Internet Congestion Control: Equilibrium and Dynamics



Kevin Tang
Cornell, School of ECE

Monday  February 2, 2009
4:00 PM, 5130 Upson Hall



Given the important role the Internet plays in our society, a fundamental understanding of how this critical resource is managed becomes crucial. In this talk, we focus on Internet congestion control. The traditional method is loss-based, i.e., congestion is measured by packet loss. A delay-based approach, in which congestion is measured by round-trip delay, has recently received much attention. We report two new general results motivated by this development.

During the first part of the talk, we study the equilibrium of networks that are shared by both loss-based and delay-based protocols. The standard utility maximization framework breaks down in such a heterogeneous network. We develop a new theory that characterizes all major equilibrium properties such as existence, uniqueness, optimality, and stability. Surprisingly we show that, unlike a homogeneous network, a heterogeneous network can have multiple equilibrium points, and we propose a slow timescale control that steers the network to the unique optimal equilibrium. In the second part, we study the accurate dynamics of window based congestion control systems. By identifying a fundamental equation which captures both the self-clocking and the link integrator effects, we are able to match both open loop and closed loop packet level simulations quantitatively. The application of this discovery to the stability analysis of FAST TCP resolves a notable discrepancy between empirical observations and previous theoretical predictions.

Throughout this talk, simulations and experiments are used to illustrate and verify our theoretical results as well as to highlight the intimate interplay between theory and practice. Connections between engineering problems and related ones in economics and mathematics will also be discussed.