Adaptive congestion protocol: A congestion control protocol with learning capability

Abstract

There is strong evidence that the current implementation of TCP will perform poorly in future high-speed networks. To address this problem many congestion control protocols have been proposed in literature which, however, fail to satisfy key design requirements of congestion control protocols, as these are outlined in the paper. In this work we develop an adaptive congestion protocol (ACP) which is shown to satisfy all the design requirements and thus outperform previous proposals. Extensive simulations indicate that the protocol is able to guide the network to a stable equilibrium which is characterized by max-min fairness, high-utilization, small queue sizes and no observable packet drops. In addition, it is found to be scalable with respect to changing bandwidths, delays and number of users utilizing the network. The protocol also exhibits nice transient properties such as smooth responses with no oscillations and fast convergence. In realistic traffic scenarios comprising of a small number of long flows and a large number of short flows, ACP outperforms both TCP and XCP, even in the presence of random packet losses. ACP does not require maintenance of per flow states within the network and utilizes an explicit multi-bit feedback signalling scheme. To maintain stability it implements at each link a novel estimation algorithm which estimates the number of flows utilizing the link. Using a simple network model, we show analytically the effectiveness of the estimation algorithm. We use the same model to generate phase portraits which demonstrate that the ACP protocol is stable for all delays.