Extended Analysis of Binary Adjustment Algorithms

Congestion control in the Internet relies on binary adjustment algorithms. For example, Transmission Control Protocol (TCP) in its congestion avoidance mode behaves similarly to Additive-Increase MultiplicativeDecrease (AIMD) algorithm. Chiu and Jain offer a theoretical justification for choosing AIMD: among stable linear algorithms, AIMD ensures the quickest convergence to maxmin-fair states. Whereas Chiu-Jain model rests on a well-known unrealistic assumption of uniform feedback, more precise analytic characterizations of TCP behavior are developed and validated. In particular, the advanced theory and experiments agree that TCP congestion control does not converge to maxmin fairness. However, despite the recent progress in TCP feedback modeling, it is still common to use Chiu-Jain model for comparison of binary adjustment algorithms. This paper argues against such practice. We provide evidence that due to the incorrect assumption of uniform feedback, Chiu-Jain model is not suitable for trustworthy conclusions about properties of an adjustment algorithm. We emphasize that until algorithms are analyzed with a more realistic feedback model, optimal choice of a binary adjustment algorithm will remain an open problem.