The effect of reverse traffic on TCP congestion control algorithms

Abstract

In recent years, several new TCP congestion control algorithms have been proposed to speed up TCP over very fast networks. All these algorithms propose their own modifications of the increasing/decreasing phases of classic Reno/NewReno TCP. On the other hand, all of them preserve the self-clocking mechanism, which is a fundamental part of the classic Van Jacobson TCP congestion control. For this reason, it is also said that the TCP dynamics is ack-clocked. An important consequence of the ack-clocking is that a TCP flow can be heavily affected by the presence of reverse traffic provoking congestion along its ACK path. In this paper we use the ns-2 simulator to investigate the effect of reverse traffic on the behavior of the new TCP stacks: BIC TCP, HSTCP, Hamilton TCP (H-TCP), STCP, and FAST TCP. Westwood+ TCP is also considered even though this stack was not designed to address the special case of very fast networks and it employs the classic slow-start and congestion-avoidance probing phases of TCP. We consider reverse traffic of the same type as the forward traffic in order to study the behavior of each single protocol with respect to stability and fairness, which are fundamental properties to be investigated before other important issues such as inter-protocol friendliness. Simulation results reported in this paper reveal that new TCP stacks designed for fast networks experience goodputs comparable to those of standard TCP SACK or TCP Westwood+ but with larger packet retransmission probabilities and timeout events in the presence of reverse and web background traffic.