Neural Network Estimation of TCP Performance

Abstract

TCP remains the protocol of choice for bulk data transfers over the Internet. A range of mathematical approaches were proposed to evaluate the performance of TCP, approaches validated through synthetic or endpoint controlled traffic, typically unsuitable for short-lived transfers or clients with unknown behaviour. This paper aims to overcome these problems by using a supervised adaptive learning approach to build the relationship between TCP performance and the influencing parameters. An earlier study indicated several advantages of the approach, as well as several issues, particularly related to the efficiency of the model on real traces. Comparison against the mathematical models showed that the proposed model provides more accurate estimates for real time traffic without losses, with tests results indicating that the average error of the connection duration, estimated using the proposed model, was 50% smaller than the value obtained using the mathematical approach.