On improving the fairness of TCP congestion avoidance

We report on a simulation-based study of the efficacy of modified congestion avoidance algorithms on TCP fairness. We first examined whether a change to that algorithm, as proposed by Floyd (1991), that results in a "constant-rate" additive increase policy can correct the bias against connections with long round trip times (RTTs). Our first main result is that although the constant-rate policy can improve fairness dramatically, we found it difficult to successfully implement such a policy in a heterogeneous network. Next, we explored whether TCP connections with long RTTs could independently improve their own throughputs by becoming slightly more aggressive during the additive increase phase of congestion avoidance. Our second main result is that the judicious use of such a policy is beneficial for long RTT connections, without significantly affecting the performance of peer connections along the same path, thereby improving our metric of network fairness without compromising the utilization of the bottleneck link.