TCP performance over end-to-end rate control and stochastic available capacity

Abstract

Motivated by TCP over end-to-end ABR, we study the performance of adaptive window congestion control, when it operates over an explicit feedback rate-control mechanism, in a situation in which the bandwidth available to the elastic traffic is stochastically time varying. It is assumed that the sender and receiver of the adaptive window protocol are colocated with the rate-control endpoints. The objective of the study is to understand if the interaction of the rate-control loop and the window-control loop is beneficial for end-to-end throughput, and how the parameters of the problem (propagation delay, bottleneck buffers, and rate of variation of the available bottleneck bandwidth) affect the performance.The available bottleneck bandwidth is modeled as a two-state Markov chain. We develop an analysis that explicitly models the bottleneck buffers, the delayed explicit rate feedback, and TCP's adaptive window mechanism. The analysis, however, applies only when the variations in the available bandwidth occur over periods larger than the round-trip delay. For fast variations of the bottleneck bandwidth, we provide results from a simulation on a TCP testbed that uses Linux TCP code, and a simulation/emulation of the network model inside the Linux kernel.We find that, over end-to-end ABR, the performance of TCP improves significantly if the network bottleneck bandwidth variations are slow as compared to the round-trip propagation delay. Further, we find that TCP over ABR is relatively insensitive to bottleneck buffer size. These results are for a short-term average link capacity feedback at the ABR level (INSTCAP). We use the testbed to study EFFCAP feedback, which is motivated by the notion of the effective capacity of the bottleneck link. We find that EFFCAP feedback is adaptive to the rate of bandwidth variations at the bottleneck link, and thus yields good performance (as compared to INSTCAP) over a wide range of the rate of bottleneck bandwidth variation. Finally, we study if TCP over ABR, with EFFCAP feedback, provides throughput fairness even if the connections have different round-trip propagation delays.