Internet Congestion Control, S.H. Low, F. Paganini, J.C. Doyle, 2002
In this article, Low et al review current implementations of TCP congestion control. The three protocol implementations (Tahoe, Reno, and Vegas) use different metrics (e.g. queuing delay and packet loss) to adjust the TCP sliding window size in an attempt to maximize throughput while avoiding congestion.
The paper formalizes the implementations, that were developed for the most part empirically, in terms of an economic model that should be optimized. In the new economic formulation, the behavior of the system is modeled at the link level. Paths are assigned prices
that are the dynamic cost of a source choosing a path and are the Lagrange multipliers for the constraint of the link flow capacity at a given state. The goal is to optimize the utility of a given source for prices of each path toward an equilibrium state. The equilibrium varies continuously (the capacities of paths change as paths are used) with traffic and more importantly packet size, as large packets saturate queues adversely impacting smaller packets in the pipeline with low latency requirements.
It is shown through analysis in the dual that the existing protocols do not perform well, particularly in cases of congestion, and do not scale. For example, in the case of TCP Reno, queue lengths and route selection have minimal impact on the performance of the network; the performance degrades with changes in the topology and the number of nodes regardless of network capacity. This is clearly undesirable.
The authors go on to show a distributed locally stable, scalable method, based on the analytical optimization approach. This method requires some additional global information, such as the presence of bottleneck links in a path and the price of a path, that can be reasonably implemented in existing networking hardware. The global parameters are used to set the local control laws, which are designed to ensure that sources choose optimal (uncongested) paths. Sources also regulate windows more gently, so as not to over-congest links and compromise global stability. The stability and efficacy of this approach is verified with packet-level numerical simulation.