On processor sharing as a model for TCP controlled HTTP-like transfers

Abstract

We explore the possibility of using the well known processor sharing (PS) model for predicting the throughput performance of TCP controlled HTTP-like transfers on a single bottleneck link. We compare two commonly used-performance measures for average session throughput. For the PS model we derive bounds for one of these. We then compute these measures for a PS queue and compare them with average session throughputs on a single bottleneck link carrying TCP controlled data traffic. Analysis and simulations are used to obtain these results. We find that for file size distributions with a finite second moment, the PS model predicts TCP throughputs quite well. For the Pareto file size distribution, however, the PS model overestimates the session throughputs, the error getting worse as the Pareto tail gets thicker. It is also seen that the PS model breaks down as the link propagation delay increases.