By Fernando Paganini, Steven H. Low, ZhikuiWang, Sanjeewa Athuraliya and John C. Doyle
We describe a new congestion avoidance system designed to maintain dynamic stability on networks of arbitrary delay, capacity, and topology. This is motivated by recent work showing the limited stability margins of TCP Reno/RED as delay or network capacity scale up. Based on earlier work establishing mathematical requirements for local stability, we develop new flow control laws that satisfy these conditions together with a certain degree of fairness in bandwidth allocation. When a congestion measure signal from links to sources is available, the system can satisfy also the key objectives of high utilization and emptying the network queues in equilibrium.
We develop a packet-level implementation of this protocol, where the congestion measure is communicated back to sources via random exponential marking of an ECN bit. We discuss parameter choices for the marking and estimation system, and demonstrate using ns-2 simulations the stability of the protocol, and the nearempty equilibrium queues, for a wide range of delays. Comparisons with the behavior of Reno /RED are provided. We also explore the situation where ECN is not used, and instead queueing delay is used as a pricing signal. This alternative protocol is also stable, but will inevitably exhibit nontrivial queues.