Free-Drop TCP

Abstract

We introduce a new class of TCP congestion control algorithms that take a non-standard approach: instead of modifying AIMD parameters or exploiting traffic measurements, the new protocols modify the rule for deciding when to cut the congestion window. The class is defined by an additional window with a packet-count parameter w; the congestion window is reduced by half when a packet loss is detected, at time t say, if and only if there has been at least one dropped packet in the last w packet transmissions prior to time t . An algorithm in the class is called Free-Drop TCP , since dropped packets are "free" (they do not cause cuts in the window size) unless they are sufficiently bursty. We propose this new class as a means to achieve high utilizations in high bandwidth-delay product networks with small buffers. We analyze a fluid model which leads to explicit estimates of the average throughput for small loss probabilities. We then give the results of experiments, which show that, relative to TCP, a family of 'shifted' response functions of the form O (1/√ p - ε) can be obtained over a wide range of p by suitably varying w . Potential costs of these increases in throughput are also examined in terms of coefficents of variation and Jain's fairness measure. The costs range from negligible to moderate.