Analyzing robust active queue management schemes: a comparative study of predictors and controllers

Active queue management (AQM) techniques are designed to detect incipient network congestion and proactively drop packets so as to avoid congestion later. Most AQM techniques use the exponentially weighted moving average (EWMA) of the queue length as a measure of congestion. More recent efforts, like adaptive virtual queue, have concentrated on rate-based packet marking. In this paper we analyze the robustness of a pure rate based packet marking scheme and compare it with our proactive AQM (PAQM) scheme. The PAQM proposed in this paper extends the calculation of the EWMA queue length by including a term that represents the future traffic intensity. To simplify the analysis, the AQM techniques were studied as a combination of a measurement module and a control module. Robustness, throughput and end-to-end delay jitter were chosen as performance metrics for various combinations of predictors and controllers. We have the following observations for a given network (i) performance of AQM schemes is sensitive to both the weight and interval of predicted future observations and prediction need not always perform better, (ii) the performance of more robust AQM schemes in terms goodput was found to be better than less robust ones.