Performance degradation of TCP Vegas in asymmetric networks and its remedies

TCP Vegas employs congestion avoidance, early detection of packet loss and conservative slow-start algorithms to improve TCP performance. With its proactive congestion detection, it utilizes network bandwidth more efficiently and achieves a higher throughput than TCP Reno. This paper shows that in asymmetric networks in which the bottleneck is on the reverse path rather than on the forward path, its performance can be significantly lower than that of Reno, in contrast to the 37% throughput improvement claimed in [4][5]. In particular, Vegas may erroneously converge to an operating region in A which the available bandwidth on the forward path is under-utilized by a large margin. Even worse, when connections running Vegas and Reno co-exist and compete on the same network, Vegas suffers a severe penalty. We propose an approach to improve Vegas' throughput in asymmetric networks.