Scalable service differentiation using purely end-to-end mechanisms: features and limitations

We investigate schemes for achieving service differentiation via weighted end-to-end congestion control mechanisms within the framework of the additive increase/multiplicative decrease (AIMD) principle, and study their performance as instantiations of the TCP protocol. Our first approach considers a class of weighted AIMD algorithms. This approach does not scale well in practice because it leads to excessive loss for flows with large weights, thereby causing early timeouts and a reduction in throughput. Our second approach considers a class of loss adaptive weighted AIMD algorithms. This approach scales by an order of magnitude compared to the previous approach, but is more susceptible to short-term unfairness and is sensitive to the accuracy of loss estimates. We conclude that adapting the congestion control parameters to the loss characteristics is critical to scalable service differentiation; on the other hand, estimating loss characteristics using purely end-to-end mechanisms is an inherently difficult problem.