Nonlinear Dynamics of TCP and its Implications to Network Performance

Abstract

TCP flows continuously intertwine in the Internet competing with each other for service capacity and buffer space in bottleneck routers. The window based flow and congestion control algorithms implemented in end-hosts control competition between traffic flows. Previous work has modeled TCP competition from a macroscopic point of view that is the average TCP flow rate have been derived as a function of delay, loss and service rate [9][15]. In this paper, we introduce a novel approach to model the competition between multiple TCP connections sharing a common bottleneck buffer. We demonstrate that the end-to-end congestion control used by the TCP protocol, while competing for networking resources, generates complex non-linear dynamics. An important message of this work is that random traffic behavior is not exclusively due to “random” effects, but also due to complex behavior of TCP. The TCP protocol, although driven by deterministic, rules may produce time-series seemingly indistinguishable from stochastic processes. On the other hand, non-linear systems have unique properties and they are able to produce a diversity of phenomena.