR and RS to get the exact
process to grid nodes or all participant nodes in grid system. knowledge with considering economic aspects. This section is
divided in three sub-sections: Rough Set Analyzer, Case-based
A general architecture of the grid scheduling system has reasoning method, and calculating some information for
been depicted in Fig.1. Since, this architecture uses an auction computing of priority.
mechanism by meta-scheduler and participant local schedulers
29 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
We used Rough Set (RS) theory [26] to generate rules in rules to classify and define training set. It consists of two steps:
order to analyze by GNM to classify proper nodes to use by 1) selecting consistent rules for the job in order to get desired
CBR method. Rough set analysis provides an effective means samples (records) to define training sets. In this case, it can
for analysis of data by synthesizing or constructing select a best training set. 2) Final processing and predicting the
approximations (upper and lower) of set concepts from the situation of a coming job by using neighboring records (in the
acquired data. It also proved to be very useful for analysis of same class).
decision problems concerning objects described in a data table
by a set of condition attributes and decision attributes. The goal After doing CBR, the obtained knowledge about job and
of using the rough set theory in this research is to generate job (executing job on this node) will be sent to scheduler. In the
useful rules for classifying similar states to apply by CBR in next sections, we will describe how local scheduler use this
order to explain the best state for node to accept or reject this extracted knowledge for resource allocation.
new task. Our purpose is to carry out resource selection for the B. Task Management
desired job based on job condition in the scheduling phase. To Since the capacity of resources in each node is changed at
do this issue, we will use Rough Set Analyzer (RSA) to the moment, new task must be processed before submitting
generate rules. It takes the nodes’ information data table as because the existing capacity may not be sufficient for a
input. The output is three Matrixes (generated rules are shown desired task in determined deadline. In this case, task is not
in matrix form). inserted to queue and rejection information is sent to local
The RSA uses three important attributes (final status of scheduler (LS). This operation is done after CBR execution and
task, completion time, and cost price) as decision attributes. the result is sent along with extracted knowledge (by CBR). In
These attributes can be acted upon as the condition attributes contrast, if the existing resources be enough for this task, it will
and decision attribute of a decision system. Desired application be successfully submitted in the queue of the grid’s task on the
only uses one of this attributes at a moment as decision provider node. Also, all information about this task is inserted
attribute and at the same time, other two attributes will be in the related data table as a new Recordset. GNM record
considered as conditional attributes. For example, if several important properties at this time such as CPU Load,
dependability and speed factors be more important, the second Free memory (RAM), Task-ID, size of the new task, priority of
and third attribute is considered as Decision attribute, the new task (we consider only 3 priority Low, Normal and
respectively. There are other conditional attributes that we have High), number of all grid tasks (in waiting status), amounts of
mentioned in next section. In addition, RSA needs to discretize Data Transmission Rate (DTR) related to this node in the grid
the input data for some attributes. Since RSA takes analysis (DTR probably has upheaval in sometimes), start time of task
time in order to perform the rough set method, though not execution, spent time for this task, completion time, status of a
considerable, it is possible that we are encountered with this task (wait, running, success, and fail). Some of this information
question: When will RSA execute rough set analysis? To (e.g. spent and completion time, task status and so on) is
answer this question, we supply two conditions for doing a updated after finishing a task. In our approach, task has four
rough set analysis: states: wait, running, fail, and success. After submit a task in
the queue, at first, it take wait state. When a task is started for
Number of currently added tasks to this node is more than executing, its state changes to running state until it is
1% of previous submitted tasks in the past days. terminated. After successfully finishing, the task state will be
Rough set analysis has not been done in last 24 hours. changed to success state. It is possible that task state is changed
to fail state due to diverse software and hardware problems. At
Case-based Reasoning is a technique that adapts past the end, the result completely is given back to LS if the task
solutions for new demands by using earlier cases to explain, successfully is executed.
criticize and interpret novel situations for a new problem [27].
The basic CBR processes are defined as a cycle and include the C. Announcer Section
following: RETRIEVE the most similar cases; REUSE the This section is the most important section in GNM. It is
cases to solve the problem; REVISE the proposed problem responsible to decide on whether the node is ready to accept a
solution; RETAIN the modified solution as a new case. These new task or deny new task. Announcer section (AS) analyzes
steps fully must be done to get the satisfied knowledge. Now, the grid tasks queue and its own status (mentioned in above) to
we encounter with this question: When will Case-based specify coming status. For example, it specifies that in next two
Reasoning be executed? For this question, first we should say hours it cannot accept any new task. This section is definitely
that when will the nodes get the new tasks (or job) analyzing its own status after every submitting. It evaluates
information? During online resource selection by local deadline and execution time of the previous submitted task
scheduler, the job information is sent to all nodes. In [28] an (waiting and running state) to determine how many processes
optimized version of Case-Based Reasoning had been proposed in the near future will finish. With the assumption of finishing
to increase accuracy in final results. This method applies CBR these processes, when would the desired node be able to accept
algorithm by using Decision Tree in order to select suitable new tasks in the future? In addition, it is possible that some
sampling. Improving accuracy criterion was a success key in high priorities local processes will join to current processes in
this method. However, due to classification of input data by near future (e.g. automatically start a Virus Scan program, Auto
data mining techniques such as decision tree, selecting training saves or backup by some application, and so forth). Thus, AS
set takes much time that is not negligible for online processes. has to consider all possible status to get the best decision. This
Therefore, to reduce of this overhead time, we use rough set
30 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
process will be done by sub-section that is called Node Status Four important measures were evaluated in the simulation:
Announcing (NSA) module. dependability or reliability, accuracy prediction, and success
ratio and iteration of job in other nodes. In GridSim, each
NSA module also computes some related information, such simulated interactive component (e.g. resource and user) is an
as Success Ratio, Average of Completion Time (ACT), entity that must inherit from the class GridSim and override a
Average of CPU-Idle (how much percent is CPU free or idle) body()method to implement its desired behavior. Class Input
and Average of free memory (RAM), about this node and and Output in GridSim is considered for interacting with other
sending it along with other obtained results to Local scheduler. entities. Both classes have their own body() method to handle
For instance, ACT measure is computed as following equation: incoming and outgoing events, respectively. Entities modeled
Success Ra
tio= Ns /N a in GridSim include the resources, users, information services,
n and network-based I/O.A resource, that in our method called
ACT k = (∑GTp ) / n
i =1
i (1) provider node, is characterized by a number of processors,
speed of processing (a specialized CPU rate for the grid task),
GTpi is completiontime for taski; and n is the number of success tasks by node k th . The tolerance of price variation for provider node, The real
data transmission rate per second, The capacity of RAM
Ns: Number of successfully completed tasks. memory, Monetary unit, Tolerance of the price variation and
Na: Number of Successful + Failed tasks time zone. Furthermore, the node’s price is computed based on
mentioned characteristics. Tolerance of price variation is a
It is mentioned that aborted tasks are different from failed parameter to give a discount over node’s price that is used for
task. Fail event can be occurred because of a nodes’ problem some low budget jobs. For each resource, the CPU speed has
such as software, hardware, deadline or budget problems. been determined by MIPS measure (million instructions per
Where abort event is done by scheduler for that canceling of second). Each property is defined in an object of the
job by consumer or other problems and executive node has not ResourceCharacteristics class. The flow of information from
any problem for continuing job execution. Therefore, aborted other entities via I/O port can be processed by overriding
task is considered as neutral tasks and those are not taken into processOtherEvent() method. We used a uniform allocation
account for measuring of the success ratio. method for all nodes [23].
Sometimes a node is encountered with unpredictable cases. For our testing, we define three different local scheduler
For example, suppose that a desired node is ready to accept (Table 1) and three groups of jobs (Table2). Each group of jobs
new tasks. If node’s resources have unexpectedly been has special features that have been mentioned in Table 2. In our
occupied by local tasks (OS processes), this node cannot accept previous work [18] the nodes’ specifications and their
a new task until to come back to normal state. In this case, performance is collected from a real local grid. In this research,
Urgent Change section, a sub-section in Announcer Section, the updated of this data table is used for supposition nodes, and
has to change its status to non-acceptance and then inform this so we do not explain about nodes’ properties.
change to scheduler. After come back to normal state, this
section has to announce it to Local Scheduler. Each group of jobs is submitted on different times all three
local schedulers. It is necessary to say that, tasks of jobs are
Another subsection is Price adjusting section. This module submitted in a parallel form on available nodes in every local
is responsible for determining the price of a node based on scheduler. For example, the Job_Group1 is composed of 250
standard factors and the current node status. For example, if the tasks and 45500 Million Instructions for every task that each
computed price based on standard parameters for one minute task averagely has 1200 second deadline to complete integrally.
become α, this module can change this price based on current Each group of jobs has been tested 15 times separately on each
status such as the number of current submitted tasks (in waiting local scheduler’s node.
state), number of success tasks/ number of failure tasks in last
day and last week and so on. Its mention that, due to respect for Since, most of the presented scheduling systems and
grid owners and grid costumers profits, the price increment or scheduling algorithms were tested and evaluated based on
decrement can be in the following range: specific assumptions and parameters of authors, therefore,
nobody cannot claim that his/her method is the best. However,
α*(1-p)< Offered Price <α*(1+p) : α is standard in this research we tried to test of our approach in GridSim
price, and 0≤p≤0.5 simulator with developing real nodes’ behavior for node
At the end, this Offered Price is sent to local scheduler. (resource) entity. The following experiments show the
Therefore, the offered price by provider node always is comparison of GNM effect to use in node selection step by
dynamic. local schedulers. In Fig. 3 the number of tasks’ completion is
compared for new approach and recent work [18].
V. EXPERIMENTAL RESULTS AND DISCUSSION
To observe the effect of GNM architecture, we used The analysis of obtained results in Fig. 3 show that, due to
GridSim simulator [29]. GridSim support hierarchical and use rough set based case base reasoning on grid node module
economic-based grid scheduling systems and it is used as a (it is not necessary online), the workload of schedulers is
reference simulation environment for most of the significant decreased and so the overhead time for selection node is
research such as [30-32] and compare our results with the job decreased. Consequently, as it is seen in Fig. 3, the overhead
scheduling algorithms proposed in [18, 33]. time for node selections, starting, and gathering results plus
execution time for the new proposed approach is less than the
31 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
nsidered deadline for each task of job_G
con Groups3 on loc cal Howev decision a
ver, he ach
accuracy for th new approa was low
heduler LS1, w
sch based scheduli
whereas this time in fuzzy b ing rather than other m methods for jobs that reliability and
met lly an
thod [18] total is more tha the task dea adline. Therefo
ore, comple early had equal priority because of especial
etion factors ne
y
any task of job_ not
_Groups3 cann be finishe in consider
ed red y y
priority computing by each method.
adline on LS1 based on fuz based sch
dea zzy heduling metho od.
T
TABLE I. THE DESCRIPTIO OF CONSIDER
T ON HEDULERS.
RED LOCAL SCH
Locaal N
Number f
of Medium node’s s
MIPS) allocated current
GMIPS (M d queue
schedduler’s a
available dep
pendability
atus
CPU MIPS for grid tasks deadline sta (sec)
)
(LS) Name N
Nodes
LS1 4
400 65 460 2
0.72
LS2 3
320 140 350 3
0.93
LS3 7
750 80 400 5
0.85
TABLE II. F
SAMPLE OF JOBS.
Groups of Number of jobs Deadline for each task Me
emory for each T
Task length by (Miillion
reliability completion time
ame
jobs na (sec)
( task (MB) I)
Instruction) (MI
Job_GGroup1 5 job (25 task)
50 1
1200 1.93 45500 0.8 0.2
Job_GGroup2 10
3 job (21 task) 2
2100 3.4 72000 0.3 0.7
Job_GGroup3 5 job (10 task)
00 900 6.25 30000 0.5 0.5
LS1 LS2 LS3
0.94 0.8 0.945 0.91
885 0.92
0
0.89 0.915 0.92
0.89 0.9 0.93 0.95
1 05
0.90
0.89 0.9
0.82
0.9
9
0.5
0 LS3
LS2
Job_Group1
0
Job_Group2
LS1
Job_Group3
Job_Group1
Job_Group2
Job_Group3
roach
New appr
eduling [18]
Fuzzy based sche
e ted n dulers for new app
Figure 3. The ration of complet tasks of jobs on three local sched work [18] based on grouped jobs in Table. 2.
proach and earlier w
oach
New appro ased scheduling [1
Fuzzy ba 18] TRF method [33]
1
0.9
0.8
Accuracy of prediction
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
LS1 LS2 L
LS3
e son rediction for the ne approach and o
Figure 4. The comparis of accuracy pr ew other two methods.
32 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
15
10
5 LS1
LS2
L
LS3
0 LS3
S2
LS
up1
Job_Grou
Job_Group2 1
LS1
Job_Group3
b_Group1
Job
Job_
_Group2
New approach Job_G
Group3
Fuzzy based scheduling [18]
valuation of tasks iteration in the new approach and ot
Figure 5. The ev w ther method.
on
The evaluatio of the Fig. 4 show that the fuzzy bas sed
sch
heduling metho has better accuracy predi
od a iction rather th
han REFERENCES
w
new approach a and TRF meth hods [33] because of onli ine 1. The Globus Project, ht ttp://www.glob bus.org/MDS/.
cision with u
dec d
up-to date and consuming more time f for 2. Ferre id h
eira, L., et al., Introduction to Gri Computing with Globus. 2002,
knoowledge extrac er, od
ction. Howeve TRF metho is acceptab ble ibm.com/redbooks.
when scheduler want to select a small nu umber of nod des 3. Frey, J., et al. Condor on
r-G: A Computatio Management A Agent for Multi-
ween large n
betw numbers of av s.
vailable nodes By the wa ay, tutional Grids. in 10th IEEE Sy
Instit n gh
ymposium on Hig Performance
ove or s
erhead time fo TRF is less than both oth methods a
her and Distr g
ributed Computing (HPDC10). 2001 San Francisco, C
1. CA.
close to our new a
approach. 4. Proje C. http://ww
ect, ww.cs.wisc.edu/ orG/
/condor/condo . [cited.
5. Doug ,
glas, T., T. Todd, and L. Miron, D Distributed comput ting in practice:
The purpose of iteration in this paper is the replace ne ew the C ce:
Condor experienc Research Arti Comput. : Pract.
icles. Concurr. C
nod instead of fa
de d
aulted node and restart job in new node. Since, er.,
Expe 2005. 17(2-4): p. 323-356.
e ed
the overhead time in fuzzy base method is m has
much, if a task h 6. Berm man, F., et al., Ad g ng
daptive computing on the grid usin apples. IEEE
not been a suffici
t i
ient deadline, it cannot be ite
erated on the neew Trannsaction on Parallel and Distributed Systems, April 2003. 14(4): p.
d
nod and so this t
de task is failed. According to obtained results in
A s 369-382.
g.
Fig 5, the new a approach has more task iter ration than fuzzzy 7. Wols R., N. Spring and J. Hayes, T Network Wea
ski, g, The ather Service: A
Distrributed Resourc e
ce Performance Forecasting Service for
heduling metho and so it will act a bit b
sch od, w better than fuzz
zy- Meta re
acomputing. Futur Generation Com mputing Systems, Metacomputing
sed sed er
bas method bas on a numbe of iterations. Spec Issue, October 1999. 15(5-6): p. 757–768.
cial r
8. Vadh hiyar, S.S.D., J.J A metaschedul for the Grid. in 11th IEEE
J. ler
VI. CON
NCLUSION rnational Symposi
Inter ium on High Perf formance Distribu uted Computing,
ier
According to the weaknesses of the earli grid resour rce 2.
2002 HPDC-11. 2002.
disccovery approaches, such as using out-of- date informati ion 9. Adzi igogov, L., J. Sold ymenakos, EMPER
datos, and L. Poly ROR: An OGSA
grid meta-scheduler b c
based on dynamic resource predict tions. Journal of
ng,
for job schedulin much elap psed time for task submissio on, Grid Computing, 2005 3(1-2): p. 19-37.
d 5. .
connsiderable over rhead commun per
nications and cost, this pap
man, an, d
10. Singh, G., C. Kesselm and E. Deelma An End-to-End Framework for
es l
use an optimal strategy to improve res source discove ery Provvisioning-Based R Resource and App pplication Management. Systems
iciency. In this approach, all information ab
effi s s
bout grid nodes is nal,
Journ IEEE, 2009. 3 3(1): p. 25-48.
t o
not necessarily aggregated on grid resou urce informatiion 11. Kerteesz, A. and P. Kacsuk, Grid Interoperability Solutions in Grid
nters. We used a local data table on each grid node to
cen d a Reso t. l, ):
ource Management Systems Journal IEEE, 2009. 3(1) p. 131-141.
mai rmation about grid node’s s
intain all infor status. Moreovver, 12. Edua recursive architect
ardo, H., et al., A r cal
ture for hierarchic grid resource
d
grid resource infoormation center maintain a su
rs ummary of up-to- mana agement. Future G Gener. Comput. Sy 2009. 25(4): p. 401-405.
yst.,
e
date information that continua d s
ally is updated if there was a 13. Radu P. and N. Vlad Prediction-based real-time resour provisioning
u, d, d rce
signnificance changing in node’s resources or pperformance. TThe for mmassively multipl es.
layer online game Future Gener. Comput. Syst.,
9.
2009 25(7): p. 785-79 93.
exp ults hat
perimental resu explain th our appro oach reduces t the
14. Ivan, R., et al., Grid broker selection str
rategies using aggrregated resource
ove nd e
erhead time an improves the resource disc covery efficien
ncy infor
rmation. Future Ge st.
ener. Comput. Sys 26(1): p. 72-86.
in a grid system.
15. Casaanova, H., et al., Adaptive Schedul ling for Task Far rming with Grid
Midddleware. Internati ional Journal of Supercomputer A Applications and
Highh-Performance Com mputing, 1999.
33 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
16. Chard, K. and K. Bubendorfer, A Distributed Economic Meta-scheduler
for the Grid, in Cluster Computing and the Grid, 2008. CCGRID '08. 8th
IEEE International Symposium on May 2008. p. 542-547
17. Bouyer, A., E. Mohebi, and A.H. Abdullah, Using Self-Announcer
Approach for Resource Availability Detection in Grid Environment, in
Fourth International Multi-Conference on Computing in the Global
Information Technology - ICCGI2009. 23-29 Aug. 2009 IEEE Computer
Society: Cannes, La Bocca , France. p. 151-156.
18. Bouyer, A., et al., A new Approach for Selecting Best Resources Nodes by
Using Fuzzy Decision Tree in Grid Resource Broker. International Journal
of Grid and Distributed Computing, 2008. Vol(1)(1): p. 49-62.
19. A. Bouyer, M.K., M. Jalali. An online and predictive Method for Grid
Scheduling based on Data Mining and Rough Set. in ICCSA2009: 9th
International Conference on Computational Science and Its Applications.
June 2009. South Korea: Springer-Verlag Berlin Heidelberg.
20. Bouyer, A. and S.M. Mousavi, A Predictive Approach for Selecting
Suitable Computing Nodes in Grid Environment by Using Data Mining
Technique, in Advances in Computational Science and Engineering. 2009,
Springer-Verlag Berlin Heidelberg. p. 190-205.
21. N.B. Gorde, S.K.A. A Fault Tolerance Scheme for Hierarchical Dynamic
Schedulers in Grids. in International Conference on Parallel Processing -
Workshops, 2008. ICPP-W '08. Sept. 2008.
22. Li, C. and L. Li, Utility-based scheduling for grid computing under
constraints of energy budget and deadline. Computer Standards &
Interfaces, 2009. 31(6): p. 1131-1142.
23. Li, Z.-j., C.-t. Cheng, and F.-x. Huang, Utility-driven solution for optimal
resource allocation in computational grid. Computer Languages, Systems
& Structures, 2009. 35(4): p. 406-421.
24. Young Choon, L. and Y.Z. Albert, Practical Scheduling of Bag-of-Tasks
Applications on Grids with Dynamic Resilience. IEEE Trans. Comput.,
2007. 56(6): p. 815-825.
25. Bouyer, A., E. Mohebi, and A.A. Hanan, Using Self-Announcer Approach
for Resource Availability Detection in Grid Environment, in Proceedings
of the 2009 Fourth International Multi-Conference on Computing in the
Global Information Technology. 2009, IEEE Computer Society.
26. Z. Pawlak, J.G.-B., R. Slowinski, W. Ziarko, Rough Sets.
Communications on the ACM, 1995: p. 89–95.
27. Kolodner, J., Case-Based Reasoning. 1993: Morgan Kaufmann.
28. Bouyer, A., B. Arasteh, and A. Movaghar, A new Hybrid Model using
Case-Based Reasoning and Decision Tree Methods for improving
Speedup and Accuracy, in 4th Iadis International Conference Applied
Computing 2007. 2007: Spain. p. 787-789.
29. Buyya, R. and M. Murshed, GridSim: a Toolkit for modeling and
simulation of grid resource management and scheduling. Concurrency
and Computation—Practice & Experience, 2002. 14 p. 1175–1220.
30. Foster, I., C. Kesselman, and S. Tuecke, The anatomy of the grid:
enabling scalable virtual organizations. International Journal of
Supercomputer Applications 2001. 15 p. 200–222.
31. Li, H. and R. Buyya, Model-based simulation and performance
evaluation of grid scheduling strategies. Future Generation Computer
Systems, 2009. 25(4): p. 460-465.
32. Leal, K., E. Huedo, and I.M. Llorente, A decentralized model for
scheduling independent tasks in Federated Grids. Future Generation
Computer Systems, 2009. 25(8): p. 840-852.
33. Brent, R. and J.L. Michael, Resource Availability Prediction for Improved
Grid Scheduling, in Proceedings of the 2008 Fourth IEEE International
Conference on eScience. 2008, IEEE Computer Society.
34 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
The New Embedded System Design Methodology
For Improving Design Process Performance
Maman Abdurohman Sarwono Sutikno
Informatics Faculty STEI Faculty
Telecom Institute of Technology Bandung Institute of Technology
Bandung, Indonesia Bandung, Indonesia
mma@ittelkom.ac.id ssarwono@gmail.com
Kuspriyanto Arif Sasongko
STEI Faculty STEI Faculty
Bandung Institute of Technology Bandung Institute of Technology
Bandung, Indonesia Bandung, Indonesia
kuspriyanto@yahoo.com asasongko@gmail.com
Abstract—Time-to-market pressure and productivity gap force hardware and software co-design environments, do not fit with
vendors and researchers to improve embedded system design the rising demands.
methodology. Current used design method, Register Transfer
Level (RTL), is no longer be adequate to comply with embedded Fortunately, the electronic design automation industry has
system design necessity. It needs a new methodology for facing prepared to face this problem by providing engineers with the
the lack of RTL. In this paper, a new methodology of hardware support for these challenging. The introduction of register
embedded system modeling process is designed for improving transfer level (RTL) as a higher abstraction layer over gate
design process performance using Transaction Level Modeling level design is a revolution step to face this challenges. The
(TLM). TLM is a higher abstraction design concept model above RTL abstraction layer is accepted as the abstraction layer for
RTL model. Parameters measured include design process time describing hardware designs. The vendor of EDA is pushing
and accuracy of design. For implementing RTL model used the abstraction layer for addressing the lack of RTL. The
Avalon and Wishbone buses, both are System on Chip bus. definition of ESL is “a level above RTL including both
Performance improvement measured by comparing TLM and hardware and software design” as suggested by The
RTL model process. The experiment results show performance International Technology Roadmap for Semiconductors
improvements for Avalon RTL using new design methodology (ITRS).
are 1,03 for 3-tiers, 1,47 for 4-tiers and 1,69 for 5-tiers.
Performance improvements for Wishbone RTL are 1,12 for 3- ESL design and verification methodology consists of a
tiers, 1,17 for 4-tiers and 1,34 for 5-tiers. These results show the broad spectrum of environments for describing formal and
trend of design process improvement. functional specifications. There are many terms used to
illustrate ESL layer such as hardware and software co-design
Keywords : Design Methodology, Transaction Level Modeling models, architectural models, RTL and software models, and
(TLM), Register Transfer level (RTL), System on Chip. cell-level models. This prescription include the modeling,
simulation, validation and verification of system level designs.
I. INTRODUCTION Models at the higher layer level are descriptions above the RTL
Design is an important step on whole embedded system abstraction layer depicting the system behavior. There are a
design process. Embedded system development process begins number of ways to define the abstraction layer may be raised
by making hardware and software specification. The growing above RTL. For example, SystemC presents transaction level
consumer demands for more functionality tools has lead to an model for modeling and simulation of embedded software
increase in complexity of the final implementation of such systems.
designs. The ability of semiconductor industry to reduce the
minimum feature sizes of chip has supported these demands. Time-to-Market Pressure
Also, show that the Moore’s Law, roughly doubling the devices The growing consumer demands for various complex
per chip every eighteen to twenty-four months, is still accurate. application led to pressure vendor to design, implement
However, even though current IC technology is following the embedded system in short time frame. The late to enter market
growing consumer demands, the effort needed in modeling, is means cost or opportunity lost.
simulating, and validating such designs is adversely affected.
This is because current modeling method and frameworks, This condition called time-to-market pressure for embedded
system vendor. It needs shorter design process approach.
35 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
amount of time is not fit with the increase in complexity. This
is referred to as the productivity gap, which is based on the
ITRS (International Technology Roadmap for
Semiconductors).
Revenues ($)
Time (months)
Figure 1. Time-to-Market and revenues [5]
Embedded system design
Design flow of embedded system begins with design
specification, its define system constraint, both cost and
processing time. System functionality is defined in behavioral
description, hardware software partitioning is done to optimize
design result and still fit the requirement. Hardware and Figure 3. Moore’s law [9]
software integration is done after hardware/software detail
design. Register transfer level design is carried out by means Increasing the complexity and functionality of electronics
hardware programming language such as, Verilog, VHDL and systems, causes the increasing of the possible design choices
Esterel. Verification and testing process is done to ensure and the alternatives to explore for optimization purposes.
embedded system design is fit to specification [1]. Therefore, design space exploration is vital when constructing
a system in order to choose the optimal alternative with respect
Fase 1 : Product specification to performance, cost, etc. The reduction of time to develop
these system-level models for optimization purposes can
improve design acceleration with acceptable performance. A
Fase 2 : HW/SW partitioning
possible way to reduce this time is to raise the abstraction layer
of design.
2 – 6 months needed
Register Transfer Level design
One of the past design revolutions in hardware design was
SW design
HW design
Fase 3 : Detailed
HW/SW desing
the introduction of RTL design layer as the entry point of the
design flow. At RT level, registers and a data-flow description
of the transfers between them replace the gate-level instantiate
of independent flip-flops and logical operators. Some hardware
Fase 4 : HW/SW description languages such as VHDL, Verilog and Esterel are
integration used for writing models at this RT level. The translation to gate
level is called synthesis. Component example at this level are
Fase 5 : Acceptance Testing
adder, multiplexer, decoder, and memory.
The complexity of hardware design combined with the lack
Fase 6 : Maintenance and Upgrade
of a revolution design approach similar to the RTL introduction
Figure 2. Embedded system design flow [1]
has induced very slow simulations and caused productivity gap.
The peak problem for system-on-chips is software development
need, co-simulating embedded software with the RTL model is
The embedded design process is not as simple as the possible, but too slow to allow its effective development.
concept. A considerable amount of iteration and optimization Designer are forced to wait the final chip to begin writing the
occurs within phases and between phases. software of the system. This results is wasted time in the
development cycle and increased time-to-market. While
efficient in terms of speed, the still require the RTL model be
Moore’s Law and Productivity gap available, they are very costly and they provide limited
Moore’s Law said that silicon capacity has been steadily debugging capabilities. Another approach to face the problem
doubling every 18-24 months. Its allow companies to build is to try to raise the abstraction level : by creating models with
more complex systems on a single silicon chip. However, less details before the RTL one, it should be possible to achieve
designer ability to develop such systems in a reasonable better simulation speeds while at the same time less accuracy.
36 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
C. VULCAN
Vulcan is designed to cut cost of ASIC Program Time. The
II. RELATED WORK cost reduction could be achieved by separating designing parts
into software.
A. Ptolemy
Initial specification is written in Hardware-V, i.e., the HS-
Ptolemy is a project developed at the University California, Description Language (HDL_. That could be synthesis trough
Berkeley [13]. The latest Ptolemy Release is Ptolemy II 7.0.1 OLYMPUS synthesis System. Specifications in C would be
that has been launched u 4 April 2008. Ptolemy is a framework mapped into representations between Control-Data Flow Graph
for simulation, prototype, and synthesis of software that has (CDFG). It is this level that Vulcan separates hardware from
been dedicated solely to digital signal processing (DSP). software.
The basic concept of Ptolemy is the use of a pre-defined The separation of hardware from software is achieved
commutation model that will regulates inter components through heuristic graphs partition algorithm that work under
interactions. The main problem address by the Ptolemy is the polynomial time. This separation algorithm has paid its
use of the mix of various commutation models. Some of the attention on different partitions on CDFG between Hardware
model domains that have been implemented are: CT and software, to minimize hardware costs but simultaneously
(continuous-time modeling), DDF(dynamic dataflow), DE maintain the predetermined deadlines.
(discrete-event modeling), FSM(finite state machines and
modal model), PN(process networks with asynchronous TABLE 1. COMPARISON OF MODELING FRAMEWORKS
message passing), Rendezvous( process), networks with
synchronous message passing, SDF (synchronous dataflow), Name Specific Modeling HW/SW SW HW
ation Part design design
SR (synchronous reactive), Wireless. Ptolemy C++ FSM GCLP C VHDL
Ptolemy II comprises supporting packages such as graphs, Cosyma C* Syntax Sim C Hardw
DAG annealing areC
provides the manipulations of Graph theory, math, provides Vulcan Hercules Vulcan Greedy DLXC HEBE
mathematical matrices and vectors and signal processing, C
plots, provides visual data display, data, provides type system, Stellar Nova Nebula Magellan GCC Asserta
data wrapping and expression parses. Ptolemy II package
comprises the following parts: D. STELLAR
STELLAR is a system level co-synthesis environment for
Ptolemy II C Code Generation: The main function is to
transformative application. A transformative transformation is
generate codes for the SDF model, FSM and HDF: the entire
an application that executes processes every time it has a
model could be converted into C Codes.
trigger such as JPEG Encoder. As an input specification,
Ptalon: Is an actor oriented designing representing the most STELLAR provides a C++ Library with ist NOVA name.
commonly designing strategy in an embedded system The inputted specifications are in forms of application
designing. This system is frequently modeled as block diagram, specifications, the architecture and performance yardsticks.
where a block presents system or ;lines or inter-block arrows The outer format of the NOVA is executable. STELLAR
representing signals. supports software estimation through profiling and using
ASSERTA synthesis device in estimating hardware.
Backtracking: This facilities serve the function to save the
previous system state values. The function is the most critical STELLAR get input specification and definitions in
in a distributed computations. NEBULA mediating format. Its designing environment
provides two devices: MAGELLAN and ULYSSES of co-
Continuous domain : Continuous Domain is a remake of synthesizing and evaluation the GW-SW. the MAGELLAN
Continuous Time domains with meticulous semantics. optimize latency retimes and the ULYSSES OPTIMIZES the
applications throughput. Its outer part comprises hardware
B. COSYMA (CO-SYnthesis for eMbedded Architectures). specification, software and interface. The exterior specification
The Cosyma is developed by the Braunschweig University/ could be translated into SystemC code. And its functionalities
The Cosyma performs operation-separation process on the would be verified through simulation.
lowest blocks to improve the speed of program execution time.
Table 1 shows the comparison between all embedded
This speed improvement is achieved by adding co- system design frameworks.
processors hard ware that will perform part of the functions
that traditionally run by software. The following Figure
indicates Cosyma flow diagram. Its inputs comprises the Cx
program (8). It is an extension of the C Program to enhance III. THE NEW DESIGN METHODOLOGY
parallel data processing. Its final output is hardware block and
the primitive of communication in hardware software. A. Transaction Level Modeling (TLM)
Transaction-level Modeling fills the gap between purely
functional descriptions and RTL model. They are crated after
hardware/software partitioning, that is, after is has been
37 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
decided which, for each processing, if it would be done using a • The type of transaction determinates the direction of the
specific hardware block or by software. The main application data exchange, it is generally read or write.
of TLM is to serve as a virtual chip (or virtual platform) on • The address is an integer determining the target
which the embedded software can be run.
component and the register or internal component
The main idea of TLM is to abstract away communication memory address.
on the buses by so-called transactions : instead of modeling all • The data that is sent to received.
the bus wires and their state change, only the logical operations • Some additional meta-data including : a return status
(reading, writing etc) carried out by the busses are considered
(error, success, etc), duration of the transaction, bus
in the model. In contrary to the RTL, where everything is
synchronized on one or more clocks (synchronous description), attributes (priority, etc).
TLM models do not use clocks. They are asynchronous by
nature, with synchronzation occuring during the The most basic functionality shared by all buses or more
communication between components. These abstractions allow generally interconnection networks is to route the transactions
simulations multiple orders of magnitude faster than RTL. to their destination depending on their address. The destination
is determined by the global memory address map which
associates a memory range to each target port.
Algorithm Model
UnTimed Functional Model
Timed Functional Model
Bus Cycle Accurate Model
Figure 5. TLM process model
Cycle Accurate Model
In order for the embedded software to execute correctly, the
address map, the offset for each register must be the same as in
Register Transfer Level Model the final chip (register accuracy). Additionally, the data
produced and exchanged by the components must also be the
same (data accuracy). Finally, the interrupts have to correspond
Figure 4. TLM Model Stack
logically to the final ones. One can view these requirements as
a contract between the embedded software and the hardware.
The other advantage of TLM models is that they require far This contract guarantees that if the embedded software runs
less modeling effort than RTL or than Cycle Accurate model. flawlessly on the virtual platform, then it will run in the same
This modeling effort is further reduced when there alreay exists way on the final chip.
a C/C++ functional code for the processing done by the
hardware block to model. For instance, one can reuse the B. New Design Flow of Hardware Embedded System Design
reference code for a video decoder or for a digital signal In this paper, the new design flow for modeling hardware
processing chain to produce a TL model. Unlike a Cycle embedded system is designed using transaction level modeling
Accurate model, which is no longer the reference after RTL is (TLM) method for early verification purpose. Verification
created, TLM is by this means an executable, “golden model” process done at the first step before detail design. Transaction
for the hardware. Various definitions of TLM exist; some of level modeling is one of new trends on embedded system
them even rely on clocks for synchronization, which looks design after the development of register transfer level
more like Cycle Accurate level. A transaction term is an atomic modeling.
data exchange between an initiator and target. The initiator has
the initiative to do the transaction whereas the target is The research scope is particularly on hardware embedded
considered as always able to receive it (at least, to indicate to system after performing separation of hardware and software
the initiator that it is busy). This corresponds to classical process. There are three stages in detailed design:
concepts in bus protocols. The initiator issues transactions 1. Hardware part definition: hardware embedded system
through an initiator port, respectively a target receives them by definition that will be implemented.
a target port. Some components only have initiator ports some
have only targets ports. Also, some components contain both 2. TLM modeling: Model construction with transaction
initiator and target ports. modeling approach and perform early verification. Model
refinement process can be generate by performing tuple
The information exchanged via a transaction depends on correction : M, S, A, PM.
the bus protocol. However, some of them are generally
common to all protocols : 3. RTL modeling: RTL model construction is the final
process of all hardware designs of embedded system. In
38 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
this process, transformation from TLM model into RTL 1) Diagram Block
model is conducted. Diagram block is a diagram that shows many inputs and
outputs of the system. Inputs of diagram block of transaction
level modeling include :
• Master : Number of master component actively perform
read () and write process as standard operation of
components
• Slave : Number of slave components considered passive
components and waiting for transaction of master.
• Arbiter : bus management system, namely mutual access
management algorithm of one slave with one master or
more.
• PM : Process taking a place in master and slave such as
read() and write() process.
• Tiers : Total the whole main components existing in a
system including Arbiter.
• Specification is system requirement explanation that
should be met by system designed.
Output of design system block is a TLM Model. Model
formulation process is conducted systematically.
Figure 6. The new design methodology
C. Procudure and Modeling Diagram Block
Basic procedure of modeling is designed as standard
process on hardware modeling. Modeling steps of new design
methodology are:
1. Define : 4-tuple input (M, S, A, PM).
2. A module with port and method is made for each master.
3. A module with port and method is made for each slave.
Figure 7. Diagram Block
4. An arbiter bus is made with algorithm in A.
5. Every method in master and slave is defined in PM.
2) Defining Master and Slave
6. Early verification of system requirement compliance
• Master and slave component definition consists of three
7. If system requirement is not satisfying, then perform tuple parts; name, port, and functions/method. Example of
refinement starting from step 1. master:
8. Adding port and RTL process Name : MicroPro
9. Port and process removal from TLM.
Port :
10. RTL arbitrary implementation. int mydata, input1, input2, input3;
int cnt = 0;
11. Port mapping unsigned int addr = m_start_address;
Stages 1 to 6 are initial stage of transaction level modeling
creation in the purpose of early verification of hardware Function/Method :
modeling. The first process output is a TLM model that fulfill Do_add();
the design requirements. Memory_Read();
Memory_write();
39 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
• Arbiter is bus management algorithm, such as: round • Sc_in_clk clock; (added)
robin.
• Sc_port<sc_signal_out_if<bool> > grantM1; (deleted)
• PM is a process in master. PM is the more detail definition
of in the form of pseudo code. 2. Process addition and deletion:
In transaction level modeling, data transfer process and • Sel_mux_master1 (added)
control from master to slave are conducted by accessing a bus • grantM1process (deleted)
controlled by an arbiter. Each master can deliver request of bus
access to send data or read data from slave. There will be 3. Total Master and Slave determination:
several possible conditions achieved by master; they are bus • Determining the amount of rows added and reduced of
condition is OK if bus is not being operated by other masters or all systems.
WAIT condition if bus is being used by other master or
ERROR condition if targeted slave is not around in slave list. 4. Arbitrary determination
• Wishbone protocol arbitrary is Round robin
3) TLM – RTL Transformation • Every master sends request of slave access. If there
After finishing early verification process and being met are several masters requesting access of one similar
with given specification, then the last stage is transformation slave, then the arbiter will give an access for the
from TLM into RTL. The purpose of the transformation is to master and send waiting signal for other masters.
generate detail model available for synthesis. Phases of TLM 5. Port mapping of all modules: master and slave
into RTL model transformation can be divided into several
general stages; those are: • Mapping of master post to all multiplexers.
• Port addition and deletion: in the process of TLM • Mapping of multiplexer post to slave and the master.
modeling, there should be ports that are required to delete, • Mapping of multiplexer post to slave and the master.
because basic principle is not needed in RTL model, such
port request. Meanwhile, it is necessary to add new ports
in RTL model for performing detail process, as the nature
of RTL modeling. D. Criteria and Measurement
There are two criteria used for measuring new system
• Process addition and deletion: In spite of ports addition experiment, they are :
and deletion, it is also necessary to add and delete process.
Example of process that must be deleted from RLM is 1. Design performance improvement (Te)
such process that tries to send request, while addition Performance improvement is characterized by the decrease
process that should be given in RTL model is process of of time required to design embedded system. Design period by
accessing multiplexer. using new method will be compared with RTL design period.
• Total Master and Slave determination: Total master and Time difference needed to design the same systems from two
slave is used to make pattern of RTL bus. Total master and different methods will be considered the success of new design.
slave can influence total multiplexers and types of New design system is considered success if design period
multiplexer. Multiplexer for 4 masters applies the first needed is shorter than the previous time design. General
mux4 while 2 masters apply the first mux2. formula of design performance improvement (Te) is Te =
TRTL/TTLM. TRTL is design process time for modeling RTL
• Determining arbitrary (according to given protocol) model and TTLM is design process time for modeling TLM.
Arbitrary is management algorithm of slave access when
the access is from one master or more. Example of 2. Target of Criteria: Accuracy level (α)
algorithm used is round robin, such as in Avalon bus. Design model difference can be conducted in the purpose of
• Port mapping : The last stage of transformation is improving performance. This can be accepted if both models
connecting all ports from all components available along can bring about the same output for the same input. The closer
with additional components, such as multiplexer, detail, the result of both systems to the same input is, the more
pin-per-pin. accurate the system is. Accuracy level (α) = P(Input) –
P(Input) ≈ 0.
4) Examples of TLM-RTL Transformation
The followings are examples of transformation from TLM IV. EXPERIMENT AND RESULTS ANALYSIS
to RTL by using RTL bus with Wishbone.
Bus target: Wishbone A. Avalon and Wishbone Bus for on Chip System (SoC)
Avalon and Wishbone bus are busses for SoC. The bus is
1. Port addition and deletion: designed for chip-based application. SoC is a compact system
40 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
with three kind components, master and slave components and On each testing scenario, testing model is generated in
bus system. transaction level modeling and RTL. Both of the models will
be compared based on the line amount required to implement.
Avalon and Wishbone buses are used in implementation Testing is conducted starting from simple system, consisting of
stage in the level of RTL. There are 5 main components in a master and a slave. Then, testing component complexity will
Avalon bus along with each function as follows: be ignored periodically by adding the amount of tiers
1. Master : active components which have initiative to continuously.
perform data access either read() or write().
2. Slave : passive component waiting for data access from
master.
3. Logic Request : components managing access requests
from master for slave. each component has one logic
request component.
4. Logic Arbitrator: component managing access of one slave
according to request of one master of more. Each slave has
one Logic Arbitrator to manage the slave access.
5. Multiplexer : component for managing access of a slave
according to request of Logic Arbitrator. There are 5
multiplexers for each slave; mux address, mux BE_n, mux
write, mux writedata and mux read. There is one mux for
data displayed for master; mux master.
Figure 9. Wishbone bus architecture
Tiers are general terms of displaying embedded system
components which communicate each other, for example, 2-
tiers means there are two components communicating each
others, 3-tiers means there are 3 components communicating
each others, and so on.
Figure 8. Avalon bus architecture
Figure 10. multi master-slave system
There are 5 main components in Wishbone bus include :
master, slave, decoder, round robin arbiter and multiplexor. C. TLM-RTL Model Testing
The function of each component the same as on Avalon bus.
a. Line of Code of TLM-RTL Model
The testing involves two, three, four and five components,
B. Testing Scenario those are master, slave and arbiter. Master component actively
Test is performed to measure performance increase of generates and sends data to slave, while slave serves as
design by using new design flow of transaction level modeling receiver.
(TLM – Transaction Level Modeling) compared to Level
Register Transfer modeling (RTL – Register Transfer Level). Based on the experiment results, it suggests that the amount
Some testing scenario will be conducted in testing process of lines needed to model system by using TLM is less than that
involving several master components, slave and arbiter. of using RTL. Such condition can take a place because master,
41 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
slave, and bus definition on RTL is more detail than that on b. Measurement is based on design time (man days).
TLM.
Design time refers to time needed by programming for code
generation and report. Standardarization used is 8 line codes
per man day.Design time can be directly decreased from the
total modeling code for each TLM and RTL modeling. Figure
12 shows comparison result of time needed to design the four
testing scenario.
Measurement of Design Process Performance (Te)
Performance is one of the important parameters to measure
the success of new method. In this dissertation, design process
performance can be measured according to the comparison
between design process times needed by using RTL model
compared with TLM model.
The measurement of performance improvement of design
process can be conducted by using the following equation:
Figure 11. Line of Code comparison T(e) = Trtl / Ttlm
Detail level of modeling on RTL can influence several parts Based on the experiment result conducted as shown in
of program, including: Figures 11 and 12, it can be concluded that performance
improvement graph as shown in Figure 13 can be obtained.
• Port definition of each master and slave component.
As shown in Figure 13 it indicates that design process
• Initial definition of top level system, including port performance increases as the increase of amount of component
addition, instantaneousness, port mapping and destruction. on the embedded system except for case study of 2-tiers whose
performance is higher than that of 3-tiers.
• New component definition called as multiplexor. On
Avalon bus of each slave addition, 6 new multiplexor shall
be added accordingly.
In TLM, component definition process and port mapping is
simpler than that of in RTL, so it does not require many
instructions as compared to RTL. New port mapping when
master addition takes a place is a mapping with bus and clock.
Figure 13. Design performance improvement
One of the advantages of TLM modeling is that transaction
will occur among components. The more components the
Figure 12. Design process time comparison system is, the higher increase of transaction level by the use of
bus is. Transaction improvement of components is very
appropriate to TLM modeling. In RTL level modeling, in the
42 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
contrary, the amount of components and transactions of a [2] Chatha, Karamvir Sigh. “System-Level Cosynthesis of
system will make the design difficulty higher. Transformative Application for Heterogeneous Hardware-
Software Architecture”. Dissertation at University of Cincinnati.
Therefore, total components and transaction in the 2001.
modeling are getting bigger and bigger, and thereby design [3] Cornet, Jerome. “Separation of Functional and Non-Functional
process will get slower. Based on the condition mentioned Aspects in Transactional Level Models of Systems-on-Chip”.
Dissertation at Institut Polytechnique De Grenoble. 2008.
above, design process on TLM modeling is getting better and
[4] Cummings, Clifford. “SystemVerilog’s priority & Unique – A
better under the circumstances that there are many components Solution to Verilog’s full_case & parallel_case Evil Twins”.
making interaction each others. Those are the advantages of SNUG. Israel. 2005.
TLM level modeling taking a place in transaction level. [5] Frank Vahid and Tony Givargis. “Embedded system A Unified
Hardware/Software Introduction”. JohnWiley & Sons, Inc., New
York, 2002.
V. CONCLUSION
[6] Genovese, Matt. ”A Quick-Start Guide for Learning SystemC”.
Based on the testing shown in the previous chapter, it can The University of Texas. Austin. 2004. 15
be concluded that there are several important things, including: [7] Gordon E. Moore. “Cramming more components onto integrated
circuits”. Electronics, 38(8):114-117, 19 April 1965.
1. The new embedded system design flow can be used to [8] Leung, Julie. Kern, Keith. Dawson, Jeremy. “Genetic
increase design process performance. It means that using Algorithms and Evolution Strategies”.
this method the design process will shorter than RTL [9] Mathaikutty, D., A. (2007) : Metamodeling Driven IP Reuse for
modeling with performance improvement compare to RTL System-on-chip Integration and Microprocessor Design,
Avalon bus are 1.03, 1.47, 1.69 for 3,4 and 5 tiers Dissertation at Virginia Polytechnic Institute and State
respectively. The performance improvement compare to University.
RTL Wishbone bus are 1.12, 1.17 and 1.34 for 3,4 and 5 [10] Mooney III, Vincent John. “Hardware/Software co-design of
tiers respectively. run-time systems”. Dissertation at Stanford University. 1998.
[11] Palnitkar, Samir. “Verilog® HDL: A Guide to Digital Design
2. TLM level modeling will be better implemented in a and Synthesis, Second Edition”. Sun Microsystems. Inc.
complex system, in the condition of more than two California. 2003.
components having interactions in which there occurs [12] Patel, Hiren D. “Ingredients for Successful System Level
arbitrary process. Automation & Design Methodology”. Dissertation at Virginia
Polytechnic Institute and State University. 2007.
Contribution of this paper are the new embedded system [13] _____, “Ptolemy II Project”. UC. Berkeley. 2008.
design flow by using transaction level modeling approach and [14] _____. (2002) : Describing Synthesizable RTL in SystemC™,
standard procedure to design hardware RTL model. Synopsys, Inc., Version 1.2, November 2002.
www.synopsys.com
In the effort of constructing an integrated framework [15] _____. (2003) : Avalon Bus Spesification : Reference Manual,
ranging from specification till model construction that is Altera. : www.altera.com
available to be synthesized, the following discussion shall be
conducted in the next research:
1. Automation process of all standard procedure
AUTHORS PROFILE
2. Design of parts of system software and application
3. Integrated process between software and hardware of Maman Abdurohman is a PhD student at STEI faculty of Bandung Institute of
Technology. He is working at Informatics faculty of Telecom Institute of
software. Techonolgy – Bandung. His primary areas of interest include embedded
system design and microcontroller. Maman has an master degree from
By adding the three parts of processes, then new framework Bandung Institute of Technology. Contact him at mma@ittelkom.ac.id
will soon be generated under embedded system design. The
initial framework includes systematic steps of embedded
Kuspriyanto is a Professor at STEI faculty of Bandung Institute of
system construction. The next process is automation of the Technology. He is a senior lecturer in computer engineering laboratory.
whole processes. His major areas of interest include digital system and electronic design.
His job is Head of Laboratory of Computer Engineering. Contact him at
kuspriyanto@yahoo.com
ACKNOWLEDGMENT
Maman Abdurohman thanks to the Faculty of Informatic IT Sarwono Sutikno is a Associate Profesor at STEI faculty of Bandung Institute
Telkom and Faculty of STEI Electro Bandung Institut of of Technology. His major areas of interest include cryptography and
Technology for their financial support and research resources embedded system design. His job is a director in PPATK on electronic
so that this research could be completed. transaction control. Contact him at ssarwono@gmail.com
Arif Sasongko is a lecturer at STEI faculty of Bandung Institute of
REFERENCES Technology. His major areas of interest include wimax design and
embedded system design. His current project is designing highspeed
[1] Berger, Arnold S. “Embedded System Design : An Introduction data link wimax system. Contact him at asasongko@gmail.com
to Processes, Tools, and Techniques”. CMP Books. 2002.
43 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Semi-Trusted Mixer Based Privacy Preserving
Distributed Data Mining for Resource Constrained
Devices
Md. Golam Kaosar Xun Yi, Associate Preofessor
School of Engineering and Science School of Engineering and Science
Victoria University Victoria University
Melbourne, Australia Melbourne, Australia
md.kaosar@live.vu.edu.au xun.yi@vu.edu.au
Abstract— In this paper a homomorphic privacy preserving Rapid development of information technology, increasing
association rule mining algorithm is proposed which can be use of advanced devices and development of algorithms have
deployed in resource constrained devices (RCD). Privacy amplified the necessity of privacy preservation in all kind of
preserved exchange of counts of itemsets among distributed transactions. It is more important in case of data mining since
mining sites is a vital part in association rule mining process. sharing of information is a primary requirement for the
Existing cryptography based privacy preserving solutions accomplishment of data mining process. As a matter of fact the
consume lot of computation due to complex mathematical more the privacy preservation requirement is increased, the less
equations involved. Therefore less computation involved privacy the accuracy the mining process can achieve. Therefore a trade-
solutions are extremely necessary to deploy mining applications
off between privacy and accuracy is determined for a particular
in RCD. In this algorithm, a semi-trusted mixer is used to unify
the counts of itemsets encrypted by all mining sites without
application.
revealing individual values. The proposed algorithm is built on
with a well known communication efficient association rule In this paper we denote Resource Constrained Device
mining algorithm named count distribution (CD). Security proofs (RCD) as any kind of device having limited capability of
along with performance analysis and comparison show the well transmission, computation, storage, battery or any other
acceptability and effectiveness of the proposed algorithm. features. Examples includes but not limited to mobile phones,
Efficient and straightforward privacy model and satisfactory Personal Digital Assistants (PDAs), sensor devices, smart
performance of the protocol promote itself among one of the cards, Radio Frequency Identification (RFID) devices etc. We
initiatives in deploying data mining application in RCD. also interpret lightweight algorithm as a simple algorithm
which requires less computation, low communication overhead
Keywords- Resource Constrained Devices (RCD), semi-trusted
and less memory and can be deployed in a RCD. Integration of
mixer, association rule mining, stream cipher, privacy, data mining.
communication devices of various architectures lead to global
I. INTRODUCTION heterogeneous network which comprises of trusted, semi-
trusted, untrustworthy, authorized, unauthorized, suspicious,
intruders, hackers types of terminals/devices supported by
Data mining sometimes known as data or knowledge fewer or no dedicated and authorized infrastructure. Sharing
discovery is a process of analyzing data from different point of data for data mining purposes among such resource constrained
views and to deduce into useful information which can be ad-hoc environment is a big challenge itself. Preservation of
applied in various applications including advertisement, privacy intensifies the problem by another fold. Therefore
bioinformatics, database marketing, fraud detection, e- privacy preserving data mining in RCD envisions facilitating
commerce, health care, security, sports, telecommunication, the mining capability to all these tiny devices which may have
web, weather forecasting, financial forecasting, etc. a major impact in the market of near future.
Association rule mining is one of the data mining techniques
which helps discovering underlying correlation among
different data items in a certain database. It can deduce some Data mining capability of RCD would flourish the future
hidden and unpredictable knowledge which may provide high era of ubiquitous computing too. Owner of the device would
interestingness to the database owners or miners. perform mining operation on the fly. Small sensor devices
would be able to optimize or extend their operations based on
the dynamic circumstance instead of waiting for time
consuming decision from the server. Scattered agents of a
44 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
security department can take instant decision of actions about a Some research approaches address the issue of hiding
crime or a criminal while in duty. To comprehend the necessity sensitive information from data repository. In [23] and [24]
of lightweight privacy preserving data mining, let us consider authors basically propose some techniques to hide sensitive
another circumstance: there are many scattered sensor devices association rules before the data is disclosed to public. A
located in a geographical location belonging to different hardware enhanced association rule mining technique is
authorities which are serving different purposes with some proposed in [25]. Data is needed to be fed into the hardware
common records about the environment. Now if it is required before the hash based association rule mining process starts.
to mine data among those sensor devices to accomplish a This approach may not be well realistic for RCD because it
common interest of the authorities in real time, then preserving requires special purpose hardware as well as it does not
privacy would be the first issue that must be ensured. Another handle privacy issue. A homomorphic encryption technique;
motivation behind developing our proposed system could be Paillier encryption is used by X. Yi and Y. Zhang [9] to
healthcare awareness. Let us assume some community preserve privacy where authors propose a privacy preserving
members or some university students want to know about the distributed association rule mining using a semi-trusted mixer.
extent of attack of some infectious diseases such as swine flu, This algorithm involves lot of computation due to the use of
bird flu, AIDS etc. Each individual is very concerned about the complex mathematical equations and big prime numbers as
privacy since the matter is very sensitive. They are equipped keys in the Paillier encryption.
with a mobile phone or similar smart device and want to know
the mining result on the fly. In such circumstances, a A heterogeneous mobile device based data collection
distributed lightweight privacy preserving data mining architecture is proposed by P.P. Jayaraman [10]. Sensor
technique would provide a perfect solution. In addition to that; devices are scattered in the environment to collect various data
relevant people can be warned or prescribed based on all whereas regular mobile phones can work as bearers of the data.
available health information including previously generated Detail architecture of the model is available in [10]. Authors
knowledge about a particular infectious diseases. did not consider the privacy issue during the transmission of
data. If the mobile devices in the environment are intended to
There is not much research work done for lightweight be utilized to work as a data bearer then privacy should be one
privacy preserving data mining but there is plenty of research of the major concerns. Therefore it would be difficult to be
on privacy preserving data mining. Essentially two main implementable in real life unless privacy is preserved. A
approaches are adapted for privacy preserving data mining lightweight privacy preserving algorithm similar like in this
solutions. First one is the randomization which is basically used paper could provide privacy preservation as well as data
for centralized data. In this approach data is perturbed using mining solution for these kinds of models.
randomization function and submitted for mining.
Randomization function is chosen such that the aggregated Main focus of CD [18] algorithm is to reduce
property of the data can be recognized in the miner side. In [1, communication overhead with the cost of redundant parallel
2, 3] authors have proposed such approaches. One of the major computation in each data site. In addition to that this algorithm
drawbacks of randomization approach is: if the precision of does not transmit the large itemset in the association rule
data mining result is increased, the privacy is not fully mining process. Rather it communicates the counts of the
preserved [4]. itemsets only, which let it reduce communication overhead
dramatically. These features make it feasible to be deployed in
Another one is the cryptographic approach in which the RCD. On the other hand semi-trusted mixer based privacy
data is encrypted before it is being shared. The miner cannot solution provided by Yi and Zhang in [9] requires lot of
decrypt individual inputs separately rather it needs to decrypt computation with managing big encryption key size. In this
unified encrypted data together. Therefore the miner cannot paper a more efficient semi-trusted mixer and homomorphic
associate particular information to a particular party. An encryption based privacy algorithm is proposed which adopts
example of such approach is Secure Multiparty Computation the rule mining technique of CD to make the solution
(SMC) proposed by Yao [5]. Another cryptography based deployable in RCD.
privacy preservation technique is proposed by M. Kantarcioglu
and C. Clifton [6] which involves enormous amount of The remainder of the paper is oriented as follows: Section 2
mathematical computation and communication between data describes some necessary background information. Section 3
sites. This is too heavy to be implemented in a RCD. Among describes proposed solution which consists of privacy
other privacy preserving data mining, [7] and [8] are ones preserving algorithm and association rule mining algorithm for
which also involve vast mathematical complex equations to be RCD. Section 4 contains security analysis and section 5
solved. There are some research works on privacy issues for discusses the proofs and performance comparison. Finally the
RCD separately too. Authors in [21] propose a technique to conclusion is presented in section 6.
hide location information of a particular device for location
based applications. A middleware LocServ is designed which
lies in between the location-based application and the location II. BACKGROUND
tracking technology. A group signature based privacy for
vehicles is proposed in [22], which addresses the issue of Privacy: According to The American Heritage Dictionary
preserving privacy in exchanging secret information such as privacy means “The quality or condition of being secluded
vehicle’s speed, location etc. from the presence or view of others”. In data mining if the
45 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
owner of the data requires the miner to preserve privacy, then Stream Cipher: It is a symmetric key cipher where
the miner gets no authority to use the data unless an acceptable plaintext bits are combined with a pseudorandom cipher bit
and trustworthy privacy preservation technique is ensured. stream typically by an XOR operation. In stream cipher a seed
Different points of views define privacy in different ways. For is used as a key to generate continuous stream of bits. This
simplicity we consider a definition which is most relevant to idea can be used in generating random keys by encrypting a
this work. According to J.Vaidya [1] privacy preserving data constant with the secret key/seed. Therefore multiple
mining technique must ensure two conditions: ‘any randomly generated keys can be shared among multiple
information disclosed cannot be traced back to an individual’ entities simply by sharing a seed. In our proposed algorithm
and ‘any information disclosed does not constitute an we need some randomly generated keys which can be
intrusion’. More technical definition of privacy can be found generated by Output Feedback Mode (OFB) of Data
in [11]. This paper also provides technical definition in Encryption Standard (DES) detail of which is available in
security analysis in section 4. [13].
Homomorphic Encryption: Homomorphic encryption
is a special form of encryption using which one can perform a
Association Rule Mining: Let us consider; in a distributed specific algebraic operation on the plaintext by performing the
data mining environment collective database DB is subdivided same or different operation on the ciphertext. Detail definition
into DB1, DB2, … , DBN in wireless data sites S1, S2, … ,SN could be found in [13]. If x1 and x2 are two plaintext and E
respectively. I= {i1, i2, … , im} is the set of items where each and D denotes encryption and decryption function
transaction T⊆I. Typical form of an association rule is X⇒Y, respectively. Let us consider y1 and y2 are two ciphertexts
where X⊆I, Y⊆I and X∩Y=φ. The support s of X⇒Y is the such that: y1=Ek(x1) and y2=Ek(x2) where, k is the encryption
probability of a transaction in DB contains both X and Y. On key. This encryption will be considered homomorphic if the
the other hand confidence c of X⇒Y is the probability of a following condition is held: y1+y2=Ek(x1+x2).
transaction containing X will contain Y too. Usually it is the
interest of the data vendor to find all association rules having III. PROPOSED SOLUTION
support and confidence greater than or equal to minimum
threshold value. For another instance of an association rule
In this paper we propose a privacy preserving secret
AB⇒C, computation protocol which is based on a homomorphic
encryption technique for distributed data sites. In this section
⇒C
first the privacy preserving frequency mining algorithm is
discussed and then the modified CD algorithm is discussed
which ensures privacy in the association rule mining process.
A. Privacy Preserving Frequency Mining
In our proposed approach, there would be a number of
⇒C participating semi honest devices or data sites (>2) which are
⇒C
connected to each other using heterogeneous media. There
would be a semi-trusted mixer which would receive encrypted
More detail on association rule mining process is
count values from sites through its private channel. It is
available in [12, 20].
assumed that the semi-trusted mixer would never collude with
Association rule mining process consists of two major any of the data site. In practice it could be assumed that it is
parts. First one is to find frequent large itemsets which have authorized by the government or semi-government agent. Data
support and confidence values more than a threshold number sites communicate to the mixer through the private channel
of times. Second part is to construct association rules from and the mixer communicates to all sites through public
those large itemsets. Due to the simplicity and straightforward channel. Necessary keys would be distributed to the sites by
nature of the second part, most of the association rule mining corresponding authorities or owner of the sites. It is also
papers do not address this. Apriori algorithm is one of the assumed that the private channel is protected by a standard
leading algorithms, which determines all frequent large secret key cryptosystem, such as DES [15] or AES [16]. Fig.1
itemsets along with their support counts from a database describes the proposed model in brief.
efficiently. This algorithm was proposed by Agrawal in [14]
which is discussed here in brief:
Let us say Li be the frequent i-itemsets. Apriori algorithm
finds Lk from Lk-1 in two stages: joining and pruning:
Joining: Generates a set of k-itemsets Ck, known as
candidate itemsets by joining Lk-1 and other possible items in
the database.
Pruning: Any (k−1)-itemsets cannot be a subset of a
frequent k –itemset which is not frequent. Therefore it should
be removed.
46 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
S
S
(1) MN (1) M1
(1) Mi (1) M2 Fig.2: Random key generation from stream cipher for each
Si Mixer S iteration.
(1) M3
It is already mentioned that, each data sites communicate
(2) є’ to the mixer through a private channel and the mixer
communicates to all sites through public channel.
S Communication stages of the algorithm are depicted in the
flow diagram of fig.3.
α1
Broadcast channel
Private channel between each site and mixer α2
Broadcast channel between all sites and mixer
…
Fig.1: Privacy preserving communication and computation
process between data sites and the mixer αi
Each Site
Mixer
In our proposed model we also assume no site would
collude with the mixer to violate other’s privacy since this
…
would reveal privacy of itself too. In this model the privacy
would be preserved if (1) the coalition of N-2 sites would not
certain a revelation of privacy of any site and (2) mixer can αN
learn nothing about the distributed database.
Let us consider; there are N resource constrained sites S1,
S2 … SN want to share the summation of a specific secret value ε’
of their own without disclosing the value itself. The secret
values are c1, c2 … cN respectively. cij denotes the value
belongs to site i for jth iteration (in case of association rule
mining it would be jth itemset).
Secret parameters: Let us assume ρ is a large prime Fig.3: Flow diagram of the algorithm
N Step 1: (Encryption)
number such that, ρ . Stream cipher seed
is µ. These ρ and µ are shared by all the sites using any key 1.1 Each site Si computes rj following above mentioned
agreement algorithm similar to one proposed in [17]. In fact constraints
there will not be any effect if ρ is disclosed to the mixer. There
are two more parameters r and n which are generated from a 1.2 Encodes its count : α
stream cipher in which the seed µ is used as key and any 1.3 Then Si sends αi using mixer’s private channel
constant (may be ρ) as a plaintext. In each repetition the values Step 2: (Mixing)
of r and n will be different due to the characteristics of the
stream cipher. Parameter r is chosen such that . If it is 2.1 The mixer receives αi in its private channel (for all
assumed that the length of r and n are l bits then total number i=1 to N).
of bits in each chunk in the stream will be: l+N.l = l(1+N). 2.2 Adds them all α together: ′
α
First l bits would be the value of r, second l bits for ni which is
a random number allocated for ith site for communication 2.3 Broadcasts ′
back to all participating sites.
purpose. In every iteration the value of ni would be different
(similar to the value of nonce used in various cryptosystems). Step 3: (Decryption)
Thus for jth site nj will be allocated from bit l+j.l to l+(j+1).l. 3.1 Each participating site Si receives ε′ .
Following figure (Fig.2) describes the allocation of values of r
and n from the stream cipher. The length of l should be chosen 3.2 Si already had computed . It gets the sum
such that following constrained is held: . of the current iteration j by computing ε′
N
Where, .
47 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Thus , the sum of the count is shared among previous iteration using Apriori algorithm (discussed in
all sites without revealing individual count. An example of the section 2).
algorithm is provided in the following section for more (2) Count computation: Si passes over all the transactions
clarification. in DBi to compute the count for all items in Lk-1.
B. Example (3) Transmission of counts: Counts of Ck is sent to the
For simplicity let us consider three sites S1, S2 and S3 mixer using privacy preserving communication techniques
want to share the sum of their count values {5, 7 and 6} discussed in subsection 3.1. Communication between the data
without revealing their own values among themselves. Other sites and the mixer is performed through the private channel.
shared and generated secret parameters are: ρ=91, r=23, The value of j in the algorithm (subsection 3.1) maps to the
n1=17, n2=11, n3=10 and (mod 91). To minimize itemsets sequence number.
complexity, values of r and ni are not calculated from the (4) Mixer functions: Mixer adds all the encrypted counts
stream cipher, rather their values are chosen spontaneously. received from all the sites and broadcasts the result back to all
Also let us assume the values of r-1 are the same instead of sites.
different for each site. Communication between sites and the
mixer is performed using private channel which is not (5) Result decryption: Each data site decrypts the result
depicted in this example too. received from the mixer as it is stated in section 3.1 to get sum
of the counts.
Exchange of count values: Each site transmits it’s to
the mixer using private channel. (6) Termination: Since all sites perform identical
operation, all of them terminate at the same iteration and end
up with generation of large itemset.
IV. SECURITY ANALYSIS
The mixer computes In this section we demonstrate that our proposed
protocol preserves privacy during the transmission of counts
′
of itemsets in association rule mining process. With the basis
′ of privacy requirement and security definition provided in [9,
is received in all sites. Sites calculate sum of counts T:
19], following formulation can be addressed.
Let us assume N≥3, since privacy preservation is
impossible for less than three parties. VIEW(Si, N) implies
view of the party Si where total number of participants is N.
Thus T is equal to the intended sum of {5, 7 and 6}. Similarly VIEW(M,N) implies the view of the mixer.
C. Association Rule Mining Therefore by definition VIEW(M,0), VIEW(Si,0), VIEW(Si,1)
and VIEW(Si,2) all equal to Φ. If X and Y are two random
Among many association rule mining algorithms we variables then,
choose the one which focuses on reduction of communication
cost; Parallel Mining of Association Rules [18]. In this paper X≈polyY = (the probability of distinguishing X and Y)
authors have proposed three algorithms for the ≤ for all polynomials Q(l) [9]. N parties want to find
accomplishment three different objectives. Count Distribution the sum of their counts of itemset c1, c2 … cN. The privacy will
(CD) is one of them which aims to reduce the communication be preserved if following conditions are satisfied [9].
cost at the cost of parallel redundant computation in each data
site. In this subsection we would integrate our proposed (a) Two random variables
privacy preserving communication technique with CD and
algorithm which would be suitable for RCD in terms of are polynomially indistinguishable
computation and communication. (AN,j≈polyBN,j) for 1≤j≤N and 0≤R<ρ.
Since frequent large itemset computation is considered as (b) Two random variables
the major task in association rule mining algorithms, we focus and
our effort for the accomplishment of the same task as it is the are polynomially
case in many other papers. Following are the notations, major indistinguishable (CN,j≈polyDN,j) for n ≥ 3, 1 ≤ j ≤ n-2
stages and actions performed in each data site in every cycle: and 0≤R<ρ.
Let, Si: Data site (site) of index i. N: Number of sites. Since all users have identical values of
DBi: Database (collection of transactions) in Si. Lk: Set of
frequent k-itemset. Ck: Set of candidate k-itemset and , …
. , they are the same.
(1) Candidate set generation: Each site Si generates a Theorem 1: The proposed protocol preserves privacy
complete candidate set Ck from Lk-1 which is computed in the based on the above mentioned privacy definition.
48 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Proof: (a) When N=1, then j=1 and can decrypt the outer encryption of the double encrypted
=(α,c1). ciphertext. It cannot decrypt or read the secret value of Si.
Mixer only adds all the ciphertexts together and broadcasts the
Since the mixer does not know the secret parameters (ρ, µ) it result to all sites in step 2. Now the sum is known to all
cannot decrypt α. Therefore parties. They all can decrypt it which is a summation of their
α α . secret values. Therefore none can reveal or relate any secret
When N>1 and 1≤j≤N value associated to any site.
Theorem 3: Security against the mixer and any other
individual participant or outsider
Proof: Unlike any other kind of regular security protocols
our proposed protocol has neither a straight forward sender
nor a receiver. Rather it involves encryption of different
contents with different keys by multiple senders, a mixer and
multiple receivers together in a complete single
communication. The senders send in the first step and receive
in the third step. Moreover each transaction in this protocol is
consists of multiple communication attempts, which make the
[Since
protocol different and more secure compared to other
] protocols. Let us study the vulnerability in following cases:
Replay attack: If an eavesdropper gets all the
(b) When n=3, j=1. Therefore communications between all sites and the mixer, he cannot
learn anything significant about the secret value of an
individual party. Because in every communication the value of
nj chosen randomly in step 1.2 of the algorithm, which would
raise the high degree of unpredictability of the data in the
channel.
With given c1 and , party S1 cannot be certain
about c2. Therefore, Brute force attack: Again due to the frequent and random
change of value of nj in each communication, brute force
= attack is unrealistic.
When N>3 and 1≤ j ≤ n-2,
V. PERFORMANCE ANALYSIS
Yi-Zhang’s [9] privacy preserving association rule mining
algorithm uses semi-trusted mixer which is similar to our
′ ′ ′ proposed model. We compare the performance of the
Let us assume proposed protocol with Yi-Zhang protocol. To measure and
′
compare the performance between these two protocols, let us
Since assume following parameters:
H= Average number of items in the large k-itemset.
′ ′ ′ ′ ′ L= Size of each entry in the list of large itemsets to store
index and count in Bytes.
N= Number of data sites.
K= Average size of each item in Byte (number of
characters as for example).
Therefore the privacy is preserved for the proposed φ=Encryption ratio ( ) in step 1.2 in the
protocol. proposed algorithm.
Theorem 2: The protocol does not reveal support count of | αi|=Size of αi in step 1.2 in the proposed algorithm.
one participant to either the mixer or to other participants. |є'|= Size of є' in step 2.3 in the proposed algorithm.
Proof: In step 1.2 of the algorithm, each site Si encrypts Proposed algorithm: Communication payload in each
the secret value using private keys which are only known to iteration is
sites. Before the ciphertext is transmitted to the private N*| αi|*H+|є'|*H*N = N*φ*L*H+φ*H*N = φHN(1+L)
channel of the mixer, it is farther encrypted using the public
key of the mixer in step 1.3. None has the private key of the In case of φ=1, Communication overhead= HN(1+L).
mixer except the mixer itself; therefore no eavesdropper can
get access to the ciphertext. On the other hand the mixer only
49 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Yi-Zhang algorithm [9]: Let us assume same encryption Measure Our Proposed Yi-Zhang
ratio (that is same φ and β') in both level of encryptions. Protocol Protocol
Communication payload in each iteration is:
Communication 3NH 6NH
N*(Cipher-text sent by each site) + Data broadcasted by the overhead (each
mixer = N*φ*H*K+N* φ*H*K=2φNHK. iteration)
If φ=1, Communication overhead= 2NHK. Number of 0 4
For farther comparison let us assume value of L=2 (two exponential
bytes to store two values: index and count) and K=3 (on an operations
average). Therefore communication payload in our proposed Key size 80 1024
algorithm and Yi-Zhang’s algorithm are 3NH and 6NH bytes
respectively. Therefore the proposed algorithm generates as Table 1: Performance comparison between Yi-Zhang and the
much as half communication payload of the Yi-Zhang proposed protocol.
algorithm.
Though there is no use of exponent operations in the
Let us now compare the number of instructions necessary proposed algorithm, it involves some other cryptographic
in encrypting and decrypting a message m. We compare only operations which would be efficient enough due to small key
the homomorphic encryption involved in both Yi-Zhang and size. Therefore the performance comparison shows that the
the proposed protocol. Basic encryption and decryption proposed algorithm is more efficient and straightforward,
equations of Yi-Zhang protocol are: which make it suitable for RCD.
Encryption: and VI. CONCLUSION
λ
Decryption: λ
Rapid development and increasing popularity of
Where, m: the message, c: the ciphertext, N: pq (p and q ubiquitous computing and RCD in the environment demands
are large prime numbers), g: public key, r: a random number. the deployment of varieties of lightweight applications. A
lightweight algorithm which would lead one step ahead to
Therefore number of operations involved for encryption deploy data mining applications in RCD is proposed in this
and decryption are: paper. All security protocols involve detail consideration of
Exponential operations: 1+1+1+1=4 various security threats. But our proposed model can avoid
many security threats such as replay attack, brute force attack
Basic operations: 1+1+1+1+1+1+1+1+1+1+1+1+1=13 etc, due to the nature of the protocol itself. This is so because
in this protocol a single communication is not consists of
In case of the proposed protocol, basic encryption and simply between a sender and a receiver rather it involves
decryption equations are (as stated in section 3.1): multiple senders, receivers and the mixer all together. All the
Encryption: and secret parameters and keys in our proposed homomorphic
encryption technique are very small in size; therefore less
Decryption: ρ computation is involved in the encryption and decryption
Where, r and n: random numbers, ρ: prime number > sum process. This feature makes the proposed algorithm more
of counts of items. For the sake of measuring the operations suitable for RCD. Performance analysis and proofs of privacy
count, we treat and r as the same. and security also imply the strength and appropriateness of the
algorithm. Therefore this effort should be considered as one of
Therefore number of basic instructions involved in the effective initiative towards the deployment of data mining
encryption and decryption are: in ubiquitous computing environment.
Exponential operations: 0 REFERENCES
Basic operations: 1+1+1+1=4
[1] R. Agrawal, R. Srikant, “Privacy-preserving data mining”, Proceedings
Finally let us compare the size of the keys used in both
the protocols: of ACM SIGMOD International Conference on Management of Data,
2000, pp. 439–450.
In the proposed protocol the key is considered as the seed
µ of the stream cipher. The size of µ can be considered as a [2] A. Evfimievski, “Randomization in privacy preserving data mining”,
typical one: 80 bits. ACM SIGKDD Explorations Newsletter, Volume 4 Issue 2, December
Yi-Zhang protocol: It is mentioned in [9] that for security 2002, pp. 43 – 48
concern the value of N should be such that . [3] A. Evfimievski, S. Ramakrishnan, R. Agrawal, J. Gehrke, “Privacy-
Therefore size of N is 1024 bits. preserving mining of association rules”, 8th ACM SIGKDD
All the performance comparisons between Yi-Zhang and International Conference on Knowledge Discovery and Data Mining,
the proposed protocol are summarized in table 1. ACM Press, 2002, pp. 217–228.
50 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
[4] H. Kargupta, S. Datta, Q. Wang, K. Sivakumar, “On the privacy [18] R. Agrawal and J.C. Shafer, “Parallel Mining of Association Rules”,
preserving properties of random data perturbation techniques” 3rd Int’l Knowledge and Data Engineering, IEEE Transactions on Volume 8,
Conference on Data Mining, 2003, pp. 99–106. Issue 6, Dec. 1996 pp. 962-969.
[5] A.C. Yao, “How to generate and exchange secrets”, 27th IEEE [19] W.G. TZeng, “A secure fault-tolerant conference key agreement
Symposium on Foundations of Computer Science, 1986, pp. 162–167. protocol”, IEEE Transactions on Computers vol. 51 issue 4, April 2002,
[6] M. Kantarcioglu, C. Clifton,”Privacy-preserving distributed mining of pp. 373 – 379.
association rules on horizontally partitioned data”, Knowledge and Data [20] P-N. Tan, M. Steinbach, V. Kumar, “Introduction to Data Mining”, 1st
Engineering IEEE Transaction Volume 16, Issue 9, Sep. 2004 pp. 1026- Edition, ISBN/ISSN: 0321321367, 2006.
1037. [21] G. Myles, A. Friday, N. Davies, “Preserving privacy in environments
[7] Y. Lindell, B. Pinkas, “Privacy Preserving Data Mining”, Journal of with location-based applications”, Pervasive Computing IEEE, vol. 2,
Cryptology, Volume 15 - Number 3, 2002. issue 1, Jan-Mar 2003, pp 56-64.
[8] Z. Yang, S. Zhong, R. N. Wright, “Privacy-Preserving Classification of [22] J. Guo, J. P. Baugh, S. Wang, “A Group Signature Based Secure and
Customer Data without Loss of Accuracy”, Proceedings of the Fifth Privacy-Preserving Vehicular Communication Framework” 2007 Mobile
SIAM International Conference on Data Mining, Newport Beach, CA, Networking for Vehicular Environments, May 2007, pp 103-108.
April 21-23, 2005. [23] V.S. Verykios, A.K. Elmagarmid, E. Bertino,Y. Saygin, E. Dasseni,
[9] X. Yi, Y. Zhang, “Privacy-preserving distributed association rule mining “Association rule hiding”, Knowledge and Data Engineering, IEEE
via semi-trusted mixer”, Data and Knowledge Engineering, vol. 63, no. Transactions, Volume 16, Issue 4, April 2004, pp. 434 – 447.
2, 2007. [24] B.J. Ramaiah, A. Reddy, M.K. Kumari,” Parallel Privacy Preserving
[10] P.P. Jayaraman, A. Zaslavsky, J. Delsing, “Sensor Data Collection Using Association rule mining on pc Clusters”, 2009 IEEE International
Heterogeneous Mobile Devices”, Pervasive Services, IEEE International Advance Computing Conference (IACC 2009), March 2009, pp. 1538 –
Conference, Istambul, 15-20 July 2007 pp. 161-164. 1542.
[11] J. Vaidya, C. Clifton, M. Zhu, “Privacy Preserving Data Mining”, [25] Y.H. Wen, J.W. Huang, M.S. Chen, “Hardware-Enhanced
Springer 2006, ISBN-13: 978-0-387-25886-8. Association Rule Mining with Hashing and Pipelining”, IEEE
[12] J. Han, M. Kamber, “Data Mining Concepts and Techniques”, Second Transactions on Knowledge and Data Engineering, vol. 20, no. 6, June
Edition, Elsevier Inc. 2006. ISBN: 13:978-1-55860-901-3. 2008.
[13] J. Katz, Y. Lindell, “Introduction to Modern Cryptography”, Taylor &
Francis Group, LLC, 2008. ISBN: 13: 978-1-58488-551-1.
[14] R. Agrawal and R. Srikant, “Fast algorithms for mining association AUTHORS PROFILE
rules,” in Proceedings of the 20th International Conference on Very
Md. Golam Kaosar is a PhD student at the School of Engineering and
Large Data Bases. Santiago, Chile: VLDB, Sept. 12-15 1994, pp. 487– Science, Victoria University, Australia. Before he starts his PhD, he used to
499. work as an engineer at Research Institute (RI) in King Fahd University of
Petroleum and Minerals (KFUPM), Saudi Arabia. Before that he got his MS
[15] NBS FIPS PUB 46, Data Encryption Standard (National Bureau of in Computer Engineering and BSC in Computer Science and Engineering
from KFUPM, and Bangladesh University of Engineering and Technology
Standards, US Department of Commerce, 1977).
(BUET), Bangladesh at the years 2006 and 2001 respectively.
[16] FIPS PUB 197, Advanced Encryption Standard (Federal Information As a young researcher, he has a good research background. He has published
Processing Standards Publications, US Department of number of conference papers including IEEE and some journals. His area of
research includes but not limited to Privacy Preserving Data Mining,
Commerce/N.I.S.T., National Technical Information Service, 2001). Ubiquitous Computing, Security and Cryptography, Ad-hoc sensor network,
[17] S. B. Wilson, A. Menezes, “Authenticated Diffie-Hellman Key Mobile and Wireless Network, Network Protocol, etc.
Agreement Protocols”, Lecture Notes in Computer Science, Springer-
Verlag Berlin Heidelberg, ISBN- 978-3-540-65894-8, January 1999, pp.
339-361
51 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
ADAPTIVE SLOT ALLOCATION AND
BANDWIDTH SHARING FOR PRIORITIZED
HANDOFF CALLS IN MOBILE NETWOKS
S.Malathy G.Sudhasadasivam K.Murugan S.Lokesh
Research Scholar Professor, CSE Department Lecturer, IT Dept Lecturer,CSE Dept
Anna University PSG College of Technology Hindusthan Institute of Tech Hindusthan Institute of Tech
Coimbatore Coimbatore Coimbatore Coimbatore
joymalathy@gmail.com
Service (QoS) is required to manage the incoming new
calls and handoff calls more efficiently. The Geographical
Abstract - Mobility management and bandwidth management
are two major research issues in a cellular mobile network.
Mobility management consists of two basic components:
area is divided into smaller areas in the share of hexagon.
location management and handoff management. To Provide
These hexagonal areas are called as cells. A Base Station
QoS to the users Handoff is a key element in wireless cellular
(BS) is located at each cell. The Mobile Terminals (MT)
networks. It is often initiated either by crossing a cell boundary
or by deterioration in the quality of signal in the current within that region is served by these BS. Before a mobile
channel. In this paper, a new admission control policy for user can communicate with other mobile user in the
cellular mobile network is being proposed. Two important QoS network, a group of channels should be assigned. The cell
parameter in cellular networks are Call Dropping Probability size plays a major role in the channel utilization. A user has
(CDP) and Handoff Dropping Probability (HDP). CDP to cross several cells during the ongoing conversation, if
represents the probability that a call is dropped due to a handoff the cell size is small. During the ongoing conversation, the
failure. HDP represents the probability of a handoff failure due
call has to be transferred from one cell to another to
to insufficient available resources in the target cell. Most of the
achieve the call continuation during boundary crossing.
algorithms try to limit the HDP to some target maximum but not
CDP. In this paper, we show that when HDP is controlled, the
CDP is also controlled to a minimum extent while maintaining Here comes the role of handoff. Transferring the active
lower blocking rates for new calls in the system. call from one cell to another without disturbing the call is
called as the process of Handoff. Hand is otherwise a
Index Terms— Wireless Cellular Networks, Handoff “make before break” process. Time slot, Frequency band,
Dropping Probability, Call Dropping Probability, Resource or code word to a new base station [1] may be the terms of
Allocation, Prioritization Schemes. call transfer from a cell to another.
1. INTRODUCTION A typical Cellular network is shown in figure 1. A
Due to the increased urge to use the wireless limited frequency spectrum is allocated. But it is very
communication in a satisfied way, a promised Quality of successfully utilized because of the frequency reuse
52 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
concept. To avoid the interference while neighboring cells
CHANNEL
are utilizing the same frequency, the group of channels ALLOCATION
STRATEGY
assigned to one cell should be different from the
neighboring cells. The MTSO scans the residence of the CAC HANDOFF POWER
CONTROL
MS and assigns the channel to that cell for the call.
MULTIPLE ACCESS
SYSTEM
If the MS is travelling while the call is in progress, the
MS need to get a new channel from the neighboring BS to FIGURE 2 RESOURCE MANAGEMENT IN CELLULAR
continue the call without dropping. The MSs located in the NETWORKS
cell share the available channels. The Multiple Access
Call Admission Control denotes the process of admitting a
Methods and channel allocation schemes govern the
fresh call or a handoff call based on the availability of
sharing and allocating the channels in a cell, respectively.
resources.
II LITERATURE SURVEY
Various handoff schemes proposed [2] are Guard
channel scheme (GCS), Handoff based on Relative Signal
Strength [4], Handoff based on Relative Signal Strength
with threshold, Handoff based on Relative Signal Strength
with Hysteresis and threshold [3], Handoff based on
MTSO
Prediction techniques [5]. When MS moves from one cell
to another, the corresponding BS hands off the MSs Call to
PSTN
the neighbor. This process is done under the control of
FIGURE 1 CELLULAR NETWORK MTSO. The handoff in initiated based on various
parameters like signal strength received from BS, travelling
The Scenario of a basic cellular network is depicted in speed of the MS etc.
Figure1.
A handoff method based on the kinds of state
The resource management in the cellular system deals information [6] that have been defined for MSs, as well as
with CAC, Utilization of Power and channel allocation the kinds of network entities that maintain the state
strategy. The channel allocation strategy may be Fixed or information has been devised. The handoff decision may be
Dynamic. The resource allocation is shown in Figure 2. made at the MS or network. Based on the decision, three
types of handoff may exist namely, Network-Controlled
Handoff, Mobile-Assisted Handoff, and Mobile-Controlled
Handoff [7]. Handoff based on Queuing is analyzed [7] for
voice calls. The Queue accommodates both the originating
calls and handoff requests [9]. Handoff schemes with two-
level priority [10] have been proposed. How the non-real-
53 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
time service has to be incorporated and its effect needs to get the service. The priority is more for the handoff calls
be taken into consideration is proposed [11]. A new two- than the originating calls.
dimensional model for cellular mobile systems with pre-
emptive priority to real time service calls [12] is proposed. The following assumptions are made over the calls.
In [13] the concept of prioritization of handoff calls over a) The arrival pattern of the calls follows the Poisson
new calls since it is desirable to complete an ongoing call process.
rather than accepting a new one is employed. b) The cell consists of N Channels. If free channels
exist, both the calls will be served. If channels are
In [14], a situation where the handoff calls are queued not available, then the originating calls will be
and no new calls are handled before the handoff calls in the dropped.
queue is presented. By combing guard channel and queue c) Priority is given to the handoff calls on based on
schemes performs better [15]. [16] developed a non- the call dwell time in the cells. The priority is low
preemptive prioritization scheme for access control in for a longer dwell time calls than the shorter calls.
cellular networks. The channel holding time is assumed to have
exponential distribution.
III. SYSTEM DESCRIPTION
New calls
If users request connection to the base station at the same
λO CHo 2 1
time, the system checks the type of origin of the call. The N
1st priorityhandoff calls channe
handoff decision may be made by the MS or the network
λh COC 2 1 ls
based on the RSS, Traffic pattern, Location management
2nd priorityhandoff calls
etc., while handoff is made the channel assignment plays λh CHhf 2 1
µ
an important role. The total channels in the BS can be
allocated to different types of calls. If the originating calls
and handoff calls are treated in the same way, then the
FIGURE 3 QUEUEING CALLS
request from both kinds are not served if there are no free
channels.
d) Two types of Queues are assumed. The queue for
In another scheme, Priority is given to the handoff call handoff calls QHC and queue for originating calls
request by reserving a minimum number of channels to the QOC respectively.
handoff call. If there is N number of channels available, the e) If no channels are available the handoff calls are
R number of channels is reserved to the handoff calls and queued in QHC, whose capacity is CHC .The
the remaining (N-R) channels are shared by the handoff originating calls are queued in QOC, only if the
and originating call requests. The handoff call request is available channels at the time of arrival are less
dropped only if there are no channels available in the cells. than (N-R). The originating call is blocked if the
To overcome this drawback of dropping the handoff calls, queue is full.
our system proposes an new model of queuing scheme in f) Queue is cleared if the call is completed or the
with the handoff calls and originating calls are queued to user moves away from the cell.
54 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
g) The capacity CHC of QHC is large enough so that
∞ ∞
blocking probability of the handoff call is ∫ e−μHt dt = ∫ ((1 −λn FTHn(t) − λn FTHh(t))dt (1)
neglected. 0 0 λ λ
where FTHn(t) and FTHh(t) are actual distribution of
The channel holding time TH can be calculated by using channel holding time for new and handoff calls. [17]
the following formula
FIGURE 3 CHANNEL ALLOCATION ALGORITHM
IV RESULTS channels for handoff calls of real time traffic gets shared
dynamically shared by handoff calls of non-real-time
In this paper, a dynamic sharing of channels for the traffic. The comparison between a normal bandwidth
handoff calls and new calls has been proposed. In the reservation scheme and the proposed model is simulated. It
proposed scheme, when there is no channels, the reserved is shown that, the call blocking probability as well as the
55 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
Channel Utilization Full Reduced
handoff dropping probability is reduced when compared Traffic Management Controlled Controlled
to the traditional algorithms even when traffic is increased. Call Dropping Reduced Reduced
TABLE 1 COMPARISON BETWEEN Probability
Call Blocking Not Decreased Decreased as
EXISTING & PROPOSED SCHEMES
Probabilty well
Parameter Existing Proposed
Scheme Scheme
The New Call Blocking Probability and the Handoff Call
Dropping Probability with an increase in call arrival rate in
a cell is reduced when compared to the traditional
algorithm.
IV CONCLUSION
In this paper, we have showed that by integrating the
concept of buffering and dwell time of the call, the New
RESULT 1
Call blocking probability and handoff call dropping
BANDWIDTH UTILIZATION VERSUS CALL
probability has been considerably reduced. In future, this
ARRIVAL RATE
work can be extended for different types of calls and
integrated services like data and images.
The above graph shows that by adopting the new
algorithm the bandwidth utilization is considerably
REFERENCES:
increased with the increase in call rate.
[1] S. Tekinay and B. Jabbari, “Handover and channel assignment
in mobile cellular networks,” IEEE Commun. Mag., vol. 29, no.
11, 1991, pp. 42-46.
[2] I. Katzela and M. Naghshineh, “Channel assignment schemes
for cellular mobile telecommunication systems: A comprehensive
survey,”
IEEE Personal Communications, pp. 10-31,June 1996.
[3] Gregory P. Pollini, “Trends in Handover Design,” IEEE
Communication Magazine, March 1996, pp. 82–90.
[4] N. Zhang, Jack M. Holtzman, “Analysis of Handoff
Algorithms Using Both Absolute and Relative Measurements,”
IEEE Trans. Vehicular Tech., vol. 45, no. 1, pp. 174-179,
February 1996.
[5] Shian-Tsong Sheu, and Chin-Chiang Wu, “Using Grey
prediction theory to reduce handoff overhead in cellular
RESULT 2 communication systems”, The 11th IEEE International
Symposium on Personal, Indoor and Mobile Radio
CALL BLOCKING PROBABILITY VERSUS CALL Communications, (PIMRC 2000), vol. 2, pp. 782-786, 2000.
[6] N. D. Tripathi, J. H. Reed, and H. F. Vanlandingham, Handoff
ARRIVAL RATE in Cellular Systems, IEEE Personal
56 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
Commun., December 1998. Professor in Department of Computer Science and
[7] Handoff in Wireless Mobile Networks, QING-AN ZENG and Engineering in PSG College of Technology, India. Her
DHARMA P. AGRAWAL, media.wiley.com areas of interest include, Distributed Systems,
[8] Guerin R, “Queuing Blocking System with Two Arrival
Distributed Object Technology, Grid and Cloud
Streams and Guard Channels”, IEEE
Transactions on Communications, 1998, 36:153-163. Computing. She has published 20 papers in referred
[9] Zeng A. A, Mukumoto K. and Fukuda A., “Performance journals and 32 papers in National and International
Analysis of Mobile Cellular Radio System with Priority Conferences. She has authored 3 books. She has
Reservation Handoff Procedure”, IEEE VTC-94, , Vol 3, 1994, coordinated two AICTE – RPS projects in Distributed
pp. 1829-1833. and Grid Computing areas. She is also the coordinator
[10] Zeng A. A, Mukumoto K. and Fukuda A., “Performance for PSG-Yahoo Research on Grid and Cloud computing.
Analysis of Mobile Cellular Radio You may contact her at sudhasadhasivam@yahoo.com
System with Two-level Priority Reservation Procedure”, IEICE
Transactions on Communication, Vol E80-B, No 4, 1997, pp.
598-607.
[11] Goodman D. J, “Trends in Cellular and Cordless
Communication”, IEEE Communications Magazine, Vol. 29, No.
6, 1991, pp.31-40.
[12] Pavlidou F.N, “Two-Dimensional Traffic Models for
Cellular Mobile Systems”, IEEE Transactions on
Communications, Vol 42, No 2/3/4, 1994, pp. 1505-1511.
[13] Jabbari B. & Tekinay S., “Handover and Channel
Assignment in Mobile Cellular Networks”, IEEE
Communications Magazine, 30 (11),1991, pp.42-46. Mr.K.Murugan is currently a research
[14] Sirin Tekinay, “A Measurement-Based Prioritization Scholar in Karpagam University Coimbatore. He has a
Scheme for Handovers in Mobile Cellular Networks”, IEEE teaching experience of 15 years.He has presented various
JSAC, Vol. 10, 1992, pp. 1343-1350. papers in national and international conferences. His
[15] Hou J and Fang Y., “Mobility-Based call Admission Control research areas include Mobile networks, Grid Computing,
Schemes for Wireless Mobile Networks”, Wireless Data Mining.
Communications Mobile Computing, 2001, 1:269-282.
[16] Novella Bartolini, Handoff and Optimal Channel Assignment
in Wireless Networks”, Mobile Networks and Applications, 6,
2001, pp. 511-524.
[17 ] Kundan Kandhway, “Dynamic Priority Queueing of
Handover Calls in Wireless Networks: An Analytical
Framework” Mr.S.Lokesh is currently a research Scholar in Anna
University Trichy.. He has a teaching experience of 5
years.He has presented nearly 6 papers in national and
Author Profile: international conferences. His research areas include
Mobile networks, Digital Image Processing, Signal
Processing
Ms.S.Malathy is currently a research
Scholar in Anna University Coimbatore. She has presented
nearly 10 paper in national and international
conferences.Her research areas include Mobile networks
and wireless communication.
Dr G Sudha Sadasivam is working as a
57 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
A New Vein Pattern-based Verification System
Mohit Soni Sandesh Gupta M.S. Rao Phalguni Gupta
DFS, UIET, CSJMU, DFS, IIT Kanpur,
New Delhi, INDIA Kanpur, UP, INDIA New Delhi, INDIA Kanpur, UP, INDIA
myk.soni@gmail.com Sandesh@iitk.ac.in msrnd@rediffmail.com pg@iitk.ac.in
Abstract— This paper presents an efficient human recognition Venal patterns, on the other hand, have the potential to
system based on vein pattern from the palma dorsa. A new surpass most such problems. Apart from the size of the pattern,
absorption based technique has been proposed to collect good the basic geometry always stays the same. Unlike fingerprints,
quality images with the help of a low cost camera and light veins are located underneath the skin surface and are not prone
source. The system automatically detects the region of interest to external manipulations. Vein patterns are also almost
from the image and does the necessary preprocessing to extract impossible to replicate because they lie under the skin surface
features. A Euclidean Distance based matching technique has [6].
been used for making the decision. It has been tested on a data set
of 1750 image samples collected from 341 individuals. The It seems, the first known work in the field of venal pattern
accuracy of the verification system is found to be 99.26% with has been found in [10]. Badawi [1] has also tried to establish
false rejection rate (FRR) of 0.03%. the uniqueness of vein patterns using the patterns from the
palma dorsa. The data acquisition technique mentioned in [1] is
Keywords- verification system; palma dorsa; region of interest; based on a clenched fist holding a handle to fixate the hand
vein structure; minutiae; ridge forkings during image capture. This method however, has limitations
with respect to orientation. Substantial works in this field have
I. INTRODUCTION been done by Leedham and Wang [11] [12] [13]. In these,
Vein pattern of the palma dorsa can be defined as a random thermal imaging of the complete non fisted hand has been done
‘mesh’ of blood carrying tubes. The back of the hand veins are using Infrared light sources. Generally, near infra-red lamps of
not deeply placed and hence these can be made visible with the intensity-value ranging from 700 to 900 nm in wavelength are
help of a good image acquisition system and technique. The used to design such a system [12]. These lamps are found to be
geometry of these veins is found to be unique and universal costly. Also infra-red light has been used to either reflect or
[14]. Hence, it can be considered as one of the good human transmit light to the desired surface [8] [11] [12] [14]. These
recognition systems. techniques have both their advantages and disadvantages. It has
been observed that the images captured through a reflection
Forensic scientists have always been the biggest reapers of based system, as proposed in [11], would never produce
successful biometric systems. User authentication, identity consistent results owing to excessive noise generated due to
establishment, access control and personal verification etc are a unnecessary surface information. The surroundings have to be
few avenues where forensic scientists employ biometrics. Over controlled at all times and the skin color or skin abnormalities
time various biometric traits have been used for the above are bound to have an effect. The best results can only be
mentioned purposes. Some of them have gained and lost expected after exposure from the near infra-red lamps which
relevance in the course of time. Therefore, constant evolution are costly. A system of capturing images from the front of the
of existing traits and acceptance of new biometric systems is hand has been proposed in [14]. The palm prints may interfere
inevitable. The existing biometric traits, with varying with the pattern of the veins, in this case.
capabilities, have proven successful over the years. Traits like
Face, Ear, Iris, Fingerprints, Signatures etc., have dominated Matching technique in a biometric system is a crucial step
the world of biometrics over the years. But each of these because the accuracy of the system alone can determine its
biometric traits has its shortcomings. Ear and iris pose a effectiveness. There exist various matching techniques for
problem during sample collection. Not only is an expensive proving the individuality of a source. Badawi has used a
and highly attended system required for iris but it also has a correlation based matching algorithm and achieved excellent
high failure to enroll rate. In case of ear data, it is hard to results. However, correlation-based techniques, though the
capture a non occluded image in real time environment. most popular, become costly on larger databases. Wang and
In case of the most well known face recognition systems there Leedham used a matching technique based on Hausdorff
exist some limitations like aging, background, etc [2]. distancing, which is limited in principle by the slightest change
Fingerprints, though most reliable, still lack automation and in orientation.
viability as they are also susceptible to wear and aging. This paper proposes an efficient absorption based technique
Signatures, are liable to forgery. for human identification through venal patterns. It makes use of
a low cost sensor to acquire images. It uses a fully automated
foreground segmentation technique based on active contouring.
Reduced manual interference and an automatic segmentation
58 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
technique guarantee uniform segmentation of all samples, • The focal length and the exposure time of the camera
irrespective of their size and position. The paper also employs a lens.
rotation and translation invariant matching technique. It is also
realized that since the collected images are very large in size In our experiment, a simple digital SLR camera combined
(owing to a high definition camera) slow techniques like with an infra-red filter has been used for data acquisition. Also
correlation-based matching would hinder the overall efficiency. it makes use of a low cost night vision lamp of wavelength 940
Therefore, the proposed system uses critical features of veins to nm. The proposed set-up is a modest wooden box with a
make the decision. The results, thus far, have been found to be hollow rod lodged in the middle accommodating the infra-red
encouraging and fulfilling. lamp. The camera is fixed at a perpendicular to the light source
pre-adjusted to a fixed height of 10 inches above the light
Section 2 presents the experimental setup used to acquire source. The camera is held on a tripod attached to the box. The
images. Next section deals with the proposed venal pattern robustness and the flat face of the night vision lamp provides
based biometric system. It tries to handle some of the critical for a sturdy plinth for the subject’s hand. The sensor here is
issues such as use of a low cost sensor for acquiring images, kept on the opposite side of the light source as shown in Fig. 2.
automatic detection of region of interest, rotation and This design has specific advantages. The subject has to place
translation invariant matching etc. This system has been tested his palm on the plinth surface, to provide image. If the camera
on the IITK database consisting of 1750 image samples is not able to pick up the pattern the attendant can immediately
collected from 341 subjects in a controlled environment, over rectify the hand position. The image can be captured only when
the period of a month. Experimental results have been analyzed the camera can pick up the veins.
in Section 3. Concluding remarks are given in the last section.
II. PROPOSED SYSTEM
Like any other biometric system, the venal pattern based
system consists of three major tasks and they are (i) image
acquisition (ii) preprocessing of acquired image data (iii)
feature extraction (iv) matching. The flow diagram of the
proposed system is given in Fig. 1.
Figure 2. Experimental Setup
The setup prepared for the proposed system is not only cost
effective but also meets the requirement for good quality data
acquisition. It is found that through this camera along with the
mentioned light the veins appear black. The light source is
placed behind the surface to be captured. This helps to make an
ideal data scheme and standard, as all parameters can be fixed.
Unlike [8], [11], [12] and [14] where infra-red light has been
reflected or transmitted to the desired surface, this paper
Figure 1. Flow Diagram of the Proposed System proposes an absorption-based technique to acquire images. The
proposed technique provides a good quality image regardless
A. Image Acquisition of the skin color or, any aberrations or discolorations, on the
surface of the hand. In this technique the veins pop out when
Performance of this type of system always depends on the the hand is fisted and it becomes much easier to capture high
quality of data. Since venal data is always captured under a contrast images. The time and cost of image processing
controlled environment, there is enough scope to obtain good therefore can be kept to a minimum. Since the veins are
quality data for veins. Parameters required for getting good illuminated from the behind and captured from the other side,
quality of data are carefully studied in making the setup and are any anomalies in the skin of the palm (including the natural
given below: palm lines) would not interfere in the pattern. The image
• The distance between the camera and the lamp. capturing however would be limited by anomalies on the palma
dorsa itself, like tattoos etc. On the other hand, skin color or the
• The position of the lamp. gradual change of it (due to diseases or sunlight, etc.) or the
• The fixed area for the placement of the hand. gain and loss of weight would not hamper the pattern collection
process.
• The orientation of the lamp once clenched within the
palm.
59 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
Since the light illuminates the entire hand, it is a common are spatial co-ordinates, typical external energy can be defined
notion that the veins in the front of the hand might interfere as follows to lead the contour towards edges:
with the pattern at the back. However, it is crucial to note, that
infra-red light does not make the hand transparent. It simply
2
illuminates the hemoglobin in the veins, which appear black. Eext = − ∇I ( x, y)
The partition of the bone between the two planes in the front
and the back of the hand, does not allow interference. And (2)
since the sensor is always facing the dorsal surface, it is the where ∇ is gradient operator. For color images, we estimate
only surface to be captured. the intensity gradient which takes the maximum of the
gradients of R, G and B bands at every pixel, using:
The only factor due to which an inconsistency can occur
during image acquisition is the size of a subject’s hand since
there is no control over the size and thickness of a subject’s ∇ I = max( ∇ R , ∇ G , ∇ B )
hand in a practical scenario. Therefore, the exact distance (3)
between the object and the camera’s lens can never be pre-
determined or fixed. To handle this situation, a necessary The gradient obtained using the above equation gives
arrangement has been made in the setup. The focal shift of the better edge information. An active contour that minimizes
camera which can be fine tuned to the order of millimeters Econtour must satisfy the following Euler equation:
ensures the relative prevalence of the desired conditions.
B. Data Pre-Processing η1v" (s) − η 2 v iv ( s) − ∇Eext = 0
(4)
The color image acquired through a camera generally
contains some additional information which is not required to where v”(s) and v””(s) are the second and fourth order
obtain the venal pattern. So there is a need to extract the region
derivatives of v(s). The above equation can also be viewed as a
of interest from the acquired image and finally to convert into a
force balancing equation, Fint + Fext = 0 where,
noise free thinned image from which one can generate the
venal tree. Badawi [1] has considered the skin component of
the image as the region of interest (ROI). Wang and Leedham Fint = η1v" ( s) − η 2 v iv ( s)
[11] used anthropometric points of a hand to segregate an ROI (5)
from the acquired images. Most similar works based on ROI
and
selection employ arbitrary and inconsistent techniques and so
end up enhancing manual intervention during processing [8]. Fext = −∇E ext
This extracted region is used for further processing to obtain (6)
the venal pattern. This section presents the method followed to
extract the ROI and then to obtain the venal tree. It consists of
Fint, the internal force is responsible for the stretching and
four major tasks and they are (i) Segmentation (ii) Image
bending and Fext, the external force, attracts the contour
Enhancement and Binarization (iii) Dilation and
Skeletonization and (iv) Venal pattern generation. towards the desired features in the image. The active contour
deforms itself with time to exactly fit around the object. It can
The segmentation technique used in this paper to segregate thus be represented as a time varying curve.
the skin area from the acquired image selects the ROI in a
systematic manner and it also, successfully gets rid of all v(s, t) = [x(s, t), y(s, t)]
manual intervention. It fuses a traditional technique based on (7)
active contouring [5] with a common cropping technique. It where s [0, 1] is arc length and t R+ is time.
works on the principle of intensity gradient, where the user
initializes a contour around the object, for it to detect the Active contouring helps the contours to settle at the object
boundary of the object easily. A traditional active contour is boundary. It is then followed by the iterative use of a cropping
defined as a parametric curve v(s) = [x(s), y(s)],s [0, 1], tool which helps extract the object automatically and al most
which minimizes the following energy functional. flawlessly (3). It can be noted, that the active contouring snake
has been modified from its regular run. Instead of initiating the
11 snake from the outside in, it is run in reverse, after initiating it
Econtour = ∫ (η1 v' (s) +η2 v" (s) ) + Eext (v(s)).ds
2 2
from the centre of the image. This initiation is always done
02
(1) automatically.
The extracted ROI of the colored image is converted into a
where η1 and η2 are weighing constants to control the relative grey scale image by the technique given in [3] as shown in Fig.
importance of the elastic and bending ability of the active 3.The segmented grey scale has been enhanced using Gaussian
contour respectively; v’(s) and v′′(s) are the first and second filtering technique and is then normalized by converting it to an
order derivatives of v(s), and Eext is derived from the image so image having a pre defined mean and variance. The resultant
that it takes smaller values at the feature of interest that is image is then binarized by mean filtering. However, it may
edges, object boundaries etc. For an image I(x, y), where (x, y) contain noises like salt and pepper, blobs or stains, etc. Median
60 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
filtering is used to remove salt and pepper type noises. again to remove all disconnected isolated components from the
Eventually a grey scale image has been denoised and binarized final skeleton.
as given in Fig. 4.
C. Feature Extraction
This section presents a technique which extracts the
forkings from the skeleton image by examining the local
neighborhood of each ridge pixel using a 3X3 window. It can
be seen from the preprocessing image that an ROI contains
some thinned lines/ridges. These ridges representing vein
patterns can be used to extract features. Features like ridge
Figure 3. Automatically Segmented Image forkings are determined by computing the number of arms
originating from a pixel. This can be represented as A. The A
The image may consist of a few edges of the vein pattern for a pixel P can be given as:
that may have been falsely eroded during filtering. These edges
are reconnected by dilation, i.e., running a disk of ascertained
8
radius over the obtained pattern. Then these obtained images
are skeletonized. Each vein is reduced to its central pixel and
A = 0.5∑ | Pi − Pi +1 |,P9 = P1
i =1
their thickness is reduced to 1 pixel size only. A skeletonized
image can hence, be obtained (see Fig. 5). (8)
For a pixel P, its eight neighboring pixels are scanned in an
anti-clockwise direction as follows:
P4 P3 P2
P5 P P1
P6 P7 P8
A given pixel P is termed as a ridge forking for a vein
Figure 4. Enhanced and Binarized Grey Scale Image pattern if the value of A for the pixel is 3 or more. This ridge
forking pixel is considered as a feature point which can be
In order to obtain only desired components amongst veins, defined by (x, y, θ) where x and y are coordinates and θ is the
all connected components are labeled and others are discarded. orientation with respect to a reference point.
The CCL (Connected Component Labeling) algorithm [6] is
modified to determine all the connected components in an
image. This modified algorithm detects and removes all
isolated and disconnected components of size less than a
specified threshold.
Figure 6. Four Arms emitting from a forking point
Figure 5. The Binarized image can be processed to give the vein skeleton in
the hand The proposed method for calculating A can accommodate
three or four or more arms emitting out of a forking point.
From the skeleton of the hand, the skeletonized veins are Cases where four arms emit from a forking point are common,
extracted. A vertical and a horizontal line (one pixel thick) are as shown in Fig. 6. Fig.7 shows the final image of the extracted
run through each coordinate of each image alternatively. The vein pattern with all forking points marked.
coordinates of the first and the last image pixels encountered by
the line, in both axes, are stored. These coordinates were later
turned black and the venal tree was extracted. The modified
connected component labeling (CCL) algorithm is executed
61 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
n
∑ A[i ]
V = i =1
X 100 %
n
(9)
If V is more than a given threshold then one can draw the
conclusion that both the patterns are matched.
III. EXPERIMENTAL RESULTS
The proposed system has been tested against the IITK
database to analyze its performance. The database consists of
Figure 7. The Final Vein Pattern with all Forking Points Marked 1750 images obtained from 341 individuals under controlled
environment. Out of these 1750 images, 341 are used as query
D. Matching Strategy samples. A graph is plotted for the achieved accuracy against
Suppose, N and M are two patterns having n and m features the various threshold values as shown in Fig.8. It is observed
respectively. Then the sets N and M are given by: that the maximum accuracy of 99.26% can be achieved at the
threshold value, T, of 25. Graphically, it is also found in Fig. 9
that the value of FRR for which the system achieves maximum
N= {(x1, y1, θ1), (x2, y2, θ2), (x3, y3, θ3), …, (xn, yn, θn)} accuracy is 0.03%. Finally, the ROC curve is taken between the
values of GAR and the FAR is given in Fig. 10.
M={(a1, b1, φ1), (a2, b2, φ2), (a3, b3, φ3),..., (am, bm, φm)}
where (xi, yi, θi) and (aj, bj, φj) are the corresponding
features in pattern N and M respectively. For a given minutiae
(xi,yi,θi) in N, it first determines a minutiae (aj, bj, φj) such that
the distance ( xi − a j ) 2 + ( yi − b j ) 2 is minimum for all j,
j=1,2,3 ...,m. Let the distance be sdi and the corresponding
difference between two directions be ddi, where
ddi = θi − ϕ j . Figure 8. Graph Accuracy
This is done for all features in N. To avoid the selection of
same feature in M for a given minutiae in N, one can follow the
following procedure. Suppose, for the ith feature in N, one gets
sdi for the jth feature in M. Then, in order to determine sdi+1,
one considers all features in M which are not selected in sd1,
sd2….sdi. Let A be a binary array of n elements satisfying
A[ i ] = {
1 if ( sdi ≤ ti ) and ( ddi ≤ t 2 )
0 otherwise
Figure 9. Graph indicating FAR and FRR
where t1 and t2 are predefined thresholds. The threshold values
defined by t1 and t2 are necessary to compensate for the
unavoidable errors made by feature extraction algorithms and
to account for the small plastic distortions that cause the
minutiae positions to change. These are thresholds determined
by averaging the different feature shifts based on intensive
testing.
Then the percentage of match is obtained for the pattern N
having n features against the pattern can be computed by
62 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
[10] Sang-Kyun Im, Hyung-Man Park, Young-Woo Kim, Sang-Chan Han,
Soo-Won Kim and Chul-Hee Kang “Improved Vein Pattern Extracting
Algorithm and Its Implementation” in Journal of the Korean Physical
Society, 38/3/pp. 268-272 (2001).
[11] L. Wang,C.G Leedham, “A Thermal Hand Vein Pattern Verification
System” in Pattern Recognition and Image Analysis, of Lecture Notes in
Computer Science, Springer 3687/pp. 58–65 (2005).
[12] L. Wang, C.G. Leedham, “Near and Far-infrared Imaging for Vein
Pattern Biometrics” in Proceedings of IEEE International Conference on
Advanced Video and Signal Based Surveillance/pp. 52 (2006).
[13] L.Wang, G. Leedham, D. Siu-Yeung Cho, “Minutiae feature analysis for
infrared hand vein pattern biometrics” in Pattern Recognition
41/3/pp.920-929 (2008).
[14] M. Watanabe, T. Endoh, M. Shiohara, and S. Sasaki “Palm Vein
Figure 10. The ROC Curve- GAR v/s FAR Authentication Technology and its Applications”, Proceedings of the
Biometric Consortium Conference, September, 2005.
IV. CONCLUSION
AUTHORS PROFILE
This paper has proposed a new absorption based vein
pattern recognition system. It has a very low cost data
acquisition set up, compared to that used by others. The system Mohit Soni graduated from the Delhi University with an honors degree in
has made an attempt to handle issues such as effects of rotation Botany and Biotechnology. He received his Masters degree in Forensic
and translation on acquired images, minimizing the manual Science from the National Institute of Criminology and Forensic Sciences,
New Delhi. Thereafter he received a research fellowship from the Directorate
intervention to decide on the verification of an individual. It of Forensic Sciences, New Delhi in 2006 and is currently pursuing his
has been tested in a controlled environment and against a Doctoral degree in Biometrics and Computer Science from the Uttar Pradesh
dataset of 1750 samples obtained from 341 subjects. The Technical University, Lucknow.
experimental results provide an excellent accuracy of 99.26%
with FRR 0.03%. This is found to be comparable to most
previous works [2] [11] [12] [13] and is achieved through a Sandesh Gupta received his Bachelors in Technology from the University
technique which is found to be much simpler. Institute of Engineering and Technology, C.S.J.M University Kanpur in 2001.
He is working currently as a lecturer for the computer science department in
the same institution and is pursuing his PhD from the Uttar Pradesh Technical
REFERENCES University, Lucknow.
[1] A.Badawi “ Hand Vein Biometric Verification Prototype: A Testing
Performance and Patterns Similarity” in Proceedings of the International
Conference on Image Processing, Computer Vision, and Pattern M S Rao is a well known forensic scientist of the country and started his
Recognition/pp. 3-9 (2006). career in Forensic Science in the year 1975 from Orissa Forensic Science
Laboratory. He carried extensive R&D work on Proton Induced X-Ray
[2] R. de Luis-Garcia, C. Alberola-Lopez, O. Aghzoutb, Ruiz-Alzola, Emission (PIXE) in Forensic Applications during 1978-1981. He was
Biometric Identification Systems, Signal Processing, 83/pp. 2539-2557 appointed as Chief Forensic Scientist to the Government of India in 2001. He
(2003). was Secretary and Treasurer for the Indian Academy of Forensic Sciences
[3] R.C Gonzalez, R.E Woods Digital Image Processing using MATLAB, from 1988 to 2000 and is now the President of the Academy. He was convener
Prentice Hall, 1st Edition, 2003. of the Forum on Forensic Science of the Indian Science Congress during 1992
[4] U. Halici, L.C Jain, A. Erol “Introduction to Fingerprint Recognition” in and 2001. He is the Chairman of the Experts Committee on Forensic Science.
Intelligent Biometric Techniques in Fingerprint and Face
Recognition/pp. 3–34 (1999).
[5] M. Kass, A. Witkin, D. Terzopoulos, “Snake: Active Contour Models” Phalguni Gupta received the Doctoral degree from Indian Institute of
in International Journal of Computer Vision, 1/5/ pp. 321-331 (1988). Technology Kharagpur, India in 1986. Currently he is a Professor in the
[6] I. Khan, “Vein Pattern Recognition – Biometrics Underneath the Skin” Department of Computer Science & Engineering, Indian Institute of
in Article 320 on www.findbiometrics.com (2006). Technology Kanpur (IITK), Kanpur, India. He works in the field of
biometrics, data structures, sequential algorithms, parallel algorithms, on-line
[7] V. Khanna, P. Gupta, C.J Hwang, “Finding Connected Components in algorithms. He is an author of 2 books and 10 book chapters. He has
Digital Images by Aggressive Reuse of Labels” in International Vision
published more than 200 papers in International Journals and International
Computing. 20/8/ pp.557-568 (2002).
Conferences. He is responsible for several research projects in the area of
[8] C. Lakshmi Deepika, A. Kandaswamy, “An Algorithm for Improved Biometric Systems, Image Processing, Graph Theory and Network Flow.
Accuracy in Unimodal Biometric Systems through Fusion of Multiple
Feature Sets” in ICGST-GVIP Journal 9/3/pp.33-40 (2009).
[9] D. Maltoni, D. Maio, A.K Jain, S. Prabhakar, “Handbook of Fingerprint
Recognition”, Springer, New York (2003).
63 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Extending Logical Networking Concepts in Overlay
Network-on-Chip Architectures
Omar Tayan
College of Computer Science and Engineering, Department of Computer Science,
Taibah University, Saudi Arabia, P.O. Box 30002
Email: otayan@taibahu.edu.sa
Abstract—System-on-Chip (SoC) complexity scaling driven by [1-7]. This study summarizes the design challenges of future
the effect of Moore’s Law in Integrated Circuits (ICs) are NoCs and reviews the literature of (some) emerging NoC
required to integrate from dozens of cores today to hundreds architectures introduced to enhance on-chip communication.
of cores within a single chip in the near future. Furthermore,
SoC designs shall impose strong requirements on scalability, An argument is then presented on the scalability and perfor-
reusability and performance of the underlying interconnection mance benefits obtained in NoCs by using overlay networks of
system in order to satisfy constraints of future technologies. particular topologies that were previously considered as logical
The use of scalable Network-on-Chip (NoC) as the underlying networks for use in optical networks.
communications infrastructure is critical to meet such stringent
future demands. This paper focuses on the state-of-the-art in NoC
development trends and seeks to develop increased understanding
II. F UTURE N O C D ESIGN R EQUIREMENTS
of how ideal regular NoC topologies such as the hypercube, The benefits introduced by employing the NoC ap-
de-bruijn, and Manhattan Street Network, can scale to meet
the needs of regular and irregular future NoC structures with proach in SoC designs can be classified as improvements in
increasing numbers of core resources. The contributions of this structure, performance and modularity [4]. The main challenge
paper are three-fold. First, the study introduces a new design for NoC designers will be to provide functionally correct,
framework for overlay architectures based on the success of the reliable operation of the interacting subsystem components.
hypercube, de-bruijn and Manhattan Street Network in NoCs, On-chip interconnection networks aim to minimize current
providing increased scalability for regular structures, as well as
support for irregular structures. Second, the study proposes how SoC limitations in performance, energy consumption and
the regular topologies may be combined to form hybrid overlay synchronization issues. In [10, 11], the globally asynchronous
architectures on NoCs. Third, the study demonstrates how such and locally synchronous (GALS) synchronization paradigm
overlay and hybrid overlay architectures can be used to extend was identified as a strong candidate for emerging ICs. GALS
benefits from logical topologies previously considered in optical eliminates the clock skew in single clock systems by using
networks for use with increased flexibility in the NoC domain.
many different clocks in a distributed manner. Thus, the
Keywords: Network-on-Chip, logical networks, overlay ar-
subsystem components become distributed systems that initiate
chitectures, hybrid architectures.
data transfers autonomously with little or no global coordina-
I. I NTRODUCTION tion [1].
A key issue in SoC design is the trade-off between
Future performance requirements of networking tech-
generality (i.e. the reusability of hardware, operating systems
nologies will be significantly different than current demands
and development techniques) and performance (delay, cost and
on performance. Consequently, ultra-fast communication net-
power consumption in application specific structures) [5]. An
work technologies such as optical networks have emerged
important issue is to consider the implications of the NoC
as a high-bandwidth communication infrastructure for multi-
design approach on the design and implementation costs. For
processor interconnection architectures and their presence as
instance, [5] emphasizes that increasing non-recurring costs of
an interconnection infrastructure is beginning to emerge in
NoC-based ICs requires that the design cost of ICs are shared
the NoC literature. Device scaling trends driven by the ef-
across applications, in which case the design methodology
fect of Moore’s Law suggests that future SoC designs must
would support product family management.
integrate from several dozen cores to hundreds of resource
cores within a single chip, thereby necessitating the need
III. R EVIEW OF O N -C HIP I NTERCONNECTION
for increased bandwidth and performance requirements. The
T ECHNOLOGIES
literature evidences that SoC designs have moved out of bus-
based approaches towards the acceptance of a variety of NoC Various NoC architectures have been proposed to meet
approaches for interconnecting resource cores. future performance requirements for intra-chip communica-
NoC approaches have progressed as the widely adopted tion. Essentially, an NoC architecture is the on-chip commu-
alternative to shared-bus architectures, with the ability to nication infrastructure consisting of the physical layer, the data
meet future performance requirements since NoCs support link layer and the network layer. In addition, the NoC architec-
reusability and network bandwidth scales with system growth ture may be characterized by its switching technique, routing
64 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
protocol, topology and node organization. These characteris-
tics comprise the design space of future on-chip networks [10].
Typically, network functionalities and data transfer properties
differ between on-chip and inter-chip networks, and hence the 110 111
design space for future SoCs must be explored. This section
reviews the literature of emerging SoC platforms, contrasting 100 101
,
differences in the design space of each architecture.
One particular architecture commonly used as the basis 010 011
of many NoC design proposals is the 2-dimensional mesh,
forming a torus or Manhattan-like topology. In [5], the NoC
000 001
architecture is an m x n mesh of switches and resources. A
simple 2-dimensional topology was selected for its scalability De Bruijn Graph Hypercube Manhattan Street Network
and simplistic layout. Consistent with the GALS paradigm,
internal communications within each resource is synchronous Fig. 1. A subset of multi-processor interconnection architectures
and resources operate asynchronously with respect to each
other. Dally et al. [4] presents a 2-dimensional folded torus
topology with the motivation to minimize the total area
communications architecture is not new and has been the focus
overhead for an on-chip network implementation. The work
of previous studies [15, 16]. Furthermore, the work presented
presented in [10] considers the development of a communica-
in [6, 13] describes the hardware emulation of a regular logical
tions protocol for a 2-dimensional mesh topology using point-
topology using Field Programmable Gate Array (FPGA) logic
to-point crossbar interconnects, with the assumption that the
and a hardware emulator. From the study [6, 13], it is noted
sole user of the network is a programmer. Hence, the network
that a number of hardware and software design issues must
must be able to handle the needs of the programmer and the
be addressed before realizing the hardware implementation of
surrounding chip environment, therefore requiring support of
a logical network as an NoC. In the literature, a number of
static and dynamic traffic. In contrast, [2] presents a compar-
comparisons were drawn with related works which have also
ison between a bus architecture and a generic NoC model.
explored the NoC implementation of similar torus-like archi-
Principally, the work demonstrates that a NoC-based system
tectures [14] and hierarchical bus-based approaches combined
yields significant improvements in performance compared with
with a crossbar NoC architecture [10] implemented on FPGAs.
a bus architecture used in SoC systems.
An interesting alternative to the 2-dimensional mesh The literature [17-20] presents the hypercube as a
topology is presented in the literature. For instance, in Hemani regular 3D architecture for SoC interconnection. Whilst several
et al. [11], the nodes are organized as a honeycomb structure, studies present arguments of the benefits of such a regular
whereby resources are organized as nodes of the hexagon structure as an NoC, other studies focus on improving on
with a local switch at the center that interconnects these disadvantages associated with the hypercube [20], yet whilst
resources. The proposed NoC architecture in [11] was generic, others emphasize on the need for irregular NoC structures.
it was not tailored to a specific application domain and was The de-bruijn network, on the other hand, presents a
required to support re-configurability at the task or process stronger case for significant performance improvements, scal-
level. The work in [11] presents arguments to justify that the ing abilities and support for optimized routing techniques [21-
area and performance penalty incurred using the honeycomb 26]. In the literature, several studies had presented variations of
architecture would be minimum. the de-bruijn network in order to emphasize its superiority in
More recently, an interesting area of research has scaling, reliability, routing, performance, power consumption
considered the use of multi-processor interconnection archi- and complexity [21 -26], whereas other studies used the de-
tectures that were previously considered for use as logical bruijn network as the benchmark for comparison with other
topologies deployed in optical networks for use in NoCs. NoCs, including the butterfly and Benes topology [24] and
Figure 1 illustrates a subset of logical topologies considered with the mesh and torus [23]. All comparative performance
here for use in NoCs. [12] provides an insight as to how metrics had demonstrated the superiority of the de-bruijn as a
any self-routing logical topology may be applied in optical- communication architecture for SoCs.
networks. This study shall focus on the design and implementation
[7, 12] had considered one particular logical topology considerations of general regular logical networks for use in
for use in light-wave physical networks because of its simple NoCs in order to extend the topological benefits and findings
routing and control mechanism. However, a NoC platform from mature work on logical topology deployment into the
would now imply that the use of logical networks would NoC domain. In particular, this study extends the work of
require to operate under different constraints and assumptions logical topology deployment onto general physical networks,
from those considered earlier in an optical-network environ- and applies an adopted and enhanced concept of overlaying
ment. The principle of applying a regular logical network, such logical topologies in optical networks for deriving flexible
as the Manhattan Street Network (MSN), for use as an NoC regular and irregular NoC architectures. Figure 2 illustrates
65 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
the concept of using logical networks in NoCs.
Micro-Level MSN Implementation Graph1 Graph1
110 111
1 2 3 4
100 101
5 6 7 8
, Graph2 Graph2
MSN
010 011
9 10 11 12
1000
3 1 4 001 15 16
De Bruijn Graph Hypercube Manhattan Street Network
co-location co-location
The Regular NoC Structure
Regular or Hybrid Irregular Networks can be of edge of internal
Removes the Embedding Problem
Physically Implemented or Mapped to the NoC nodes nodes
Regular
Underlying
NoC Structure
Network
(a) (b)
Fig. 3. Network Overlay concept applied to the De-Bruijn
Fig. 2. Logical Network Implementation on NoCs
IV. M ETHODOLOGY AND I MPLEMENTATION
This paper introduces a new design framework for
the overlay of multi-processor interconnection architectures
that supports regular and irregular core-numbers on-chip. A
rich source of literature exists on the use of multi-processor
interconnection architectures as regular logical networks de-
ployed on host optical networks. The motivation here is to
apply such logical networking concepts and benefits, through
the use of the hypercube, de-bruijn and MSN in the SoC Graph1 Graph2
domain, whilst removing the restriction of the highly course-
granular regular structure associated with NoC topologies as
NoC sizes scale. Therefore, a mechanism for applying overlay Fig. 4. The topology produced by overlaying two de-bruijn graphs
architectures to support regular and irregular scalable NoCs
is introduced as follows. When considering the connectivity
and node functionality of each network (see Figure 1), we a hybrid of overlay architectures. An example in Figure 5
find that the in-degree and out-degree for each node is similar demonstrates one hybrid of the MSN and the de-bruijn.
throughout within each network. Hence, from Figure 1, the The significance of this novel approach to NoC architec-
de-bruijn, hypercube and MSN have a node degree of 2, 3, ture design is that it supports performance-intensive tasks to be
and 2 respectively (where the in-degree equals the out-degree). mapped onto the particularly ideal (high-performance) network
Following an overlay of the de-bruijn (as in Figure 3a), for segments, such as graph-1 or graph-2, whilst other highly-
example, we find two-instances where the functionality of two localized traffic-generating tasks are mapped onto the MSN
nodes are ’co-located’ onto a single node (e.g. at the interface portion for instance. Hence, this framework also advances
between two overlay topologies). The additional functionality optimization techniques for application-mapping of tasks onto
at the co-located nodes may be supported/accommodated NoCs, providing an insight into further opportunities for
by providing additional buffers that separate network-traffic progress of key significance in the NoC application-mapping
destined to graph-1 from traffic destined to graph-2 then literature. This section has demonstrated how overlay networks
routing as normal. Different nodes in graph-1 may be co- and hybrid overlay networks may be applied to extend the
located to yield various degrees of (comparatively-granular) benefits of logical networks from the optical networks domain
NoC architectures, therefore providing support for regular (as evident in the mature literature in this topic) to the NoC
and irregular structures (Figure 3b). Figure 4 illustrates the domain, while providing support for regular and irregular
new NoC architecture after co-location of edge nodes of the structures with comparatively granular flexibility in design.
de-bruijn network. 1 This paper also extends the concept Additionally, the proposed hybrid overlay design enables opti-
of overlay networks to different network-types, producing mization of application-task-mapping onto particular segments
1 A similar concept of overlays can also be applied to larger sizes of the
of the network architecture based on the relative (and signif-
de-bruijn, hypercube and MSN. However, this paper has applied the concept icance) of properties for each segment and the corresponding
to one-size of each logical topology. task-constraints.
66 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
[11] Hemani A., Jantsch A., Kumar S., Postula A., Oberg J., Millberg M.,
Lindqvist D., Network on a Chip: An architecture for billion transistor
era, Source: http://www.ele.kth.se/ axel/papers/2000/norchip-noc.pdf.
De-Bruijn MSN De-Bruijn
Graph 1 Graph 2 [12] Komolafe, O., High Speed Optical Packet Switching over Arbitrary
Physical Topologies using the Manhattan Street Network, Ph.D Thesis,
Univ. of Strathclyde, U.K., 2001.
[13] Oommen K., Harle D., Evalation of a network on chip architecture
based on the clockwork routed Manhattan street network using hardware
, emulation, Proceedings of the 48th Midwest Symposium on Circuits and
Systems, U.S.A, August, 2005.
[14] Lusala A.K., Manet P., Rousseau B., Legat J.-D., NoC Implementation in
FPGA using Torus Topology, Proceedings of the International Conference
on Field Programmable Logic and Applications, 2007.
[15] Tayan O., Harle D., A Manhattan Street Network Implementation for
Networks on Chip, Proceedings of the First International Conference on
Information and Communication Technologies from Theory to Applica-
tions, Damascus, Syria, April 2004.
[16] Tayan O., Networks-on-Chip: Challenges, Trends and Mechanisms for
Enhancements, Proceedings of the Third International Conference on
Fig. 5. A hybrid overlay network that combines the de-bruijn and MSN Information and Communication Technologies, Karachi, Pakistan, August
2009.
[17] Li Ping Sun, El Mostapha Aboulhamid, David J.-P, Networks on
chip using a reconfigurable platform, Proceedings of the International
V. D ISCUSSION AND C ONCLUSION Symposium on Micro-NanoMechatronics and Human Science, 2003.
[18] Derutin J.P., Damez L., Desportes A., Lazaro Galilea J.L., Design
This paper has presented state-of-the-art techniques in of a Scalable Network of Communicating Soft Processors on FPGA,
NoC developments in order to increase understanding of how Proceedings of the International Workshop on Computer Architecture for
ideal regular NoC topologies, including the hypercube de- Machine Perception and Sensing, 2007.
[19] Martinez Vallina F., Jachimiec N., Saniie J., NOVA interconnect for
bruijn and MSN can scale to satisfy the demands of regular and dynamically reconfigurable NoC systems, Proceedings of the International
irregular NoC structures. Significant contributions of the paper Conference on Electro/Information Technology, 2007.
include; introduction of a new design framework for overlay [20] Damaj S., Goubier T., Blanc F., Pottier B., A Heuristic (delta D) Digraph
to Interpolate between Hypercube and de Bruijn Topologies for Future
NoC architectures,a proposed design framework for combining On-Chip Interconnection Networks, Proceedings on the International
regular topologies to form various hybrid overlay architectures Conference on Parallel Processing Workshops, 2009.
that can be optimized for a particular application-workload [21] Moussa H., Baghdadi A., Jezequel M., Binary de Bruijn on-chip network
for a flexible multiprocessor LDPC decoder, Proceedings of the 45th
scenario, and finally, by demonstrating how such overlay and ACM/IEEE Design Automation Conference, 2008.
overlay hybrid architectures can be used to extend benefits [22] Yiou Chen, Jianhao Hu, Xiang Ling, De Briujn graph based 3D Network
adopted from logical networking, previously considered in on Chip architecture deign, Proceedings of the International Conference
on Communications, Circuits and Systems, 2009.
optical networks, for use as on-chip network architectures. [23] Hosseinabady M., Kakoee M.R., Mathew J., Pradhan D.K., Reliable
network-on-chip based on generalized de Bruijn graph, Proceedings of
R EFERENCES the IEEE International High Level Design Validation and Test Workshop,
2007.
[1] Benini L., De Micheli G., Networks on chips: a new SoC paradigm, [24] Moussa H., On-chip communication network for flexible multiprocessor
Computer , Volume: 35 , Issue: 1, Jan. 2002. turbo coding, Proceedings of the Third International Conference on Infor-
[2] Zeferino C.A., Kreutz M.E., Carro L., Susin A.A., A study on commu- mation and Communication Technologies: From Theory to Applications,
nication issues for systems-on-chip, Proceedings of the 15th Symposium 2008.
on Integrated Circuits and Systems Design, 9-14 Sept., 2002. [25] Sabbaghi-Nadooshan R., Modarressi M., Sarbazi-Azad H., The 2D
[3] Ali M., Welzl M., Zwicknagl M., Networks on Chips: Scalable inter- DBM: An Attractive alternative to the simple 2D mesh topology for
connects for future systems on chips, Proceedings of the 4th European on-chip networks, Proceedings of the IEEE International Conference on
conference on Circuits and Systems for Communications, 2008. Computer Design, 2008.
[4] Dally W.J., Towles B., Route Packets, Not Wires: On-Chip Interconnection [26] Hosseinabady M., Kakoee M.R., Mathew J., Pradhan D.K., De Bruijn
Networks, Proceedings of DAC, June 18-22, 2001, Las Vegas, U.S.A. Graph as a Low Latency Scalable Architecture for Energy Efficient
[5] Kumar S., Jantsch A., Soininen J.-P., Forsell M., Millberg M., Oberg Massive NoCs, Proceedings of the Design, Automation and Test in Europe
J., Tiensyrja K., Hemani A., A network on chip architecture and design Conference, 2008.
methodology, Proceedings of the IEEE Computer Society Annual Sym-
posium on VLSI, 25-26 April 2002.
[6] Oommen K., Tayan O., Harle D., Network on Chip Architecture using a
Clockwork Routed Manhattan Street Network, Proceedings of the System
on Chip Design and Test Conference, Loughborough, U.K, Sept. 2004.
[7] Tayan, O., Exploring the Scalability of a High-Speed Networking Topol-
ogy in Macro- and Micro-Networks, Ph.D Thesis, Univ. of Strathclyde,
U.K., December 2005.
[8] Pontes J., Moreira M., Soares R., Calazans N., Hermes-GLP: A GALS
Network on Chip Router with Power Control Techniques, Proceedings of
the IEEE Computer Society Annual Symposium on VLSI, 2008.
[9] GaoMing Du., DuoLi Zhang., YongSheng Yin., Liang Ma., LuoFeng
Geng., YuKung Song., MingLun Gao/. FPGA prototype design of Network
on Chips, Proceedings of the 2nd International Conference on Anti-
counterfeiting, Security and Identification, 2008.
[10] Heo S., Kim J., Ma A., Next Generation On-chip Communication
Networks, www.cag.lcs.mit.edu/6.893f2000/project/heo check1.pdf.
67 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No1,April 2010
Effective Bandwidth Utilization in IEEE802.11 for
VOIP
S.Vijay Bhanu Dr.RM.Chandrasekaran Dr.V.Balakrishnan
Research Scholar, Anna University, Coimbatore Registrar, Anna University, Trichy Research Co-Supervisor
Tamilnadu, India, Pincode-641013 Tamilnadu, India, Pincode: 620024. Anna University, Coimbatore
E-Mail: vbhanu22@yahoo.in E-mail: aurmc@hotmail.com E-Mail :profdrvb@gmail.com
Abstract -Voice over Internet protocol (VoIP) is one of the most properly) [2]; they are used for data transmission, and a
important applications for the IEEE 802.11 wireless local area network only designed for data transmission is not ideal for
networks (WLANs). For network planners who are deploying voice transmission. Compare to data packet, voice packets are
VoIP over WLANs, one of the important issues is the VoIP small in size. Due to the large overhead involved in
capacity. VoIP bandwidth consumption over a WAN is one of the transmitting small packets, the bandwidth available for VoIP
most important factors to consider when building a VoIP traffic is far less than the bandwidth available for data traffic.
infrastructure. Failure to account for VoIP bandwidth This overhead comprises transmitting the extra bytes from
requirements will severely limit the reliability of a VoIP system various networking layers (packet headers) and the extra time
and place a huge burden on the WAN infrastructure. Less (backoff and deferral time) imposed by the Distributed
bandwidth utilization is the key reasons for reduced number of
Coordination Function (DCF) of 802.11b.
channel accesses in VOIP. But in the QoS point of view the free
bandwidth of atleast 1-5% will improve the voice quality. This This paper experimentally study the relationship
proposal utilizes the maximum bandwidth by leaving 1-5% free between bandwidth utilization in the wireless LAN and the
bandwidth. A Bandwidth Data rate Moderation (BDM) quality of VoIP calls transmitted over the wireless medium.
algorithm has been proposed which correlates the data rate On an 802.11 b WLAN, frames are transmitted at up to 11
specified in IEEE802.11b with the free bandwidth. At each time Mbps. There is a lot of overhead before and after the actual
BDM will calculate the bandwidth utilization before sending the transmission of frame data, however, and the real maximum
packet to improve performance and voice quality of VoIP. The end-to-end throughput is more on the order of 5 Mbps. So, in
bandwidth calculation in BDM can be done by using Erlang and theory, 802.11b should be able to support 50-plus
VOIP bandwidth calculator. Finally, ns2 experimental study
simultaneous phone calls[1]. But practically it support only 5
shows the relationship between bandwidth utilization, free
bandwidth and data rate. The paper concludes that marginal
calls. This proposal improves bandwidth utilization in order to
VoIP call rate has been increased by BDM algorithm. achieve maximum channel access and improved QoS by using
BDM algorithm. The number of channel access can be
Keywords: WLAN ,VOIP ,MAC Layer, Call Capacity, Wireless improved by changing the data rate frequently.
Network This paper is structured as follows: Section IA
describes about basic history of 802.11 MAC and previous
I. INTRODUCTION related work. Section III introduces a method for predicting
VoIP services have been significantly gaining VoIP bandwidth utilization. Section IV shows the BDM
prominence over the last few years because of a number of algorithm and its functionalities and Section V&VI discuss
impressive advantages over their traditional circuit-switched about the simulation topology, parameters and results. Final
counterparts including but not limited to high bandwidth part contains conclusion and future enhancement.
efficiency, low cost, and flexibility of using various
compression strategies. In contrast to wired networks, the A. Basic Theory of 802.11 MAC
bandwidth of wireless network is limited. Furthermore, a The basic 802.11 MAC protocol is the Distributed
wireless channel is error-prone and packets can be discarded Coordination Function (DCF), which is based on the Carrier
in transmission due to wireless errors such as signal fading or Sense Multiple Access/Collision Avoidance (CSMA/CA)
interference. Thus, the efficiency of a wireless channel access mechanism [3] [4]. A mobile station (STA) is allowed to send
becomes a critical issue. packets after the medium is sensed idle for the duration greater
than a Distributed Inter-Frame Space (DIFS). If during
Currently, the most popular WLAN standard is the anytime in between the medium is sensed busy, a back-off
IEEE 802.11b, which can theoretically support data rates up to procedure should be invoked. Specifically, a random variable
11 Mb/s, however, this data rate is for optimal conditions [1]. uniformly distributed between zero and a Contention Window
On the other hand, 802.11a and 802.11g networks have data (CW) value should be chosen to set a Back-off Timer. This
rates up to 54 Mb/s and they are not designed to support voice Back-off Timer will start to decrement in units of slot time,
transmission (because of the APs are not distributed in the provided that no medium activity is indicated during that
most optimum way, communication can be established particular slot-time. The back-off procedure shall be
68 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No1,April 2010
suspended anytime the medium is determined to be busy and plain IEEE802.11 MAC protocol, and adopting an additional
will be resumed after the medium is determined to be idle for application aware module, logically placed above the MAC
another DIFS period. The STA is allowed to start transmission layer. In reference [9] proposes two feedback-based bandwidth
as soon as the Back-off Timer reaches zero. A mobile station allocation algorithms exploiting HCCA to provide service
(STA) shall wait for an ACK when a frame is sent out. If the with guaranteed bounded delays: (1) the Feedback Based
ACK is not successfully received within a specific ACK Dynamic Scheduler (FBDS) and (2) the Proportional Integral
timeout period, the STA shall invoke back-off and (PI)-FBDS. They have been designed using classic discrete-
retransmission procedure. The CW value shall be increased time feedback control theory. We will assume that both
exponentially from a CWmin value until up to a CWmax value algorithms, running at the HC, allocate the WLAN channel
during each retransmission. bandwidth to wireless stations hosting real-time applications,
An additional Request to Send/ Clear To Send using HCCA functionalities. This allows the HC to assign
(RTS/CTS) mechanism is defined to solve a hidden terminal TXOPs (transmission opportunity) to ACs by taking into
problem inherent in Wireless LAN. The successful the account their specific time constraints and transmission queue
exchange of RTS/CTS ensures that channel has been reserved levels. We will refer to a WLAN system made of an Access
for the transmission from the particular sender to the particular Point and a set of quality of service enabled mobile stations
receiver. The use of RTS/CTS is more helpful when the actual (QSTAs). Each QSTA has up to 4 queues, one for each AC in
data size is larger compared with the size of RTS/CTS. When the 802.11e proposal. FBDS require a high computational
the data size is comparable with the size of RTS/CTS, the overhead at the beginning of each service period, due to the
overhead caused by the RTS/CTS would compromise the queue length estimation.
overall performance.
By [8] Wireless Timed Token Protocol (WTTP)
II. PREVIOUS WORKS provides traffic streams with a minimum reserved rate, as
required by the standard, and it accounts for two types of
This section, reviews the existing literature related to traffic streams simultaneously, depending on the
enhancing voip call capacity. In reference [5] Aggregation corresponding application: constant bit rate, which are served
with fragment Retransmission (AFR) scheme, multiple according to their rate, and variable bit rate traffic streams.
packets are aggregated into and transmitted in a single large Additionally, WTTP shares the capacity which is not reserved
frame. If errors happen during the transmission, only the for QoS traffic streams transmissions among traffic flows with
corrupted fragments of the large frame are retransmitted. no specific QoS requirements. This VAD [10] algorithm is
Clearly, new data and ACK frame formats are a primary capable of removing white noise as well as frequency selective
concern in developing a practical AFR scheme. Optimal frame nose and maintaining a good quality of speech.
and fragment sizes are calculated using this model, and an
algorithm for dividing packets into near-optimal fragments is III. CALCULATING BANDWIDTH CONSUMPTION FOR
designed. Difficulties for new formats include 1) respecting VOIP
the constraints on overhead noted previously and 2) ensuring Bandwidth is defined as the ability to transfer data
that, in an erroneous transmission, the receiver is able to (such as a VoIP telephone call) from one point to another in a
retrieve the correctly transmitted fragments—this is not fixed amount of time.The bandwidth needed for VoIP
straightforward because the sizes of the corrupted fragments transmission will depends on a few factors: the compression
may be unknown to the receiver. technology, packet overhead, network protocol used and
whether silence suppression is used. Voice streams are first
Extended dual queue scheme (EDQ) provides a QoS encapsulated into RTP packets, and they are carried by
for the VoIP service enhancement over 802.11 WLAN. It UDP/IP protocol stack [3]. A single voice call consists of two
proposes a simple software upgrade based solution, called an opposite RTP/UDP flows. One is originated from the AP to a
Extended Dual queue Scheme (EDQ), to provide QoS to real- wireless station, and the other oppositely flows. There are two
time services such as VoIP [6]. The extended dual queue primary strategies for improving IP network performance for
scheme operates on top of the legacy MAC. The dual queue voice: several techniques were proposed for QoS provisioning
approach is to implement two queues, called VoIP queue and in wireless networks[11] [12]. Allocate more VoIP bandwidth
data queue. Especially, these queues are implemented above and implement QoS.
the 802.11 MAC controllers, i.e., in the device driver of the How much bandwidth to allocate depends on:
802.11 network interface card (NIC), such that a packet
scheduling can be performed in the driver level. Packets from Packet size for voice (10 to 320 bytes of digital
the higher layer or from the wire line port (in case of the AP) voice)
are classified to transmit into VoIP or data types. Packets in
the queues are served by a simple strict priority queuing so CODEC and compression technique (G.711, G.729,
that the data queue is never served as long as the VoIP queue G.723.1, G.722, proprietary)
is not empty. But the hardware upgrade is undesirable.
Header compression (RTP + UDP + IP), which is
The cross-layer scheme of [7] [8] is named as optional
Vertical Aggregation (VA) since it works along the same flow.
The main advantage is that it enhances voice capacity using a
69 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No1,April 2010
Layer 2 protocols, such as point-to-point protocol IV. PROPOSED ALGORITHM
(PPP), Frame Relay and Ethernet The main reason for the extra bandwidth usage is IP
and UDP headers. VoIP sends small packets and so many
Silence suppression / voice activity detection times, the headers are actually much larger than the data part
of the packet. The proposed algorithm based on the following
Calculating the bandwidth for a VoIP call is not two factors. 1) A small frame is in error then there is a high
difficult once you know the method and the factors to include. probability of error for a large frame as well. Similarly when a
The chart below, "Calculating one-way voice bandwidth," large frame is successful, there is a very high probability of
demonstrates the overhead calculation for 20 and 40 byte success for small frames as well. 2) The amount of free
compressed voice (G.729) being transmitted over a Frame bandwidth decreases as the number of VoIP calls increases. As
Relay WAN connection [13]. Twenty bytes of G.729 well as the call quality decreases as the number of VoIP calls
compressed voice is equal to 20 ms of a word. increases. Free Bandwidth (BWfree) that corresponds to the
remaining unused idle time that can be viewed as spare or
Voice digitization and compression: available capacity. In BDM algorithm, at each frame
transmission will calculates the free bandwidth availability.
G .711: 64,000 bps or 8000 bytes per second
Variables
G.729: 8000 bps or 1000 bytes per second
BWfree: Unused idle bandwidth viewed as spare or available
Protocol packet overhead: capacity
BWload: Specifies the bandwidth used for transmission of the
IP = 20 bytes, UDP = 8 bytes, RTP =12 bytes data frames
Total:40 bytes Drate: It specifies the data rate
Incr: Increment operation
If one packet carries the voice samples representing Decr: Decrement operation
20 milliseconds, the 50 such samples are required to be
transmitted in every second. Each sample carries an Functions:
IP/UDP/RTP header overhead of 320 bits [14]. Therefore, in UpperLevel (): upper level according to table 1, 2
each second, 16,000 header bits are sent. As a general rule of LowerLevel (): lower level according to the table 1, 2
‘thumb’, it can be assumed that header information will add
16kbps to the bandwidth requirement for voice over IP. For A. BDM ALGORITHM:
example, if an 8kbps algorithm such as G.729 is used, the total Initial level:
bandwidth required to transmit each voice channel would be Drate: LowerLevel ()
24kbps. BWfree: UpperLevel ()
S: Previous transmission
The voice transmission requirements are, If (S = Success)
{
Bandwidth requirements reduced with compression, Incr Drate to next UpperLevel ()
G.711, G.729 etc. Decr BWfree to next LowerLevel ()
}
Bandwidth requirements reduced when longer Else
packets are used, thereby reducing overhead. {
Decr Drate to next LowerLevel ()
Even though the voice compression is an 8 to 1 ratio, Incr BWfree to next UpperLevel ()
the bandwidth reduction is about 3 or 4 to 1. The }
overhead negates some of the voice compression According to IEEE 802.11b only four types of data
bandwidth savings. rates are available, which are 1, 2, 5.5, 11mbps. When the data
rate is high then the throughput increases at the same time the
Compressing the RTP, UDP and IP headers is most chance for occurring error also increases [1] [15]. To avoid
valuable when the packet also carries compressed this situation BDM allocates some free bandwidth to improve
voice. the QoS. This free bandwidth allocation should be at the
minimum level otherwise again quality degradation occurs.
A. Packet Overhead
To support voice over WLANs, it is important to TABLE1: LEVELS OF DATA RATE
reduce the overhead and improve the transmission efficiency
over the radio link. Recently, various header compression Levels Data Rate
techniques for VoIP have been proposed [14]. The Level 0 1 mbps
RTP/UDP/IP headers can be compressed to as small as 2 Level 1 2 mbps
bytes. Level 2 5.5 mbps
Level 3 11 mbps
70 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No1,April 2010
TABLE2: LEVELS OF BANDWIDTH FREE
Levels % of Free Bandwidth TABLE 3: PARAMETERS USED FOR SIMULATION
Level 0 1 Parameter Value
Level 1 2
Level 2 3 DIFS 50 µsec
Level 3 4 SIFS 10 µsec
Level 4 5
Slot time 20 µsec
Number of calls = Correc_Fac ( RB-RBT ) / Codec CWmin 32
Where, CWmax 1023
Correc_Fac: Correction factor of real network performance
RB: Real bandwidth usage Data Rate 1,2,5.5,11 Mbps
RBT: Real bandwidth used for data transmission
Codec: Bandwidth used by the codec to establish a call Basic rate 1 Mbps
PHY header 192 µsec
End-to-end (phone-to-phone) delay needs to be
limited. The shorter the packet creation delay, the more MAC header 34 bytes
network delay the VoIP call can tolerate. Shorter packets ACK 248 µsec
cause less of a problem if the packet is lost. Short packets
require more bandwidth, however, because of increased packet
overhead (this is discussed below). Longer packets that
contain more speech bytes reduce the bandwidth requirements
but produce a longer construction delay and are harder to fix if
lost. By BDM the data rate and free bandwidth will improve
the number of VOIP calls as well as performance. TABLE4: DATA RATES AND DISTANCE FOR VOIP
V. SIMULATION TOPOLOGY
Data rate in Mbps Distance in meters
The simulation study is conducted using the ns-2 54 0-27
simulator. The simulation result will be compared with IEEE
802.11 specifications. Any node can communicate with any 48 27-29
other node through base station. The number of stations can be 36 29-30
varied from 5 to 50. Wireless LAN networks are set up to 24 30-42
provide wireless connectivity within a finite coverage area of 18 42-54
20 to 30m. The network simulator will be used to form an
appropriate network topology under the Media Access Control 11 0 - 48
(MAC) layer of the IEEE 802.11b. According to the IEEE
802.11b protocol specifications [16], the parameters for the
WLAN are shown in Table 3. When calculating bandwidth,
one can't assume that every channel is used all the time.
Normal conversation includes a lot of silence, which often
means no packets are sent at all. So even if one voice call sets
up two 64 Kbit RTP streams over UDP over IP over Ethernet
(which adds overhead), the full bandwidth is not used at all
times.
Based on [2] the data rate and coverage area will be changed.
802.11b standard can cover up to 82 meters of distance,
considering that only the first 48 meters are usable for voice,
the other 34 meters are not usable, therefore cellular area must
be fit only to the 48 meters from the AP in order to avoid
interferences, which is depicted in table 4.
Fig 1: Free bandwidth analysis
71 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No1,April 2010
Free bandwidth= Total bandwidth-bandwidth utilized
=100-87.5
=13.5
In this sample calculation the free bandwidth is
13.5%. From this 8.5% of bandwidth can be utilized for frame
transmission to achieve maximum throughput and leave 5% to
obtain Qos. Fig 1 shows the difference between bandwidth
utilization and free bandwidth. When the amount of free
bandwidth dropped below 1% call quality became
unacceptable for all ongoing calls. The amount of free
bandwidth is a good indicator for predicting VoIP call quality,
but in the throughput point of view it should be reduced. This
contradiction can be solved by using BDM algorithm.
VI. SIMULATION RESULTS
Fig 3: Variations in packet loss when number of frames
A. Throughput increases
The throughput (measured in bps) corresponds to the C. Delay
amount of data in bits that is transmitted over the channel per Investigating our third metric, average access delay
unit time. In the following Fig 2 X-axis specifies timeslot and for high priority traffic, Fig 4 shows that has very low delays
Y-axis specifies the throughput. Consider for each time slot in most cases, even though the delays increases when the load
the channel receives 10 frames. When the time slot is 4ms, the gets very high. However, all the schemes have acceptable
throughput is 4200kbps, when it increases into 8ms it is delays [6], even though EDCA in most cases incur a longer
7000kbps. The graph shows the gradual improvement and the delay than the other schemes. Even if a scheme can give low
overall throughput is increased upto 87.5%. average access delay to high priority traffic, there might still
be many packets that get rather high delays. With the number
of data stations increasing, the delay performance of the voice
stations degrades. This tells that the VoIP performance is
sensitive to the data traffic.
In fig-4 graph X-axis specifies the number of frames
and Y-axis specifies the delay in ms. When the number of
frames is in between 5-10 the delay is gradually increased
after that there is no change in the delay. Nearly 30% of the
delay is reduced by BDM algorithm.
Fig 2: Variations in throughput with respect to timeslot
B. Frame Loss
Frame loss is expressed as a ratio of the number of
frames lost to the total number of frames transmitted. Frame
loss results when frames send are not received at the final
destination. Frames that are totally lost or dropped are rare in
WLANs. In the fig-3 X-axis shows the number of frames and
Y-axis shows the frame loss percentage. The frame loss value
increased upto 0.5 when it reaches a threshold value it slowly
decreasing . These frame loss degradation will improve our
VOIP performance in a great manner. Fig 4: Variations in delay when number of frames
increases
72 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No1,April 2010
D. Bandwidth Utilization 10 stations). This calculation shows that 16% of overall voip
call rate is increased by BDM algorithm.
In most cases, the normal VoIP telephone call will
use up 90Kbps. When calculating bandwidth, one can't assume
that every channel is used all the time. Normal conversation
includes a lot of silence, which often means no packets are
sent at all. So even if one voice call sets up two 64 Kbit RTP
streams over UDP over IP over Ethernet (which adds
overhead), the full bandwidth is not used at all times. Note that
for the graphs where the priority load is low, the utilization
first increases linearly.
Fig 6: Variations in throughput when number of frames
increases
Fig5: Variations in bandwidth utilization
In fig-5 X-axis specifies the stations and Y-axis
specifies the bandwidth utilization percentage. From the F. Comparison of Bandwidth Utilization
number of stations 5 the curve starts incrementing, when the VoIP bandwidth consumption over a WAN (wide
stations become 15 the bandwidth utilization percentage is
area network) is one of the most important factors to consider
beyond 80%. In the above specified graph 30% of overall When building a VoIP infrastructure [9]. Failure to account
bandwidth utilization is increased.
for VoIP bandwidth requirements will severely limit the
reliability of a VoIP system and place a huge burden on the
E. Throughput Comparison
WAN infrastructure. Short packets require more bandwidth,
however, because of increased packet overhead (this is
Here the throughput performance of the EDCA
discussed below). Longer packets that contain more speech
algorithm and our proposed BDM algorithms are compared bytes reduce the bandwidth requirements but produce a longer
[17]. By using values of maximum achievable throughput
construction delay and are harder to fix if lost.
from simulation, VoIP capacity in WLAN can also be
evaluated. The following formula is used for getting the In Fig.7 X-axis shows the number of stations and Y-
average packets sent from AP and all VoIP nodes in one axis shows the bandwidth utilization percentage. When the
second. number of frame is 25 the EDCA algorithm gives only 65% of
Capacity = Maximum Throughput / Data Rate bandwidth utilization and it start to decrease if the number of
stations exceeds 40. But BDM algorithm gives 85% of
In BDM algorithm the data rate is changed frequently in an bandwidth utilization also the curve is gradually increases
effective manner, So that the overall capacity will be when the number of stations increased. By BDM algorithm
improved. When data rate increases, automatically the 20% of bandwidth utilization is increased.
throughput will increase. Due to the low level of data transfer
rate the throughput seldom reach 600 kbps in EDCA. Due to
moderate data rate the maximum throughput in BDM is 1000
kbps, it shows that 15% improvement in overall throughput.
From the graph, the EDCA algorithm can support only 7.5
calls but BDM algorithm can support 9.1 voip calls (only for
73 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No1,April 2010
[7]. F. Maguolo, F. De Pellegrini, A. Zanella and M. Zorzi,”Cross-Layer
Solutions to Performance Problems In Vein Over WLANs”,14th European
Signal Processing Conference (EUSIPCO 2006), Florence, Italy, September 4-
8, 2006.
[8]. Claudio Cicconetti, Luciano Lenzini, Enzo Mingozzi, Giovanni Stea “An
Efficient Cross Layer Scheduler for Multimedia Traffic in Wireless Local
Area Networks with IEEE 802.11e HCCA” Mobile Computing and
Communications Review, Vol 11, No. 3, Nov 2005.
[9]. Gennaro Boggia, Pietro Camarda, Luigi Alfredo Grieco, and Saverio
Mascolo, “Feedback-Based Control for Providing Real-Time Services with
the 802.11e MAC”, IEEE/ACM Transactions On Networking, Vol. 15, No. 2,
April 2007.
[10]. T. Ravichandran and Durai Samy, “Performance Enhancement on Voice
using VAD Algorithm and Cepstral Analysis”, Journal of Computer Science-
2006.
[11]. Antonio grilo, Mario macedo and Mario nunes, “A Scheduling algorithm
for Qos support in 802.11e networks”, IEEE wireless communication, June
2003.
[12]. JengFarn Lee, Wanjiun Liao, Jie-Ming Chen, and Hsiu-Hui Lee, “A
Practical QoS Solution to Voice over IP in IEEE 802.11 WLANs”, IEEE
Communications Magazine, April 2009.
[13]. C. Mahlo, C. Hoene, A. Rostami, A. Wolisz, “Adaptive Coding and
Packet Rates for TCP-Friendly VoIP Flows” International Symposium on
Telecommunication, September 2005.
[14]. Ping Wang, Hai Jiang, and Weihua Zhuang, “Capacity Improvement and
Fig 7: Variations in bandwidth utilization Analysis for Voice/Data Traffic over WLANs”, IEEE Transactions on
Wireless Communications, Vol. 6, No. 4, April 2007.
[15]. Sangho Shin and Henning Schulzrinne, “Measurement and Analysis of
VII. CONCLUSION the VoIP Capacity in IEEE 802.11 WLAN”, IEEE Transactions On Mobile
Computing, Vol. 8, No. 9, September 2009.
This proposal discusses a new association algorithm [16]. Ahmed Shawish, Xiaohong Jiang, Pin-Han Ho and Susumu Horiguchi,
called bandwidth and data rate moderation algorithm for IEEE “Wireless Access Point Voice Capacity Analysis and Enhancement Based on
802.11 stations. BDM is designed to enhance the performance Clients’ Spatial Distribution”, IEEE Transactions On Vehicular Technology,
of an individual station by calculating the free bandwidth Vol. 58, No. 5, June 2009.]
[17]. Kei Igarashi, Akira Yamada and Tomoyuki Ohya, “Capacity
availability. When several users are working simultaneously, Improvement of Wireless LAN Voip Using Distributed Transmission
the real bandwidth is divided among the whole users. Through Scheduling”, 18th Annual IEEE International Symposium on Personal, Indoor
experimentation with a number of VoIP calls and various data and Mobile Radio Communications (PIMRC'07), 2007
rates in an 802.11b WLAN shoes a close relationship between
wireless bandwidth utilization and call quality. When the AUTHORS PROFILE
amount of free bandwidth dropped below 1% call quality
become unacceptable for all ongoing calls. Simulated result
shows that marginal improvement in VOIP call rate is realized
by BDM algorithm. The future work will consider the
Mr. S. Vijay Bhanu is currently
different type of codec techniques and coverage area to working as Lecturer (senior Scale) in
increase the bandwidth utilization. the Computer Science & Engineering
Wing, Directorate of Distance
Education, Annamalai University. He
is an co-author for a monograph on
REFERENCES Multimedia. He served as wing Head,
DDE, Annamalai University,
Chidambaram for nearly five years. He
[1]. Wei Wang, Soung Chang Liew, and Victor O. K. Li, “Solutions to served as Additional Controller of
Performance Problems in VoIP Over a 802.11 Wireless LAN” IEEE Examination at Bharathiar University,
transactions on vehicular technology, Vol. 54, No. 1, January 2005. Coimbatore for two years. He
[2]. Lizzie Narvaez, Jesus Perez, Carlos Garcia and Victor Chi, “Designing conducted a workshop on Business
802.11 WLANs for VoIP and Data”, IJCSNS International Journal of intelligence in the year 2004. He is a
Computer Science and Network Security, Vol.7 No.7, July 2007. life Member in Indian Society for Technical Education. Email:
[3]. Jong-Ok Kim, Hideki Tode and Koso Murakami, “Friendly Coexistence vbhanu22@yahoo.in
of Voice and Data Traffic in IEEE 802.11 WLANs”, IEEE Transactions on
Consumer Electronics, Vol. 52, No. 2, May 2006.
[4]. Hongqiang Zhai, Xiang Chen, and Yuguang Fang, “How Well Can the
IEEE 802.11 Wireless LAN Support Quality of Service?”, IEEE transactions
on wireless communications, Vol. 4, No. 6, November 2005.
[5]. Tianji Li, Qiang Ni, David Malone, Douglas Leith, Yang Xiao, and
Thierry Turletti, “Aggregation with Fragment Retransmission for Very High-
Speed WLANs”, IEEE/ACM Transactions on Networking, Vol. 17, No. 2,
April 2009.
[6]. S.Vijay Bhanu and Dr.RM.Chandrasekaran, “Enhancing WLAN MAC
Protocol performance using Differentiated VOIP and Data Services Strategy”,
IJCSNS International Journal of Computer Science and Network Security,
Vol.9 No.12, December 2009.
74 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No1,April 2010
Dr.RM.Chandrasekaran is currently working as a Registrar
of Anna University,
Tiruchirappalli and
Professor at the
Department of Computer
Science and
Engineering, Annamalai
University, Annamalai
Nagar, Tamilnadu, India.
From 1999 to 2001 he
worked as a software
consultant in Etiam, Inc,
California, USA. He received his Ph.D degree in 2006 from
Annamalai University, Chidambaram. He has conducted
workshops and conferences in the area of Multimedia,
Business Intelligence, Analysis of Algorithms and Data
Mining. Ha has presented and published more than 32 papers
in conferences and journals and is the co-author of the book
Numerical Methods with C++ Program( PHI,2005). His
research interests include Data Mining, Algorithms and
Mobile Computing. He is life member of the Computer
Society of India, Indian Society for Technical Education,
Institute of Engineers and Indian Science Congress Assciation.
Email: aurmc@hotmail.com
Dr.V.Balakrishnan, formerly Director, Anna University,
Coimbatore has got 35 years of service to his credit in
teaching, research, training, extension, consultancy and
administration. He has guided
30 M.Phil Scholars, guiding
15 Ph.D Scholars. He guided
around 1000 MBA projects.
He got 13 international and
national awards including,
‘Management Dhronocharaya’
and ‘Manithaneya Mamani’.
At Anna University he
introduced 26 branches in
MBA. He published around 82
articles in international and
national Journals. He partook
in about 100 international and national seminars. He served in
various academic bodies like Academic Council, Faculty of
Arts, Board of Selection, and Board of Examiners in most of
the Universities in South India.
75 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
ECG Feature Extraction Techniques - A Survey
Approach
S.Karpagachelvi, Dr.M.Arthanari, Prof. & Head, M.Sivakumar,
Doctoral Research Scholar, Dept. of Computer Science and Engineering, Doctoral Research Scholar,
Mother Teresa Women's University, Tejaa Shakthi Institute of Technology for Women, Anna University – Coimbatore,
Kodaikanal, Tamilnadu, India. Coimbatore- 641 659, Tamilnadu, India. Tamilnadu, India
email : karpagachelvis@yahoo.com email: arthanarimsvc@gmail.com email : sivala@gmail.com
Abstract—ECG Feature Extraction plays a significant role in ECG is essentially responsible for patient monitoring and
diagnosing most of the cardiac diseases. One cardiac cycle in an diagnosis. The extracted feature from the ECG signal plays a
ECG signal consists of the P-QRS-T waves. This feature vital in diagnosing the cardiac disease. The development of
extraction scheme determines the amplitudes and intervals in the accurate and quick methods for automatic ECG feature
ECG signal for subsequent analysis. The amplitudes and extraction is of major importance. Therefore it is necessary
intervals value of P-QRS-T segment determines the functioning
that the feature extraction system performs accurately. The
of heart of every human. Recently, numerous research and
techniques have been developed for analyzing the ECG signal. purpose of feature extraction is to find as few properties as
The proposed schemes were mostly based on Fuzzy Logic possible within ECG signal that would allow successful
Methods, Artificial Neural Networks (ANN), Genetic Algorithm abnormality detection and efficient prognosis.
(GA), Support Vector Machines (SVM), and other Signal
Analysis techniques. All these techniques and algorithms have
their advantages and limitations. This proposed paper discusses
various techniques and transformations proposed earlier in
literature for extracting feature from an ECG signal. In addition
this paper also provides a comparative study of various methods
proposed by researchers in extracting the feature from ECG
signal.
Keywords—Artificial Neural Networks (ANN), Cardiac Cycle,
ECG signal, Feature Extraction, Fuzzy Logic, Genetic Algorithm
(GA), and Support Vector Machines (SVM).
I. INTRODUCTION
The investigation of the ECG has been extensively used for
diagnosing many cardiac diseases. The ECG is a realistic
record of the direction and magnitude of the electrical
commotion that is generated by depolarization and re- Figure.1 A Sample ECG Signal showing P-QRS-T Wave
polarization of the atria and ventricles. One cardiac cycle in an
ECG signal consists of the P-QRS-T waves. Figure 1 shows a In recent year, several research and algorithm have been
sample ECG signal. The majority of the clinically useful developed for the exertion of analyzing and classifying the
information in the ECG is originated in the intervals and ECG signal. The classifying method which have been
amplitudes defined by its features (characteristic wave peaks proposed during the last decade and under evaluation includes
and time durations). The improvement of precise and rapid digital signal analysis, Fuzzy Logic methods, Artificial Neural
methods for automatic ECG feature extraction is of chief Network, Hidden Markov Model, Genetic Algorithm, Support
importance, particularly for the examination of long Vector Machines, Self-Organizing Map, Bayesian and other
recordings [1]. method with each approach exhibiting its own advantages and
The ECG feature extraction system provides fundamental disadvantages. This paper provides an over view on various
features (amplitudes and intervals) to be used in subsequent techniques and transformations used for extracting the feature
automatic analysis. In recent times, a number of techniques from ECG signal. In addition the future enhancement gives a
have been proposed to detect these features [2] [3] [4]. The general idea for improvement and development of the feature
previously proposed method of ECG signal analysis was based extraction techniques.
on time domain method. But this is not always adequate to
study all the features of ECG signals. Therefore the frequency The remainder of this paper is structured as follows. Section
representation of a signal is required. The deviations in the 2 discusses the related work that was earlier proposed in
normal electrical patterns indicate various cardiac disorders. literature for ECG feature extraction. Section 3 gives a general
Cardiac cells, in the normal state are electrically polarized [5]. idea of further improvements of the earlier approaches in ECG
76 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
feature detection, and Section 4 concludes the paper with one complex is used to trace the peaks of the individual
fewer discussions. waves, including onsets and offsets of the P and T waves
which are present in one cardiac cycle. Their experimental
II. LITERATURE REVIEW results revealed that their proposed approach for ECG feature
ECG feature extraction has been studied from early time and extraction achieved sensitivity of 99.18% and a positive
lots of advanced techniques as well as transformations have predictivity of 98%.
been proposed for accurate and fast ECG feature extraction.
This section of the paper discusses various techniques and A Mathematical morphology for ECG feature extraction
transformations proposed earlier in literature for extracting was proposed by Tadejko and Rakowski in [8]. The primary
feature from ECG. focus of their work is to evaluate the classification
performance of an automatic classifier of the
Zhao et al. [6] proposed a feature extraction method using electrocardiogram (ECG) for the detection abnormal beats
wavelet transform and support vector machines. The paper with new concept of feature extraction stage. The obtained
presented a new approach to the feature extraction for reliable feature sets were based on ECG morphology and RR-intervals.
heart rhythm recognition. The proposed system of Configuration adopted a well known Kohonen self-organizing
classification is comprised of three components including data maps (SOM) for examination of signal features and clustering.
preprocessing, feature extraction and classification of ECG A classifier was developed with SOM and learning vector
signals. Two diverse feature extraction methods are applied quantization (LVQ) algorithms using the data from the records
together to achieve the feature vector of ECG data. The recommended by ANSI/AAMI EC57 standard. In addition
wavelet transform is used to extract the coefficients of the their work compares two strategies for classification of
transform as the features of each ECG segment. Concurrently, annotated QRS complexes: based on original ECG
autoregressive modeling (AR) is also applied to get hold of the morphology features and proposed new approach - based on
temporal structures of ECG waveforms. Then at last the preprocessed ECG morphology features. The mathematical
support vector machine (SVM) with Gaussian kernel is used to morphology filtering is used for the preprocessing of ECG
classify different ECG heart rhythm. The results of computer signal.
simulations provided to determine the performance of the
proposed approach reached the overall accuracy of 99.68%. Sufi et al. in [9] formulated a new ECG obfuscation method
for feature extraction and corruption detection. They present a
A novel approach for ECG feature extraction was put forth new ECG obfuscation method, which uses cross correlation
by Castro et al. in [7]. Their proposed paper present an based template matching approach to distinguish all ECG
algorithm, based on the wavelet transform, for feature features followed by corruption of those features with added
extraction from an electrocardiograph (ECG) signal and noises. It is extremely difficult to reconstruct the obfuscated
recognition of abnormal heartbeats. Since wavelet transforms features without the knowledge of the templates used for
can be localized both in the frequency and time domains. They feature matching and the noise. Therefore, they considered
developed a method for choosing an optimal mother wavelet three templates and three noises for P wave, QRS Complex
from a set of orthogonal and bi-orthogonal wavelet filter bank and T wave comprise the key, which is only 0.4%-0.9% of the
by means of the best correlation with the ECG signal. The original ECG file size. The key distribution among the
foremost step of their approach is to denoise (remove noise) authorized doctors is efficient and fast because of its small
the ECG signal by a soft or hard threshold with limitation of size. To conclude, the experiments carried on with
99.99 reconstructs ability and then each PQRST cycle is unimaginably high number of noise combinations the security
decomposed into a coefficients vector by the optimal wavelet strength of the presented method was very high.
function. The coefficients, approximations of the last scale
level and the details of the all levels, are used for the ECG Saxena et al in [10] described an approach for effective
analyzed. They divided the coefficients of each cycle into feature extraction form ECG signals. Their paper deals with an
three segments that are related to P-wave, QRS complex, and competent composite method which has been developed for
T-wave. The summation of the values from these segments data compression, signal retrieval and feature extraction of
provided the feature vectors of single cycles. ECG signals. After signal retrieval from the compressed data,
it has been found that the network not only compresses the
Mahmoodabadi et al. in [1] described an approach for ECG data, but also improves the quality of retrieved ECG signal
feature extraction which utilizes Daubechies Wavelets with respect to elimination of high-frequency interference
transform. They had developed and evaluated an present in the original signal. With the implementation of
electrocardiogram (ECG) feature extraction system based on artificial neural network (ANN) the compression ratio
the multi-resolution wavelet transform. The ECG signals from increases as the number of ECG cycle increases. Moreover the
Modified Lead II (MLII) were chosen for processing. The features extracted by amplitude, slope and duration criteria
wavelet filter with scaling function further intimately similar from the retrieved signal match with the features of the
to the shape of the ECG signal achieved better detection. The original signal. Their experimental results at every stage are
foremost step of their approach was to de-noise the ECG steady and consistent and prove beyond doubt that the
signal by removing the equivalent wavelet coefficients at composite method can be used for efficient data management
higher scales. Then, QRS complexes are detected and each and feature extraction of ECG signals in many real-time
applications.
77 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
region and suppressed in non-QRS region. The proposed
A feature extraction method using Discrete Wavelet method has detection rate and positive predictivity of 98.56%
Transform (DWT) was proposed by Emran et al. in [11]. They and 99.18% respectively.
used a discrete wavelet transform (DWT) to extract the
relevant information from the ECG input data in order to Xu et al. in [16] described an algorithm using Slope Vector
perform the classification task. Their proposed work includes Waveform (SVW) for ECG QRS complex detection and RR
the following modules data acquisition, pre-processing beat interval evaluation. In their proposed method variable stage
detection, feature extraction and classification. In the feature differentiation is used to achieve the desired slope vectors for
extraction module the Wavelet Transform (DWT) is designed feature extraction, and the non-linear amplification is used to
to address the problem of non-stationary ECG signals. It was get better of the signal-to-noise ratio. The method allows for a
derived from a single generating function called the mother fast and accurate search of the R location, QRS complex
wavelet by translation and dilation operations. Using DWT in duration, and RR interval and yields excellent ECG feature
feature extraction may lead to an optimal frequency resolution extraction results. In order to get QRS durations, the feature
in all frequency ranges as it has a varying window size, broad extraction rules are needed.
at lower frequencies, and narrow at higher frequencies. The
DWT characterization will deliver the stable features to the A method for automatic extraction of both time interval and
morphology variations of the ECG waveforms. morphological features, from the Electrocardiogram (ECG) to
classify ECGs into normal and arrhythmic was described by
Tayel and Bouridy together in [12] put forth a technique for Alexakis et al. in [17]. The method utilized the combination of
ECG image classification by extracting their feature using artificial neural networks (ANN) and Linear Discriminant
wavelet transformation and neural networks. Features are Analysis (LDA) techniques for feature extraction. Five ECG
extracted from wavelet decomposition of the ECG images features namely RR, RTc, T wave amplitude, T wave skew
intensity. The obtained ECG features are then further ness, and T wave kurtosis were used in their method. These
processed using artificial neural networks. The features are: features are obtained with the assistance of automatic
mean, median, maximum, minimum, range, standard algorithms. The onset and end of the T wave were detected
deviation, variance, and mean absolute deviation. The using the tangent method. The three feature combinations used
introduced ANN was trained by the main features of the 63 had very analogous performance when considering the
ECG images of different diseases. The test results showed that average performance metrics.
the classification accuracy of the introduced classifier was up
to 92%. The extracted features of the ECG signal using A modified combined wavelet transforms technique was
wavelet decomposition was effectively utilized by ANN in developed by Saxena et al. in [18]. The technique has been
producing the classification accuracy of 92%. developed to analyze multi lead electrocardiogram signals for
cardiac disease diagnostics. Two wavelets have been used, i.e.
Alan and Nikola in [13] proposed chaos theory that can be a quadratic spline wavelet (QSWT) for QRS detection and the
successfully applied to ECG feature extraction. They also Daubechies six coefficient (DU6) wavelet for P and T
discussed numerous chaos methods, including phase space and detection. A procedure has been evolved using
attractors, correlation dimension, spatial filling index, central electrocardiogram parameters with a point scoring system for
tendency measure and approximate entropy. They created a diagnosis of various cardiac diseases. The consistency and
new feature extraction environment called ECG chaos reliability of the identified and measured parameters were
extractor to apply the above mentioned chaos methods. A new confirmed when both the diagnostic criteria gave the same
semi-automatic program for ECG feature extraction has been results. Table 1 shows the comparison of different ECG signal
implemented and is presented in this article. Graphical feature extraction techniques.
interface is used to specify ECG files employed in the
extraction procedure as well as for method selection and A robust ECG feature extraction scheme was put forth by
results saving. The program extracts features from ECG files. Olvera in [19]. The proposed method utilizes a matched filter
to detect different signal features on a human heart
An algorithm was presented by Chouhan and Mehta in [14] electrocardiogram signal. The detection of the ST segment,
for detection of QRS complexities. The recognition of QRS- which is a precursor of possible cardiac problems, was more
complexes forms the origin for more or less all automated difficult to extract using the matched filter due to noise and
ECG analysis algorithms. The presented algorithm utilizes a amplitude variability. By improving on the methods used;
modified definition of slope, of ECG signal, as the feature for using a different form of the matched filter and better
detection of QRS. A succession of transformations of the threshold detection, the matched filter ECG feature extraction
filtered and baseline drift corrected ECG signal is used for could be made more successful. The detection of different
mining of a new modified slope-feature. In the presented features in the ECG waveform was much harder than
algorithm, filtering procedure based on moving averages [15] anticipated but it was not due to the implementation of the
provides smooth spike-free ECG signal, which is appropriate matched filter. The more complex part was creating the
for slope feature extraction. The foremost step is to extort revealing method to remove the feature of interest in each
slope feature from the filtered and drift corrected ECG signal, ECG signal.
by processing and transforming it, in such a way that the
extracted feature signal is significantly enhanced in QRS
78 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Jen et al. in [20] formulated an approach using neural performed other conventional methods of ECG feature
networks for determining the features of ECG signal. They extraction.
presented an integrated system for ECG diagnosis. The
integrated system comprised of cepstrum coefficient method III. FUTURE ENHANCEMENT
for feature extraction from long-term ECG signals and The electrocardiogram (ECG) is a noninvasive and the record
artificial neural network (ANN) models for the classification. of variation of the bio-potential signal of the human
Utilizing the proposed method, one can identify the heartbeats. The ECG detection which shows the information
characteristics hiding inside an ECG signal and then classify of the heart and cardiovascular condition is essential to
the signal as well as diagnose the abnormalities. To explore enhance the patient living quality and appropriate treatment.
the performance of the proposed method various types of ECG The ECG features can be extracted in time domain [23] or in
data from the MIT/BIH database were used for verification. frequency domain [24]. The extracted feature from the ECG
The experimental results showed that the accuracy of signal plays a vital in diagnosing the cardiac disease. The
diagnosing cardiac disease was above 97.5%. In addition the development of accurate and quick methods for automatic
proposed method successfully extracted the corresponding ECG feature extraction is of major importance. Some of the
feature vectors, distinguished the difference and classified features extraction methods implemented in previous research
ECG signals. includes Discrete Wavelet Transform, Karhunen-Loeve
Transform, Hermitian Basis and other methods. Every method
Correlation analysis for abnormal ECG signal feature has its own advantages and limitations. The future work
extraction was explained by Ramli and Ahmad in [21]. Their primarily focus on feature extraction from an ECG signal
proposed work investigated the technique to extract the using more statistical data. In addition the future enhancement
important features from the 12 lead system eye on utilizing different transformation technique that
(electrocardiogram) ECG signals. They chose II for their provides higher accuracy in feature extraction. The parameters
entire analysis due to its representative characteristics for that must be considered while developing an algorithm for
identifying the common heart diseases. The analysis technique feature extraction of an ECG signal are simplicity of the
chosen is the cross-correlation analysis. Cross-correlation algorithm and the accuracy of the algorithm in providing the
analysis measures the similarity between the two signals and best results in feature extraction.
extracts the information present in the signals. Their test
results suggested that the proposed technique could effectively Table I. Comparison of Different Feature Extraction Techniques from an ECG
extract features, which differentiate between the types of heart Signal where H, M, L denotes High, Medium and Low respectively
diseases analyzed and also for normal heart signal.
Approach Simplicity Accuracy Predictivity
Ubeyli et al. in [22] described an approach for feature
Zhao et al. H H H
extraction from ECG signal. They developed an automated
diagnostic systems employing dissimilar and amalgamated Mahmoodabadi
M H H
features for electrocardiogram (ECG) signals were analyzed et al.
and their accuracies were determined. The classification Tadejko and
accuracies of mixture of experts (ME) trained on composite L M M
Rakowski
features and modified mixture of experts (MME) trained on Tayel and
diverse features were also compared in their work. The inputs M M H
Bouridy
of these automated diagnostic systems were composed of Jen et al.
diverse or composite features and these were chosen based on H H H
the network structures. The achieved accuracy rates of their Alexakis et al.
proposed approach were higher than that of the ME trained on H M M
composite features. Ramli and
M M M
Ahmad
Fatemian et al. [25] proposed an approach for ECG feature Xu et al.
extraction. They suggested a new wavelet based framework M H H
for automatic analysis of single lead electrocardiogram (ECG) Olvera
for application in human recognition. Their system utilized a H M M
robust preprocessing stage, which enables it to handle noise Emran et al
and outliers. This facilitates it to be directly applied on the raw H M L
ECG signal. In addition the proposed system is capable of
managing ECGs regardless of the heart rate (HR) which
IV. CONCLUSION
renders making presumptions on the individual's stress level
unnecessary. The substantial reduction of the template gallery The examination of the ECG has been comprehensively
size decreases the storage requirements of the system used for diagnosing many cardiac diseases. Various
appreciably. Additionally, the categorization process is techniques and transformations have been proposed earlier in
speeded up by eliminating the need for dimensionality literature for extracting feature from ECG. This proposed
reduction techniques such as PCA or LDA. Their experimental paper provides an over view of various ECG feature extraction
results revealed the fact that the proposed technique out techniques and algorithms proposed in literature. The feature
79 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
extraction technique or algorithm developed for ECG must be [17] C. Alexakis, H. O. Nyongesa, R. Saatchi, N. D. Harris, C. Davies, C.
Emery, R. H. Ireland, and S. R. Heller, “Feature Extraction and
highly accurate and should ensure fast extraction of features
Classification of Electrocardiogram (ECG) Signals Related to
from the ECG signal. This proposed paper also revealed a Hypoglycaemia,” Conference on computers in Cardiology, pp. 537-540,
comparative table evaluating the performance of different IEEE, 2003.
algorithms that were proposed earlier for ECG signal feature [18] S. C. Saxena, V. Kumar, and S. T. Hamde, “Feature extraction from
ECG signals using wavelet transforms for disease diagnostics,”
extraction. The future work mainly concentrates on
International Journal of Systems Science, vol. 33, no. 13, pp. 1073-1085,
developing an algorithm for accurate and fast feature 2002.
extraction. Moreover additional statistical data will be utilized [19] Felipe E. Olvera, “Electrocardiogram Waveform Feature Extraction
for evaluating the performance of an algorithm in ECG signal Using the Matched Filter,” 2006.
[20] Kuo-Kuang Jen, and Yean-Ren Hwang, “ECG Feature Extraction and
feature detection. Improving the accuracy of diagnosing the
Classification Using Cepstrum and Neural Networks,” Journal of
cardiac disease at the earliest is necessary in the case of patient Medical and Biological Engineering, vol. 28, no. 1, 2008.
monitoring system. Therefore our future work also has an eye [21] A. B. Ramli, and P. A. Ahmad, “Correlation analysis for abnormal ECG
on improvement in diagnosing the cardiac disease. signal features extraction,” 4th National Conference on
Telecommunication Technology, 2003. NCTT 2003 Proceedings, pp.
232-237, 2003.
REFERENCES [22] Ubeyli, and Elif Derya, “Feature extraction for analysis of ECG signals,”
[1] S. Z. Mahmoodabadi, A. Ahmadian, and M. D. Abolhasani, “ECG Engineering in Medicine and Biology Society, 2008. EMBS 2008. 30th
Feature Extraction using Daubechies Wavelets,” Proceedings of the fifth Annual International Conference of the IEEE, pp. 1080-1083, 2008.
IASTED International conference on Visualization, Imaging and Image [23] Y. H. Hu, S. Palreddy, and W. Tompkins, “A Patient Adaptable ECG
Processing, pp. 343-348, 2005. Beat Classifier Using A Mixture Of Experts Approach”, IEEE
[2] Juan Pablo Martínez, Rute Almeida, Salvador Olmos, Ana Paula Rocha, Transactions on Biomedical Engineering vol. 44, pp. 891–900, 1997.
and Pablo Laguna, “A Wavelet-Based ECG Delineator: Evaluation on [24] Costas Papaloukas, Dimitrios I. Fotiadis, Aristidis Likas, Lampros K.
Standard Databases,” IEEE Transactions on Biomedical Engineering Michalis, “Automated Methods for Ischemia Detection in Long Duration
Vol. 51, No. 4, pp. 570-581, 2004. ECGs”, 2003.
[3] Krishna Prasad and J. S. Sahambi, “Classification of ECG Arrhythmias [25] S. Z. Fatemian, and D. Hatzinakos, “A new ECG feature extractor for
using Multi-Resolution Analysis and Neural Networks,” IEEE biometric recognition,” 16th International Conference on Digital Signal
Transactions on Biomedical Engineering, vol. 1, pp. 227-231, 2003. Processing, pp. 1-6, 2009.
[4] Cuiwei Li, Chongxun Zheng, and Changfeng Tai, “Detection of ECG
Characteristic Points using Wavelet Transforms,” IEEE Transactions on AUTHORS PROFILE
Biomedical Engineering, Vol. 42, No. 1, pp. 21-28, 1995.
[5] C. Saritha, V. Sukanya, and Y. Narasimha Murthy, “ECG Signal
Analysis Using Wavelet Transforms,” Bulgarian Journal of Physics, Karpagachelvi.S: She received the BSc degree in
vol. 35, pp. 68-77, 2008. physics from Bharathiar University in 1993 and
[6] Qibin Zhao, and Liqing Zhan, “ECG Feature Extraction and Masters in Computer Applications from Madras
Classification Using Wavelet Transform and Support Vector University in 1996. She has 12 years of teaching
Machines,” International Conference on Neural Networks and Brain, experience. She is currently a PhD student with the
ICNN&B ’05, vol. 2, pp. 1089-1092, 2005. Department of Computer Science at Mother Teresa
[7] B. Castro, D. Kogan, and A. B. Geva, “ECG feature extraction using University.
optimal mother wavelet,” The 21st IEEE Convention of the Electrical
and Electronic Engineers in Israel, pp. 346-350, 2000.
[8] P. Tadejko, and W. Rakowski, “Mathematical Morphology Based ECG
Feature Extraction for the Purpose of Heartbeat Classification,” 6th
International Conference on Computer Information Systems and
Industrial Management Applications, CISIM '07, pp. 322-327, 2007.
[9] F. Sufi, S. Mahmoud, I. Khalil, “A new ECG obfuscation method: A Dr.M.Arthanari: He has obtained Doctorate in
joint feature extraction & corruption approach,” International Mathematics in Madras University in the year 1981.
Conference on Information Technology and Applications in He has 35 years of teaching experience and 25 years
Biomedicine, 2008. ITAB 2008, pp. 334-337, May 2008. of research experience. He has a Patent in Computer
[10] S. C. Saxena, A. Sharma, and S. C. Chaudhary, “Data compression and Science approved by Govt. of India.
feature extraction of ECG signals,” International Journal of Systems
Science, vol. 28, no. 5, pp. 483-498, 1997.
[11] Emran M. Tamil, Nor Hafeezah Kamarudin, Rosli Salleh, M. Yamani
Idna Idris, Noorzaily M.Noor, and Azmi Mohd Tamil, “Heartbeat
Electrocardiogram (ECG) Signal Feature Extraction Using Discrete
Wavelet Transforms (DWT).” Sivakumar M : He has 10+ years of experience in the
[12] Mazhar B. Tayel, and Mohamed E. El-Bouridy, “ECG Images software industry including Oracle Corporation. He
Classification Using Feature Extraction Based On Wavelet received his Bachelor degree in Physics and Masters
Transformation And Neural Network,” ICGST, International in Computer Applications from the Bharathiar
Conference on AIML, June 2006. University, India. He holds patent for the invention
[13] Alan Jovic, and Nikola Bogunovic, “Feature Extraction for ECG Time- in embedded technology. He is technically certified
Series Mining based on Chaos Theory,” Proceedings of 29th by various professional bodies like ITIL, IBM
International Conference on Information Technology Interfaces, 2007. Rational Clearcase Administrator, OCP - Oracle
[14] V. S. Chouhan, and S. S. Mehta, “Detection of QRS Complexes in 12- Certified Professional 10g and ISTQB.
lead ECG using Adaptive Quantized Threshold,” IJCSNS International
Journal of Computer Science and Network Security, vol. 8, no. 1, 2008.
[15] V. S. Chouhan, and S. S. Mehta, “Total Removal of Baseline Drift from
ECG Signal”, Proceedings of International conference on Computing:
Theory and Applications, ICTTA–07, pp. 512-515, ISI, March, 2007.
[16] Xiaomin Xu, and Ying Liu, “ECG QRS Complex Detection Using Slope
Vector Waveform (SVW) Algorithm,” Proceedings of the 26th Annual
International Conference of the IEEE EMBS, pp. 3597-3600, 2004.
80 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Implementation of the Six Channel Redundancy
to achieve fault tolerance in testing of satellites.
H S Aravinda Dr H D Maheshappa Dr Ranjan Moodithaya
Dept of ECE, REVA ITM, Director & Principal, E PC ET Head, KTM Division, NAL,
Bangalore-64, Karnataka, India Bangalore- 40, Karnataka, India Bangalore-17, Karnataka, India
aravindhs1@gmail.com hdmappa@gmail.com ranjanmk@hotmail.com
Abstract:-This paper aims to implement the six channel The Indian Space Research Organization launches number
redundancy to achieve fault tolerance in testing of satellites of satellites for application in communication [5], remote
with acoustic spectrum. We mainly focus here on achieving sensing, meteorology etc. The powerful launch vehicles are
fault tolerance. An immediate application is the microphone used to accelerate the satellite through the earth’s
data acquisition and to do analysis at the Acoustic Test Facility atmosphere and to make it an artificial earth satellite. The
(ATF) centre, National Aerospace Laboratories. It has an 1100 Launch Vehicles [6] used will generate high levels of sound
cubic meter reverberation chamber in which a maximum during lift-off and Tran’s atmospheric acceleration. The
sound pressure level of 157 dB is generated. The six channel
payload satellites experiences mechanical loads of various
Redundancy software with fault tolerant operation is devised
and developed. The data are applied to program written in C frequencies and load on the vehicle from acoustic sources
language. The program is run using the Code Composer Studio due to two factors. One is Rocket vehicle generated noise at
by accepting the inputs. This is tested with the TMS 320C 6727 lift-off, and the other is an aerodynamic noise caused by
DSP, Pro Audio Development Kit (PADK). turbulence, particularly at frontal area transition. The
Key words: Fault Tolerance, Redundancy, Acoustics. acoustic field thus created is strong enough to damage the
delicate payload. The sources of acoustics, its combined
I. INTRODUCTION spectrum are shown in fig.2 and fig.3.
Acoustic Test Facility is a national facility for acoustic
environmental qualification of satellites, launch vehicle
stages and their subsystems for the ISRO [1]. The ATF
has a reverberation chamber (RC) for simulating the
acoustic environment experienced by spacecraft and launch
vehicles during launch [2]. The RC has a diffused uniform
sound pressure level distribution. Its wall surface ensures
reflectance of 99% of the sound energy. It is used for
simulating the acoustic environment experienced by
spacecraft and launch vehicles during the launch. The one
such facility is shown in Fig.1.
Fig.2. the load on the vehicle from two acoustic sources.
Fig.1. View of the Reverberation Chamber
pdfMachine
A pdf writer that produces quality PDF files with ease!
Produce quality PDF files in seconds and preserve the integrity of your original documents. Compatible across 81
nearly all Windows platforms, simply open the document you want to convert, click “print”, select the
“Broadgun pdfMachine printer” and that’s it! Get yours now!
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
realized in the RC, like the US Delta, Atlas Centaur, Titan
IIIC, Space Shuttle, Ariane 4 & 5 of the ESA, Vostok,
Soyuz of Russia and ASLV, PSLV and GSLV of India.
A. Requirements For acoustic Testing of Satellite are
Noise generation unit, Spectrum Shapers, Power
amplifiers and horns (two), Reverberation Chamber, Micro
phones and Multiplexer, Real time frequency (spectrum)
analyzer, Graphic recorder and display, Tape recorder for
recording, Accelerometers, Charge amplifier and data
recorder, Dual channel analyzer and plotter.
Fig.3. The combined spectra from the two acoustic sources.
Faults and failures are not acceptable due to high cost of
satellites. Hence, all payload satellites should undergo an
acoustic test before launching under simulated conditions
and are tested for their endurance of such dynamic loads.
The satellite is subjected to maximum overall sound
pressure level to ensure the functional aspects of all the test
setup. Acoustic test is a major dynamic test for qualification
of space systems and components. The purpose of the tests
are, Search for weak elements in the subsystem with respect
to acoustic failure. The Qualification tests to demonstrate
spacecraft performance in meeting design goals set.
Acceptance tests to uncover workmanship nature of defects.
II. ACOUSTIC TESTING
The acoustic environment inside the Reverberation Chamber
is created by modulating a stream of clean and dry air at
about 30 PSI pressure using electro pneumatic transducers. The satellite is kept in RC and the high frequency high level
The drive signal is derived from a random noise generator spectrum characteristics of the launch vehicle are generated
and modified by a spectrum shaper. The microphone data and its dynamic behavior is studied. It is essential that the
from the RC is observed on a real time analyzer and the acoustic noise generated is a true simulation of the launch
spectrum shaper outputs are adjusted to achieve the target vehicle acoustic spectrum, and it is the input acoustic load to
spectrum. There are two sets of modulators, one delivering be simulated in the RC and is specified as SPL (sound
an acoustic power of 60KW in the 31.5 Hz to 500 Hz and pressure level) in dB verses frequency. The spectrum of
the others delivering 20 KW in the 200 to 1200 Hz range, various launch vehicles like delta, atlas centaur, titan-IIIC,
the spectrum beyond 1200 HZ is controlled to some extent Arianne, vostok, ASLV, PSLV, GSLV.., and Indian
using the effects of the higher harmonics by changing the satellites like IRS, INSAT.., are realizable in the
spectral contents of the drive to the modulators. The Reverberation Chamber. Each launch vehicle has unique
acoustic excitation is coupled to the RC through optimally spectral features and is drawn in Octave Band Centre
configured exponential horns to achieve efficient transfer of Frequencies (OBCF), in the range from 31.5 Hz to 16 kHz.
the acoustic energy into the chamber. The chamber wall
surface treatment design ensures reflectance of 99% of the
sound energy incident on them. The chamber has a diffused
uniform, sound pressure level distribution with in 1dB in
the central ten percent of volume of the chamber where the
test specimen is located. The spectrum for almost all
contemporary launch vehicles around the world can be
pdfMachine
A pdf writer that produces quality PDF files with ease!
Produce quality PDF files in seconds and preserve the integrity of your original documents. Compatible across 82
nearly all Windows platforms, simply open the document you want to convert, click “print”, select the
“Broadgun pdfMachine printer” and that’s it! Get yours now!
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
The Three levels of acceptance of acoustic spectrum are ,
first is Full level or Qualification test, it is normally for 120
seconds and Maximum of 156dB, second is Acceptance
level test, it is normally for 60 or 90 seconds and Maximum
of 153dB, third is a Low level test, it is normally for 30
seconds and Maximum of 150dB.
III. IMPLEMENTATION
Fault tolerant application software to ensure data integrity will
be developed. This paper is implemented by taking the six Fig.6. input data 1
channel data from reverberation chamber and is applied as the
input to the program. The six microphones data are connected to Comment: The six channel input data, indicating channel 5
the TMS 320C 6727 DSP, Pro Audio Development Kit (PADK) is going bad (low) from duration 4 -10.
after signal conditioning via analog to digital converters. All six
microphone data is fed to DSP processor as shown in Fig.5. The
FFT is taken for all the six channel data and are compared with
each other to find out which channel microphone data is good or
bad. A threshold level is maintained to check the validity of the
microphone. If the data is well with in the threshold it is
accepted or else it will be rejected. Here if the two channel
microphone data is bad then it will only be identified.
Sc-1 ADC
Sc-2 ADC TMS Fig.7. output data 1
320C
6727 Comment: The six channel output data, indicating all are
Sc-3 ADC good except channel 5 is going bad, which is reflected as
DSP
PAD low from the duration 4-10.
Sc-4 ADC K,
KIT
Sc-5 ADC
Sc-6 ADC
Fig.5. Block Diagram representation for six channel redundancy
management technique.
IV. TEST AND RESULTS
The data is extracted using the six microphone channels. It
Fig.8. input data 2
is fed to the TMS320C 6727 DSP, Pro Audio Development
Kit (PADK) for further processing using the Code Comment: The six channel input data, indicating channel 2
Composer Studio for different cases. The data is applied as is going bad (high) from duration 2.5 -10.
an input to the program written in C language and is run
using the code composer studio. The results are obtained for
different cases are shown below.
pdfMachine
A pdf writer that produces quality PDF files with ease!
Produce quality PDF files in seconds and preserve the integrity of your original documents. Compatible across 83
nearly all Windows platforms, simply open the document you want to convert, click “print”, select the
“Broadgun pdfMachine printer” and that’s it! Get yours now!
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Fig.9. output data 2 Fig.12. input data 4
Comment: The six channel output data, indicating all are Comment: The six channel input data, indicating channel 2
good except channel 2 is going bad, which is reflected as is going bad (high) from duration 2.5 -10, and channel 5
low from the duration 2.5-10. going bad (low) from duration 4-10.
Fig.10. input data 3
Fig.13. output data 4
Comment: The six channel input data, indicating channel 6
is going bad (low) from duration 3.5 -10. Comment: The six channel output data, indicating all are
good except channel 2 and channel 4 is going bad, which is
reflected as low from the duration 2.5-10 for channel 2 and
from the duration 4-10 for channel 4.
Fig.11. output data 3
Comment: The six channel output data, indicating all are
good except channel 6 is going bad, which is reflected as
low from the duration 3.5-10.
Fig.14. input data 5
Comment: The six channel input data, indicating channel 2
is going bad ( low ) from duration 1-10, and channel 6
going bad ( low) from duration 3.5-10.
pdfMachine
A pdf writer that produces quality PDF files with ease!
Produce quality PDF files in seconds and preserve the integrity of your original documents. Compatible across 84
nearly all Windows platforms, simply open the document you want to convert, click “print”, select the
“Broadgun pdfMachine printer” and that’s it! Get yours now!
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
AUTHORS PROFILE
Sri. H.S. Aravinda Graduated in
Electronics & Communication
Engineering from University of
Bangalore, India in 1997. He has a Post
Graduation in Bio- Medical
Instrumentation from University of My
sore, India in 1999. His special interests
in research are Fault tolerance, signal
processing. He has been teaching engineering for UG & P G for last 12
years. He served various engineering colleges as a teacher and at present
he is an Assistant professor in the Department of Electronics &
Communication in Reva Institute of Technology & Management,
Bangalore, India. He has more than 13 research papers in various National
and International Journals & Conferences. He is a member of ISTE Also
Fig.15. output data 5 has served on the advisory and technical national conferences.
Comment: The six channel output data, indicating all are
Sri. Dr. H.D. Maheshappa Graduated
good except channel 4 and channel 6 is going bad, which is in Electronics & Communication
reflected as low from the duration 1-10 for channel 4 and Engineering from University of
from the duration 3.5-10 for channel 6. Mysore, India in 1983. He has a Post
Graduation in Industrial Electronics
from University of My sore, India in
IV. CONCLUSION 1987. He holds a Doctoral Degree in
Engineering from Indian Institute of
Science, Bangalore, India, since 2001.
The data is compared with the results obtained. If we compare He is specialized in Electrical contacts, Micro contacts, Signal integrity
the output results with that of the data input, the output is interconnects etc. His special interests in research are Bandwidth
becoming zero whenever there is less amplitude data in the input Utilization in Computer Networks. He has been teaching engineering for
and also high amplitude in the input indicating the wrong data. UG & P G for last 25 years. He served various engineering colleges as a
teacher and at present he is a Professor & Head of the Department of
The wrong data is identified and are displayed in the plots. Here Electronics & Communication in Reva Institute of Technology &
if the two channel microphone data is bad then it will only be Management, Bangalore , India. He has more than 35 research papers in
identified. The results are matching with the expected output. It various National and International Journals & Conferences . He is a
proves that the algorithm implemented in C language is member of IEEE, ISTE, CSI & ISOI. He is a member of Doctoral
Committee of Coventry University UK. He has been a Reviewer of many
effectively working for the given data. It is successfully Text Books for the publishers McGraw-Hill Education (India) Pvt., Ltd.,
identifying and detecting the correct and wrong data. Hence we Chaired Technical Sessions, and National Conferences and also has served
could verify and prove the redundancy software works better for on the advisory and technical national conferences.
achieving fault tolerance for testing of satellites with acoustic
Sri. Dr.. Ranjan Moodithaya did his MSc from
spectrum. Mysore University in 1970 and got his Ph D
V. REFERENCES from Indian Institute of Science in 1986. He
joined NAL in 1973 after 2 yrs of teaching at
[1] Aravinda H S, Dr H D Maheshappa, Dr Ranjan Moodithaya ,
“Verification of the Six Channel Quad Redundancy Management Software Post Graduate centre of Mysore University. At
with the Fault Tolerant Measurement Techniques of Acoustic Spectrum of present, he is Scientist G and he is incharge of
Satellites” International conference on PDPTA’09 WORLDCOMP 2009, the NAL-ISRO Acoustic Test Facility and
NAL’s Knowledge and Technology
held at lasvegas, Nevada, USA, Vol II, PP 553 to 558, July 13-16 2009.
Management Division. He is a Life member of
Acoustic Society of India and Aeronautical
[2] R.K. Das, S.Sen & S. Dasgupta, “ Robust and fault tolerant controller Society of India. His innovative products are
for attitude control of a satellite launch vehicle ” IET Control theory & sold to prestigious institutions like Westinghouse, Lockheed, Boeing and
applications, PP 304 to 312, Feb 2007. Mitsubishi through M/s. Wyle Laboratories, USA, the world leaders in the
design of acoustic facilities. Dr. Ranjan has more than a dozen publications
[3] Jan-lung sung “A dynamic slack management for real time distributed
systems” IEEE Transaction on computers, PP 30 to 39, Feb 2008. in international and national journals and more than 30 internal technical
publications. He is a Life member of Acoustic Society of India and
[4] Louis p Bolduc “Redundancy management system for the x-33 vehicle Aeronautical Society of India. He is also a Member of Instrument Society
and mission computer”. IEEE Transaction on computers, PP 31 to 37, May of India.
2000.
[5] Hilmer, H. , Kochs, H.-D. and Dittmar, E. “A Fault tolerant
communication architecture for real time control systems”. IEEE
Transaction on computers, PP 111 to 118, Jun1997.
[6] Oleg Sokolsky, Mohamed Younisy, Insup Leez, Hee-Hwan Kwakz and
Jeff Zhouy “Verification of the Redundancy Management System for Space
Launch Vehicle” IEEE Transaction on computers, PP 42 to 52, Sep 1998.
pdfMachine
A pdf writer that produces quality PDF files with ease!
Produce quality PDF files in seconds and preserve the integrity of your original documents. Compatible across 85
nearly all Windows platforms, simply open the document you want to convert, click “print”, select the
“Broadgun pdfMachine printer” and that’s it! Get yours now!
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No.1, 2010
PERFORMANCE ORIENTED QUERY PROCESSING IN GEO
BASED LOCATION SEARCH ENGINES
1 2
Dr.M.Umamaheswari S.Sivasubramanian
Dept. of Computer Science, Bharath University Dept. of Computer Science, Bharath University
Chennai-73, Tamil Nadu, India Chennai-73, Tamil Nadu, India
druma_cs@yahoo.com sivamdu2001@Yahoo.Com
ABSTRACT increasingly challenging to satisfy certain
Geographic location search engines information needs. While search engines are still
allow users to constrain and order search able to index a reasonable subset of the (surface)
results in an intuitive manner by focusing a web, the pages a user is really looking for are
query on a particular geographic region. often buried under hundreds of thousands of less
Geographic search technology, also called interesting results. Thus, search engine users are
location search, has recently received in danger of drowning in information. Adding
significant interest from major search engine additional terms to standard keyword searches
companies. Academic research in this area has often fails to narrow down results in the desired
focused primarily on techniques for extracting direction. A natural approach is to add advanced
geographic knowledge from the web. In this features that allow users to express other
paper, we study the problem of efficient query constraints or preferences in an intuitive
processing in scalable geographic search manner, resulting in the desired documents to be
engines. Query processing is a major bottleneck returned among the first results. In fact, search
in standard web search engines, and the main engines have added a variety of such features,
reason for the thousands of machines used by often under a special “advanced search”
the major engines. Geographic search engine interface, but mostly limited to fairly simple
query processing is different in that it requires conditions on domain, link structure, or
a combination of text and spatial data modification date. In this paper we focus on
processing techniques. We propose several geographic web search engines, which allow
algorithms for efficient query processing in users to constrain web queries to certain
geographic search engines, integrate them into geographic areas. In many cases, users are
an existing web search query processor, and interested in information with geographic
evaluate them on large sets of real data and constraints, such as local businesses, locally
query traces. relevant news items, or Permission to make
Key word: location, search engine, query digital or hard copies of all or part of this work
processing for personal or classroom use is granted without
I.INTRODUCTION The World-Wide Web fee provided that copies are not made or
has reached a size where it is becoming distributed for profit or commercial advantage
86 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No.1, 2010
and that copies bear this notice and the full the bridge to Tambaram or nearby in Gundy
citation on the first page. To copy otherwise, tore might also be acceptable). Second, geographic
publish, to post on servers or to redistribute to search is a fundamental enabling technology for
lists, requires prior specific tourism information “location-based services”, including electronic
about a particular region. For example, when commerce via cellular phones and other mobile
searching for yoga classes, local yoga schools devices. Third, geographic search supports
are of much higher interest than the web sites of locally targeted web advertising, thus attracting
the world’s largest yoga schools. We expect that advertisement budgets of small businesses with a
‘geographic search engine’s, that is, search local focus. Other opportunities arise from
engines that support geographic preferences, will mining geographic properties of the web,
have a major impact on search technology and example, for market research and competitive
their business models. First, geographic search intelligence. Given these opportunities, it comes
engines provide a very useful tool. They allow as no surprise that over the last two years leading
users to express in a single query what might search engine companies such as Google and
take multiple queries with a standard Yahoo have made significant efforts to deploy
search engine. their own versions of geographic web search.
A. LOCATION BASED There has also been some work by the academic
A user of a standard search engine looking for a research community, to mainly on the problem
yoga school in or close to Tambaram, Chennai, of extracting geographic knowledge from web
might have to try queries such as pages and queries. Our approach here is based on
• yoga ‘‘Delhi’’ a setup for geographic query processing that we
• yoga “Chennai” recently introduced in [1] in the context of a
• yoga‘‘Tambaram’’ (a part of Chennai) geographic search engine prototype. While there
are many different ways to formalize the query
processing problem in geographic search
engines, we believe that our approach results in a
very general framework that can capture many
scenarios.
but this might yield inferior results as there are
many ways to refer to a particular area, and since B. QUERY FOOTPRINT
a purely text-based engine has no notion of We focus on the efficiency of query processing
geographical closeness (example, a result across in geographic search engines, example, how to
87 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No.1, 2010
maximize the query throughput for a given approaches it is assumed that this information is
problem size and amount of hardware. Query provided via meta tags or by third parties. The
processing is the major performance bottleneck resulting page footprint is an arbitrary, possibly
in current standard web search engines, and the noncontiguous area, with an amplitude value
main reason behind the thousands of machines specifying the degree of relevance of each
used by larger commercial players. Adding location. Footprints can be represented as
geographic constraints to search queries results polygons or bitmap-based structures; details of
in additional challenges during query execution the representation are not important here. A geo
which we now briefly outline. In a nutshell, search engine computes and orders results based
given a user query consisting of several on two factors.
keywords, a standard search engine ranks the C. KEYWORDS AND GEOGRAPHY.
pages in its collection in terms of their relevance Given a query, it identifies pages that contain the
to the keywords. This is done by using a text keywords and whose page footprint intersects
index structure called an inverted index to with the query footprint, and ranks these results
retrieve the IDs of pages containing the according to a combination of a term-based
keywords, and then evaluating a term-based ranking function and a geographic ranking
ranking function on these pages to determine the function that might, example, depend on the
k highest-scoring pages. (Other factors such as volume of the intersection between page and
hyperlink structure and user behavior are also query footprint. Page footprints could of course
often used, as discussed later). Query processing be indexed via standard spatial indexes such as
is highly optimized to exploit the properties of R∗-trees, but how can such index structures be
inverted index structures, stored in an optimized
integrated into a search engine query processor,
compressed format, fetched from disk using
which is optimized towards inverted index
efficient scan operations, and cached in main
structures? How should the various structures be
memory. In contrast, a query to a geographic
laid out on disk for maximal throughput, and
search engine consists of keywords and the
how should the data flow during query execution
geographic area that interests the user, called
in such a mixed engine? Should we first execute
“query footprint”.
the textual part of the query, or first the spatial
Each page in the search engine also has a
part, or choose a different ordering for each
geographic area of relevance associated with it,
query? These are the basic types of problems that
called the ‘geographic footprin’t of the page.
we address in this paper. We first provide some
This area of relevance can be obtained by
background on web search engines and
analyzing the collection in a preprocessing step
geographic web search technology. We assume
that extracts geographic information, such as city
that readers are somewhat familiar with basic
names, addresses, or references to landmarks,
spatial data structures and processing, but may
from the pages and then maps these to positions
have less background about search engines and
using external geographic databases. In other
their inner workings. Our own perspective is
88 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No.1, 2010
more search-engine centric: given a high- number of search term occurrences and their
performance search engine query processor positions and contexts, to compute a score for
developed in our group, our goal is to efficiently each document containing the search terms. We
integrate the types of spatial operations arising in now formally introduce some of these concepts.
geographic search engines A. DOCUMENTS, TERMS, AND QUERIES:
II. BASICS OF SEARCH ENGINE ARCHITECTURE We assume a collection D = {d0, d1, . . . dn−1}
The basic functions of a crawl-based web search of n web pages that have been crawled and are
engine can be divided into ‘crawling, data stored on disk. Let W = {w0,w1, . . . , wm−1} be
mining, index construction, and query all the different words that occur anywhere in D.
processing’. During crawling, a set of initial seed Typically, almost any text string that appears
pages is fetched from the web, parsed for between separating symbols such as spaces,
hyperlinks, and then the pages pointed to by commas, etc., is treated as a valid word (or term).
these hyperlinks are fetched and parsed, and so A query
on, until a sufficient number of pages has been q = {t0, t1, . . . , td−1} (1)
acquired. Second, various data mining operations is a set1 of words (terms).
are performed on the acquired data, example, B. INVERTED INDEX:
detection of web spam and duplicates, link An inverted index I for the collection consists of
analysis based on Page rank [7], or mining of a set of inverted lists
word associations. Third, a text index structure is Iw0, Iw1, . . . , Iwm−1 (2)
built on the collection to support efficient query Where list Iw contains a posting for each
processing. Finally, when users issue queries, the occurrence of word w. Each posting contains the
top-10 results are retrieved by traversing this ID of the document where the word occurs, the
index structure and ranking encountered pages position within the document, and possibly some
according to various measures of relevance. context (in a title, in large or bold font, in an
Search engines typically use a text index anchor text). The postings in each inverted list
structure called an inverted index, which allows are usually sorted by document IDs and laid out
efficient retrieval of documents containing a sequentially on disk, enabling efficient retrieval
particular word (term). Such an index consists of and decompression of the list. Thus, Boolean
many inverted lists, where each inverted list Iw queries can be implemented as unions and
contains the IDs of all documents in the intersections of these lists, while phrase searches
collection that contain a particular word w, C. TERM-BASED RANKING:
usually sorted by document ID, plus additional The most common way to perform ranking is
information about each occurrence. Given, based on comparing the words (terms) contained
example, a query containing the search terms in the document and in the query. More
“apple”,“ orange”, and “pear”, a search engine precisely, documents are modeled as unordered
traverses the inverted list of each term and uses bags of words, and a ranking function assigns a
the information embedded therein, such as the score to each document with respect to the
89 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No.1, 2010
current query, based on the frequency of each highly efficient scan operations, without any
query word in the page and in the overall random lookups.
collection, the length of the document, and III. BASICS OF GEOGRAPHIC WEB SEARCH
maybe the context of the occurrence (example, We now discuss the additional issues that arise in
higher score if term in title or bold face). a geographic web search engine. Most details of
Formally, given “a query (1) is”, a ranking the existing commercial systems are proprietary;
function F assigns to each document D a score F our discussion here draws from the published
(D, q). The system then returns the k documents descriptions of academic efforts in [1, 3] the first
with the highest score. One popular class of task, crawling, stays the same if the engine aims
ranking functions is the cosine measure [44], for to cover the entire web. In our systems we focus
example on Germany and crawl the de domain; in cases
where the coverage area does not correspond
well to any set of domains, focused crawling
(3)
strategies [4 may be needed to find the relevant
In the equation (3) Where fD,ti and fti are the
pages.
frequency of term ti in document D and in the
A. GEO CODING: Additional steps are performed
entire collection, respectively. Many other
as part of the data mining task in geographical
functions have been proposed, and the
search engines, in order to extract geographical
techniques in this paper are not limited to any
information from the collection. Recall that the
particular class. In addition, scores based on link
footprint of a page is a potentially noncontiguous
analysis or user feedback are often added into the
area of geographical relevance. For every
total score of a document; in most cases this does
location in the footprint, an associated integer
not affect the overall query execution strategy if
value expresses the certainty with which we
these contributions can be pre computed offline
believe the page is actually relevant to the
and stored in a memory-based table or embedded
location. The process of determining suitable
into the index. For example, the ranking function
geographic footprints for the pages is called ‘geo
might become something like F (D, q) =
coding’ [3] In [1], geo coding consists of three
pr(D)+F(D, q) where pr(D) is a pre computed
steps, geo extraction, geo matching, and geo
and suitably normalized Page rank score of page
propagation. The first step extracts all elements
D. The key point is that the above types of
from a page that indicate a location, such as city
ranking functions can be computed by first
names, addresses, landmarks, phone numbers, or
scanning the inverted lists associated with the
company names. The second step maps the
search terms to find the documents in their
extracted elements to actual locations (that is,
intersection, and then evaluating the ranking
coordinates), if necessary resolving any
function only on those documents, using the
remaining ambiguities, example, between cities
information embedded in the index. Thus, at
of the same name. This results in an initial set of
least in its basic form, query processing with
footprints for the pages. Note that if a page
inverted lists can be performed using only a few
90 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No.1, 2010
contains several geographic references, its example, as polygons, would also work. All of
footprint may consist of several noncontiguous our algorithms approximate the footprints by sets
areas, possibly with higher certainty values of bounding rectangles; we only assume the
resulting, say, from a complete address at the top existence of a black-box procedure for
of a page or a town name in the URL than from a computing the precise geographical score
single use of a town name somewhere else in the between a query footprint and a document
page text. footprint. During index construction, additional
spatial index structures are created for document
R*TREE Spatial Index Structures India footprints as described later.
B. GEOGRAPHIC QUERY PROCESSING: As in
North East Middle
zone
West South zone [1], each search query consists of a set of
zone zone zone
(textual) terms, and a query footprint that
Kerala Karnataka AP TamilNadu
specifies the geographical area of interest to the
user. We assume a geographic ranking function
Chennai Coimbat Madurai Salem that assigns a score to each document footprint
with respect to the query footprint, and that is
Tambaram Puthamallai Vellachary T.Nagar
zero if the intersection is empty; natural choices
Tambaram Puthamalla Vellachary T.Nagar
Local Data i Local Data Local Data
are the inner product or the volume of the
intersection. Thus, our overall ranking function
Figure 1.1shows an example of a page and its might be of the form F (D, q) = g(fD, fq) + pr(D)
split footprint. + F(D, q), with a term-based ranking function
The third step, geo propagation, improves F(D, q), a global rank pr(D) (example,
quality and coverage often initial geo coding by Pagerank), and a geographic score g(fD, fq)
analysis of link structure and site topology. Thus, computed from query footprint fq and document
a page on the same site as many pages relevant footprint fD (with appropriate normalization of
to Chennai City, or with many hyperlinks to or the three terms). Our focus in this paper is on
from such pages, is also more likely to be how to efficiently compute such ranking
relevant to Chennai and should inherit such a functions using a combination of text and spatial
footprint (though with lower certainty). In index structures. Note that the query footprint
addition, geo coding might exploit external data can be supplied by the user in a number of ways.
sources such as whois data, yellow pages, or For mobile devices, it seems natural to choose a
regional web directories. The result of the data certain area around the current location of the
mining phase is a set of footprints for the pages user as a default footprint. In other cases, a
in the collection. In [30], footprints were footprint could be determined by analyzing a
represented as which were stored in a highly textual query for geographic terms, or by
compressed quad-tree structure, but this decision allowing the user to click on a map. This is an
is not really of concern to us here. Other
reasonably compact and efficient representations,
91 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No.1, 2010
interface issue that is completely orthogonal to fetches the remaining footprints from disk, in
our approach. order to score documents precisely.
IV. ALGORITHMS C. K-SWEEP ALGORITHM
A. TEXT-FIRST BASELINE: This algorithm first The main idea of the first improved algorithm is
filters results according to textual search terms to retrieve all required toe print data through a
and thereafter according to geography. Thus, it fixed number k of contiguous scans from disk. In
first accesses the inverted index, as in a standard particular, we build a grid-based spatial structure
search engine, retrieving a sorted list of the in memory that contains for each tile in a 1024
docIDs (and associated data) of documents that 1024 domain a list of m toe print ID intervals.
contain all query terms. Next, it retrieves all For example, for m = 2 a tile T might have two
footprints of these documents. Footprints are intervals [3476, 3500] and [23400, 31000] that
arranged on disk sorted by docID, and a indicate that all toe prints that intersect this tile
reasonable disk access policy is used to fetch have toe print IDs in the ranges [3476, 3500] and
them: footprints close to each other are fetched
[23400, 31000]. In the case of a 1024 1024 grid,
in a single access, while larger gaps between including about 50% empty tiles, the entire
footprints on disk are traversed via a forward auxiliary structure can be stored in a few MB.
seek. Note that in the context of a DAAT text This could be reduced as needed by compressing
query processor, the various steps in fact overlap. the data or choosing slightly larger tiles (without
The inverted index access results in a sorted changing the resolution of the actual footprint
stream of docIDs for documents that contain all data). Given a query, the system first fetches the
query terms, which is directly fed into the interval information for all tiles intersecting the
retrieval of document footprints, and precise query footprint, and then computes up to
scores are computed as soon as footprints arrive k ≥ m larger intervals called sweeps that cover
from disk. the union of the intervals of these tiles. Due to
B. GEO-FIRST BASELINE: This algorithm uses a the characteristics of space filling curves, each
spatial data structure to decrease the number of interval is usually fairly small and intervals of
footprints fetched from disk. In particular, neighboring tiles overlap each other
footprints are approximated by MBRs that substantially. As a result, the k generated sweeps
(together with their corresponding docIDs) are are much smaller than the total toe print data.
kept in a small (memory-resident) R∗-tree. As The system next fetches all needed toe print data
before, the actual footprints are stored on disk, from disk, by means of k highly efficient scans.
sorted by docID. The algorithm first accesses the The IDs of the encountered toe prints are then
translated into doc IDs and sorted. Using the
R∗-tree to obtain the docIDs of all documents
sorted list of docIDs, we then access the inverted
whose footprint is likely to intersect the query
index to filter out documents containing the
footprint. It sorts the docIDs, and then filters
textual query terms. Finally we evaluate the
them by using the inverted index. Finally, it
geographic score between the query footprint
92 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No.1, 2010
and the remaining documents and their geographic query processing can be performed at
footprints. The algorithm can be summarized as about the same level of efficiency as text-only
follows: queries. There are a number of open problems
K-SWEEP ALGORITHM: that we plan to address. Moderate improvements
(1) Retrieve the toe print ID intervals of all tiles in performance should be obtainable by further
intersecting the tuning of our implementation. Beyond these
Query footprint. optimizations, we plan to study pruning
(2) Perform up to k sweeps on disk, to fetch all techniques for geographic search engines that
toe prints in the union of intervals from Step (1). can produce top-k results without computing the
(3) Sort the doc IDs of the toe prints retrieved in precise scores of all documents in the result set.
Step (2) and access the inverted index to filter Such techniques could combine early termination
these doc IDs. approaches from search engines with the use of
(4) Compute the geo scores for the remaining approximate (lossy-compressed) footprint data.
doc IDs using the toe prints retrieved in Step (2). Finally, we plan to study parallel geographic
One limitation of this algorithm is that it fetches query processing on clusters of machines. In this
the complete data of all toe prints that intersect case, it may be preferable to assign documents to
the query footprint (plus other close by toe participating nodes not at random, as commonly
prints), without first filtering by query terms. done by standard search engines, but based on an
Note that this is necessary since our simple appropriate partitioning of the underlying
spatial data structure does not contain the actual Draw back Advantage of
Searching of
docIDs for toe prints intersecting the tile. Storing of old Proposed
Data
a list of docIDs in each tile would significantly system system
increase the size of the structure as most docIDs Accuracy of Very less Accuracy and
would appear in multiple tiles. Thus, we have to Local data local data more local data
first access the toe print data on disk to obtain 0.65
Processing time 0.34 seconds
candidate docIDs that can be filtered through the seconds
inverted index Splitting
Regional --------NIL-
CONCLUSIONS different type
specification ----
We discussed a general framework for ranking of Region
search results based on a combination of textual No link
and spatial criteria, and proposed several Good link
Spatial data between
algorithms for efficiently executing ranked between text
structure text and
queries on very large collections. We integrated and spatial data
spatial data
our algorithms into an existing high-performance
search engine query processor and evaluated REFERENCES
[1]. A. Markowetz, Y.-Y. Chen, T. Suel, X. Long, and B.
them on a large data set and realistic geographic
Seeger. Design and implementation of a geographic search
queries. Our results show that in many cases
93 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No.1, 2010
engine. In 8th Int. Workshop on the Web and Databases projects,23 M.Tech projects and 6 PhD
(WebDB), June 2005. research works.153 -
[2]. V. Anh, O. Kretser, and A. Moffat.Vector-space ranking
with effective early termination. In Proc. of the 24th Annual 2)
SIGIR Conf. on Research and Development in Information
Retrieval, pages 35–42, September 2001.
[3]. K. McCurley. Geospatial mapping
and navigation of the web. In Proc. of the 10th Int. World
Wide Web Conference, pages 221–229,May 2001.
[4]. S. Chakrabarti, M. van den Berg, and
B. Dom. Focused crawling: Anew approach to topic-specific Mr.S.Sivasubramanian received his
web resource discovery. In Proc. of the 8th Int. World Wide Diploma in Hardware Software
Web Conference, May 1999. installing in ECIL-BDPS, Govt of India,
[5]. Y. Zhou, X. Xie, C. Wang, Y. Gong, and W. Ma. Hybrid and Advanced Diploma in computer
index structures for location-based web search. In Proc. of Application in UCA, Madurai, and
the 14th Conf.on Information and Knowledge Management Bachelor of Science in Physics from
(CIKM), pages 155–162, November 2005. Madurai Kamaraj University in 1995,
[6]. Reference for query processing in web search engine Master of Science in Physics from
based on the Journal for Yen-Yu Chen Polytechnic Madurai Kamaraj University in 1997,
University Brooklyn, NY 11201, USA Torsten Suel Post Graduate Diploma in Computer and
Polytechnic University Brooklyn, NY 11201, USA June 2006 Application in Government of India
AUTHOR’S PROFILE: 2000. Master of Technology in
Computer Science and engineering from
1)
Bharath University Chennai 2007.
Pursing PhD in Bharath University
Chennai. He has more than 5 years of
teaching experience and guided 20
B.Tech projects, 11 M.Tech projects
Dr.M.Uma Maheswari received her
Bachelor of science in Computer
Science from Bharathidasan university
in 1995, Master of Computer
Applications in Computer Science from
Bharathidasan University in1998,M.Phil
in Computer Science from Alagappa
University, Karaikudi, Master of
Technology in Computer Science from
Mahatma Gandhi Kasi Vidyapeeth
university in 2005 and Ph.D in
Computer Science from Magadh
Universty, Bodh Gaya in 2007.She has
more than 10 years of teaching
experience and guided 150 M.C.A
94 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
Tunable Multifunction Filter Using
Current Conveyor
Manish Kumar M.C. Srivastava Umesh Kumar
Electronics and Communication Electronics and Communication Electrical Engineering Department
Engineering Department Engineering Department Indian Institute of Technology
Jaypee Institute of Information Jaypee Institute of Information Delhi, India
Technology Technology drumeshkumar98@rediffmail.com
Noida, India Noida, India
manishkumar.jiit@gmail.com mc.srivastava@jiit.ac.in
Abstract—The paper presents a current tunable multifunction single input and three multiple outputs using three OTAs and
filter using current conveyor. The proposed circuit can be current follower (CF) [6].
realized as on chip tunable low pass, high pass, band pass and
elliptical notch filter. The circuit employs two current conveyors, In the recent years there has been emphasis on
one OTA, four resistors and two grounded capacitors, ideal for implementation of the voltage mode/current mode active filters
integration. It has only one output terminal and the number of using second generation current conveyors (CCIIs) which
input terminals may be used. Further, there is no requirement provide simple realization with higher bandwidth, greater
for component matching in the circuit. The resonance frequency linearity and larger dynamic range. Kerwin-Huelsman-
(ω0) and bandwidth (ω0 /Q) enjoy orthogonal tuning. The cutoff Newcomb (KHN) biquad realization of low-pass, band-pass
frequency of the filter is tunable by changing the bias current, and high-pass filters with single input and three outputs,
which makes it on chip tunable filter. The circuit is realized by employing five current conveyor (CCII), two capacitors and six
using commercially available current conveyor AD844 and OTA resistors was proposed by Soliman in 1995 [7]. A universal
LM13700. A HSPICE simulation of circuit is also studied for the voltage-mode filter proposed by Higasimura et al. employs
verification of theoretical results. seven current conveyors, two capacitors and eight resistors [8].
Realization of high-pass, low-pass and band-pass filters using
Keywords- Active filter; Current Conveyor; Voltage- mode filter three positive current conveyor and five passive components
was reported by Ozoguz et. al.[9]. Chang and Lee [10] and
I. INTRODUCTION subsequently Toker et. al. [11] proposed realization of low-
Active filters with current/voltage controllable frequency pass, high-pass and band-pass filters employing current
have a wide range of applications in the signal processing and conveyors and passive components with specific requirements.
instrumentation area. Tsividis et. al. employed the realization Manish et. al. [12] proposed the realization of multifunction
of on chip MOSFET as voltage controlled resistor [1]. Their filter (low-pass, high-pass, band-pass and notch filters) with
contributions and several other research papers may be minimum current conveyors and passive components. The
considered to be motivation for the VLSI industry to make on central/cutoff frequency of these realizations could be changed
chip tunable filters [2],[3]. These realizations have small range by changing the passive components.
of variation in the frequency. The OTA-C structure is highly In 2001 Wang and Lee implemented insensitive current
suitable for realizing electronically tunable continuous time mode universal biquad MIMO realization using three balanced
filters. A number of voltage mode/current mode OTA-C biquad output current conveyors and two grounded capacitors [13]. In
have been reported in the literature. Multiple-input multiple- 2004 Tangsrirat and Surakampontorn proposed electronically
output (MIMO), multiple-input single-output (MISO) and tunable current mode filters employing five current controlled
single-input multiple-output (SIMO) type circuits have also current conveyors and two grounded capacitors [14]. A tunable
appeared in the literature. In 1996, Fidler and Sun proposed current mode multifunction filter was reported in 2008 using
realization of current mode filter with multiple inputs and two five universal current conveyors and eight passive components
outputs at different nodes using four dual output OTA’s and [15]. Recently Chen and Chu realized universal electronically
two grounded capacitors [4]. Later, Chang proposed controlled current mode filter using three multi-output current
multifunctional biquadratic filters, using three operational controlled conveyors and two capacitors, however the
transconductance amplifiers and two grounded capacitors [5]. frequency and quality factor of their realizations are not
In 2003, Tsukutani et. al proposed current mode biquad with independent [16].
95 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
The proposed realization in this paper employs two current The current Iabc and parameter gm may be expressed as
conveyors, one OTA, five resistors and two grounded follows
capacitors with one output terminal and three input terminals.
All the basic low-pass, high-pass, band-pass and notch filters I abc
may be realized by the proposed circuit by selecting proper
gm =
2VT
input terminals. The frequency of the filter can be changed by
changing the control voltage of the OTA.
The following section presents circuit description of the
current conveyor. The sensitivity analysis, simulation results Where VT = KT is the thermal voltage. The routine
q
and conclusion are discussed in the subsequent sections.
analysis yields the following transfer function:
II. CIRCUIT DESCRIPTION (2)
1 ⎛ s C 2 C5 R1 R3 R4 R6 g mV2 + R3V1 + ⎞
2
Vout = ⎜ ⎟
The first and second generation current conveyors were D( s ) ⎜ sC5 g m R1 R4 R6V3
⎝
⎟
⎠
introduced by Sedra and Smith in 1968, 1970 respectively;
these are symbolically shown in fig 1 and are characterized by
the port relations given by “(1)” Where
D( s ) = s 2 C 2 C 5 g m R1 R3 R4 R6 + sC 5 g m R1 R4 R6 + R3 (3)
Thus by using “(2)” we can realize low-pass, band-pass,
high-pass and notch filter responses at the single output
terminal by applying proper inputs at different nodes as shown
in table1.
TABLE I. VARIOUS FILTER RESPONSES
Filter\Input V1 V2 V3
Low-pass 1 0 0
Figure 1. Symbol of Current Conveyor II
High-pass 0 1 0
⎡V x ⎤ ⎡ 0 B 0⎤ ⎡ I x ⎤ (1) Band-pass 0 0 1
⎢I ⎥ = ⎢ 0 0 0⎥ ⎢V y ⎥
Notch 1 1 0
⎢ y⎥ ⎢ ⎥⎢ ⎥
⎢ I z ⎥ ⎢± K
⎣ ⎦ ⎣ 0 0⎥ ⎢V z ⎥
⎦⎣ ⎦ The denominators for the all filter responses are same. The
The values of B and K are frequency dependent and ideally filtering parameter cutoff frequency (ωo), bandwidth (ωo/Q) and
B=1 and K=1. The ±K indicates the nature of current conveyor. quality factor (Q) are given by
+ve sign indicates positive type current conveyor while –ve (4)
1
sign indicates negative. ω0 =
The proposed circuit shown in fig 2 employs only two R1 R 4 R 6 C 2 C 5 g m
current conveyors, five resistors and two capacitors. The
grounded capacitors are particularly very attractive for the ω0 1 (5)
integrated circuit implementation. =
Q R3 C 2
C2 (6)
Q = R3
g m R1 R 4 R 6 C 5
It can be seen from a perusal of “(4)” - “(6)” that the center
frequency and bandwidth ω0/Q can be controlled independently
through R6 and /or C5 and R3. The transconductance of the OTA
is independently tuned by varying the bias current of the OTA.
Figure 2. Proposed Voltage Mode Multifunction Filter
96 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
III. SENSITIVITY ANALYSIS IV. SIMULATION RESULT
The complete circuit is simulated using commercially
The sensitivity analysis of the proposed circuit is presented available AD844 and LM13700. The AD844 is used for the
in terms of the sensitivity of ω0 and Q with respect to the realization of CCII+ and CCII-. Figure 3 displays the
variation in the passive components as follows: simulation result for the proposed filter. The circuit is designed
for ω0 = 8.7 KHz and Q=0.12 by considering R1 = R4 = R6 =
ω 1 (7) 10KΩ, C2 = C5 = 10nF, R3 = 14KΩ and gm=13.2mS. The
SC20,C5 , R1 , R4 , R6 , g m = −
2 theoretical results have been are verified to match with
(8) simulation result. Figure 3 shows that the quality factor of the
S Q
R3 =1 notch filter is very high. It is due to the transfer function of the
notch filter is having complex conjugate zeros with zero real
1 (9) values. Figure 4 shows the cutoff/center frequency of the filter
S g m , R1 , R4 , R6 ,C5 = −
Q
2 with respect to the changes in bias current of the OTA. The
(10) response is showing the when the bias current is higher than the
1
S C2 =
Q output current of the OTA than the frequency variation is linear
2 and circuit will be stable.
As per these expressions, both the ω0 and Q sensitivities are V. CONCLUSION
less than ± ½ with a maximum value of S
Q
R3 =1. The circuit proposed in this paper generates low-pass, high-
pass, band-pass and notch filter using two current conveyors,
four resistors and two capacitors. The circuit provides more
number of filter realizations at the single output terminal. In
addition of this proposed circuit does not have any matching
constraint/cancellation condition. The circuit employs’
grounded capacitor, suitable for IC fabrication. The circuit
enjoys the othogonality between the cutoff frequency and the
bandwidth. The OTA is linearly tunable when the bias current
is higher than the output current. It has low sensitivities figure
of both active and passive components.
REFERENCES
[1] Y. Tsividis, M. Banu and J. Khoury,” Continious –time MOSFET-C
filters in VLSI,” IEEE journal of solid-state circuits, vol. sc-21, no.1, pp.
15-30, 1986.
[2] M. Ismail, S. V. Smith and R. G. Beale, “.A new MOSFET-C universal
filter structure for VLSI,” IEEE journal of solid-state circuits, vol. sc-23,
no.1, pp. 182-194, 1988.
Figure3: Multifunction Filter Response
[3] Jaap van der Plas“MOSFET-C Filter with Low Excess Noise and
Accurate Automatic Tuning”,IEEE Journal of Solid State Circuits, vol.
26, no. 7, pp.922-929, 1991.
[4] Yichuang Sun and J. K. Fidler ,”Structure Generation of Current-Mode
Two Integrator Loop Dual Output-OTA Grounded Capacitor Filters,”
IEEE trans. on cas-II: analog and digital signal processing, vol.43, no.9,
pp. 659-663, 1996.
[5] Chun-Ming Chang,” New Multifunction OTA-C Biquads,” IEEE trans.
on cas-II: analog and digital signal processing, vol..46, no.6, pp. 820-824,
1999.
[6] T. Tsukutani, Y. Sumi, M. Higashimura and Y. Fukui,” Current-mode
biquad using OTAs and CF,” Electronics letters, vol. 39, no-3, pp 262-
263, 2003.
[7] A. M. Soliman, “Kerwin–Huelsman–Newcomb circuit using current
conveyors,” Electron. Lett., vol. 30, no. 24, pp. 2019–2020, Nov. 1994.
[8] M. Higasimura and Y. Fukui, “Universal filter using plus-type CCII’s,”
Electron. Lett. vol. 32, no. 9, pp. 810-811, Apr. 1996.
Figure4: Frequency Vs Control Bias Current
[9] S. Ozoguz, A. Toker and O. Cicekoglu, “High output impedance current-
mode multifunction filter with minimum number of active and reduced
number of passive elements,” Electronics Letters, vol. 34, no 19, pp.
1807-1809, 1998
97 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, 2010
[10] Chun-Ming Chang and Ming- Jye Lee, “Voltage mode multifunction
filter with single input and three outputs using two compound current
conveyors,” IEEE Trans. On Circuits and Systems-I: vol. 46, no. 11, AUTHORS PROFILE
pp.1364-1365, 1999.
[11] A. Toker, O. Çiçekoglu, S. Özcan and H. Kuntman ,” High-output- Manish Kumar was born in India in 1977. He received
impedance transadmittance type continuous-time multifunction filter with his B.E. in electronics engineering from S.R.T.M.U.
minimum active elements,” International Journal of Electronics, Volume Nanded in 1999 and M.E. degree from Indian Institute
88, Number 10, pp. 1085-1091, 1 October 2001. of Science, Bangalore in 2003. . He is perusing Ph.D.
He is working as faculty in Electronics and
[12] Manish Kumar, M.C. Srivastava and Umesh Kumar,” Current conveyor
Communication Engineering Department of Jaypee
based multifunction filter,” International Journal of Computer Science
and Information Security, vol. 7 no. 2, pp. 104-107, 2009. Institute of Information Technology, Noida He is the
author of 10 papers published in scientific journals and
[13] Hung-Yu Wang and Ching-Ting Lee, “Versatile insensitive current-mode conference proceedings. His current area of research
universal biquad implementation using current conveyors,” IEEE Trans. interests includes analogue circuits, active filters and
On Circuits and Systems-I: vol. 48, no. 4, pp.409-413, 2001. fuzzy logic.
[14] Worapong Tangsrirat and Wanlop Surakampontorn,” Electronically
tunable current-mode universal filter employing only plus-type current- M. C. Srivastava received his B.E. degree from
controlled conveyors and grounded capacitors, “Circuits systems signal Roorkee University (now IIT Roorkee), M.Tech.
processing, circuits systems signal processing, vol. 25, no. 6, pp. 701– from Indian Institute of Technology, Mumbai and
713, 2006. Ph.D from Indian Institute of Technology, Delhi in
1974. He was associated with I.T. BHU, Birla
[15] Norbert Herencsar and Kamil Vrba,”Tunable current-mode multifunction Institute of Technology and Science Pilani, Birla
filter using universal current conveyors,” IEEE third international Institute of Technology Ranchi, and ECE Dept. JIIT
conference on systems, 2008. Sector-62 Noida. He has published about 60 research
[16] H.P. Chen and P.L.Chu,”Versatile universal electronically tunable papers. His area of research is signal processing and
current-mode filter using CCCIIs,” IEICE Electronics Express, vol. 6, no. communications. He was awarded with Maghnad
2 pp. 122-128, 2009. Saha Award for his research paper.
[17] A. M. Soliman, “Current mode universal filters using current conveyors: Umesh Kumar is a senior member, IEEE. He
classification and review,” Circuits syst Signal Process, vol. 27, pp. 405- received B.Tech and Ph.D degree from IIT Delhi. He
427, 2008. has published about 100 research papers in various
journals and conferences. He is working as faculty in
[18] P. V. Anada Mohan, Current Mode VLSI Analog Filters, Birkhauser,
Electrical Engineering Department, IIT Delhi.
Boston, 2003.
98 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Artificial Neural Network based Diagnostic
Model for Causes of Success and Failures
Bikrampal Kaur, Dr.Himanshu Aggrawal,
Deptt.of Computer Science & Engineering, Deptt.of Computer Engineering,
Chandigarh Engineering College, Punjabi Unniversity,
Mohali,India. Patiala,India.
dhaliwal_bikram@yahoo.com himanshu@pbi.ac.in
Abstract— Resource management has always been an area of management of the human resources. In this paper an
prime concern for the organizations. Out of all the resources attempt have been made to identify and suggest HR factors
human resource has been most difficult to plan, utilize and and propose a model to determine the influence of HR
manage. Therefore, in the recent past there has been a lot of factors leading to failure. It is particularly important as the
research thrust on the managing the human resource. Studies
neural networks have proved their potential in several fields
have revealed that even best of the Information Systems do fail
due to neglect of the human resource. In this paper an attempt such as Industry, transport, dairy sectors etc.. India has
has been made to identify most important human resource distinguished IT strength in global scenario and using
factors and propose a diagnostic model based on the back- technologies like neural networks is extremely important
propagation and connectionist model approaches of artificial due to their decision making capabilities like human brain.
neural network (ANN). The focus of the study is on the mobile
-communication industry of India. The ANN based approach is In this paper a Neuro-Computing approach has been
particularly important because conventional approaches (such proposed with some metrics collected through pre
as algorithmic) to the problem solving have their inherent acquisition step from the communication industry. In this
disadvantages. The algorithmic approach is well-suited to the
study, a coding of backpropagation algorithium have been
problems that are well-understood and known solution(s). On
the other hand the ANNs have learning by example and used to predict success or failure of company and also a
processing capabilities similar to that of a human brain. ANN comparison is made with the connectionist model for
has been followed due to its inherent advantage over predicting the results. The back-propagation learning
conversion algorithmic like approaches and having algorithm based on gradient descent method with adaptive
capabilities, training and human like intuitive decision making learning mechanism.. The configuration of the connectionist
capabilities. Therefore, this ANN based approach is likely to approach has also been designed empirically. To this effect,
help researchers and organizations to reach a better solution to several architectural parameters such as data pre-processing,
the problem of managing the human resource. The study is data partitioning scheme, number of hidden layers, number
particularly important as many studies have been carried in
of neurons in each hidden layer, transfer functions, learning
developed countries but there is a shortage of such studies in
developing nations like India. Here, a model has been derived rate, epochs and error goal have been empirically explored
using connectionist-ANN approach and improved and verified to reach an optimum connectionist network.
via back-propagation algorithm. This suggested ANN based
model can be used for testing the success and failure human II. REVIEW OF LITERATURE
factors in any of the communication Industry. Results have
been obtained on the basis of connectionist model, which has The review of IS literature suggests that for the past 15
been further refined by BPNN to an accuracy of 99.99%. Any years, the success and the failure HR factors in information
company to predict failure due to HR factors can directly systems have been major concern for the academics,
deploy this model.
practitioners, business consultants and research
organizations.
Keywords— Neural Networks, Human resource factors, Company
success and failure factors. A number of researchers and organizations throughout
the world have been studying that why information systems
I. INTRODUCTION do fail, some important IS failure factors identified by [6,7]
Achieving the information system success is a major are:
issue for the business organizations. Prediction of a • Critical Fear-based culture.
company’s success or failure is largely dependent on the • Technical fix sought.
management of human resource (HR). Appropriate • Poor reporting structures
utilization of human resource may lead to the success of the • Poor consultation.
company and their underutilization may lead to its failure.
• Over commitment.
• Changing requirements.
In most of the organizations management makes use of • Political pressures.
conventional Information System (IS) for predicting the
• Weak procurement.
.
99 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
• Technology focused. III. OBJECTIVES OF STUDY
• Development sites split.
• Leading edge system (i) To design a HR model of factors affecting
• Project timetable slippage success and failure in Indian Organisations of
• Complexity underestimated Telecom sectors.
• Inadequate testing. (ii) To propose a diagnostic ANN based model of
the prevailing HR success/failure factors in
• Poor training
these organizations.
Six major dimensions of IS viz. superior quality (the
measure of IT itself), information quality (the measure of
A model depicting important human resources factors
information quality), information use (recipient
has been designed on the basis of literature survey and
consumption of IS output), user satisfaction (recipient
researchers experiences in the industry under this study
response to use of IS output), individual impact (the impact
has been in figure1.
of information on the behavior of the recipient) and
organizational impact (the impact of information on
organizational performance) had already been proposed [8]
All these dimensions directly or indirectly are related to HR
of IS.
Cancellation of IS projects [11] are usually due to a
combination of:
• Poorly stated project goals
• Poor project team composition
• Lack of project management and control
• Little technical know-how
• Poor technology base or infrastructure
• Lack of senior management involvement
• Escalating project cost and time of completion
Some of the other elements of failure [12] identified were:
• Approaches to the conception of systems;
• IS development issues (e.g. user involvement)
• Systems planning
• Organizational roles of IS professionals
• Organizational politics
• Organizational culture
• Skill resources
• Development practices (e.g. participation)
• Management of change through IT
• Project management
• Monetary impact of failure
• “Soft” and Hard” perceptions of technology
Fig.-1 Exhaustive View of HR Model
• Systems accountability
• Project risk
• Prior experience with IT
IV. EXPERIMENTAL SCHEME
• Prior experience with developing methods
• Faith in technology A. Using ANN
• Skills, attitude to risk
Neural networks differ from conventional approach of
All the studies predict that during the past two decades, problem solving in a way similar to the human brain. The
investment in Information technology and Information network is composed of a large number of highly
system have increased significantly in the organization. But interconnected processing elements (neurons) working in
the rate of failure remains quite high. Therefore an attempt parallel to solve a specific problem. Neural networks learn
is made to prepare the HR model for the prediction of the by example. Differences in ANN and conventional systems
success or failure of the organization. have been given below in TABLE I.
.
100 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
a) Universe of study : All managers working at the
three levels of the selected organizations.
TABLE-1 b) Sample Selection: A number of respondents based
COMPARISON OF ANN AND CONVENTIONAL SYSTEM on proportional stratified sampling from all of
these organizations will be selected. The
S.No ANN Conventional Systems respondents will be identified from various levels
1 Learn by examples Solve problems by in each organization. The sample size from a
algorithmic approach
2 Unpredictable Highly predictable & well
stratum was determined on the basis of the
defined following criterion:
3 Better decision making due to No decision making 50% of the population where sample size > 5
human like intelligence 100% of the population where sample size < 5.
4 Trial and error method of No learning method
learning
5 Combination of IT & Human Only deal with IT B. Data collection tools
Brain
6 Cannot be programmed Can be programmed Primary data has been collected through a
questionnaire-cum-interview method from the selected
respondents (Appendix C). The questionnaire was designed
Henceforth from TABLE I it can be seen that ANN are based on the literature survey, and detailed discussion with
better suited for the problem that are not so well defined and many academicians, professionals and industry experts. The
predictable. Further ANN’s advantage is due to its detailed sampling plan of both the organizations has been
clustering unlike other conventional systems .Hence ANN is shown in Table II.
betted suited for the problems that are not so well defined
and predictable.
Applying ANN to HR factors graphically has been shown TABLE II
DETAILS OF THE SAMPLING PLAN
in fig 2. PUNCOM, MOHALI AND RELIANCE, CHANDIGARH
Lev Designation Universe Sample % age Total
el Sampl
e
I Executive
Director 2 2 100
(MD)
General 17
Manager 7 7 100
Fig. 2 Levels of HR of IS with ANN
Deputy
General 6 8 50
Manager
V RESEARCH METHODOLOGY
II Assistant 10 7 70
A. Sampling scheme General
Manager
17
The research involves the collection of data from the Senior 10 5 50
managers working at various levels within the selected Manager
10 5 50
enterprises. The total number of respondents in these Manager
enterprises, the sample size selection and application of the
Deputy 30 15 50
neural network approaches has been followed. The study III
Manager
comprises of survey of employees of two
Senior 90 45 50 135
telecommunication companies. With this aim, two
Officer
prestigious companies (first one is Reliance 150 75 50
Communications, Chandigarh and the other one is Puncom, Officer
Mohali) have been considered in this study. The details of
the research methodology adopted in this research are given
below.
C. Processing of data
1) For the Organisation: The responses of the 169 managers of the selected
a) Universe of study: Telecommunication industry organizations under study were recorded on five-point likert
comprises of Reliance InfoCom Vodafone, Essar, scale with scores ranging from 1 to 5. The mean scores of
Idea, and Bharti-Airtel. the managers of the two organizations and those over the
b) Sample Selection: Reliance InfoCom, Chandigarh whole industry considering all the managers included in the
and Punjab Communication Ltd(Puncom) Mohali. study.
2) For the Respondents
.
101 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
The valid responses were entered in Microsoft Excel data(Appendix B). The N/W used was
software. Thus, this data formed has been the basis for the backpropagtion with training function,traingda
corresponding files on the ANN software. The 70% and adaptation learing function, learngdm. The
responses of total inputs scores along with their known mean square error MSE was found to be
target from MS-Excel sheet were fed for training the neural 0.096841. The accuracy of connectionist model
network. The remaining scores of 30% responses were fed for the prediction of success and/or failure of
during the testing. Then the error_min of testing found to be company results out to be 99.90%
less than the error_min of training data. The accuracy of
99.90% is shown in Table III.
Before analysis it is important to define:
VI. EXPERIMENT RESULT AND DISCUSSION HL: Hidden Layer (e.g. HL1: first Hidden Layer; HL2:
second hidden Layer)
A. Dataset Epoch: During iterative training of Neural Network, an
The investigations have been carried out on the data epoch is a single pass through the entire training
obtained from telecommunication sector industry. This set, Followed by the testing of the verification
industry comprises of Reliance Communication, Vodafone, set.
Essay, Idea, and Bharti-Airtel. But the data has been
MSE: Mean Square Error Learning Rate Coefficient η -It
undertaken at Reliance Communication, Chandigarh and
determines the size of the weight adjustments
Punjab Communication Ltd. (Puncom) Mohali. The specific
made at each iteration which influence the rate
choice has been made because:
of convergence.
• The Telecom sector is very dynamic and fast The description of the Simulation Results of the Table III
growing. India is the second largest country of the has been explained as
world in mobile usage.
• The first industry is the early adopters of IT and Col-1 It includes the configuration of the network having
has by now, gained a lot of growth and experience hidden layers 1(HL1) with 1 neuron and training
in IS development and whereas the other one lag function tansigmoidal, which remain same from
behind and leads to its failure. 35-1000 epochs. Then 2/logsig tried for HL1 in the
network, it has 2 neurons and HL2 i.e. hidden layer
2 having training faction tansigmodal tried for 35-
One industry is considered for the study because of the 1000 epoch. In this way the no. of neurons, training
fact that the working constraints of various organizations functions and hidden layers have been changed
under one industry are similar and hence adds to the during trial and error method.
reliability of the study finding. The input and output Col-2 No. of epoch (defined earlier) varies from 35-1000
variables, considered in the study, include strategic per cycle
parameter(x1), tactical parameter (x2), operational Col-3 Error goal is predefined for its tolerance.
parameter (x3), employee outcome (y). The dataset Col-4 Learning Coefficient
comprises of 52 patterns has been considered for the Col-5 Mean Square Error for which the network is trained
training purpose of ANN and the remaining 23 patterns for
testing the network.
Supervised feed-forward back propagation
B. Connectionist model connectionist models based on viz., gradient descent
The well versed ‘trial and error’ approach has been used algorithm has been used. The network was investigated
throughout in this study. The Neural Network Toolbox empirically with a single hidden layer containing different
under MATLAB R2007a is used for all training as well as numbers of hidden neurons and gradually more layers has
simulation experiments. added as depicted in the Table III. Several combinations of
different parameters such as the data partitioning strategy;
the number of epochs; the performance goal; transfer
1) Collected scores for both input data and known functions in hidden layers are explored on trial and error
target were entered in MS-Excel as an input. basis so as to reach the optimum combination.
2) The input data of 70% of total data were
imported to MAT lab’s workspace for training
the ANN as depicted in Appendix A. The performance of the models developed in this study
3) The known target data has been also imported to is evaluated in terms of mean square error (MSE) for the
Mat lab’s workspace. connectionist model using the neural tool kit. The mean
4) Then both the input data & the target data were square error indicates the accuracy for the prediction of
entered in the ANN toolbox and network is success and/or failure of the organization comes out to be
created using back propagation neural network. 99.90% through this model. The experimental results of
5) The training has been done using 70% of the simulation of data of success or Failure Company through
input data and then testing (simulation) has been this model are summarized in Table III.
done on the rest of the 30% of the available
.
102 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
TABLE-III epoc 1405 and its weights has been saved for feeding to
SIMULATION RESULTS (CONNECTIONIST MODEL)
testing algorithm.
USING ANN TOOLKIT
The algorithm has been tested with 30% of data
Network Epoc Error Learnin MSE selected randomly from the given data which results in
Configuration hs Goal g error=0.009174, no. of epochs=13 by doing programming of
rate BPNN algorithium using Mat lab as shown in Table IV.
HL1 HL2 The results from the programming code have been
shown through Matlab.
35 0.738111 TABLE-IV
1/ tansig - 40 0.01 0.01 0.617595 BPNN CODE TESTING RESULTS
` MSE
1/ tansig - 45 0.01 0.01 0.634038
↓
1/ tansig - 50 0.01 0.01 1.33051 error=1.664782 no.of epoches=1
error=1.496816 no.of epoches=2
1/ tansig - 80 0.01 0.01 0.580224 error=1.093136 no.of epoches=3
1/ tansig - 400 0.01 0.01 0.466721 error=0.547380 no.of epoches=4
error=0.476718 no.of epoches=5
1/ tansig - 1000 0.01 0.01 0.421348 error=0.429089 no.of epoches=6
error=0.370989 no.of epoches=7
2/ logsig 1/ tansig 35 0.01 0.01 1.22608
error=0.303513 no.of epoches=8
2/ logsig 1/ tansig 100 0.01 0.01 0.735229 error=0.225575 no.of epoches=9
error=0.142591 no.of epoches=10
2/ logsig 1/ tansig 200 0.01 0.01 0.402351 error=0.071286 no.of epoches=11
error=0.027775 no.of epoches=12
2/ logsig 1/ tansig 500 0.01 0.01 0.282909
error=0.009174 no.of epoches=13
2/ logsig 1/ tansig 1000 0.01 0.01 0.138904
3/ logsig 1/ tansig 35 0.01 0.01 0.81653
During testing the BPNN coding, error_minima has
3/ logsig 1/ tansig 1000 0.01 0.01 0.143183 found to be less than error_minima of training, which
validates the algorithm .It, has been further added that this
4/ logsig 1/ tansig 1000 0.01 0.01 0.096841
accuracy of BPNN algorithium is found to be 99.99%
whereas it was 99.90 in the connectionist model. Therefore
this result is better than result obtained through hit and trail
HL1: First hidden layer method (connectionist model) using neural network toolkit
HL2: Second hidden layer and hence BPNN algorithm’s coding has fast performance
and better results i.e.better prediction on low number of
The table-III shows when first hidden layer has 4 epochs at the time of testing could be achieved. During
neurons and second hidden layer has 1 neuron with 1000 testing error_minima is less than error_minima of training
epochs, error goal 0.01, learning rate 0.01 mean square root for remaining 30% data, which validates algorithm. It
is 0.096841, therefore accuracy of connectionist model for comes out to be error=0.009174, at no. of
the prediction of failure company becomes 99.90%. epochs=13.Therefore the accuracy of the coding of the
For further improvement Back propagation Approach BPNN algorithium for the failure model comes out to be
has been deployed to reach better results. BPNN Code was 99.99%.
written that generates error value for 1 to 2000 epochs and
has shown the change in mean square error value. VII CONCLUSION
HR factors have strong influence over company success
C. Back Propagation Algorithm
and failure. Earlier HR factors were measured through
For each input pattern do the following steps.
variance estimation and statistical software’s, Due to the
Step 1. Set parameters eata η (…………),
inherent advantages of the Artificial Neural networks ,they
emax(maximum error value) and e(error between
are being used to replace the existing statistical models.
output and desired output).
Here, the ANN based model has been proposed that can be
Step2. Generate weights for hidden to output and input
used for testing the success and/or failure of human factors
to hidden layers.
in any of the communication Industry. Results have been
Step 3. Compute Input to Output nodes
obtained on the basis of connectionist model, which has
Step4. Compute error between output and desired
been further refined by BPNN to an accuracy of 99.99%.
output
Any company on the basis of this model can diagnose
Step5. Modify weights from hidden to output and input
failure due to HR factors by directly deploying this model.
to hidden nodes.
The limitation of the study is that it only suggests a
diagnostic model of success/failure of HR factors but it
Result: This gives us error value of error=0.014289, no.of
does not pin point them.
epochs=1405 during training It showed error_minima at
.
103 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
REFERENCES [2] Dr. Himanshu Aggarwal, is
[1] OASIG Report:The Organizational Aspects of Information Associate Professor (Reader)
Technology(1996),The report entitled: “The Performance of in Computer Engineering at
Information Technology and Role of Organizational University College of
Factors”, Engineering, Punjabi
www.shef.ac.uk/~iwp/publications/reports/itperf.html. University, Patiala. He had
[2] Millan Aikem, University of Mississippi, USA-1999, “Using completed his Bachelor’s
a Neural Network to forecast inflation, Industrial degree in Computer Science
Management & Data Systems 99/7, 1999, 296-301”. from Punjabi University
Patiala in 1993. He did his
M.E. in Computer Science in
[3] G. Bellandi, R. Dulmin and V. Mininno,“Failure rate neural
1999 from Thapar Institute of
analysis in the transport sector,” University of Pisa, Italy,
Engineering & Technology,
International Journal of Operations & Production
Patiala. He had completed his
Management, Vol. 18 No. 8, 1998, pp. 778-793,© MCB
Ph.D. in Computer
University Press, 0144-3577,New York, NY. Engineering from Punjabi
[4] Sharma, A. K., Sharma, R. K., Kasana, H. S., 2006., University Patiala in 2007.He has more than 16 years of teaching
“Empirical comparisons of feed-forward connectionist and experience. He is an active researcher who has supervised 15 M.Tech.
conventional regression approaches for prediction of first Dissertations and guiding Ph.D. to seven scholars and has contributed
lactation 305-day milk yield in Karan Fries dairy cows”. more than 40 articles in International and National Conferences and
Neural Computing and Applications 15(3–4), 359–365. 22 papers in research Journals. Dr. Aggarwal is member of many
[5] Sharma, A. K., Sharma, R. K., Kasana, H. S .2007., professional societies such as ICGST, IAENG. His areas of interest
“Prediction of first lactation 305-day milk yield in Karan are Information Systems, ERP and Parallel Computing. He is on the
Fries dairy cattle using ANN approaching”, Applied Soft review and editorial boards of several refereed Research Journals.
Computing 7(3), 1112–1120.
[6] J Jay Liebowitz , “A look at why information systems fail
Department of Information Systems,” Kybernetes, Vol. 28 APPENDIX A
No. 1, 1999,pp. 61-67, © MCB University Press,0368-492X, Table for training data is as following(70% data used for TRAINING)
University of Maryland-BaltimoreCounty, Rockville,
Maryland, USA . Strategic Tactical Operational
[7] Flowers, S. (1997), “Information systems failure: identifying Emp1 1 2 1
the critical failure factors,” Failure and Lessons Learned in Emp2 2 3 1.9
Information Technology Management: An International Emp3 4 1.5 1.5
Journal,Cognizant Communication Corp., Elmsford, New Emp4 2 3 4
York, NY, Vol. 1 No. 1, pp. 19-30. Emp5 1.7 1.6 2.5
[8] DeLone, W.H., and McLean, E.R. 2004. "Measuring E- Emp6 1 1 1
Commerce Success: Applying the DeLone & McLean
Emp7 1.2 1.3 1.4
Information Systems Success Model," International Journal
Emp8 1.7 1.8 3
of Electronic Commerce (9:1), Fall, pp 31-47.
Emp9 1.8 2 4
[9] Bruce Curry and Luiz Moutinho, “Neural networks in
marketing: Approaching consumer responses to advertising Emp10 4 1.8 2
stimuli”, European Journal of Marketing, Vol 27 No 7, 1993 Emp11 2 5 1
pp 5- 20. Emp12 2.5 2.2 2
[10] Demuth, H. B., Beale, M., 2004. User’s Guide for Neural Emp13 2.5 2 1.6
Network Toolbox( version 4) for use with MATLAB 6.1. Emp14 1.6 2 2.5
The MathWorks Inc., Natick, MA. Emp15 1 -1 1
[11] Kweku Ewusi Mensah, “Critical issues in the abandoned Emp16 -1 -1 1
information system development projects”, Loyola Emp17 -1 1 -1
Marymount University,Los Angeles, CA, Volume 40,Issue Emp18 1 1 -1
9(September 1997)pages 74-80,1997,ISSN :0001-7082. Emp19 1.2 -1 1
[12] Angeliki Poulymenakou1 and Vasilis Emp20 -1 1.2 1.5
Serafeimidis2,Volume1, number 3, 1997, “Failure & Emp21 3.6 1.2 4
Lessons Learned in Information Technology Management”, Emp22 3.6 3.6 3.6
Vol. 1, pp. 167-177. Emp23 4 4 5
Emp24 5 4 2
Emp25 5 5 1
Emp26 4 5 -1
AUTHORS PROFILE Emp27 3 2 -1
Emp28 0.5 1.5 0.5
[1] Bikram Pal Kaur is an Assistant Emp29 2.1 3.1 4.1
Professor in the Deptt. of Computer Emp30 5 5 5
Science & Information Technology and
Emp31 0.1 0.2 0.5
is also heading the Deptt. Of Computer
Emp32 0.5 0.7 1.5
Application in Chandigarh Engineering
College,Landran,Mohali. She holds the Emp33 4.1 4.2 4.3
degrees of B.tech.,M.Tech,M.Phil.. and is Emp34 5 0.1 0.2
currently pursuing her Ph.D.in the field of Emp35 0.1 2 0.1
Infornation Systems from Punjabi Emp36 1.4 5 -1
University,Patiala. She has more than 11 Emp37 1.5 4 1
years of teaching experience and served Emp38 1.6 3 2
many academic institutions. She is an Emp39 2.1 2 3
Active Researcher who has supervised Emp40 2.1 1 4
many B.Tech.Projects and MCA Dissertations and also contributed Emp41 2.3 5 5
12 research papers in various national & international conferences. Emp42 2.5 4 -1
Her areas of interest are Information System, ERP. Emp43 3.3 3 1
Emp44 3.5 2 2
.
104 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Emp45 4 1 3 The balance between cost and benefit of computer based
Emp46 4.9 5 4 information product/services
Emp47 4.1 4 5 User training
Emp48 4.3 3 -1 The flexibility in system for corrective action in case of
Emp49 3.01 2 -1 problematic output
Emp50 2.01 -1 1 Testing of system before implementation
Emp51 2.03 5 1 Operational factors
Emp52 5 4 1 Professional standard maintenance (H/W, S/W, O.S, User
Accounts, Maintenance of system)
The response of staff to the changes in existing system
APPENDIX B Trust of staff in the change for the betterment of the system
Table shows data used for testing neural network (30% data The way users input data and receive output
used for TESTING) The accuracy (Correctness) of the output
The completeness (Comprehensiveness) of the information
Strategic Tactical Operational The well defined language for interaction with computers
The volume of output generated by the system for a user
Empt1 1.3 1.2 1.1 Faith in technology/system by the user
Empt2 1.5 1.5 1.5
Empt3 1.7 1.5 1.6
Empt4 2 1 0
Empt5 3 2 2
Empt6 1.6 1.6 1.6
Empt7 4 1 2
Empt8 1 4 1.6
Empt9 2 4 4
Empt10 3.3 3.1 3.4
Empt11 2.5 3.5 2
Empt12 4.1 3.5 2.1
Empt13 1 1 1
Emptl4 1.3 1.1 1.9
Emptl5 1.8 2.3 2.1
Emptl6 0 0 0
Emptl7 0 1 0
Emptl8 0 0 1
Emptl9 3.5 4.5 5
Emptl20 1.8 1.6 2.9
Emptl21 1 0 0
Emptl22 2.5 1 1
Emptl23 3.5 1.6 1.7
APPENDIX -C
Questionnaire used for survey (containing scores from 1-5)
1-not important,2-slightly important,3-moderately
important,4-fairly important,5-most important
Factors Score
Strategic Factors
Support of the Top management
Working relationship in a team(Users & Staff)
Leadership
project goals clearly defined to the team
Thorough Understanding of business environment
User involvement in development issues
Attitude towards risk (Changes in the job profile due to the
introduction of the computers)
Adequacy of computer facility to meet functional
requirements(quality and quantity both)
Company technology focused
Over commitment in the projects
Tactical Factors
Communication
Organizational politics
Priority of the organizational units to allocate resources to
projects
Organizational culture
Skilled resources (Ease in the use of system by users)
The consistency and reliability of information
To obtain highest returns on investment through system usage
Realization of user requirements
Security of data and models from illegal users
Documentation ( formal instructions for the usage of IS)
.
105 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Detecting Security threats in the Router using
Computational Intelligence
J.Visumathi Dr. K. L. Shunmuganathan
Research Scholar Professor & Head, Department of CSE
Sathyabama University, R.M.K. Engineering College
Chennai-600 119 Chennai-601 206
jsvisu@gmail.com Kls_nathan@yahoo.com
Abstract
system storage, operating system data structures, protocol
data structures and software vulnerabilities. DoS can be a
Information security is an issue of global concern. As the
Internet is delivering great convenience and benefits to the single source attack, originating at a single host, or can be a
modern society, the rapidly increasing connectivity and multi-source source attack, where multiple hosts and
accessibility to the Internet is also posing a serious threat to networks are involved. The DoS attacks can take an
security and privacy, to individuals, organizations, and nations advantage form the distributed nature of the Internet by
alike. Finding effective ways to detect, prevent, and respond to launching a multiplicative effect, resulting in distributed DoS.
intrusions and hacker attacks of networked computers and Due to the use of dynamic protocols and address spoofing,
information systems. This paper presents a knowledge discovery detecting distributed and automated attacks still remains a
frame work to detect DoS attacks at the boundary controllers challenge.
(routers). The idea is to use machine learning approach to
discover network features that can depict the state of the
network connection. Using important network data (DoS Efforts on how to define and characterize denial of
relevant features), we have developed kernel machine based and service attacks through a collection of different perspectives
soft computing detection mechanisms that achieve high detection such as bandwidth, process information, system information,
accuracies. We also present our work of identifying DoS user information and IP address is being proposed by several
pertinent features and evaluating the applicability of these researchers [1,6]. Using the defined characteristics a few
features in detecting novel DoS attacks. Architecture for signature-based and anomaly based detection techniques are
detecting DoS attacks at the router is presented. We proposed [2,9]. Recent malware and distributed DoS attacks
demonstrate that highly efficient and accurate signature based proved that there exists no effective means to detect, respond
classifiers can be constructed by using important network
and mitigate availability attacks.
features and machine learning techniques to detect DoS attacks
at the boundary controllers.
In this paper we propose a router based approach to
Keywords: Denial of service attacks, information assurance, detect denial of service attacks using intelligent systems. A
intrusion detection, machine learning, feature ranking, data comparative study of support vector machines (SVMs), Multi
reduction adaptive regression splines (MARSs) and linear genetic
programs (LGPs) for detecting denial of service attacks is
1 Introduction performed through a variety of experiments performed on a
By nature Internet is public, connected, distributed, well know Lincoln Labs data set that consists of more than
open, and dynamic. Phenomenal growth of computing 80% of different denial of service attacks described in section
devices, connectivity speed, and number of applications 2. We address the use of machine learning approach to
running on networked systems posed engendering risk to the discover network features that can depict the state of the
Internet. Malicious usage, attacks, and sabotage have been on network connection. We also present our work of identifying
the rise as more and more computing devices are put into use. DoS pertinent features from a publicly available intrusion
Connecting information systems to networks such as the detection data set and evaluating the applicability of these
Internet and public telephone systems further magnifies the features in detecting novel DoS attacks on a live performance
potential for exposure through a variety of attack channels. network. Architecture for detecting DoS attacks at the routers.
These attacks take advantage of the flaws or omissions that
exist within the various information systems and software that In the rest of the paper, a brief introduction to the data
run on many hosts in the network. used and DoS attacks is given in section 2. An overview of
soft computing paradigms used is given in section 3.
In DoS attacks the adversary mainly targets a few Experiments for detecting DoS attacks using MARs, SVMs
services like network bandwidth, router or server CPU cycles, and LGPs are given in section 4. Significant feature
identification techniques are presented in section 5. In section
106 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
6 we present the architecture and the applicability of DoS Slows down
Udpstrom Echo
significant features in detecting DoS attacks at the routers. the network
Conclusions are presented in section 7.
2 Intrusion detection data 3 Soft computing paradigms
A sub set of the DARPA intrusion detection data set is Soft computing was first proposed by Zadeh to construct
used for off-line analysis. In the DARPA intrusion detection new generation computationally intelligent hybrid systems
evaluation program, an environment was set up to acquire raw consisting of neural networks, fuzzy inference system,
TCP/IP dump data for a network by simulating a typical U.S. approximate reasoning and derivative free optimization
Air Force LAN. The LAN was operated like a real techniques. It is well known that the intelligent systems,
environment, but being blasted with multiple attacks [5,11]. which can provide human like expertise such as domain
For each TCP/IP connection, 41 various quantitative and knowledge, uncertain reasoning, and adaptation to a noisy and
qualitative features were extracted [16]. The 41 features time varying environment, are important in tackling practical
extracted fall into three categorties, “intrinsic” features that computing problems. In contrast with conventional Artificial
describe about the individual TCP/IP connections; can be Intelligence (AI) techniques which only deal with precision,
obtained form network audit trails, “content-based” features certainty and rigor the guiding principle of hybrid systems is
that describe about payload of the network packet; can be to exploit the tolerance for imprecision, uncertainty, low
obtained from the data portion of the network packet, “traffic- solution cost, robustness, partial truth to achieve tractability,
based” features, that are computed using a specific window. and better rapport with reality
2.1 Denial of service attacks 3.1 Support vector machines
Attacks designed to make a host or network incapable of The SVM approach transforms data into a feature space
providing normal services are known as denial of service F that usually has a huge dimension. It is interesting to note
attacks. There are different types of DoS attacks: a few of that SVM generalization depends on the geometrical
them abuse the computers legitimate features; a few target the characteristics of the training data, not on the dimensions of
implementations bugs; and a few exploit the the input space [3,4]. Training a support vector machine
misconfigurations. DoS attacks are classified based on the (SVM) leads to a quadratic optimization problem with bound
services that an adversary makes unavailable to legitimate constraints and one linear equality constraint. Vapnik shows
users. A few examples include preventing legitimate network how training a SVM for the pattern recognition problem leads
traffic, preventing access to services for a group or to the following quadratic optimization problem .
individuals. DoS attacks used for offline experiments and Minimize:
identifying significant features are presented in table 1 [5,11]. l l l
∑ ∑ ∑ y i y jα iα j k ( xi , x j ) (1)
1
W (α ) = − αi +
2
i =1 i =1 j =1
TABLE 1: DoS Attack Description l
Attack Type
Servic
e
Effect of the
attack Subject to ∑ y iα i (2)
i =1
Apache2 http Crashes httpd ∀i : 0 ≤ α i ≤ C
Freezes the
Land http
machine Where l is the number of training examples α is a vector of l
Mail bomb N/A Annoyance variables and each component α i corresponds to a training
example (xi, yi). The solution of (1) is the vector α for which
Denies service *
SYN Flood TCP on one or
(1) is minimized and (2) is fulfilled.
more ports
Ping of Death Icmp None
3.2 Linear genetic programs
Denies new
Process table TCP LGP is a variant of the Genetic Programming (GP)
processes
Slows down technique that acts on linear genomes . The linear genetic
Smurf Icmp programming technique used for our current experiment is
the network
Kills the based on machine code level manipulation and evaluation of
Syslogd Syslog programs. Its main characteristics in comparison to tree-
Syslogd
based GP lies is that the evolvable units are not the
Reboots the
Teardrop N/A expressions of a functional programming language (like
machine
LISP), but the programs of an imperative language (like C)
107 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
are evolved. In the Automatic Induction of Machine Code by
Genetic Programming, individuals are manipulated directly as
binary code in memory and executed directly without passing
an interpreter during fitness calculation. The LGP tournament
selection procedure puts the lowest selection pressure on the TABLE 2: Classifier Evaluation for Offline DoS Data
individuals by allowing only two individuals to participate in Classifier Accuracy (%)
a tournament. A copy of the winner replaces the loser of each Class
tournament. The crossover points only occur between SVM LGP MARS
instructions. Inside instructions the mutation operation
randomly replaces the instruction identifier, a variable or the Normal 98.42 99.64 99.71
constant from valid ranges. In LGP the maximum size of the
program is usually restricted to prevent programs without
DoS 99.45 99.90 96
bounds. As LGP could be implemented at machine code
level, it will be fast to detect intrusions in a near real time
mode.
5 Significant feature identification
3.3 Multi adaptive regression splines Feature selection and ranking is an important issue in
intrusion detection. Of the large number of features that can
be monitored for intrusion detection purpose, which are truly
Splines can be considered as an innovative mathematical useful, which are less significant, and which may be useless?
process for complicated curve drawings and function The question is relevant because the elimination of useless
approximation. To develop a spline the X-axis is broken into a features enhances the accuracy of detection while speeding up
convenient number of regions. The boundary between regions the computation, thus improving the overall performance of
is also known as a knot. With a sufficiently large number of an IDS. In cases where there are no useless features, by
knots virtually any shape can be well approximated. While it concentrating on the most important ones we may well
is easy to draw a spline in 2-dimensions by keying on knot improve the time performance of an IDS without affecting the
locations (approximating using linear, quadratic or cubic accuracy of detection in statistically significant ways.
polynomial etc.), manipulating the mathematics in higher
dimensions is best accomplished using basis functions. The • Having a large number of input variables x = (x1, x2, …,
MARS model is a regression model using basis functions as xn) of varying degrees of importance to the output y; i.e.,
predictors in place of the original data. The basis function some elements of x are essential, some are less important,
transform makes it possible to selectively blank out certain some of them may not be mutually independent, and
regions of a variable by making them zero, and allows MARS some may be useless or irrelevant (in determining the
to focus on specific sub-regions of the data. It excels at value of y)
finding optimal variable transformations and interactions, and
the complex data structure that often hides in high- • Lacking an analytical model that provides the basis for a
dimensional data . mathematical formula that precisely describes the input-
output relationship, y = F (x)
4 Offline evaluation
• Having available a finite set of experimental data, based
on which a model (e.g. neural networks) can be built for
We partition the data into the two classes of “Normal” simulation and prediction purposes
and “DoS” patterns, where the DoS attack is a collection of
six different attacks (back, neptune, ping of death, land, 5.1 Support vector decision function ranking
smurf, and teardrop). The objective is to separate normal and Information about the features and their contribution
DoS patterns. The (training and testing) data set contains towards classification is hidden in the support vector decision
11982 randomly generated from data described in section 3, function. Using this information one can rank their
with the number of data from each class proportional to its significance, i.e., in the equation
size, except that the smallest class is completely included. A
different randomly selected set of 6890 points of the total data
set (11982) is used for testing different soft computing F (X) = ΣWiXi + b (3)
paradigms. Results of SVM, MARS and LGP classifications
are given in Table 2.
The point X belongs to the positive class if F(X) is a
positive value. The point X belongs to the negative class if
F(X) is negative. The value of F(X) depends on the
contribution of each value of X and Wi. The absolute value of
108 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Wi measures the strength of the classification. If Wi is a large
positive value then the ith feature is a key factor for positive
class. If Wi is a large negative value then the ith feature is a
key factor for negative class. If Wi is a value close to zero on TABLE 3: Classifier Evaluation for Offline DoS Data
either the positive or the negative side, then the ith feature
does not contribute significantly to the classification. Based Rank
on this idea, a ranking can be done by considering the support Feature Description
ing
vector decision function. count:
service count:
5.2 Linear genetic ranking algorithm dst ost rv_serror_rate
SVDF
serror_rate:
In the feature selection problem the interest is in the dst_host_same_src_port_rate
representation of the space of all possible subsets of the given dst_host_serror_rate: %
input set. An individual of length d corresponds to a d- count:
dimensional binary feature vector Y, where each bit compromised conditions:
represents the elimination or inclusion of the associated wrong fragments:
feature. Then, yi = 0 represents elimination and yi = 1 LGP
land:
indicates inclusion of the ith feature. Fitness F of an individual logged in:
program p is calculated as the mean square error (MSE) hot:
pred
between the predicted output ( Oij ) and the desired output count:
des service count:
( Oij ) for all n training samples and m outputs .
dst_host_srv_diff_host_rate:
MARS
n m source bytes:
F ( p) =
1
n⋅m ∑ ∑ (O
i =1 j =1
pred
ij − Oij ) 2 + CE = MSE + w ⋅ MCE
des
n
w
(4) destination bytes:
hot:
Classification Error (CE) is computed as the number of
misclassifications. Mean Classification Error (MCE) is added TABLE 4: Significant feature evaluation
to the fitness function while its contribution is proscribed by Features
an absolute value of Weight (W). Classifier Normal DoS Normal DoS
41 41 6 6
5.3 Multi adaptive regression spines ranking SVM 98.42 99.45 99.23 99.16
Generalized cross-validation is an estimate of the actual
LGP 99.64 99.90 99.77 99.14
cross-validation which involves more computationally
intensive goodness of fit measures. Along with the MARS
MARS 99.71 96 99.80 95.47
procedure, a generalized cross-validation procedure is used to
determine the significant input features. Non-contributing
input variables are thereby eliminated .
6 Real time router based DoS detection
1 N
y − f ( xi ) 2
GCV =
N
∑ 1 − k ] (5)
i =1
[ i A passive sniffer can be placed at the router to collect
N data for detecting DoS attacks.
where N is the number of records and x and y are
independent and dependent variables respectively. k is the The architecture comprises of three components: a
effective number of degrees of freedom whereby the GCV packet parser, classifier and a response module. The network
adds penalty for adding more input variables to the model. packet parser uses the WINPCAP library to capture packets
and extracts the relevant features required for DoS detection.
The output of the parser includes the twelve DoS-relevant
5.4 Significant feature off line evaluation features as selected by our ranking algorithm [7,8].
Description of most important features as ranked by
three feature-ranking algorithms (SVDF, LGP, and MARS) is The output summary of the parser includes the eleven
given in table 3. Classifier performance using all the 41 features of duration1 of the connection to the target machine,
features and most important 6 features as inputs to the protocol2 used to connect, service type3, status of the
classifier is given in table 4. connection4 (normal or error), number of source bytes5,
number of destination bytes6, number of connections to the
109 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
same host as the current one during a specified time window7 538 2141 79.91
(in our case .01seconds), number of connections to the same DoS 153 2526 94.28
host as the current one using same service during the past 0 2679 100
0.01 seconds8, percentage of connections that have SYN Accuracy
errors during the past .01 seconds9, percentage of connections 83.77 80.44
SVM
that have SYN errors while using the same service during the 99.08 99.06
LGP
past .01 seconds10, and percentage of connections to the same 63.9 73.2
MARS
service during the past .01 seconds11.
We experimented with more than 24 types of DoS The top-left entry of Table 5 shows that 2692, 2578, and
attacks, including 6 types of DoS described in section 3 and 1730 of the actual “normal” test set were detected to be
17 additional types. In the experiments performed we used normal by SVM, LGP and MARS respectively; the last
different types of DoS attacks: SYN flood, SYN full, column indicates that 99.46, 95.26 and 63.9 % of the actual
MISFRAG, SYNK, Orgasm, IGMP flood, UDP flood, “normal” data points were detected correctly. In the same
Teardrop, Jolt, Overdrop, ICMP flood, FIN flood, and way, for “DoS” 2141, 2526 and 2679 of the actual “attack”
Wingate crash, with different service and port options. test set were correctly detected; the last column indicates that
Normal data included multiple sessions of http, ftp, telnet, 79.91, 94.28 and 100 % of the actual “DoS” data points were
SSH, http, SMTP, pop3 and imap. Network data originating detected correctly. The bottom row shows that 83.77, 99.08
from a host to the server that included both normal and DoS is and 63.0 % of the test set said to be “normal” indeed were
collected for analysis; for proper labeling of data for training “normal” and 83.77, 99.06 and 73.2 % of the test set
the classifier normal data and DoS data are collected at classified, as “DoS” indeed belongs to DoS as classified by
different times. SVM, LGP and MARS respectively.
7 Conclusions
DoS Monitor Mail Server
IIS
A number of observations and conclusions are drawn
from the results reported:
•
MAIL
ROUTER
WEB
Firewal
Apache A comparison of different soft computing techniques is
SSL/VPN
given. Linear genetic programming technique
Authentication
outperformed SVM and MARS with a 94.28 % detection
Communication Link Server
rate on the real time test dataset.
INTERNET
Regarding significant feature identification, we observe
Applicatio
n
that
Communication Link
•
User
User
The three feature ranking methods produce largely
consistent results.
• The most significant features for the two classes of
User
Attacker
‘Normal’ and ‘DOS’ heavily overlap.
• Using the 6 important features for each class gives the
remarkable performance.
Regarding real time router based DoS detection, we
Figure 2. Architecture for detecting DoS attacks at the routers
observe that
• DoS attacks can be detected at the router, thus pushing
the detection as far out as possible in the network
TABLE 5: Router based detection accuracies
perimeter
Normal DoS Accuracy
Class/
SVM SVM SVM • “Third generation worms” can be detected by tuning the
Learning time based features.
LGP LGP LGP
Machine
MARS MARS MARS • “Low and slow” DoS attacks can be detected by
2692 14 99.48 judiciously selecting the time based and connection based
Normal 2578 128 95.26 features.
1730 976 63.9
Acknowledgement
110 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
We take immense pleasure in thanking our Chairman Dr. [10] J. Stolfo, F. Wei, W. Lee, A. Prodromidis, and P. K.
Jeppiaar M.A, B.L, Ph.D, the Directors of Jeppiaar Chan, “Cost-based Modeling and Evaluation for Data Mining
Engineering College Mr. Marie Wilson, B.Tech, with Application to Fraud and Intrusion Detection”, Results
MBA.,(Ph.D) Mrs. Regeena Wilson, B.Tech, MBA., (Ph.D) from the JAM Project by Salvatore, 1999.
and the Principal Dr. Sushil Lal Das M.Sc(Engg.), Ph.D for
their continual support and guidance. We would like to extend [11] S. E. Webster, “The Development and Analysis of
our thanks to my guide, our friends and family members Intrusion Detection Algorithms”, S.M. Thesis, Massachusetts
without whose inspiration and support our efforts would not Institute of Technology, 1998.
have come to true. Above all, we would like to thank God for
making all our efforts success.
.
J.Visumathi B.E.,M.E.,(Ph.D) works as
Assistant Professor in Jeppiaar
References Engineering College , Chennai and She
[1] W. J. Blackert, D. C. Furnanage, and Y. A. Koukoulas, has more than 10 years of teaching
“Analysis of Denial of service attacks Using An address experience and her areas of
Resolution Protocol Attack”, Proc. of the 2002 IEEE specializations are Networks, Artificial
Workshop on Information Assurance, US Military Academy, Intelligence, and DBMS.
pp. 17-22, 2002.
[2] D. W. Gresty, Q. Shi, and M. Merabti, “Requirements
for a general framework for response to distributed denial of Dr. K.L. Shunmuganathan B.E,
service,” Proc. Of Seventeenth Annual Computer Security M.E.,M.S.,Ph.D works as the
Applications Conference, pp. 422-229, 2001. Professor & Head of CSE Department
of RMK Engineering College,
[3] T. Joachims, “Making Large-Scale SVM Learning Chennai, TamilNadu, India. He has
Practical”, LS8-Report, University of Dortmund, LS VIII- has more than18 years of teaching
Report, 2000. experience and his areas of
specializations are Networks, Artificial
[4] T. Joachims, “SVMlight is an Implementation of Intelligence, and DBMS.
Support Vector Machines (SVMs) in C”, University of
Dortmund. Collaborative Research Center on Complexity
Reduction in Multivariate Data (SFB475), 2000.
[5] K. Kendall, “A Database of Computer Attacks for the
Evaluation of Intrusion Detection Systems”, Master's Thesis,
Massachusetts Institute of Technology, 1998.
[6] J. Mirkovic, J. Martin, and P. Reiher, “A Taxonomy of
DDoS Attacks and DDoS Defense Mechanisms”, Technical
Report # 020017, Department of Computer Science, UCLA,
2002.
[7] S. Mukkamala, and A. H. Sung, “Identifying Key
Features for Intrusion Detection Using Neural Networks”,
Proc. ICCC International Conference on Computer
Communications, pp. 1132-1138, 2002.
[8] S. Mukkamala, and A.H. Sung “Feature Selection for
Intrusion Detection Using Neural Networks and Support
Vector Machines”, Journal of the Transportation Research
Board of the National Academics, Transportation Research
Record No 1822, pp. 33-39, 2003.
[9] C. Shields, “What do we mean by network denial of
service?”, Proc. of the 2002 IEEE workshop on Information
Assurance. US Military Academy, pp. 196-203, 2002.
111 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
A Novel Algorithm for Informative Meta
Similarity Clusters Using Minimum
Spanning Tree
S. John Peter S. P. Victor
Assistant Professor Associate Professor
Department of Computer Science and Department of Computer Science and
Research Center Research Center
St. Xavier’s College, Palayamkottai St. Xavier’s College, Palayamkottai
Tamil Nadu, India. Tamil Nadu, India.
jaypeeyes@rediffmail.com victorsp@rediffmail.com
ABSTRACT - The minimum spanning tree clustering number of vertices. More efficient algorithm for
algorithm is capable of detecting clusters with constructing MSTs have also been extensively
irregular boundaries. In this paper we propose two researched [13, 5, 8]. These algorithms promise
minimum spanning trees based clustering close to linear time complexity under different
algorithm. The first algorithm produces k clusters
assumptions. A Euclidean minimum spanning tree
with center and guaranteed intra-cluster similarity.
The radius and diameter of k clusters are computed (EMST) is a spanning tree of a set of n points in a
to find the tightness of k clusters. The variance of metric space (En), where the length of an edge is
the k clusters are also computed to find the the Euclidean distance between a pair of points in
compactness of the clusters. The second algorithm the point set.
is proposed to create a dendrogram using the k The hierarchical clustering approaches are related
clusters as objects with guaranteed inter-cluster to graph theoretic clustering. Clustering
similarity. The algorithm is also finds central cluster algorithms using minimal spanning tree takes the
from the k number of clusters. The first algorithm advantage of MST. The MST ignores many
uses divisive approach, where as the second
possible connections between the data patterns, so
algorithm uses agglomerative approach. In this
paper we used both the approaches to find the cost of clustering can be decreased. The MST
Informative Meta similarity clusters. based clustering algorithm is known to be capable
of detecting clusters with various shapes and size
Key Words: Euclidean minimum spanning tree, [24]. Unlike traditional clustering algorithms, the
Subtree, Eccentricity, Center, Tightness. Hierarchical MST clustering algorithm does not assume a
clustering, Dendrogram, Radius, Diameter, Central spherical shapes structure of the underlying data.
clusters, Compactness. The EMST clustering algorithm [20, 24] uses the
Euclidean minimum spanning tree of a graph to
produce the structure of point clusters in the n-
I INTRODUCTION
dimensional Euclidean space. Clusters are
Given a connected, undirected graph G=(V,E), detected to achieve some measures of optimality,
where V is the set of nodes, E is the set of edges such as minimum intra-cluster distance or
between pairs of nodes, and a weight w (u , v) maximum inter-cluster distance [2]. The EMST
specifying weight of the edge (u, v) for each edge algorithm has been widely used in practice.
(u, v) E. A spanning tree is an acyclic subgraph
of a graph G, which contain all vertices from G. Clustering by minimal spanning tree can be
The Minimum Spanning Tree (MST) of a viewed as a hierarchical clustering algorithm
weighted graph is minimum weight spanning tree which follows a divisive approach. Using this
of that graph. Several well established MST method firstly MST is constructed for a given
algorithms exist to solve minimum spanning tree input. There are different methods to produce
problem [21, 15, 17]. The cost of constructing a group of clusters. If the number of clusters k is
minimum spanning tree is O (m log n), where m is given in advance, the simplest way to obtain k
the number of edges in the graph and n is the clusters is to sort the edges of minimum spanning
112 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
tree in descending order of their weights and An important objective of hierarchical cluster
remove edges with first k-1 heaviest weights [2, analysis is to provide picture of data that can
23]. easily be interpreted. A picture of a hierarchical
clustering is much easier for a human being to
Geometric notion of centrality are closely linked comprehend than is a list of abstract symbols. A
to facility location problem. The distance matrix dendrogram is a special type of tree structure that
D can computed rather efficiently using Dijkstra’s provides a convenient way to represent
algorithm with time complexity O( | V| 2 ln | V | ) hierarchical clustering. A dendrogram consists of
[22]. layers of nodes, each representing a cluster.
The eccentricity of a vertex x in G and radius
ρ(G), respectively are defined as In this paper we propose two EMST based
e(x) = max d(x , y)and ρ(G) = min e(x) clustering algorithm to address the issues of
y V x V undesired clustering structure and unnecessary
The center of G is the set large number of clusters. Our first algorithm
C (G) = {x V | e(x) = ρ (G)} assumes the number of clusters is given. The
C (G) is the center to the “emergency facility algorithm constructs an EMST of a point set and
location problem” which is always contain single removes the inconsistent edges that satisfy the
block of G. The length of the longest path in the inconsistence measure. The process is repeated to
graph is called diameter of the graph G. we can create a hierarchy of clusters until k clusters are
define diameter D (G) as obtained. In section 2 we review some of the
D (G) = max e(x) existing works on graph based clustering
x V algorithm and central tree in a Minimum
The diameter set of G is Spanning Trees. In Section 3 we propose
Dia (G) = {x V | e(x) = D (G)} EMSTRD algorithm which produces k clusters
with center, radius, diameter and variance. We
All existing clustering Algorithm require a also propose another algorithm called
number of parameters as their inputs and these EMSTUCC for finding cluster of clusters using
parameters can significantly affect the cluster the k clusters which are from previous EMSTRD
quality. In this paper we want to avoid algorithm. The algorithm also finds central
experimental methods and advocate the idea of cluster. Hence we named this new cluster as
need-specific as opposed to care-specific because Informative Meta similarity clusters. The radius,
users always know the needs of their applications. diameter and variance of sub tree(cluster) is used
We believe it is a good idea to allow users to to find tightness and compactness of clusters.
define their desired similarity within a cluster and Finally in conclusion we summarize the strength
allow them to have some flexibility to adjust the of our methods and possible improvements.
similarity if the adjustment is needed. Our
Algorithm produces clusters of n-dimensional II. RELATED WORK
points with a given cluster number and a naturally
approximate intra-cluster distance. Clustering by minimal spanning tree can be
viewed as a hierarchical clustering algorithm
Hierarchical clustering is a sequence of partitions which follows the divisive approach. Clustering
in which each partition is nested into the next in Algorithm based on minimum and maximum
sequence. An Agglomerative algorithm for spanning tree were extensively studied. Avis [3]
hierarchical clustering starts with disjoint found an O (n2 log2 n) algorithm for the min-max
clustering, which places each of the n objects in diameter-2 clustering problem. Asano,
an individual cluster [1]. The hierarchical Bhattacharya, Keil and Yao [2] later gave
clustering algorithm being employed dictates how optimal O (n log n) algorithm using maximum
the proximity matrix or proximity graph should be spanning trees for minimizing the maximum
interpreted to merge two or more of these trivial diameter of a bipartition. The problem becomes
clusters, thus nesting the trivial clusters into NP-complete when the number of partitions is
second partition. The process is repeated to form a beyond two [12]. Asano, Bhattacharya, Keil and
sequence of nested clustering in which the number Yao also considered the clustering problem in
of clusters decreases as a sequence progress until which the goal to maximize the minimum inter-
single cluster containing all n objects, called the cluster distance. They gave a k-partition of point
conjoint clustering, remains[1]. set removing the k-1 longest edges from the
113 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
minimum spanning tree constructed from that simply removes k-1 longest edges so that the
point set [2]. The identification of inconsistent weight of the subtrees is minimized. The second
edges causes problem in the MST clustering objective function is defined to minimize the total
algorithm. There exist numerous ways to divide distance between the center and each data point in
clusters successively, but there is not suitable the cluster. The algorithm removes first k-1 edges
choice for all cases. from the tree, which creates a k-partitions.
Zahn [24] proposes to construct MST of point set The clustering algorithm proposed by
and delete inconsistent edges – the edges, whose S.C.Johnson [11] uses proximity matrix as input
weights are significantly larger than the average data. The algorithm is an agglomerative scheme
weight of the nearby edges in the tree. Zahn’s that erases rows and columns in the proximity
inconsistent measure is defined as follows. Let e matrix as old clusters are merged into new ones.
denote an edge in the MST of the point set, v1 and The algorithm is simplified by assuming no ties in
v2 be the end nodes of e, w be the weight of e. A the proximity matrix. Graph based algorithm was
depth neighborhood N of an end node v of an proposed by Hubert [7] using single link and
edge e defined as a set of all edges that belong to complete link methods. He used threshold graph
all the path of length d originating from v, for formation of hierarchical clustering. An
excluding the path that include the edge e. Let N1 algorithm for single-link hierarchical clustering
and N2 be the depth d neighborhood of the node v1 begins with the minimum spanning tree (MST) for
and v2. Let ŴN1 be the average weight of edges in G (∞), which is a proximity graph containing n(n-
N1 and σN1 be its standard deviation. Similarly, let 1)/2 edge was proposed by Gower and Ross [9].
ŴN2 be the average weight of edges in N2 and σN2 Later Hansen and DeLattre [6] proposed another
be its standard deviation. The inconsistency hierarchical algorithm from graph coloring.
measure requires one of the three conditions hold:
Given n d-dimensional data objects or points in a
1. w > ŴN1 + c x σN1 or w > ŴN2 + c x σN2 cluster, we can define the centroid x0, radius R,
diameter D and variance of the cluster as
2. w > max(ŴN1 + c x σN1 , ŴN2 + c x σN2)
3. w >f
max (c x σN1 , c x σN2)
where c and f are preset constants. All the edges
of a tree that satisfy the inconsistency measure are
considered inconsistent and are removed from the
tree. This result in set of disjoint subtrees each
represents a separate cluster. Paivinen [19]
proposed a Scale Free Minimum Spanning Tree
(SFMST) clustering algorithm which constructs
scale free networks and outputs clusters
containing highly connected vertices and those
connected to them. where R is the average distance from member
objects to the centroid, and D is the average
The MST clustering algorithm has been widely pairwise distance within a cluster. Both R and D
used in practice. Xu (Ying), Olman and Xu reflect the tightness of the cluster around
(Dong) [23] use MST as multidimensional gene centroid[25].
expression data. They point out that MST- based
clustering algorithm does not assume that data The cospanning tree of a tree spanning tree T is
points are grouped around centers or separated by edge complement of T in G. Also the rank ρ(G)
regular geometric curve. Thus the shape of the of a Graph G with n vertices and k connected
cluster boundary has little impact on the components is n-k.
performance of the algorithm. They described A central tree[4] of a graph is a tree T0 such that
three objective functions and the corresponding the rank r of its cospanning tree To is minimum.
cluster algorithm for computing k-partition of
spanning tree for predefined k > 0. The algorithm r = ρ (T0) ≤ ρ (T), T G.
114 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Deo[4] pointed out that , if r is the rank of the clusters, with center, radius, diameter and
cospanning tree of T, then there is no tree in G at a variance of each cluster. We also present another
distance greater than r from T and there is at least algorithm to find the hierarchy of k clusters and
one tree in G at distance exactly r from T. A direct central cluster.
consequence of this is following characterization
of central tree. A. Cluster Tightness Measure and Compactness
A Spanning tree T0 is a central tree of G if and
only if the largest distance from T0 to any other In order to measure the efficacy of clustering, a
tree in G is minimum[4], ie, measure based upon the radius and diameter of
each subtree (cluster) is devised. The radius and
max d(T0,Ti) ≤ max d(T,Ti), Ti G diameter values of each cluster are expected low
value for good cluster. If the values are large that
The maximally distant tree problem, for instance , the points (objects) are spread widely and may
which ask for a pair of spanning tree(T 1,T2) such overlap. The cluster tightness measure is a within
that d(T1,T2) ≥ d(Ti,Tj), Ti,Tj G, can be solved – cluster estimate of clustering effectiveness ,
in polynomial time[14] . Also, as pointed out in however it is possible to devise inter- cluster
[16], the distance between tree pairs a graph G are measure also, to better measure the separation
in a one-to-one correspondence with the distance between the various clusters.
between vertex pairs in the tree-graph G. Thus
finding a central tree in G is equivalent to finding The Cluster compactness measure is based on the
a central vertex in a tree of G. However, while variance of the data points distributed in the
central vertex problem is known to have a subtrees (clusters). The variance of cluster T is
polynomial time algorithm (in number of computed as
vertices), such an algorithm can not be used
efficiently find a central tree, since the number of
vertices in a tree of G can be exponential.
III. OUR CLUSTERING ALGORITHM
A tree is a simple structure for representing binary Where d(xi, xj) is distance metric between two
relationship, and any connected components of points(objects) xi and xj, where n is the number of
tree is called subtree. Through this MST objects in the subtree Ti., and x0 is the mean of the
representation, we can convert a multi- subtree T. A smaller the variance value indicates,
dimensional clustering problem to a tree a higher homogeneity of the objects in the data
partitioning problem, i.e., finding particular set of set, in terms of the distance measure d ( ). Since
tree edges and then cutting them. Representing a d ( ) is the Euclidean distance, v (Ti) becomes the
set of multi-dimensional data points as simple tree statistical variance of data set σ (Ti). The cluster
structure will clearly lose some of the inter data compactness for the out put clusters generated by
relationship. However many clustering algorithm the algorithm is the defined as
proved that no essential information is lost for the
purpose of clustering. This is achieved through
rigorous proof that each cluster corresponds to
one subtree, which does not overlap the
representing subtree of any other cluster.
Clustering problem is equivalent to a problem of
identifying these subtrees through solving a tree Where k is the number of clusters generated on
partitioning problem. The inherent cluster the given data set S, v (Ti) is the variance of the
structure of a point set in a metric space is closely clusters Ti and V(S) is the variance of data set S.
related to how objects or concepts are embedded [10]
in the point set. In practice, the approximate
number of embedded objects can sometimes be The cluster compactness measure evaluates how
acquired with the help of domain experts. Other well the subtrees (clusters) of the input is
times this information is hidden and unavailable redistributed in the clustering process, compared
to the clustering algorithm. In this section we with the whole input set, in terms of data
preset clustering algorithm which produce k homogeneity reflected by Euclidean distance
115 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
metric used by the clustering process. Smaller the 10. ST = ST U {T’} // T’’ is new disjoint
cluster compactness value indicates a higher subtree
average compactness in the out put clusters. 11. nc = nc+1
12. Compute the center Ci of Ti using
B. EMSTRD Clustering Algorithm eccentricity of points
13. Compute the diameter of Ti using
Given a point set S in En and the desired number eccentricity of points
of clusters k, the hierarchical method starts by 14. Compute the variance of Ti
constructing an MST from the points in S. The 15. C = UTi ST {Ci}
weight of the edge in the tree is Euclidean 16 Until nc = k
distance between the two end points. Next the 17. Return k clusters with C
average weight Ŵ of the edges in the entire
EMST and its standard deviation σ are computed; Euclidean Minimum Spanning Tree is constructed
any edge with (W > Ŵ + σ) or (current longest at line 1. Average of edge weights and standard
edge) is removed from the tree. This leads to a set deviation are computed at lines 2-3. The variance
of disjoint subtrees ST = {T1, T2 …} (divisive of input data set S is computed at line 4. Using the
approach). Each of these subtrees Ti is treated as average weight and standard deviation, the
cluster. Oleksandr Grygorash et al proposed inconsistent edge is identified and removed from
algorithm [18] which generates k clusters. We Euclidean Minimum Spanning Tree (EMST) at
modified the algorithm in order to generate k lines (8-9). Subtree (cluster) is created at line 10.
clusters with centers. The algorithm is also find Radius, diameter and variance of subtree are
radius, diameter and variance of subtrees, which is computed at lines 12-14. Lines 9-15 in the
useful in finding tightness and compactness of algorithm are repeated until k number of subtrees
clusters. Hence we named the new algorithm as (clusters) are produced. The radius and diameter
Euclidean Minimum Spanning Tree for Radius are good measure to find the tightness of clusters.
and Diameter (EMSTRD). Each center point of k The radius and diameter values of each cluster are
clusters is a representative point for the each expected low value for good cluster. If the values
subtree ST. A point ci is assigned to a cluster i if ci are large that the points (objects) are spread
Ti. The group of center points is represented as widely. However if the value of k increases the
C = {c1, c2……ck} radius and diameter decreases.
Algorithm: EMSTRD(k) The variance for each subtree (cluster) is
computed to find the compactness of clusters. A
Input : S the point set smaller the variance value indicates, a higher
Output : k number of clusters with C (set homogeneity of the objects in the data set. The
of center points) cluster compactness measure evaluates how well
the subtrees (clusters) of the input is redistributed
Let e be an edge in the EMST constructed in the clustering process, compared with the
from S whole input set, in terms of data homogeneity
Let We be the weight of e reflected by Euclidean distance metric used by the
Let σ be the standard deviation of the edge clustering process. Smaller the cluster
weights compactness value indicates a higher average
Let ST be the set of disjoint subtrees of the compactness in the out put clusters.
EMST
Let nc be the number of clusters Figure 1 illustrate a typical example of cases in
1. Construct an EMST from S which simply remove the k-1 longest edges does
2. Compute the average weight of Ŵ of all not necessarily output the desired cluster
the edges structure. Our algorithm finds the center, radius,
3. Compute standard deviation σ of the edges diameter and variance of the each cluster which
4. Compute variance of the set S will be useful in many applications. Our algorithm
5. ST = ø; nc = 1; C= ø; will find 7 cluster structures (k=7). Figure 2
6. Repeat shows the possible distribution of the points in the
7. For each e EMST two cluster structures with their radius; diameter
8. If (We > Ŵ + σ) or (current longest edge e) and also theirs center points 5 and 3. Figure 3
9. Remove e from EMST shows a graph which shows the relationship
116 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
between radius and diameter with subtrees
(clusters). Lower the radius and diameter value
means higher the tightness. The compactness of
the subtrees is shown in the figure 4. Lower the
values of variance means higher the homogeneity
of the objects in subtree (cluster).
Figure 4. Compactness of clusters
Figure 1. Clusters connected through a point
Figure 5. EMST From 7 cluster center points
with central cluster 3
C. EMSTUCC Algorithm for Central cluster
The distance between two sub trees (clusters) of
an EMST T is defined as the number of edges
present in one sub tree (cluster) but not present in
the other.
d (T1, T2) = |T1-T2| = |T2-T1|
Figure 2. Two Clusters with radius and diameter (5 and Definition 1: A sub tree (cluster) is a tree T0 is a
3 as center point) central sub tree (central cluster) of EMST T if and
only if the largest distance from T0 to any other
sub tree (cluster) in the EMST T is minimum.
The result of the EMSTRD algorithm consists of
k number clusters with their center, radius,
diameter and variance. These center points c1, c2
….ck are connected and again minimum spanning
tree is constructed is shown in the Figure 5. A
Euclidean distance between pair of clusters can be
represented by a corresponding weighted edge.
Our Algorithm is also based on the minimum
spanning tree but not limited to two-dimensional
Figure 3: Tightness of clusters using radius and points. There were two kinds of clustering
diameter problem; one that minimizes the maximum intra-
cluster distance and the other maximizes the
minimum inter-cluster distances. Our Algorithms
117 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
produces clusters with both intra-cluster and inter- central cluster. It places the entire disjoint cluster
cluster similarity. We propose Euclidean at level 0 (line 3). It then checks to see if T still
Minimum Spanning Tree Updation Central contains some edge (line 4). If so, it finds
Cluster algorithm (EMSTUCC) converts the minimum edge e (line 5). It then finds the vertices
minimum spanning tree into dendrogram, which i, j of an edge e (line 6). It then merges the
can be used to interpret about inter-cluster vertices (clusters) and forms a new vertex
distances. This new algorithm is also finds central (agglomerative approach). At the same time the
cluster from set of clusters. sequence number is increased by one and the level
of the new cluster is set to the edge weight (line
The algorithm is neither single link clustering 7). Finally, the Updation of minimum spanning
algorithm (SLCA) nor complete link clustering tree is performed at line 8. The lines 4-8 in the
algorithm (CLCA) type of hierarchical clustering, algorithm are repeated until the minimum
but it is based on the distance between centers of spanning tree T has no edge. The algorithm takes
clusters. This approach leads to new O (| E | 2) time.
developments in hierarchical clustering. The
level function, L, records the proximity at which
each clustering is formed. The levels in the
dendrogram tell us the least amount of similarity
that points between clusters differ. This piece of
information can be very useful in several medical
and image processing applications.
Algorithm: EMSTUCC
Input: the point set C = {c1, c2……ck}
Output: central cluster and dendrogram
1. Construct an EMST T from C Figure 6. Dendrogram for Meta cluster
2. Compute the radius of T using
eccentricity of points // for central
cluster IV. CONCLUSION
3. Begin with disjoint clusters with
level L (0) = 0 and sequence number Our EMSTRD clustering algorithm assumes a
m=0 given cluster number. The algorithm gradually
4. If (T has some edge) finds k clusters with center for each cluster. These
5. e = get-min-edge(T) // for least k clusters ensures guaranteed intra-cluster
dissimilar pair of clusters similarity. The algorithm finds radius, diameter
6. (i, j) = get-vertices (e) and variance of clusters using eccentricity of
7. Increment the sequence number points in a cluster. The radius and diameter value
m = m + 1, merge the clusters (i) and gives the information about tightness of clusters.
(j), into single cluster to form next The variance value of the cluster is useful in
clustering m and set the level of this finding the compactness of cluster. Our algorithm
cluster to L(m) = e; does not require the users to select and try various
8. Update T by forming new vertex by parameters combinations in order to get the
combining the vertices i, j; desired output. All of these look nice from
go to step 4. theoretical point of view. However from practical
9. Else Stop. point of view, there is still some room for
improvement for running time of the clustering
We use the graph of Figure 5 as example to algorithm. This could perhaps be accomplished by
illustrate the EMSTUCC algorithm. The using some appropriate data structure. Our
EMSTUCC algorithm construct minimum EMSTUCC clustering algorithm generates
spanning tree T from the points c1, c2, c3….ck (line dendrogram which is used to find the relationship
1) and convert the T into dendrogram is shown in between the k number clusters produced from the
figure 6. The radius of the tree T is computed at EMSTRD algorithm. The inter-cluster distance
line 2. This radius value is useful in finding the between k clusters are shown in the Dendrogram.
118 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
The algorithm is also finds central cluster. This
[14] G.Kishi, Y.Kajitani, “Maximally distant trees and
information will be very useful in many
principal partition of linear graph”. IEEE Trans. Circuit
applications. In the future we will explore and test Theory CT-16(1969) 323-330.
our proposed clustering algorithm in various
domains. The EMSTRD algorithm uses divisive [15] J.Kruskal. “On the shortest spanning subtree and the
travelling salesman problem”. In Proceedings of the American
approach, where as the EMSTUCC algorithm
Mathematical Society, Pages 48-50, 1956.
uses agglomerative approach. In this paper we
used both the approaches to find Informative Meta [16] N.Malik, Y.Y.Lee, “Finding trees and singed tree pairs by
similarity clusters. We will further study the rich the compound method”, in Tenth Midwest Symposium on
Circuit Theory, 1967, pp. V1-5-1-V1-5-11.
properties of EMST-based clustering methods in
solving different clustering problems. [17] J.Nesetril, E.Milkova and H.Nesetrilova. Otakar boruvka
“on minimum spanning tree problem: Translation of both the
1926 papers, comments, history. DMATH:” Discrete
Mathematics, 233, 2001.
REFERENCE
[18] Oleksandr Grygorash, Yan Zhou, Zach Jorgensen.
[1] Anil K. Jain, Richard C. Dubes “Algorithm for Clustering
“Minimum spanning Tree Based Clustering Algorithms”.
Data” Michigan State University, Prentice Hall, Englewood
Proceedings of the 18th IEEE International conference on tools
Cliffs, New Jersey 07632.1988.
with Artificial Intelligence (ICTAI’06) 2006.
[2] T.Asano, B.Bhattacharya, M.Keil and F.Yao. “Clustering
[19] N.Paivinen. “Clustering with a minimum spanning of
Algorithms based on minimum and maximum spanning trees”.
scale-free-like structure”.Pattern Recogn. Lett.,26(7): 921-930,
In Proceedings of the 4th Annual Symposium on
2005.
Computational Geometry,Pages 252-257, 1988.
[20] F.Preparata and M.Shamos. “Computational Geometry:
[3] D.Avis “Diameter partitioning”. Discrete and
An Introduction”. Springer-Verlag, Newyr, NY ,USA, 1985
Computational Geometry, 1:265-276, 1986
[21] R.Prim. “Shortest connection networks and some
[4] N.Deo, “A central tree”, IEEE Trans. Circuit theory CT-13
generalization”. Bell systems Technical Journal,36:1389-1401,
(1966) 439-440, correspondence.
1957.
[5] M.Fredman and D.Willard. “Trans-dichotomous
[22] Stefan Wuchty and Peter F. Stadler. “Centers of Complex
algorithms for minimum spanning trees and shortest paths”. In
Networks”. 2006
Proceedings of the 31st Annual IEEE Symposium on
Foundations of Computer Science,pages 719-725, 1990.
[23] Y.Xu, V.Olman and D.Xu. “Minimum spanning trees for
gene expression data clustering”. Genome Informatics,12:24-
[6] P. Hansen and M. Delattre “ Complete-link cluster analysis
33, 2001.
by graph coloring” Journal of the American Statistical
Association 73, 397-403, 1978.
[24] C.Zahn. “Graph-theoretical methods for detecting and
describing gestalt clusters”. IEEE Transactions on Computers,
[7] Hubert L. J “ Min and max hierarchical clustering using
C-20:68-86, 1971.
asymmetric similarity measures ” Psychometrika 38, 63-72,
1973.
[25] T.Zhang, R.Ramakrishnan and M.Livny. “BIRCH: an
efficient data clustering method for very large databases”. In
[8] H.Gabow, T.Spencer and R.Rajan. “ Efficient algorithms
Proc.1996 ACM-SIGMOD Int. Conf. Management of Data (
for finding minimum spanning trees in undirected and directed
SIGMOD’96), pages 103-114, Montreal, Canada, June 1996.
graphs”. Combinatorica, 6(2):109-122, 1986.
[9] J.C. Gower and G.J.S. Ross “Minimum Spanning trees and BIOGRAPHY OF AUTHORS
single-linkage cluster analysis” Applied Statistics 18, 54-64,
1969. S. John Peter is working as Assistant
professor in Computer Science,
[10] Ji He, Ah-Hwee Tan, Chew- Lim Tan, Sam- Yuan Sung. St.Xavier’s college (Autonomous),
“On Quantitative Evaluation of clustering systems. Palayamkottai, Tirunelveli. He earned his
Information Retrieval and Clustering”. W.Wu and H.Xiong M.Sc degree from Bharadhidasan
(Eds.). Kluwer Academic Publishers. 2002 University, Trichirappli. He also earned
his M.Phil from Bhradhidasan University,
[11] S. C. Johnson “Hierarchical clustering schemes” Trichirappli. Now he is doing Ph.D in
Psychometrika 32, 241-254, 1967. Computer Science at Manonmaniam
Sundranar University, Tirunelveli. He has publised research
[12] D,Johnson. “The np-completeness column: An ongoing papers on clustering algorithm in international journals.
guide”. Journal of Algorithms,3:182-195, 1982.
E-mail: jaypeeyes@rediffmail.com
[13] D.Karger, P.Klein and R.Tarjan. “A randomized linear-
time algorithm to find minimum spanning trees”. Journal of
the ACM, 42(2):321-328, 1995.
119 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Dr. S. P. Victor earned his M.C.A.
degree from Bharathidasan
University, Tiruchirappalli. The M.
S. University, Tirunelveli, awarded
him Ph.D. degree in Computer
Science for his research in Parallel
Algorithms. He is the Head of the
department of computer science, and
the Director of the computer science
research centre, St. Xavier’s college (Autonomous),
Palayamkottai, Tirunelveli. The M.S. University, Tirunelveli
and Bharathiar University, Coimbatore have recognized him as
a research guide. He has published research papers in
international, national journals and conference proceedings.
He has organized Conferences and Seminars at national and
state level.
E-mail: victorsp@rediffmail.com
120 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Adaptive Tuning Algorithm for Performance tuning of
Database Management System
S.F.Rodd1, Dr. U.P.Kulkrani2
1
Asst. Prof., Gogte Institute of Technology, Belgaum, Karnataka, INDIA
Email: sfrodd@rediffmail.com
2
Prof., SDMCET Dharwar, Karnataka, INDIA.
Email: upkulkarni@yahoo.com
system resource bottlenecks. The performance of these
Abstract - Performance tuning of Database systems is affected by several factors. The important among
them include database size which grows with its usage over
Management Systems(DBMS) is both complex and
a period of time, increased user base, sudden increase in the
challenging as it involves identifying and altering several user processes, improperly or un-tuned DBMS. All of these
tend to degrade the system response time and hence call for
key performance tuning parameters. The quality of
a system that anticipates performance degradation by
tuning and the extent of performance enhancement carefully monitoring the system performance indicators and
auto tune the system.
achieved greatly depends on the skill and experience of
the Database Administrator(DBA). As neural networks Maintaining a database of an enterprise involves
have the ability to adapt to dynamically changing inputs considerable effort on part of a Database Administrator
(DBA) as, it is a continuous process and requires in-depth
and also their ability to learn makes them ideal knowledge, experience and expertise. The DBA has to
candidates for employing them for tuning purpose. In monitor several system parameters and fine tune them to
keep the system functioning smoothly in the event of
this paper, a novel tunig algorithm based on neural reduced performance or partial failure. It is therefore
network estimated tuning parameters is presented. The desirable to build a system that can tune itself and relieve
the DBA of the tedious and error prone task of tuning.
key performance indicators are proactively monitored Oracle 9i and 10g have built in support for tuning in the
and fed as input to the Neural Network and the trained form of tuning advisor. The tuning advisor estimates the
optimal values of the tuning parameters and recommends
network estimates the suitable size of the buffer cache, them to the DBA. A similar advisor is also available in SQL
shared pool and redo log buffer size. The tuner alters Server 2005 which is based on what-if analysis. In this
approach, the DBA provides a physical design as input and
these tuning parameters using the estimated values using the Tuning Advisor performs the analysis without actually
a rate change computing algorithm. The preliminary materializing the physical design. However, the advisor
available in 2005 recommends the changes needed at the
results show that the proposed method is effective in physical level such as creation of index on tables or views,
improving the query response time for a variety of restructuring of tables, creation of clustered index etc. which
are considered to be very expensive in terms of Database
workload types. Server down time and the effort on part of the DBA.
Keywords : DBA, Buffer Miss Ratio, Data Miner, Neural
Network, Buffer Cache. II. RELATED WORK
Several methods have been proposed that proactively
I. INTRODUCTION monitor the system performance indicators analyze the
Database Management Systems are an integral part of symptoms and auto tune the DBMS to deliver enhanced
any corporate house, the online systems, and e-commerce performance. Use of Materialized views and Indexes,
applications. For these systems, to provide reliable services Pruning table and column sets[1-2], Use of self healing
with quick query response times to their customers, the techniques[3-4], use of physical design tuning are among
Database Management Systems(DBMS) must be the proposed solutions. The classical control is modified and
functioning efficiently and should have built-in support for a three stage control involving Monitor, Analyze and
quick system recovery time in case of partial failure or Tune[6] is employed to ensure system stability. The
121 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
2010
architecture presented in [5] for self healing database forms IV NEURAL NETWORK
the basis for the new architecture presented here in this
paper. This paper presents a new DBMS architecture based As neural networks are best suited to handle complex
on modular approach, where in each functional module can systems and also have ability to learn based on the trained
be monitored by set of monitoring hooks. These monitoring data set, the same is used in the proposed architecture. As
hooks are responsible for saving the current status shown in Fig. 1, Neural Network will have p inputs, a
information or a snapshot of the server to the log. This specified number of nodes in the hidden layer and one or
architecture has high monitoring overhead, due to the fact more output nodes. The neural network used in this control
that when large number of parameters to be monitored, architecture is a feed forward network. The activation
almost every module's status information has to be stored on function used is sigmoid function. It is this function that
to the log and if done frequently may eat up a lot of CPU gives the neural network the ability to learn and produce an
time. Moreover, this architecture focuses more on healing output for which it is not trained. However, the neural
the system and does not consider tuning the DBMS for networks need a well defined training data set for their
performance improvement. proper functioning.
Ranking of various tuning parameters based on statistical
analysis is presented in[6]. The ranking of parameters is
based on the amount of impact they produce on the system
performance for a given workload. A formal knowledge Db_cache_size
framework for self tuning database system is presented in[7] Buffer Hit Ratio
that defines several knowledge components. The knowledge
components include Policy knowledge, Workload Avg_table_size
knowledge, Problem diagnosis knowledge, Problem
Resolution Knowledge, Effector knowledge, and
Dependency knowledge. The architecture presented in this
paper involves extracting useful information from the system Shared_pool_size
log and also from the DBMS using system related queries. Avg_table_size
This information gathered over a period of time is then used
to train a Neural Network for a desired output response time.
The neural network would then estimate the extent of
correction to be applied to the key system parameters that
Figure 1. Basic Neural Network Structure
help scale up the system performance.
The neural networks work in phases. In the first phase,
the network is trained using a well defined training set for a
III. PERFORMANCE TUNING
desired output. In the second phase a new input is presented
to the network that may or may not be part of the training
Calibrating the system for desired response time is called data set and network produces a meaningful output. For the
performance tuning. The objective of this system is to proper working of the neural network, it is important to
analyze the DBMS system log file and apply information choose a proper activation function, learning rate, number of
extraction techniques and also gather key system parameters training loops and sizeable number of nodes in the hidden
like buffer miss ratio, number of active processes and the layer.
tables that are showing signs of rapid growth. The control
architecture presented in this paper, only one parameter V. PROPOSED ARCHITECTURE
namely, the buffer cache is tuned. Using the statistical
information of these three parameters to train the Neural Fig. 2 Shows the architecture employed for identifying
Network and generate an output that gives an estimate of the the symptoms and altering key system parameters. The
optimal system buffer size. Since, the DBMS are dynamic DBMS system log file will be the primary source of
and continuously running around the clock, the above information that helps in checking the health of the system.
information must be extracted without causing any Since, the log file contains huge of amount of data, the data
significant system overhead. is first compressed into smaller information base by using a
data mining tool. The architecture has Data Miner, Neural
Extracting information from system log ensures that Network aggregator and Tuner as the basic building blocks.
there is no overhead. The queries that are used to estimate After, extracting meaningful information, the next step is to
buffer miss ratio, table size and number of user processes estimate the extent of correction required.
are carefully timed and their frequency is adjusted so that it
does not add to the overhead in monitoring the system.
122 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
New Shared Pool_Size Miss Ratio so that query execution time is reduced and the
memory is used efficiently.
DBMS
New Buff_Cache_Size
No. of Users
Tab. Size Buff.Miss Shared Pool Buff. Size
Buff_Miss_ratio Ratio size (in MB)
(in no. of (in MB)
records)
Neural 1000 0.9824 32 4
Data
Miner
Network
based
Tuner 1000 0.9895 32 4
Log
Aggregator 1000 0.9894 32 8
File
Avg_Table_size 1000 0.9505 32 8
2000 0.947 32 8
2000 0.9053 40 8
DBA 2000 0.8917 40 16
Tuning
Rules 2500 0.875 40 16
Figure 2. Neural Network based Tuning Architecture Table I. Sample Training Data Set
As suggested in[2] physical tuning should be avoided as The experiment was carried on Oracle 9i with a 3-input
it is expensive. Most importantly the internal corrective 2-output feed forward neural network with 100 internal
measure such as altering the buffer size of the DBMS used nodes. The training is carried with an epoch value of 100,
in query processing is explored in this architecture. learning rate of 0.4 and with a training dataset of size 100.
However, several parameters can be altered simultaneously The estimated buffer size generated by the Neural Network is
for better performance gain. The Neural network estimates based on the dynamic values of the above three parameters
the required buffer size based on the current DBMS input as input. The tuner takes this input and alters the buffer size.
parameters and the tuner applies the necessary correction to The results obtained are really promising. As can be seen
the buffer size based on the tuning rules. The tuner triggers a from the output in Fig. 4 the execution time is significantly
control action to fine tune the performance of the dbms lower for the increasing value of the buffer size. The query
based on the following algorithm used takes join of three tables and generate huge dataset as
result.
ALGORITHM dbTune(ESTMTD_DB_CACHE_SZ) Fig. 3 shows the effect of buffer cache size on the query
Begin response time. TPC-C type benchmark load was used which
represents an OLTP type load. As number of users grow
Compute the change in response time since beyond 12, the query response time starts rises rapidly. This
last modification (∆Rtime) is sensed by the neural network and calculates an appropriate
If ( ∆Rtime >0 and ∆Rtime > Rth) size of the new buffer size. The tuner uses the tuning rules to
Increase the new buffer_size to next apply the required correction. The tuning rules indicate when
higher granule size and at what interval of the buffer size, the correction is to be
Issue a command to alter the dbcache size applied. Tuning the DBMS frequently may affect the
performance and also lead to instability.
to the new value
Else
140
If(∆Rtime <0 and ∆Rtime < Rth) Without Tuning
120
Decrease the new buffer size to next lower
Response Time(msecs)
100
granule size.
Issue a command to alter the dbcache size 80
to the new value With Tuning
End 60
40
VI. EXPERIMENTAL RESULT
20
Table I shows the sample training data. A training data
set of size 100 was used to train the Neural Network. As can 0
be seen from the table, the buffer size is adjusted for 0 5 10 15 20 25
No. of Users
increased table size, Number of user processes and Buffer
Figure 3. Effect of Buffer size on Query Response Time
123 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
VII. CONCLUSION [7] Wiese, David; Rabinovitch, Gennadi, “Knowledge
Management in Autonomic Database Performance
A new tuning algorithm is presented in this paper. The
Neural Network estimates the buffer cache size based on the Tuning”, 20-25 April 2009.
trained data set. The correction is applied in accordance with
[8] B. DageVille and K. Dias, “Oracle’s self tuning
the tuning algorithm so as to scale up system performance.
This architecture learns from a training set to fine tune the architecture and solutions”, IEEE Data Engg. Bulletin,
system and thus it relieves the DBA of the tedious process of
tuning the DBMS and also need for an expert DBA. Vol 29, 2006.
Monitoring the macroscopic performance indicators ensures [9] S. Choudhuri and G. Weikum, “Rethinking database
that there is little monitoring overhead. However, the system
needs further refinement that takes into account sudden surge system architecture: Towards a self tuning risc style
in work loads and also the neural network training dataset database system”, in VLDB, 2000, pp 1-10.
must be derived based on proper database characterization. It
is also important to ensure that the system remains stable and [10] S. W. Cheng, D. Garlan et. al, “Architecture based self
gives consistent performance over a long period of time.
adaptation in the presence of multiple objectives”,
Proceedings of 2006 International journal of Computer
ACKNOWLEDGEMENTS
Systems and Engineering., 2006.
[11] Benoit Dageville and Karl Dias, “Oracle’s Self Tuning
We would like to thank Prof. D.A.Kulkarni for
scruitimizing the paper and for his valueable suggestions. Architecture and Solutions”., Bulletin of IEEE, 2006.
Special thanks to Prof. Santosh Saraf for his help in learning [12] Sanjay Agarwal, Nicolas Bruno, Surajit Chaudhari,
Neural Network implementation in MATLAB. We extend
our thanks to Computer Center, GIT, for providing “AutoAdmin: Self Tuning Database System
laboratory facilities. We thank our esteemed Management for Technology”, IEEE Data Engineering Bulletin, 2006.
their financial support.
[13] Soror, A.A.; Aboulnaga, A.; Salem, K., “Database
Virtualization: A New Frontier for Database Tuning
REFERENCES
[14] Gerhar Weikum, Axel Moenkerngerg et. al., Self-tuning
[1] S. Agarwal and et.al., “Automated selection of
Database Technology and Information Services : From
materialized views and indexes”, VLDB, 2007.
wishful thing to viable Engineering”, Parallel and
[2] Surjit Choudhuri, Vivek Narasayya, “Self tuning
Distributed Information System 1993.
database systems : A Decade progress”, Microsoft
[15] Satish, S.K.; Saraswatipura, M.K.; Shastry, S.C, “DB2
Research. 2007.
performance enhancements using Materialized Query
[3] Philip Koopman, “Elements of the Self-Healing System
Table for LUW Systems”, 2007. ICONS '07. Second
Problem Space”, IEEE Data Engineering Bulletin.
International Conference, April 2007.
2004.
[16] Chaudhuri, S.; Weikum G, “Foundations of Automated
[4] Peng Liu, “Design and Implementation of Self healing
Database Tuning”, Data Engineering, April 2006.
Database system”, IEEE Conference, 2005.
[17] Gennadi Rabinovitch, David Wiese, “Non-linear
[5] Rimma V. Nehme, “Dtabase, Heal Thyself”, Data Engg.
Optimization of Performance functions Autonomic
Workshop April 2008.
Database Performance Tuning”, IEEE Conference,
[6] Debnath, B.K.; Lilja, D.J.; Mokbel, M.F., “SARD: A
2007.
statistical approach for ranking database tuning
[19] Weikum G, Monkenberg A, “Self-tuning database
parameters”, Data Engineering Workshop, 2008.
technology: from wishful thinking to viable
ICDEW 2008. IEEE 24th International Conference,
engineering”, VLDB Conference, pages, 20-22.
April 2008 .
124 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
A Survey of Mobile WiMAX IEEE 802.16m
Standard.
Mr. Jha Rakesh Mr. Wankhede Vishal A. Prof. Dr. Upena Dalal
Deptt. Of E & T.C. Deptt. Of E & T.C. Deptt. Of E & T.C.
SVNIT SVNIT SVNIT
Surat, India Surat, India Surat, India
Jharakesh.45@gmail.com wankhedeva@gmail.com upena_dalal@yahoo.com
Abstract— IEEE 802.16m amends the IEEE 802.16 Wireless earlier this year has added mobility support. This is generally
MAN-OFDMA specification to provide an advanced air referred to as mobile WiMAX [1].
interface for operation in licenced bands. It will meet the
cellular layer requirements of IMT-Advanced next generation Mobile WiMAX adds significant enhancements:
mobile networks. It will be designed to provide significantly • It improves NLOS coverage by utilizing advanced
improved performance compared to other high rate antenna diversity schemes and hybrid automatic repeat
broadband cellular network systems. For the next generation
request (HARQ).
mobile networks, it is important to consider increasing peak,
sustained data reates, corresponding spectral efficiencies, • It adopts dense subchannelization, thus increasing
system capacity and cell coverage as well as decreasing latency system gain and improving indoor penetration.
and providing QoS while carefully considering overall system
complexity. In this paper we provide an overview of the state- • It uses adaptive antenna system (AAS) and multiple
of-the-art mobile WiMAX technology and its development. We input multiple output (MIMO) technologies to improve
focus our discussion on Physical Layer, MAC Layer, coverage [2].
Schedular,QoS provisioning and mobile WiMAX specification.
• It introduces a downlink subchannelization scheme,
Keywords-Mobile WiMAX; Physical Layer; MAC Layer; enabling better coverage and capacity trade-off [3-4].
Schedular; Scalable OFDM. This paper provides an overview of Mobile
WiMAX standards and highlights potential problems arising
I. INTRODUCTION from applications. Our main focuses are on the PHY layer,
IEEE 802.16, a solution to broadband wireless MAC layer specifications of mobile WiMAX. We give an
access (BWA) commonly known as Worldwide overview of the MAC specification in the IEEE 802.16j and
Interoperability for Microwave Access (WiMAX), is a recent IEEE802.16m standards, specifically focusing the discussion
wireless broadband standard that has promised high on scheduling mechanisms and QoS provisioning. We
bandwidth over long-range transmission. The standard review the new features in mobile WiMAX, including
specifies the air interface, including the medium access mobility support, handoff, and multicast services. We discuss
control (MAC) and physical (PHY) layers, of BWA. The key technical challenges in mobile WiMAX deployment. We
development in the PHY layer includes orthogonal then conclude the paper.
frequency-division multiplexing (OFDM), in which multiple
access is achieved by assigning a subset of subcarriers to II. PHYSICAL LAYER OF IEEE 802.16M.
each individual user [1]. This resembles code-division This section contains an overview of some Physical
multiple access (CDMA) spread spectrum in that it can Layer enhancements that are currently being considered for
provide different quality of service (QoS) for each user; users inclusion in future systems. Because the development of the
achieve different data rates by assigning different code 802.16m standard is still in a relatively early stage, the focus
spreading factors or different numbers of spreading codes. In is on presenting the concepts and the principles on which the
an OFDM system, the data is divided into multiple parallel proposed enhancements will be based, rather than on
substreams at a reduced data rate, and each is modulated and providing specific implementation details. Although the
transmitted on a separate orthogonal subcarrier. This exact degree of sophistication of the new additions to the
increases symbol duration and improves system robustness. standard cannot be safely predicted, it is expected that the
OFDM is achieved by providing multiplexing on user’s data additions will make some use of the concepts described
streams on both uplink and downlink transmissions. below.
Lack of mobility support seems to be one of the major
hindrances to its deployment compared to other standards A. Flexibility enhancements to support heterogeneous
such as IEEE 802.11 WLAN, since mobility support is users in IEEE 802.16m:
widely considered as one of the key features in wireless Because the goal of future wireless systems is to cater to
networks. It is natural that the new IEEE 802.16e released needs of different users, efficient and flexible designs are
125 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
needed. For some users (such as streaming low-rate B. Extending use of MIMO transmission
applications) link reliability may be more important than Multiple-Input Multiple-Output (MIMO) communication
high data rates, whereas others may be interested in is already a reality in wireless systems. It will be supported
achieving the maximum data rate even if a retransmission, by the IEEE 802.11n amendment to the 802.11 WLAN
and, therefore, additional delay may be required [4-6]. standards that is expected to be ratified in the near future.
Moreover, the co-existence of different users should be Similarly, 802.16e includes support for MIMO downlink and
achieved with relatively low control overhead. For these uplink transmission. As MIMO technology matures and
reasons, the frame format, the subcarrier mapping schemes implementation issues are being resolved, it is expected that
and the pilot structure are being modified for 802.16m with MIMO will be widely used for wireless communication.
respect to 802.16e. Each 802.16e frame consists of a Current Mobile WiMAX profiles include support for up to 2
downlink (DL) and an uplink (UL) part separated in time by transmit antennas even though the IEEE 802.16e standard
an OFDMA symbol and is of variable size [3,7]. The does not restrict the number of antennas, and allows up to 4
(downlink or uplink) frame begins by control information spatial streams. The current aim for Next Generation
that all users employ to synchronize and to determine if and WiMAX systems is to support at least up to 8 transmit
when they should receive or transmit in the given frame. antennas at the base station, 4 streams and Space-Time
Control information is followed by data transmission by the Coding [2]. Moreover, although some other MIMO features
base station (in the downlink subframe) or the mobile of 802.16e, such as closed-loop MIMO, have not appeared in
stations (in the uplink subframe). For each mobile station, Mobile WiMAX profiles yet, it is expected that they will be
transmission or reception happens in blocks that are included in new 802.16m-based systems. More specifically,
constructed from basic units called slots. Each slot can be it has been already decided to support closed-loop MIMO
thought of as a two-dimensional block, one dimension being using Channel Quality Information, Precoding Matrix Index
the time, the other dimension being the frequency. In and rank feedback in future systems.
general, a slot extends over one subchannel in the frequency
direction and over 1 to 3 OFDMA symbols in the time In 802.11 systems, as well as in the 802.16e standard,
direction, depending on the permutation scheme. The MIMO transmission is used to increase the data rate of the
subchannels are groups of OFDMA subcarriers. The number communication between a given transmitter-receiver pair
of subcarriers per subchannel and the distribution of the and/or improve the reliability of the link. It is expected that
subcarriers that make up a subchannel in the OFDMA 802.16m and future 3GPP systems will extend MIMO
symbol are determined based on the permutation scheme. As support to Multi-user (MU-) MIMO. More specifically, use
explained in more detail below, the subcarriers of a given of multiple antennas can improve the achievable rates of
subchannel are not always consecutive in frequency. users in a network with given frequency resources. In
Downlink and uplink subframes can be divided into different information theoretic terms, the capacity region of the uplink
zones where different permutation schemes are used [9-10]. and the downlink increases, in general, when MIMO
transmission is employed [2]. In many cases, a large portion
In the Partial Usage of Subchannels (PUSC) zone that is of this capacity increase can be achieved using relatively
mandatory, the priority is to improve diversity and to spread simple linear schemes (transmit beamforming at the
out the effect of inter-cell interference. Each slot extends downlink and linear equalizers at the uplink). Therefore, the
over 2 OFDMA symbols, and a subchannel consists of 24 achievable rates can be increased without the need for
data subcarriers that are distributed over the entire signal sophisticated channel coding. If larger complexity can be
bandwidth (OFDMA symbol). Thus, each subchannel has afforded, even higher gains can be attained using successive
approximately the same channel quality in terms of the decoding at the uplink and Dirty Paper Coding schemes at
channel gain and the inter-cell interference. To reduce the the downlink. An overview of the projected MIMO
effect of the inter-cell interference, when PUSC is used, the architecture for the downlink of 802.16m systems is given in
available subchannels are distributed among base stations so the System Description Document (SDD), and is repeated in
that adjacent base stations not use the same subchannels. Fig. 1 for convenience.
When the inter-cell interference is not significant, as in the
case of mobile stations located closely to a base station, it
may be advantageous to employ Full Usage of Subchannels
(FUSC). The goal of the FUSC permutation scheme is
similar to PUSC, i.e, to improve diversity and to spread out
the effect of inter-cell interference. However, as the name
suggests, in the FUSC zone all subchannels are used by a
base station. For this reason, the design of the pilot pattern
for the FUSC zone is slightly more efficient compared to
PUSC. A subchannel in the FUSC permutation zone consists
of 48 data subcarriers and the slot only comprises one
OFDMA symbol.
126 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
User1:data From legacy serving BS
Encoder 1.Resource IFFT to legacy target BS
Mapping From 802.16m serving
OFDM Handover with other BS to legacy target BS
Encoder 2. MIMO Symbol IFFT Not Specified
technologies From legacy serving BS
Encoder Constr- to 802.16m target BS
Sched -uction
ular
From 802.16m serving
3.Beam
Precoder BS to 802.16m target
BS
IEEE 802.11, 3GPP2,
GSM/EDGE, (E-
Encoder IFFT
Mobility Speed Vehicular: 120 km/h )UTRA (LTE TDD)
Using IEEE 802.21
Media Independent
Handover (MIH)
Feedback Precoding
Vector/ Indoor: 10 km/h
CSI
Matrix Basic Coverage Urban:
ACK/NAK
Not Specified 120 km/h
Mode/Link Position accuracy
High Speed: 350 km/h
Location Determination
Figure 1. MIMO architecture for the downlink of 802.16m systems.
Latency: 30 s
WiMAX and 3GPP networks employing MU- C. Resource allocation and multi-cell MIMO
MIMO will need to calculate which users should transmit In cellular networks careful frequency planning is
and receive during each frame, as well as the best required in order to achieve communication with small
achievable rate that corresponds to each user based on their outage probability and, at the same time, minimize
QoS requirements, the number of users in each cell and their interference among users of neighboring cells. Users near the
position. Although the information-theoretic capacity has cell edges are particularly vulnerable, because they receive
been characterized, this is not an easy task, even for signals of comparable strength from more than one base
narrowband systems, and it is even more challenging when stations [2]. For this reason, different parts of the frequency
all subcarriers of the OFDMA system are considered. spectrum are typically assigned to neighboring cells. The
Therefore, efficient algorithms will be needed at the base assignment in current systems is static and can only be
station for user selection that will also determine the changed by manual re-configuration of the system. Changes
beamforming filters for the downlink, the receiver filters for to the frequency allocation can only be performed
the uplink and the required power allocation at the base periodically and careful cell planning is required in order not
station and each mobile station. to affect other parts of the system. Frequencies are reused by
cells that are sufficiently far away so that the interference
TABLE I. MOST IMPORTANT FEATURES AND SYSTEM caused by transmissions on the same frequencies is small
REQUIREMENTS OF MOBILE WIMAX STANDARDS enough to guarantee satisfactory Signal- to-Interference and
Noise Ratios (SINRs). Although static frequency reuse
IEEE 802.16e IEEE802.16m schemes greatly simplify the design of cellular systems, they
Requirement
incur loss in efficiency because parts of the spectrum in some
100 Mbps for mobile cells may remain unused while, at the same time, other cells
Aggregate Data Rate 63 Mbps stations, 1 Gbps for may be restricting the rates of their mobile stations or even
fixed
denying admission to new users. Moreover, the handover
Operating Radio 2.3 GHz, 2.5-2.7 process is more complicated for mobile stations since
< 6 GHz
Frequency GHz, 3.5 GHz
communication in more than one frequencies is required.
Duplexing Schemes TDD and FDD TDD and FDD
D. Interoperability and coexistence.
up to 4 streams, no 4 or 8 streams, no limit
MIMO support In order for the standard to be able to support either
limit on antennas on antennas
legacy base and mobile stations or other technologies (e.g.
Coverage 10 km
3 km, 5-30 km and 30- LTE), the concept of the time zone, an integer number
100 km (greater than 0) of consecutive subframes, is introduced.
Handover Inter-
frequency 35-50 ms depending on scenario Interoperability among IEEE 802.16 standards [11]: The
Interruption Time 802.16m Network Reference Model permits interoperability
of IEEE 802.16m Layer 1 and Layer 2 with legacy 802.16
Handover Intra-
frequency
Not Specified 30 ms standards. The motivation for ensuring interoperability
Interruption Time comes from the fact that WiMAX networks have already
Handover between From 802.16e been deployed, and it is more realistic to require
802.16 standards serving BS to 100 ms interoperability instead of an update of the entire network.
(for corresponding 802.16e target BS Another advantage is that each 802.16 standard provides
mobile station) specific functionalities in a WiMAX network. The goal in
802.16m is to enable coexistence of all these functionalities
127 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
in a network without the need to create a new standard that units (SDUs) for the MAC CPS. This includes classification
contains all of them. The supported connections and frame of external data with the proper MAC service flow identifier
structure are summarized in Fig. 2 and Fig. 3. The legacy (SFID) and connection identifier (CID). An SDU is the basic
standard can transmit during the legacy zones (also called data unit exchanged between two adjacent protocol layers.
LZones), whereas 802.16m-capable stations can transmit [11,14] The MAC CPS provides the core functionality for
during the new zones. The Uplink (UL) portion shall start system access, allocation of bandwidth, and connection
with the legacy UL zone, because legacy base stations, establishment and maintenance. This sublayer also handles
mobile stations or relays expect IEEE 802.16e UL control the QoS aspect of data transmission. The security sublayer
information to be sent in this region. When no stations using provides functionalities such as authentication, secure key
a legacy 802.16 standard are present, the corresponding zone exchange, and encryption. For the PHY layer, the standard
is removed. The zones are multiplexed using TDM in the supports multiple PHY specifications, each handling a
downlink, whereas both TDM and FDM can used in the particular frequency range. The MAC CPS contains the
uplink. In each connection, the standard that is in charge is essential functionalities for scheduling and QoS provisioning
showcased. The Access Service Network can be connected in the system.
with other network infrastructures (e.g. 802.11, 3GPP etc.) or
IEEE 802.16d MAC provides two modes of operation:
to the Connectivity Service Network in order to provide
Internet to the clients. point-to-multipoint (PMP) and multipoint-to-multipoint
(mesh) [13]. The functionalities of the MAC sublayer are
related to PHY control (cross-layer functionalities, such as
HARQ ACK/NACK etc). The Control Signaling block is
responsible for allocating resources by exchanging messages
such as DL-MAP and UL-MAP. The QoS block allocates the
input traffic to different traffic classes based on the
scheduling and resource block, according to the SLA
guarantees. The name of other blocks, such as
fragmentation/packing,multi-radio coexistence and MAC
PDU formation, clearly describes their function. The MAC
sublayer also deploys state-of-the-art power saving and
handover mechanisms in order to enable mobility and make
connections available to speeds up to 350 km/h. Since newer
mobile devices tend to incorporate an increasing number of
functionalities, in WiMAX networks the power saving
Figure 2. Supported 802.16 connections implementation incorporates service differentiation on power
classes. A natural consequence of any sleeping mechanism is
the increase of the delay. Thus, delay-prone and non delay-
prone applications are allocated to different classes, such that
the energy savings be optimized, while satisfying the
appropriate QoS (e.g those that support web page
downloading or emails). MAC addresses play the role of
identification of individual stations. IEEE 802.16m
introduces two different types of addresses in the MAC
sublayer. 1) The IEEE 802 MAC address that has the generic
48-bit format and 2) two MAC logical addresses that are
assigned to the mobile station by management messages
from the base station. These addresses are used for resource
allocation and management of the mobile station and are
called “Station Identifiers” (assigned during network entry)
and “Flow Identifiers” (assigned for QoS purposes).
Figure 3. IEEE 802.16m frame structure with TDM Downlink and FDM
Uplink
III. BASIC FUNCTIONALITY OF MAC LAYER IN
WIMAX
Figure 4 presents the reference model in IEEE 802.16.
The MAC layer consists of three sublayers: the service-
specific convergence sublayer (CS), MAC common part
sublayer (MAC CPS), and security sublayer. The main
functionality of the CS is to transform or map external data
from the upper layers into appropriate MAC service data
128 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
schemes for various subcarriers and decides the number of
slots allocated. In systems with OFDMA PHY, the scheduler
needs to take into consideration the fact that a subset of
subcarriers is assigned to each user. Scheduler designers
need to consider the allocations logically and physically.
Logically, the scheduler should calculate the number of slots
based on QoS service classes. Physically, the scheduler
needs to select which subchannels and time intervals are
suitable for each user. The goal is to minimize power
consumption, to minimize bit error rate and to maximize the
total throughput. There are three distinct scheduling
processes: two at the BS - one for downlink and the other for
uplink and one at the MS for uplink as shown in Fig. 5. At
the BS, packets from the upper layer are put into different
queues, which ideally is per-CID queue in order to prevent
head of line (HOL) blocking. However, the optimization of
queue can be done and the number of required queues can be
reduced. Then, based on the QoS parameters and some extra
information such as the channel state condition, the DL-BS
scheduler decides which queue to service and how many
service data units (SDUs) should be transmitted to the MSs.
Since the BS controls the access to the medium, the second
scheduler - the UL-BS scheduler - makes the allocation
decision based on the bandwidth requests from the MSs and
the associated QoS parameters. Several ways to send
Figure 4. IEEE 802.16 reference model. bandwidth requests were described earlier in Section I.F.
Finally, the third scheduler is at the MS. Once the UL-BS
IV. SCHEDULER grants the bandwidth for the MS, the MS scheduler decides
Scheduling is the main component of the MAC layer that which queues should use that allocation. Recall that while
helps assure QoS to various service classes [12,13,14,16]. the requests are per connections, the grants are per subscriber
The scheduler works as a distributor to allocate the resources and the subscriber is free to choose the appropriate queue to
among MSs. The allocated resource can be defined as the service. The MS scheduler needs a mechanism to allocate the
number of slots and then these sots are mapped into a bandwidth in an efficient way. Fig. 6 classification of
number of subchannels (each subchannel is a group of scheduler is given.
multiple physical subcarriers) and time duration (OFDM
symbols). In OFDMA, the smallest logical unit for
bandwidth allocation is a slot. The definition of slot depends
upon the direction of traffic (downlink/uplink) and
subchannelization modes. For example, in PUSC mode in
downlink, one slot is equal to twenty four subcarriers (one
subchannel) for three OFDM symbols duration. In the same
mode for uplink, one slot is fourteen subcarriers (one uplink
subchannel) for two OFDM symbols duration. The mapping
process from logical subchannel to multiple physical
subcarriers is called a permutation. PUSC, discussed above is
one of the permutation modes. Others include Fully Used
Subchannelization (FUSC) and Adaptive Modulation and
Coding (band-AMC). The term band-AMC distinguishes the
permutation from adaptive modulation and coding (AMC)
MCS selection procedure. Basically there are two types of
permutations: distributed and adjacent. The distributed
subcarrier permutation is suitable for mobile users while
adjacent permutation is for fixed (stationary) users. After the
scheduler logically assigns the resource in terms of number
of slots, it may also have to consider the physical allocation,
e.g., the subcarrier allocation. In systems with Single Carrier
PHY, the scheduler assigns the entire frequency channel to a
MS. Therefore, the main task is to decide how to allocate the
number of slots in a frame for each user. In systems with
OFDM PHY, the scheduler considers the modulation Figure 5. Component Schedulers at BS and MSs
129 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Most of data traffic falls into this category. This service
Schedulars class guarantees neither delay nor throughput. The
bandwidth will be granted to the MS if and only if there is a
left-over bandwidth from other classes. In practice most
implementations allow specifying minimum reserved traffic
rate and maximum sustained traffic rate even for this class.
Key Objective
Channel Unaware Channel Aware Fairness Note that for non-real-time traffic, traffic priority is
QoS Guarantee also one the QoS parameters that can differentiate among
System different connections or subscribers within the same service
class. Consider bandwidth request mechanisms for uplink.
UGS, ertPS and rtPS are real-time traffic. UGS has a static
Intra Class Inter Class
allocation. ertPS is a combination of UGS and rtPS. Both
UGS and ertPS can reserve the bandwidth during setup.
FIFO RR,WRR
Unlike UGS, ertPS allows all kinds of bandwidth request
RR,WRR,DRR Priority,DTPQ
EDW,LWDF including contention resolution. rtPS can not participate in
Avg. Bw/Frame contention resolution. For other traffic classes (non real-time
wFQ traffic), nrtPS and BE, several types of bandwidth requests
are allowed such as piggybacking, bandwidth stealing,
unicast polling and contention resolution. These are further
Figure. 6. Classification of WiMAX schedulers
discussed in Section I.F. Thus mobile WiMAX brings
V. WIMAX QOS SERVICE CLASSES potential benefits in terms of coverage, power consumption,
self-installation, frequency reuse, and bandwidth efficiency.
IEEE 802.16 defines five QoS service classes: One of the key complications is that the incompatibility in
Unsolicited Grant Scheme (UGS), Extended Real Time the newly introduced scalable OFDM (SOFDM) in IEEE
Polling Service (ertPS), Real Time Polling Service (rtPS), 802.11e with the original OFDM scheme forces equipment
Non Real Time Polling Service (nrtPS) and Best Effort manufacturers to come up with mechanisms to ease the
Service (BE). Each of these has its own QoS parameters such transition
as minimum throughput requirement and delay/jitter
constraints. Table II presents a comparison of these classes
[15-16]. TABLE II. COMPARISON OF WIMAX QOSSERVICE CLASSES
UGS: This service class provides a fixed periodic QoS Pros Cons
bandwidth allocation. Once the connection is setup, there is Bandwidth may not be
no need to send any other requests. This service is designed UGS
No overhead. Meet guaranteed utilized fully since
for constant bit rate (CBR) real-time traffic such as E1/T1 latency for real- time service allocations are granted
regardless of current need
circuit emulation. The main QoS parameters are maximum
Need to use the polling
sustained rate (MST), maximum latency and tolerated jitter mechanism(to meet the
(the maximum delay variation). Optimal latency and data delay guarantee) and a
ertPS
overhead efficiency mechanism to let the BS
ertPS: This service is designed to support VoIP with know when the traffic
silence suppression. No traffic is sent during silent periods. starts during silent perios
ertPS service is similar to UGS in that the BS allocates the Require the overhead of
maximum sustained rate in active mode, but no bandwidth is rtPS
Optimal data transport bandwidth request and the
allocated during the silent period. There is a need to have the efficiency polling latency(to meet the
delay guarantee)
BS poll the MS during the silent period to determine if the Provide efficient service for
silent period has ended. The QoS parameters are the same as nrtPS non-real-time traffic with N/A
those in UGS. minimum reserved rate
No service guarantee,
rtPS: This service class is for variable bit rate (VBR) Provide efficient service for BE some connections may
realtime traffic such as MPEG compressed video. Unlike BE
traffic starve for long period of
UGS, rtPS bandwidth requirements vary and so the BS needs time
to regularly poll each MS to determine what allocations need
to be made. The QoS parameters are similar to the UGS but VI. CONCLUSION
minimum reserved traffic rate and maximum sustained traffic This paper presents an overview of the IEEE 802.16m
rate need to be specified separately. For UGS and ertPS PHY layer issues ,MAC protocol, specifically issues
services, these two parameters are the same, if present. associated with scheduling and QoS provisioning. It also
nrtPS: This service class is for non-real-time VBR traffic discusses the main features of the newly standardized mobile
with no delay guarantee. Only minimum rate is guaranteed. WiMAX, IEEE 802.16e to IEEE 802.16m. With the
File Transfer Protocol (FTP) traffic is an example of introduction of mobile WiMAX technology, it can be
applications using this service class. expected that future work will focus on the mobility aspect
and interoperability of mobile WiMAX with other wireless
130 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
technologies. For high quality voice and video, Internet and
mobility, demand for bandwidth is more. To address these
needs IEEE 802.16m appears as a strong candidate for
providing aggregate rates to high-speed mobile users at the
range of Gbps.
REFERENCES
[1] Baines R,“The roadmap to mobile WiMAX” , Communications
Engineer, vol. 3 , Issue 4 p.30-34, 2005.
[2] Piggin P,“Emerging mobile WiMAX antenna technologies”,
Communications Engineer, vol. 4 p.29-33,2006.
[3] Koon Hoo Teo, Zhifeng Tao, Jinyun Zhang,”The Mobile Broadband
WiMAX standard[Standards in Nutshell]”, Signal Processing
Magazine, IEEE, p144-148, 2007.
[4] Gozalvez, J, “WiMAX recognized as an IMT-2000 3G technology
[Mobile Radio],vol. 2, Issue 4, p 53-59,
[5] Tao Jiang, Weidong Xiang, Hsiao-Hwa Chen, Qiang Ni, “Multicast
Broadcast Services Support in OFDMA-Based WiMAX Systems
[Advances in Mobile Multimedia], Communications Magazine,IEEE,
vol. 45, p 76-86, 2007.
[6] Etemad,K., Lai, M, “Mobile WiMAX:a technology update[Guest
Editorial], vol. 46, p 26-28, 2008.
[7] Etemad, K., “Overview of mobile WiMAX technology and
evolution”, Communications magazine, IEEE, vol. 46, p31-40,2008.
[8] Riegel,M., “Ethernet services over mobile WiMAX”,
Communications Magazine, vol. 46, Issue 10, p 86-91, 2008.
[9] Jain R., Chakchai So-In, Al Tamimi, A.-K.,”System-level modeling
of IEEE 802.16E mobile wimax networks:Key issues”,Wireless
Communications, IEEE, vol. 15, Issue 5, p73-79, 2008.
[10] Garber, L, “Mobile WiMAX:The Next Wireless Battle Ground”,
IEEE Computer Society, vol. 41, Issue 6, p16-18,2008.
[11] Fan Wang, Ghosh A, Sankaran C, Fleming, P. Hsieh, F Benes, S,
“Mobile WiMAX systems:performance and
evolution”,Communications Magazine, IEEE, vol. 46, Issue 10, p41-
49,2008.
[12] Hongfei Du, Jiangchuan Liu, Jie Liang, “Downlink scheduling for
multimedia Multicast/broadcast over mobile wimax: connection-
oriented multistate adaptation.”,vol.16, Issue 4, p72-79, 2009.
[13] Kim,W, “Mobile WiMAX, the leader of the mobile Internet
era[WiMAX Report]”, Communications Magazine,IEEE, vol. 47,
Issue 6, p10-12, 2009.
[14] Chakchai So-In, Jain R, Tamimi, A.-k, “Scheduling in IEEE 802.16e
mobile WiMAX networks: key issues and a survey.”, vol. 27, Issue 2,
p156-171, 2009/
[15] Etemad K, Wang L.,”Multicast and broadcast multimedia services in
mobile WiMAX networks”,Communications Magazine,IEEE, vol. 47,
Issue 10, p84-91, 2009.
[16] Papapanagiotou I., Toumpakaris D, Jungwon Lee, Devetsikiotis M.,
“A survey on next generation mobile WiMAX networks: objectives,
features and technical challenges.”, Communications Surveys &
Tutorial, IEEE, vol. 11, Issue 4, p3-18, 2009.
131 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
An Analysis for Mining Imbalanced Datasets
T.Deepa ,Dr.M.Punithavalli,
DeepaRaman12@gmail.com ,mpunitha_srcw@yahoo.co.in
†
Faculty of Computer Science Department, Sri Ramakrishna College of Arts and Science for
Women,Coimbatore,Tamilnadu,India.
††
Director & Head, Sri Ramakrishna College of Arts & Science for Women, Coimbatore ,Tamil
Nadu,India
Summary: The class imbalance problem occurs when,
in a classification problem, there are many more
Mining Imbalanced datasets in a real world domain is instances of some classes than others.
an obstacle where the number of one (majority) class The class imbalance problem is pervasive
greatly outnumbers the other class (minority). This and ubiquitous, causing trouble to a large
paper traces some of the recent progress in the field of segment of the data mining community [N.
learning of Imbalanced data. It reviews approaches
adopted for this problem and it identifies challenges
Japkowicz.2000]. To better understand this
and points out future directions in this field. A problem the situation is illustrated in Figure 1. In
systematic study is developed aiming to question 1) Fig 1(a) there is a large imbalance between the
what type of Imbalance hinders the accuracy majority class (-) and the minority class(+).Fig
performance? 2) Whether the Imbalances are always 1(b) the classes are balanced.
damaging and to what extent? 3) Whether Down-
sizing approach and Over-sampling approaches can be Figure 1: (a) Many negative cases against some
proposed to deal with the problem? Finally this paper spare positive cases .(b) balanced data set with
leads to a profitable discussion of what the problem is well-defined clusters.
and how it might be addressed most effectively.
Keywords: Imbalanced Datasets, Undersampling,
Oversampling
Introduction:
The field of machine learning when transited
from the status of “academic displine” to
“applied science” a myriad of new issues arised,
one such issue is the class imbalance problem.
The Class Imbalance problem address the case
where the training sets of one class (majority) It is prevalent in many applications,
outnumbers the other class (minority). It is including: fraud/intrusion detection, risk
amply used in the world of business, industry management, text classification, and medical
and scientific research. Its importance grew as diagnosis/monitoring, and many others. It is
more and more researchers realized that it has a worth noting that in certain domains (like those
significant bottleneck in the performance by just mentioned) the class imbalance is intrinsic to
standard learning methods. On the other hand, it the problem. For example, within a given setting,
is observed that many real world domains there are very few cases of fraud
available datasets are imbalanced. as compared to the large number of honest use of
Through literature it is analyzed that the offered facilities.
imbalanced datasets is also dealt with rare However, class imbalances sometimes occur
classes or skewed data. in domains that do not have an intrinsic
imbalance. This will happen when the data
2. The Class Imbalance Problem: collection process is limited (e.g., due to
economic or privacy reasons), thus creating
\artificial Imbalances. Conversely, in certain
cases, the data abounds and it is for the scientist
132 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
to decide which examples to select and in what Visa,Anca Ralescu ,2005] [N. V. Chawla, N.
quantity [ G. Weiss and F. Provost,2003]. In Japkowicz, and A. Ko lcz, 2003].
addition, there can also be an imbalance in costs At the data level, these solutions include
of making different errors,things could vary per different forms of re-sampling such as random
case [N. V. Chawla, N. Japkowicz, and A. Ko over sampling with replacement, random under
lcz, 2003]. sampling, directed over sampling (in which no
new examples are created, but the choice of
3. Reasons for Imbalanced Datasets: samples to replace is informed rather than
1) Imbalanced ratio: The data are random), directed under sampling (where, again,
naturally imbalanced (e.g. credit card the choice of examples to eliminate is informed),
frauds and rare disease) over sampling with informed generation of new
(i.e.)IR=Number of minority samples, and combinations of the above
techniques.
Number of majority At the algorithmic level, solutions include
adjusting the costs of the various classes so as to
2) Lack of Information: counter the class imbalance, adjusting the
The data are not naturally imbalanced but it probabilistic estimate at the tree leaf (when
is too expensive to obtain data for learning working with decision trees), adjusting the
the minority class. decision threshold, and recognition-based (i.e.,
learning from one class) rather than
discrimination-based (two class) learning.
4.1Solution based on Data level for handling
Imbalanced datasets
Data level solutions include many different
forms of re-sampling such as random over
sampling with replacement, random under
Figure 2: Lack of positive data sampling, directed over sampling, directed under
sampling, over sampling with informed
3) Complexity:When the complexity generation of new samples, and combinations of
raises, learning the datasets is crucial. the above techniques.
4.1.1 Under sampling
Random under-sampling [Sofia Visa, Anca
Ralescu,2005] is a non-heuristic method that
aims to balance class distribution through the
Figure 3: High complexity data.
random elimination of majority class examples.
The logic behind this is to try to balance out the
4) Overlapping classes: where the data dataset in an attempt to overcome the
points belong to both the classes. idiosyncrasies of the machine learning algorithm.
The major drawback of random under sampling
is that this method can discard potentially useful
data that could be important for the induction
process.
Figure 3: Overlapping data. 4.1.2 Over sampling
Random over-sampling is a non-heuristic
4. Empirical Methods dealing with
method that aims to balance class distribution
Imbalanced Datasets: through the random replication of minority class
examples. Several authors[Sotiris Kotsiantis,
A number of solutions to the class- Dimitris Kanellopoulous, Panayiotis, 2006], [N.
imbalance problem were previously proposed V. Chawla, L. O. Hall, K. W. Bowyer, and W. P.
both at the data and algorithmic levels [Sofia
133 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Kegelmeyer,2002]. agree that random over- costs between classes [6]. Cost model takes the
sampling can increase the likelihood of occurring form of a cost matrix, where the cost of
overfitting, since it makes exact copies of the classifying a sample from a true class j to class i
minority class examples. corresponds to the matrix entry λij. This matrix
SMOTE generates synthetic minority examples is usually expressed in terms of average
to over-sample the minority class .Its main idea misclassification costs for the problem. The
is to form new minority class examples by diagonal elements are usually set to zero,
interpolating between several minority class meaning correct classification has no cost. We
examples that lie together. For every minority define conditional risk for making a decision αi
example, its k (which is set to 5 in SMOTE) as:
nearest neighbors of the same class are
calculated, then some examples are randomly
selected from them according to the over-
sampling rate. A new synthetic examples are
The equation states that the risk of choosing
generated along the line between the minority
class i is defined by fixed misclassification costs
example and its selected nearest neighbors. Thus,
and the uncertainty of our knowledge about the
the over fitting problem is avoided and causes
true class of x expressed by the posterior
the decision boundaries for the minority class to
probabilities. The goal in cost-sensitive
spread further into the majority class space.
classification is to minimize the cost of
misclassification, which can be realized by
4.2 Solution based on Algorithm level choosing the class (vj) with the minimum
for handling imbalance conditional risk.
4.2.1. One-class learning 4.2.3 Feature Selection
An interesting aspect of one-class (recognition- Feature selection is an important and relevant
based) learning is that, under certain conditions step for mining various data sets [I.Guyon &
such as multi-modality of the domain space, one A.Elisseef, 2003]. Learning from high
class approaches to solve the classification dimensional spaces can be very expensive and
problem superior to discriminative (two-class) usually not very accurate.
approaches (such as decision trees or Neural It is particularly relevant to various real-
Networks) is a rule induction system that world problems such as bioinformatics, image
utilizes a separate-and-conquer approach to processing, text classification, Web
iteratively build rules to cover previously categorization, etc. High dimensional real-world
uncovered training examples. Each rule is grown datasets are often accompanied by another
by adding conditions until no negative examples problem: high skew in the class distribution, with
are covered. It normally generates rules for each the class of interest being relatively rare. This
class from the rarest class to the most common makes it particularly important to select features
class. Given this architecture, it is quite that lead to a higher separability between the two
straightforward to learn rules only for the classes. It is important to select features that can
minority class one-class learning is particularly capture the high skew in the class distribution.
useful when used on extremely unbalanced data The majority of work in feature selection for
sets composed of a high dimensional noisy imbalanced data sets has focused on text
feature space. The one-class approach is related classification or Web categorization domain
to aggressive feature selection methods, but is [D.Mladenic & M.Grobelink, 1999].A couple of
more practical since feature selection can often papers in this paper concentrates at feature
be too expensive to apply. selection in the area of imbalanced data sets,
albeit in text classification or Web
4.2.2 Cost-sensitive learning categorization. [Zheng and Srihari, 2004] suggest
that existing measures used for feature selection
Changing the class distribution is not the only are not very appropriate for imbalanced data sets.
way to improve classifier performance when They propose a feature selection framework,
learning from imbalanced datasets. A different which selects features for positive and negative
approach to incorporate costs in decision-making classes separately and then explicitly combines
to define fixed and unequal misclassification them. The authors show simple ways of
134 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
converting existing measures so that they such as the small disjunct [ N. Japkowicz, 2003]
separately consider features for negative and and the rare cases problems, data duplication
positive classes. [Castillo and Serrano,2004] [Kolez, A. Chowdhury, and J. Alspector, 2003]
used a multi-strategy classifier system to and overlapping classes [S. Visa and A.
construct multiple learners, each doing its own Ralesc,2003]. It was found that in certain cases,
feature selection based on genetic algorithm. addressing the small disjunct problem with no
regard for the class imbalance problem was
5. Issues on Imbalanced Datasets: sufficient to increase performance. The method
for handling rare case disjuncts was found to be
In this section we analyze the imbalance similar to the m-estimation Laplace smoothing,
factor from various directions and we will focus but it requires less tuning. It was also found that
on answering questions such as, data duplication is generally harmful, although
Question1: What type of Imbalance hinders the for classifiers such as Naive Bayes and
accuracy performance? Perceptrons with Margins, high degrees of
Question 2: Whether the Imbalances are always duplication are necessary to harm classification
damaging and to what extent? [Kolez, A. Chowdhury, and J. Alspector,2003].
Question 3: Whether Down-sizing approach and [ Taeho Jo and N. Japkowicz (2004),]
Over-sampling approaches can be proposed to experiments suggest that the problem is not
deal with the problem? directly caused by class imbalances, but rather,
that class imbalances may yield small disjuncts
5.1 Solution for the issues: which, in turn, will cause degradation. The
resampling strategy proposed by [Taeho Jo and
Solution 1: (In N.Japkowicz ,2000) it is proved N. Japkowicz (2004] consists of clustering the
that when the datasets are learned linearly on training data of each class (separately) and
separable domains the accuracy performance is performing random oversampling cluster by
minimized. cluster. Its idea is to consider not only the
Solution 2: Class overlapping, Lack of between-class imbalance (the imbalance
information, distance between the classes also occurring between the two classes) but also the
hinders the accuracy performance. within-class imbalance (the imbalance occurring
Solution 3: Over-sampling the minority class between the subclusters of each class) and to
and Down-sizing the majority class are effective oversample the dataset by rectifying these two
when used along with recognition-based method. types of imbalances simultaneously.
Before performing random oversampling,
6. Other problems related with imbalance the training examples in the minority and the
majority classes must be clustered. Once the
However, it has also been observed that the class training examples of each class have been
imbalance is not the only problem responsible clustered, oversampling starts. In the majority
for the decrease in performance of learning class, all the clusters, except for
algorithms: the distribution of the data within the largest one, are randomly oversampled so as
each class is also relevant (between-class versus to get the same number of training examples as
within-class imbalance) [N. Japkowicz, B. the largest cluster. Let maxclasssize be the
Zadrozny and C. Elkan.,2001]. overall size of the large class.In the minority
[Prati et al,2004 ] developed a systematic class, each cluster is randomly oversampled until
study aiming to question whether class each cluster contains maxclasssize/Nsmallclass
imbalances hinder classifier induction or whether where Nsmallclass represents the number of
these deficiencies might be explained in other subclusters in the small class.
ways. Their study was developed on a series of Altogether, the experiments support the
artificial data sets in order to control all the hypothesis that cluster based
variables they wanted to analyze. The results of oversampling works better than simple
their experiments, using a discrimination-based oversampling or other methods for handling
inductive scheme, suggested that the problem is class imbalances or small disjuncts, especially
not solely caused by class imbalance, but is also when the number of training examples is small
related to the degree of data overlapping among and the problem, complex. The reason is that
the classes. cluster-based resampling identifies rare cases and
A number of papers discussed interaction re-samples them individually, so as to avoid
between the class imbalance and other issues the creation of small
135 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
disjuncts in the learned hypothesis. [3] N. V. Chawla, N. Japkowicz, and A. Ko
lcz, editors,Proceedings of the ICML'2003
7. Conclusion & Future Direction: Workshop on Learning
This paper was focused on the researches from Imbalanced Data Sets. 2003.
conducted in area of Imbalanced Datasets,
through a systematic study three issues was
considered and the solutions were found that [4] N. V. Chawla, L. O. Hall, K. W.
i)Accuracy performance hinders only if the data Bowyer, and W. P. Kegelmeyer
are learned linearily. ii) Class overlapping, Lack “SMOTE: Synthetic Minority Oversampling
of information, distance between the classes also Technique”. Journal of Artificial Intelligence
hinders the accuracy performance.iii)Over- Research, 16:321-357, 2002.
sampling the minority class and Down-sizing the
majority class are effective when used along with [5] Drummond, C., and Holte, R. C. “C4.5, Class
recognition-based method. Imbalance, and Cost Sensitivity: Why Under-
It also directs to new research directions such sampling beats Over-sampling”. In Workshop
as: on Learning from Imbalanced Data Sets II
i) Since many learning algorithms are (2003).
dealt with handling imbalanced
datasets an investigation must be [6]I. Guyon and A. Elissee_. An “Introduction to
carried out to find which method is variable and feature selection”. Journal of
appropriate. Machine Learning Research, 3:1157{1182, 2003.
ii) What is the best criterion to formulate
the accuracy performance, whether [7]N. Japkowicz, editor, Proceedings of the
confusion matrix can be considered AAAI'2000 Workshop on Learning from
or not? Imbalanced Data Sets ,AAAI Tech Report WS-
iii) All the learning methods are dealt 00-05. AAAI, 2000.
with bi-class problems, when
multiple classes are considered what [8] N. Japkowicz. “Concept-learning in the
is the best learning strategy? [Yanmin presence of between-class and within-class
Sun,2006]. imbalances”.In Proceedings of the Fourteenth
iv) Whether Expert agents is a fruitful Conference of the Canadian Society for
solution for handling imbalanced Computational Studies of Intelligence, pages 67-
datasets? [S. Kotsiantis, P. Pintelas, 77, 2001.
2003].
[9]N. Japkowicz. “Class imbalance: Are we
The above problems can be concentrated in focusing on the right issue?” In Proceedings of
future research to solve the imbalanced datasets the ICML'03 Workshop on Learning from
in real world domains. Imbalanced Data Sets, 2003.
References: [10] Kolez, A. Chowdhury, and J. Alspector.
. “Data duplication: An imbalance problem?” In
Proceedings of the ICML'2003 Workshop on
[1]G. E. A. P. A. Batista, R. C. Prati, and M.
Learning from Imbalanced Datasets, 2003.
C. Monard. “A study of the behavior of
several methods for balancing machine [11]S. Kotsiantis, P. Pintelas, “Mixture of Expert
learning training data.”SIGKDD Agents for Handling Imbalanced Data
Explorations, 6(1):20-29,2004. Sets”,Annals of Mathematics, Computing &
TeleInformatics, Vol 1, No 1 (46-55), 2003.
[2]In N. V. Chawla, N. Japkowicz, and A.
Ko lcz, “Special Issue on learning from
Imbalanced dataset”,Sigkdd Explorations [12]M. Kubat and S. Matwin. “Addressing the
Vol 6,issue 1,Pages1-6, 2006. curse of imbalanced training sets: One sided
selection.” In Proceedings of the Fourteenth
International Conference on Machine Learning,
pages 179-186, Nashville, Tennesse, 1997.
Morgan Kaufmann.
136 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
[13]Maloof, M. 2003. “Learning when data [21]Weiss, G. “Mining with rarity: A unifying
sets are imbalanced and when costs are framework.” SIGKDD Explorations 6(1):7–
unequal and unknown”. In Proc. of 19.2004.
the ICML-2003Workshop:Learning with
[22]Yanmin Sun,Mohammed S.Kamel,Yang
Imbalanced Data Sets II, 73–80.
Wang, “Boosting for Learning Multiple Classes
with imbalanced Class Distributions”
[14]D. Mladenic and M. Grobelnik,
IEEE,2006
“Feature selection for unbalanced class
distribution and naive bayes”. In
Proceedings of the 16th International [23]Z. Zheng, X. Wu, and R. Srihari. “Feature
Conference on Machine Learning, pages selection for text categorization on imbalanced
258{267, 1999.) data”. SIGKDD Explorations, 6(1):80{89, 2004.
[15]Prati, R. C., Batista, G. E. A. P. A., and [24] B. Zadrozny and C. Elkan.,”Learning and
Monard, M. C. “Class Imbalances versus making decisions when costs and probabilities
Class Overlapping: an Analysis of a are both unknown”. In Proceedings of the Sixth
ACM SIGKDD International Conference on
Learning System Behavior”. In MICAI
Knowledge Discovery and Data Mining, pages
(2004), pp. 312–321. LNAI 2972
204-213, 2001.
[16] Sofia Visa,Anca Ralescu “Issues in T.Deepa graduated with M.Sc in 2007
from Sri Ramakrishna college of arts &
Mining Imbalanced Data Sets- A Review Science for Women, India and completed
Paper”2005. M.Phil from Bharathiar University, India
during 2007-2008. Her areas of Interest
include Software Engineering & Data
[17]S. Visa and A. Ralescu. “Learning Mining. She has about 3 years of teaching
imbalanced and overlapping classes using experience. Currently she is working as a
Lecturer in Sri Ramakrishna college of Arts
fuzzy sets”. In Proceedings of the ICML'03 & Science for Women,India.
Workshop on Learning from Imbalanced
Data Sets, 2003
[18]Sotiris Kotsiantis, Dimitris
Kanellopoulous, Panayiotis, ”Handling
Imbalanced datasets:A Review “GESTS Dr. M.Punithavalli is presently working as
Director & Head of the Dept of Computer
International Transactions on Computer Science, Sri Ramakrishna College of arts and
Science and Engineering VOL..30,2006. science for women College, India.She has
published more than twenty papers in
national/International journals. Her areas of
[19]Taeho Jo and N. Japkowicz (2004), interest includes E-Learning, Software
“Class Imbalances versus Small Disjuncts,” Engineering, Data Mining, Networking and etc.
She has about 16 years of teaching experience.
Sigkdd Explorations. Volume 6, Issue 1 - She is guiding many research scholars
Page 40-49 and has published many papers in national and
international conference and in many
international journals
[20] G. Weiss and F. Provost. “Learning
when training data are costly: The effect of
class distribution on tree induction”.
Journal of Artificial Intelligence Research,
19:315-354, 2003.
137 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
QoS Routing For Mobile Adhoc Networks And
Performance Analysis Using OLSR Protocol
K.Oudidi A.Hajami M.Elkoutbi
Si2M Laboratory Si2M Laboratory Si2M Laboratory
National School of Computer Science National School of Computer Science National School of Computer Science
and Systems Analysis, Rabat, Morocco and Systems Analysis, Rabat, Morocco and Systems Analysis, Rabat, Morocco
k_oudidi@yahoo.fr Abdelmajid_hajami@yahoo.fr elkoutbi@ensias.ma
categories
Abstract-- This paper proposes a novel routing metrics based on of MANET routing protocols: Proactive (table-driven),
the residual bandwidth, energy and mobility index of the nodes. Reactive (on-demand) and Hybrid. Proactive protocols build
Metrics are designed to cope with high mobility and poor
their routing tables continuously by broadcasting periodic
residual energy resources in order to find optimal paths that
guarantee the QoS constraints. A maximizable routing metric routing updates through the network; reactive protocols build
theory has been used to develop a metric that selects, during thetheir routing tables on demand and have no prior knowledge
protocol process, routes that are more stable, offer a maximum of the route they will take to get to a particular node. Hybrid
throughput and prolong network life time. The OLSR protocols create reactive routing zones interconnected by
(Optimized Link State Routing) protocol, which is an proactive routing links and usually adapt their routing strategy
optimization of link state protocols designed for MANETs
to the amount of mobility in the network.
(Mobile Ad hoc Networks) is used as a test bed in this work. We
prove that our proposed composite metrics (based on mobility, In this paper we reiterate our proposed mobility metric.
energy and bandwidth) selects a more stable MPR set than the Based on the use of this mobility metric we propose a new
QOLSR algorithm which is a well known OLSR QoS extension. composite metric, to find the optimal path given the QoS
By mathematical analysis and simulations, we have shown the constraints. The objective of the composite metric is to find an
efficiency of this new routing metric in term of routing load, optimal stable path with maximum available bandwidth and to
packet delivery fraction, delay and prolonging the network
lifetime. prolong network life time.
Using the OLSR Protocol, we show that our proposed
Index Terms— Mobile Ad hoc networks, quality of service, routing metric selects stable MPR Set rather than the QOLSR
protocol, routing metric, mobility, residual energy. algorithm which is a well known OLSR QoS algorithm for
MANETs.
I. INTRODUCTION This paper is organized as follows. Section 2 gives an
A Mobile Ad hoc Network (MANET) is a collection of overview of the original OLSR protocol. Section 3 summarizes
mobile nodes working on a dynamic autonomous network. the state of the art dealing with QoS support in MANETs and
Nodes communicate with each other over the wireless medium describes the QoS routing problems Section 4 presents our
without need of a centralized access points or a base station. proposed composite metric based on mobility, residual energy
Since there is no existing communication infrastructure, and bandwidth as QoS parameters. In Section 5, simulations
adhoc networks cannot rely on specialised routers for path
and results are discussed. The last part of this paper concludes
discovery and routing. Therefore, nodes in such a network are
expected to act cooperatively to establish routes instantly. and presents some future work.
Such a network is also expected to route traffic, possibly over
multiple hops, in distributed manner, and to adapt itself to the II. OPTIMIZED LINK STATE ROUTING PROTOCOL
highly dynamic changes of its links , mobility and residual
energy patterns of its constituent nodes. A. Overview
OLSR (Optimized Link State Routing) protocol [2-3] is a
Providing QoS in MANETs [1] is a tedious task. It’s known
proactive table driven routing protocol for mobile ad hoc
that combining multiple criteria in the routing process is a
networks and it is fully described on RFC 3626 (Thomas
Hard problem (NP-Complet) A complete QoS model in
Clausen & Philippe Jacquet, (October 2003)). As a link state
MANETs will span multiple layers, however the network
routing protocol, OLSR periodically advertises the links
layer plays a vital role in providing the required support
building the network. However, OLSR optimizes the topology
mechanisms. The goal of QoS routing is to obtain feasible
information flooding mechanism, by reducing the amount of
paths that satisfy end-system performance requirements. Most
links that are advertised and by reducing the number of nodes
QoS routing algorithms present an extension of existing
forwarding each topology message to the set of MPRs only.
classic best effort routing algorithms. There are three main
Information topology is called Topology Control (TC) message
138 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
and exchanged using broadcasted into the network. TC Based on the above notations, the standard algorithm for
messages are only originated by nodes selected as Multipoint MPR selection is defined as follows (figure 2-b):
Relays (MPRs) by some other node in the network. MPRs are OLSR uses hop count to compute the shortest path to an
selected in such a way that a minimum amount of MPRs, arbitrary destination using the topology map consisting of all
located one-hop away from the node doing the selection its neighbours and of MPRs of all other nodes. Number of hop
(called MPR Selector), are enough to reach every single criterion as a routing metric is not suitable for QoS support as
neighbour located two-hops away of MPR selector. By a path selected based on the least number of hops may not
applying this selection mechanism only a reduced amount of satisfy the required QoS constraints.
nodes (depending on the network topology) will be selected as
MPRs[18]. Every node in the network is aware of its one-hop
and two-hop neighbours by periodically exchanging HELLO
messages containing the list of its one-hop neighbours. On the
other hand, TC messages will only advertise the links between
the MPRs and their electors. Then, only a partial amount of
the network links (the topology) will be advertised, also MPRs
are the only nodes allowed to forward TC messages and only
if messages come from a MPR Selector node. These
forwarding constrains considerably decrease the amount of
flooding retransmissions (Figure 1). This example shows the
efficiency of the MPR mechanism because only eight
transmissions are required to reach all the 23 nodes building
the network, which is a significant saving when compared to
traditional flooding mechanism where every node is asked to
Figure 2-b: MPR Selection Algorithm
retransmit to all neighbours.
III. RELATED WORK
A. Qos Support in a Manet
In this section we discuss the recent work done to provide
QoS functionality in Manets.
INSIGNIA, [7], is an adaptation of the IntServ Model to the
mobile ad hoc networks. QoS guarantee is done by per-flow
Figure 1: Flooding with MPR mechanism
information in each node that is set up by the
B. MPR Selection Algorithm signalling/reservation protocol. The destination statistically
measures QoS parameters (e.g. packet loss, delay, average
The computation of the MPR set with minimal size is a NP-
complet problem [14-16]. For this end, the standard MPR throughput,etc.) and periodically sends QoS reports to the
selection algorithm currently used in the OLSR protocol source. Based on those reports, the source node can adapt real-
time flows to avoid congestion.
implementations is as follows:
SWAN, [13], Service differentiation in stateless Wireless
Ad-hoc Network, is an adaptation of the DiffServ Model to the
mobile ad-hoc networks. Nodes do not need to keep per-flow
information in order to handle packets. QoS guarantee is
provided according to the class of the flow once it has been
accepted.
FQMM, [11], Flexible Qos Model for MANET, has been
introduced to offer a better QoS guarantee to a restricted
Figure 2-a- Example of MRRset calculation. number of flows whereas a class guarantee is offered to the
For a node x, let N(x) be the neighborhood of x. N(x) is the set other flows. FQMM is a hybrid approach combining per-flow
of nodes which are in the range of x and share with x a granularity of IntServ for high priority classes and perclass
bidirectional link. We denote by N2(x) the two-neighborhood granularity of DiffServ for low priority classes.
of x, i.e, the set of nodes which are neighbors of at least one G. Ying et al [8] have proposed enhancements that allow
node of N(x) but that do not belong to N(x) (see Figure 2-a). OLSR to find the maximum bandwidth path. The heuristics
139 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
are based on considering only bandwidth as a QoS routing • Concave {
: M(P) = min M i; j , M i;k ,..., M l;m }
constraint and revisions to the MPR selection criteria. They The proof of NP-Completeness relies heavily on the
identify that MPR selection is vital in optimal path selection. correlation of the link weight metrics. QoS Routing is NP-
The key concept in the revised MPR selection algorithm is Complete when the QoS metrics are independent, real
that a “good bandwidth” link should never be omitted. Based numbers or unbounded integers.
on this three algorithms were proposed: OLSR_R1, R2 and In general, QoS routing focuses on how to find feasible and
R1. optimal paths that satisfy QoS requirements of various voice,
The research group at INRIA [9],[10] proposed a QoS video and data applications. However, based on maximizable
routing scheme over OLSR. Their technique used delay and routing metrics theory [16], it is shown that two or more
bandwidth metric for routing table computation. Such metrics routing metrics can be combined to form a composite metric if
are included on each routing table entry corresponding to each the original metrics are bounded and monotonic.
destination. Before we proceed to the mathematical proof, we give
QOLSR [11] and work presented in [9] enhance OLSR with definitions of maximal metric tree and the properties desired
QoS support. Both propose a solution providing a path such for combining metrics i.e. bounded- ness and monotonicity.
that the bandwidth available at each node on the path is
higher than or equal to the requested bandwidth. Furthermore, Definition 1: Routing Metric
A routing metric for a network N is six-tuple (W,Wf, M, mr, met,
QOLSR considers delay as a second criterion for path
R ) where:
selection.
1. M is a set of metric values
However, all of these solutions do not take into account at
2. Wf is a function that assigns to each edge {i, j} in N a
all mobility and energy parameters induced by the nature of
Manet Network. weight Wf( {i, j}) in W
3. W is a set of edge weights
B. Qos Routing Problems 4. mr is a metric value in M assigned to the root.
One of the key issues in providing end-to-end QoS in a 5. met is a metric function whose domain is MxW and
given network is how to find a feasible path that satisfies the whose range is M (it takes a metric value and an edge
QoS constraints. The problem of finding a feasible path is NP- value and returns a metric value).
Complete if the number of constraints is more than two, it 6. R is a binary relation over m, the set of metric values that
cannot be exactly solved in polynomial time and mostly dealt satisfy the following four conditions of irreflexivity,
with using heuristics and approximations. The network layer Definition 2: Maximum Metric Tree
has a critical role to play in the QoS provision process. The A spanning tree of N is called a maximum metric tree with
approaches used by the QoS routing algorithms follow a trade- respect to an assigned metric iff every rooted path in T is
off between the optimality of paths and the complexity of maximum metric with respect to the assigned metric. In
algorithms especially in computing multiconstrained path. A simple words every node obtains its maximum metric through
survey on such solutions can be found in [14]. its path along a maximum metric tree.
The computation complexity is primarily determined by the
Definition 3: Boundedness
composition rules of the metrics [16]. The three basic
A routing metric (W, Wf, M, mr, met, R ) is bounded iff the
composition rules are: additive (such as delay, delay jitter,
following condition holds for every edge weight w in W and
logarithm of successful transmission, hop count and cost),
every metric value m in M.
multiplicative (like reliability and probability of successful met (m,w) R m ∨ met(m,w) = m
transmission) and concave/min-max (e.g. bandwidth). The
additive and multiplicative metric of a path is the sum and Definition 4: Monotonicity
multiplication of the metric respectively for all the links A routing metric (W,Wf, M, mr, met, R ) is monotonic iff the
constituting the path. The concave metric of a path is the following condition holds hue for every edge weight w in W
maximum or the minimum of the metric over all the links in and every pair of metric values m and m’ in M:
the path. m R m’ ⇒ (met (m,w) R met (m’,w)
Otherwise, if M i; j is the metric for link {i, j} and P is the ∨ met (m,w) = met (m’,w))
path between (i, j, k,..1,m) nodes, the QoS metric M(P) is (W,Wf, M, mr, met, R ) is called strict monotonic iff
defined as [14-15]: m R m’ ⇒ met (m,w) R met (m’,w)
• Additive : M(P) = M i ; j + M i;k +…+ M l ;m
Theorem 1 (Necessity condition of Boundedness)
• Multiplicative : M(P) = M i ; j x M i;k x…x M l ;m
If a routing metric is chosen for any network N, and if N
has maximal spanning tree with respect to the metric, then the
140 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
routing metric is bounded Mob QOLSR
To the best of our knowledge, this work is amongst the first
Theorem 2 (Necessity condition of Monotonicity) efforts to consider nodes with mobility and energy
If a routing metric is chosen for any network N, and if N constraints in Manets.
has maximal spanning tree with respect to the metric, then the
routing metric is monotonic B. Proposed criterion
Theorem 3 (sufficiency of Boundedness and Monotonicity) Our goal is to select the metric to maximize network
If a routing metric is chosen for any network N, and if N throughput taking into account taking into account the key
has maximal spanning tree with respect to the metric, then the constraints of MANET environment (mobility, energy). The
routing metric is monotonic. idea behind the composite metric is that a cost function is
computed locally at each node during the topology
IV. OUR IMPROVEMENT information dissemination during the flooding process.
Once the network converges, each node runs a shortest path
A. Presentation of the solution algorithm based on the calculated composite metric to find the
Our solution can be summarized as follows. Bandwidth is optimal route to the destination. An underlying implication of
one of the most important factors required and requested by this is that each node should also be able to measure or gather
customer’s applications. Mobility and energy are crucial the information required. Bandwidth, mobility and remaining
problem in MANETs, and up to now, the majority of routing energy information’s are available and could simply be
protocols have shown some weaknesses to face a high mobility gathered from lower layers. This paper is mainly focused on
and poor energy resources in the network. solving the routing issues based on the assumption that an
Our objective consists in positively manage the network underlying mechanism is there to gather the necessary
bandwidth taking into account the constraints of energy and information about the individual metrics.
mobility, in order to adapt and improve the performance of We suggest the simple solutions already proposed in [7] can
manet routing protocol and prolog network life time. be used to get bandwidth. Mobility estimation will be based on
Initially, we start by giving the results of comparing our our lightweight proposed mobility measure cited [4-6] due to
approach based solely on mobility parameter. Thus we its simplicity and lightweight. Energy information is derived
evaluate the modified OLSR (Mob-OLSR) that uses our from the energy model used in NS2 simulator at MAC Layer
proposed mobility metric [4]. Mob-OLSR is then compared to [4].
the standard version of the OLSR protocol (without QoS Individual metrics must be combined according to the
extension) and QOLSR (The well known OLSR QoS following dependencies:
extension for Manets). • Nodes with no energy must be rejected in the process of
Simulations results conduct us to think to use mobility route discovery and maintenance
parameters to fulfil QoS requirements. So, we focus on • Nodes with a high degree of mobility should be avoided
maximizing the bandwidth based on the parameters of in the process of routes construction.
mobility. In this regard, two metrics are proposed. The first is • Tolerate a slight decrease in throughput in order to
based on the sum criteria and the second is based on the maximize other performance parameters (delay,
product criteria. collisions, NRL)
We have processed in the performance comparison between • Nodes start with a maximum energy and bandwith
the OLSR protocol using the MPR standard algorithm, and ressources. The residual energy decreases over time
the two modified OLSR protocols: SUM-OLSR and PRD- depending on node’s states (transmitting/receiving, in
OLSR protocols. The SUM-OLSR protocol is related to the idle/transition mode, etc.).
sum criteria, and the PRD-OLSR protocol is related to the Based on these results, the proposed relationship for the
product criteria. By the end we have eliminated the sum composite metric is given below:
criteria for his hard cost in terms of PDR (comparing to
product critéria). However, it is important to mention that the
eliminated criteria (the sum) also perform well comparing to
QOLSR protocol.
In a second step, and in order to maximize bandwidth while
taking into account the constraints of energy, a new
generalized metric is presented.
The proposed metric (EN-OLSR) will be compared to
different proposed metrics so called PRD-OLSR and OLSR-
141 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
starting from root, the metric is non-increasing. The metric
Where relation is given by: met {m,W(i,j)}.
BW : Available Bandwidth in kilobits per second
E : residual energy of node (number in range 0 to 5; 0 Given m is the metric of the root. It is evident that this
refers no energy for node to perform) meets the boundedness and that monotonicity conditions hold
for the selected metric. The available bandwidth is always
The constants K0, K1, K2, will be set by the administrator positive, hence for any node located at distance “d” from the
based on network nature. For example, in a very dynamic root W(i,j) would always be less than or equal to the metric
environment, and to give more importance to the mobility of value at the root. Since the bandwidth is always positive and
nodes, we can fix K0 to 0, K1 to 1 and K2 to 10. greater than zero hence it satisfies the boundedness and
Constant k3 is not null and is used to indicate if the monotonicity conditions.
environment takes into account the energy or not. Thus, an Mobility & energy:
important value of K3 indicates that energy is important in
the process of routing. The mobility metric represents the rate of changes in the
Composite metric scales Bandwidth metric with following neighbouring of a node at time t compared to the previous
calculations: state at time t − ∆t . In a previous work [18], We define the
BW = 106 / available bandwidth mobility degree of a mobile node i at a time t by the following
The proposed metric reflects a real dynamic environment formula:
where nodes have limited energy resources, and bandwidth NodesOut( t ) NodesIn( t )
M iλ ( t ) = λ + (1 − λ ) (7)
constraints are crucial (streaming application). The idea Nodes( t − ∆t ) Nodes( t )
behind the proposed metric is that in Manets environments, Where:
durable-stable link with optimal bandwidth should never be NodesIn( t ) : The number of nodes that joined the
omitted.
communication range of i during the interval [t − ∆t,t ] .
C. Proprieties of the proposed metrics NodesOut( t ) : The number of nodes that left the
communication range of i during the interval [t − ∆t,t ] .
In this subsection we prove that each of the individual
metrics satisfies the conditions of houndness and monotonicity Nodes( t ) : The number of nodes in the communication
conditions then we prove the proposed metric. range of i at time t.
λ : The mobility coefficient between 0 and 1 defined in
Node and Link Available Bandwidth: advance. For example, in an environment where the number
of entrants is large relative to the number of leavers, we can
The bandwidth metric represents the available bandwidth at encourage entrants taking λ = 0.25
the link. A simple technique proposed in [17], which Many simulations have been done for different values of λ
computes available bandwidth based on throughput can be ( λ =0, 0.25, 0.5, 0.75, 1). Simulation result [4] shows that for
used to measure the bandwidth on any given node (respect. λ =0.75 the network performs well (in term of delay, Packet
link L(i,j)). delivery fraction and throughput). For this reason, we consider
Available bandwith “ α ” for each node could be estimated λ =0,75 in the rest of this work.
by calculating the percentage of free time TL which is then
multiplied by the maximum capacity of the medium Cmax as Let Wij = M be the edge weight on the link L(i,j). The
L (i, j )
follows [17]: link mobility between two nodes A and B is defined as the
α = TL * C max (4) average mobility of the involved nodes (see Figure 4), as
Let Bav (i,j) represent available bandwidth of the link then, showed in following equation:
λ λ
M A (t ) + M B (t )
Bav (i , j) = min{Bav (i ); Bav ( j )}
λ
(5) M L ( A,B ) = (8)
2
Where Bav (i ) is the available bandwidth of the node i
Also let Wi,j be the edge weight on the link L(i,j). Wi,j can be
estimated from the following relationship given below.
1 Figure 4. Link mobility estimation example: M L ( A; B ) = 45%
Wi , j = (6) As node’s mobility reflects how likely it is to either corrupt
Bav (i, j )
or drop data. It could be considered as reliability metrics [15].
The condition of boundness implies that along any path
142 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Because the reliability metric is bounded and strictly By exchanging Hello messages, every node is aware of its
monotonic, it may be sequenced with the partial metric while neighbor nodes and can simply compute its Cost-to-Forward
preserving boundedness and monotonicity. value (i.e. to forward packet).
Moreover, residual energy function is monotonic and The Cost-to-Forward function (F(i)) for each of the four
bounded its value decreases (depending on the state of the models can be defined as shown in figure 6.
node: transmission / reception, transition/sleep mode,etc.) To motivate the nodes to reveal their exact Cost-to-Forward
from a maximum value (ex 200) to 0. it also reflects how value during the cluster head election and the MPR selection,
likely it is to either corrupt or drop data. Consequently, it can a reputation-based incentive mechanism using the VCG
be sequenced with the partial metric while preserving mechanism could be used [20]. The nodes participating
boundedness and monotonicity. truthfully to these processes see their reputation to increase.
Since the network services are offered according to the
Energy consumption parameters are derived from the reputation of the nodes, they would benefit to participate
energy model defined in NS2 [19] as follows : honestly.
Pt_consume= 1.320 (~ 3.2W drained for packet transmission);
V. SIMULATIONS AND RESULTS
Pr_consume= 0.8 (2.4W drained for reception); P_idle=0.07,
P_sleep =06; P_transition=0.5 In this section we have compared the performance of the
original OLSR protocol based on the MPR selection standard
The edge weight E ij for the link L(i,j) (see figure 5) can be algorithm, and the two modified OLSR protocols related to
estimated from the following relationship: different proposed model: : bandwidth model (QOLSR) ,
Eij = Min ( E i , E j ) . mobility model (MobOLSR), sum_bandwidth-mobility Model
(Sum-OLSR), prd_bandwidth-mobility model (prd-OLSR)
Where Ei : the remaining energy for the node i and Ei =0 and bandwidth-energy-mobility model(EN-OLSR).
means that the node i have drained out its energy. Thus,
routing protocol should omit such node in the process of
learning routes. A. Performance metrics
For comparison process, we have used the most important
metrics for evaluating performance of MANET routing
protocols during simulation. These considered metrics are:
Figure 5. Link energy estimation example: E L (i ; j ) = 200 Normalized Routing Overhead (NRL): It represents the ratio
of the control packets number propagated by every node in the
To validate the robustness and efficiency of the proposed
network to the data packets number received by the
Metrics , we use four models: bandwidth model , mobility
destination nodes. This metric reflect the efficiency of the
model, sum_bandwidth-mobility Model, prd_bandwidth-
implemented routing protocols in the network.
mobility model and bandwidth-energy-mobility model.
Packet Delivery Fraction (PDF): This is a total number of
delivered data packets divided by total number of data packets
transmitted by all nodes. This performance metric will give us
an idea of how well the protocol is performing in terms of
packet delivery by using different traffic models.
Average End-to-End delay (Avg-End-to-End): This is the
average time delay for data packets from the source node to
the destination node. This metric is calculated by subtracting
(9)
”time at which first packet was transmitted by source” from
”time at which first data packet arrived to destination”. This
(10) includes all possible delays caused by buffering during route
discovery latency, queuing at the interface queue,
retransmission delays at the MAC layer, propagation and
transfer times.
(11)
Collision: It represents the number of interfered packets
during simulation time. It occurs when two or more stations
Figure 6: the proposed metrics for QoS attempt to transmit a packet across the network at the same
time. This is a common phenomenon in a shared medium .
Metrics serves as Cost-to-Forward function. In OLSR, Packet collisions can result in the loss of packet integrity or
metrics will be used as criterion in MPR selection algorithm. can impede the performance of a network
143 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
Avg_throughput: is the average rate of successful message ensure a good enhancement in terms of delay when compared
delivery over a communication channel. Throughput and to the original OLSR protocol for all maximum speeds.
quality-of-service (QoS) over multi-cell environments are two Precisely, the QOLSR and MobOLSR protocols delay is
of the most challenging issues that must be addressed when around 1.25 seconds (enhancement by 0.4s comparing to he
developing next generation wireless network standards original OLSR) with higher mobility rate (maximum speed
equal to 140km/h) and decreases to almost 1.25 seconds
(enhancement by 0.1sec comparing to he original OLSR) with
B. Simulation environment
static topology conditions.
For simulating the original OLSR protocol and the modified For the original OLSR protocol the delay gets more than
OLSR protocols related to our proposed criterions, we have twice as large being almost 2.1 sec for high mobility and
used the OLSR protocol implementation which runs in version surprisingly increasing to over 1.4 seconds when the mobility
2.33 of Network Simulator NS2 [19-22]. is decreased.
We use a network consisting of 50 mobile nodes to simulate For the intermediate speed (from 40m/s to 100m/s) a
a high-density network. These nodes are randomly moved in lightweight difference between MobLSR and QOLSR is
an area of 800m by 600m according to the Random Waypoint noticed (enhancement by 0.1sec for MobOLSR when
(RWP) mobility model [21]. Moreover, to simulate a high compared to QOLSR for maximum speeds (0m/s and 30m/s))
dynamic environment (the worst case), we have consider the . This allows us to conclude that MobOLSR performs well
RWP mobility model with a pause time equal to 0. nodes can than QOLSR for intermediate speed. According to the
move arbitrarily with a maximum velocity of 140km/h. All Figure7-b, the original OLSR and MobOLSR protocols ensure
simulations run for 100s. in the whole the same packet delivery fraction for all
A random distributed CBR (Constant Bit Rate) traffic maximum speeds with a slight improvement for the original
model is used which allows every node in the network to be a OLSR for all maximum speed.
potential traffic source and destination. The CBR packet size
is fixed at 512 bytes. The application agent is sending at a rate OLSR
QOLSR De lay
of 10 packets per second whenever a connection is made. All MOBOLSR
peer to peer connections are started at times uniformly 2.5
distributed between 5s and 90s seconds. The total number of
connections and simulation time are 8 and 100s, respectively.
For each presented sample point, 40 random mobility 2
delay (s)
scenarios are generated. The simulation results are thereafter
statistically presented by the mean of the performance metrics.
This reduces the chances that the observations are dominated 1.5
by a certain scenario which favors one protocol over another.
As we are interested in the case of high mobility (i.e. high link
status and topology changes) we have reduced the HELLO 1
0 20 40 60 80 100
interval and TC interval at 0.5s and 3s, respectively, for quick pause time(s)
updates of the neighbors and topology data bases. Figure 7-a. Comparison of the three versions of the OLSR
protocol in term of delay.
C. Results and discussion Indeed, it can be seen that the number of packets dropped
To show how the modified versions of the OLSR protocol along the path is quite similar for all maximum speed being
are more adapted to the link status and topology changes approximately 45% at worst for the original OLSR and
comparing to the original OLSR protocol, we have made MobOLSR and 35% for QOLSR.
several performance comparison based on the five Moreover, the ratio is worse for a continuously changing
performance metrics cited in Section 5-A. Moreover, with the network (i.e. high maximum speed) than for the static path
supposed configuration cited above, we have run simulations conditions, because the number of link failures grows along
in different mobility levels by varying maximum speed of with the mobility. However, it is interesting to notice that even
nodes between 0km/h (no mobility) to 140km/h (very high with static topology conditions, sending nodes do not achieve
mobility) in steps of 10km/h. To maximize performances we 100% packet delivery but only 85%-89%. This clearly shows
have chosen the mobility coefficient equal to λ =0.75. the impact of the network congestion and packet interference
as the load on the network increases. Moreover, when
a) Comparing MobOLSR to OLSR and QOLSR comparing MobOLSR and original OLSR to QOLSR, QOLSR
Figure 7-a shows that Mob-OLSR and QOLSR protocols protocol presents a remarkable degradation in PDF for all
144 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 1, April 2010
maximum speeds. This is because QOLSR does not take into maximum speed.
account the state of links in MPR selection process. In
Figure7-d illustrates the normalized routing load (NRL)
summary, we can say that the MobOLSR protocol is more
introduced into the network for the three versions of OLSR
adapted to all levels of mobility from 0m/s (no mobility) to
protocol, where the number of routing packets is normalized
40m/s (very high mobility).
against sent data packets. A fairly stable normalized control
OLSR
QOLSR Pdf message overhead would be a desirable property when
MOBOLSR
considering the performance as it would indicate that the
actual control overhead increases linearly with maximum
90
speed of nodes due to the number of messages needed to
establish and maintain connection. The original OLSR
75
protocol and MobOLSR protocol produces the lowest amount
of NRL when compared to QOLSR protocol during all
rate (%)
60
maximum speed values. Moreover, original OLSR and
MobOLSR protocol produce the sme routing load for all the
45
maximum speed.
30
In the worst case (at the maximum speed value equal to
0 20 40 60 80 100 40m/s), the NRL increases to 2.1% for QOLSR protocol and
pause time
1.3% for the original OLSR. Precisely, comparing to QOLSR
Figure 7-b. Comparison of the three versions of the OLSR protocol, the MobOLSR and original OLSR protocols produce
protocol in terms of packet delivery fraction. twice less routing packets. This explains that our proposed
Figure7-c, shows the average throughput for the three criterion based on mobility parameter request less routing
version of protocols. The original OLSR and MobOLSR packets to establish and maintain routes in the network.
protocols ensure in the whole the same average throughput
OLSR
for all maximum speeds being approximately 125 kbps at QOLSR NRL
MOBOLSR
worst.
2.5
The ratio is worse for a continuously changing network
2
than for the static conditions. Moreover, it is interesting to
notice that even with static topology conditions, the network
1.5
average throughput does not reach the channel capacity
rate (%)
(5Mbps) but only 230 kbps. This clearly shows the impact of 1
the network congestion and packet interference as the load on
the network increases. 0.5
OLSR
QOLSR Avg Throughput 0
MOBOLSR 0 20 40 60 80 100
pause time (s)
Figure 7-d. Comparison of the three versions of the OLSR protocol in term of
230 NRL
Collision is a common phenomenon in a shared medium.
200
Packet collisions can result in the loss of packet integrity or