|
|
MIT Electrical Engineering and Computer Science
EECS Event |
Monday, May 14, 2001
2:00 PM (refreshments 1:45)
Grier Room, Room 34-401A
EECS Special Seminar
Abstract
Title: Network Traffic Modeling and the Transmission Control Protocol
Abstract:
The discovery of the self-similar nature of network traffic has led to radical changes in traffic modeling and our understanding of network behavior. With the long-range dependence and bursty nature associated with self-similar traffic, network performance degrades significantly and traditional network design and management techniques become insufficient. Accurate characterization of network traffic and understanding the causes behind its self-similarity is thus of utmost importance to understand network dynamics and formulate steps to alleviate the undesirable effects of long-range dependence. In this talk we look at various aspects of network traffic modeling with a particular emphasis on TCP traffic. We first derive two convergence results which show the weak convergence of MMPP and FARIMA based long range dependent sources to fractional Brownian motion. We then model the steady-state throughput and latency of various versions of TCP under both correlated and independent losses. Using these derivations as the basis to model the packet transmission pattern of TCP sources, we then explore the effect of TCP on the self-similar and multifractal nature of network traffic. Our results show that TCP's congestion control mechanism, in particular its exponential backoff and slow start mechanism can lead to long range dependence and the degree of self-similarity is directly proportional to the loss rates experienced by the flow. We also develop a cascade based multifractal model for TCP traffic which relates network parameters like loss rates and round trip times to the scaling exponents. Finally, we look at the implications of this work and suggest ways through efficient buffering mechanisms to eliminate the long-range dependence in TCP traffic.