Evaluation and application of available bandwidth estimation techniques to improve TCP performance

Abstract

Several tools are available to estimate the available bandwidth in an end-to-end path. They use either the packet gap model or the packet rate model whereby probe packets are sent and analyzed at the destination to infer the available bandwidth in the tight link. Available bandwidth estimations can be very useful in many different scenarios and applications. We use available bandwidth estimation techniques (ABET) for flow and congestion control in TCP. We first evaluate the performance of Pathload, IGI and pathChirp in terms of their accuracy, intrusiveness, and overhead. Using simulations, we perform a 2/sup k/ factorial design to analyze the importance of the packet size, number of trains, number of packets per train and frequency of run in these performance metrics. Then we utilize these results to embed the most appropriate tool with the best set of parameters in a modified version of TCP SACK to solve the "blindness" of TCP when changing its congestion window and threshold values. Our simulations show that, using the available bandwidth estimates provided by IGI instead of the "by half" reduction rule of TCP, the throughput of the proposed ABET-based TCP version is improved compared to the plain TCP SACK.