Microscopic Behavior of Internet Congestion Control

Author: Wei, Xiaoliang (David)

Year: 2007

Degree: Dissertation (Ph.D.)

Advisor: Low, Steven H.

Committee Members: Low, Steven H.; Hickey, Jason J.; Doyle, John Comstock; Chandy, K. Mani; Cao, Pei

Option: Computer Science

DOI: 10.7907/W5E3-9N04

Abstract

The Internet research community has focused on the macroscopic behavior of Transmission Control Protocol (TCP) and overlooked its microscopic behavior for years. This thesis studies the microscopic behavior of TCP and its effects on performance. We go into the packet-level details of TCP control algorithms and explore the behavior in short time scales within one round-trip time. We find that the burstiness effects in such small time scales have significant impacts on both delay-based TCP and loss-based TCP.

For delay-based TCP algorithms, the micro-burst leads to much faster queue convergence than what the traditional macroscopic models predict. With such fast queue convergence, some delay-based congestion control algorithms are much more stable in reality than in the analytical results from existing macroscopic models. This observation allows us to design more responsive yet stable algorithm which would otherwise be impossible.

For loss-based TCP algorithms, the sub-RTT burstiness in TCP packet transmission process has significant impacts on the loss synchronization rate, an important parameter which affects the efficiency, fairness and convergence of loss-based TCP congestion control algorithms.

Our findings explain several long-standing controversial problems and have inspired new algorithms that achieve better TCP performance.

Files