A simulation study on the throughput fairness of TCP Vegas

Abstract

Throughput fairness is an important criterion for evaluating TCP performance. Fairness is especially important for best effort service, which is the dominant type of service in the Internet still and, predictably, in the years to come. However, the TCP protocols prevailing in the Internet, including TCP Tahoe and TCP Reno, are known to be unfair, especially to connections with larger round-trip delays. Using the ns simulator, we have throughly examined the fairness of TCP Vegas focusing on three issues: (1) is TCP Vegas really fair to connections with larger propagation delays? (2) what is the impact on fairness of the thresholds, /spl alpha/ and /spl beta/, used in TCP Vegas' congestion avoidance algorithm? (3) when there is a mixture of TCP Vegas and TCP Reno connections, are TCP Vegas and TCP Reno fair to each other? The simulation results support that TCP Vegas is still unfair to connections with larger propagation delays, for example, when /spl alpha/ = 1 and /spl beta/ = 3. However, unlike TCP Reno, the delay bias does not necessarily increase as the delay difference increases. The unfairness problem can be resolved by an enhanced TCP Vegas that sets /spl alpha/ = /spl beta/ = 2 or 3 but not 1. When /spl alpha/ = /spl beta/ = 1, the fairness is unstable and may be worse than that when /spl alpha/ = 1 and /spl beta/ = 3. Considering a trade-off among fairness, stability and aggressiveness, a value of 3 seems to be an acceptably good choice. Finally, fairness between TCP Reno and TCP Vegas connections depends on the RED (random early detection) gateway thresholds, the number of active flows and TCP Vegas parameters /spl alpha/ and /spl beta/.