Smoothing the TCP Rate by Learning the Delay versus Window Size Dependency

Abstract

We propose TCP-L, an improved version of TCP, equipped with a learning algorithm whose purpose is to avoid probing for additional bandwidth when the network conditions are known to be unfavourable. TCP-L learns the relationship between its current (average) one-trip delay and its current window size when congestion occurs, leading to packet loss. After the learning phase, TCP-L will only probe for bandwidth by increasing its window if, under the current network conditions (measured by the one-trip delay), this inflated window has not previously created congestion. Simulations show that after the learning phase, TCP-L reaches a much more stable throughput, while remaining TCP-friendly, which makes it usable for a larger class of applications, including some multimedia applications that will benefit from that stability. TCP-L is a simple backward compatible extension of TCP which can thus be deployed progressively. We show that there is a benefit for the Internet to deploy TCP-L, because the overall traffic becomes smoother when the proportion of TCP-L flows increases. Finally, our learning component can also be easily embedded in other unicast or multicast transport protocols.