Improving Bandwidth Utilization and Fairness between TCP Flows based on a Machine-learning Approach

Abstract

TCP, a current de facto standard transport-layer protocol of the Internet, cannot fully utilize the available bandwidth. Fairness between TCP flows is another important measure of TCP performance. We proposed a method for predicting the optimal size of the congestion window to avoid network congestion by using a machine learning approach. In this paper, based on the machine learning approach, we further improve the congestion algorithm with respect to utilization of the available bandwidth and fairness between TCP flows. The improvement includes bringing a size of the congestion windows closer to the optimum value, realizing fairness against congestion algorithms that aggressively use bandwidth, and adapting to the network where the available bandwidth abruptly changes. The proposed method is evaluated with respect to utilization of bandwidth and fairness between TCP flows including flows aggressively using bandwidth by simulation using NS-2.