A Self-Tuning Method for Router Throughput Shaping

A great deal of research has been conducted over the last several years into optimizing the performance of Internet communications by way of controlling the size of router queues for packets waiting for transmission over multiplexed links. The optimization algorithm is invariably framed in terms of either minimizing the number of queued packets waiting for transmission, maximizing the outgoing link utilization, or both. More recently, a separate group of researchers have put forward a short flow/long flow classification architecture, arguing in effect that latency is the key criterion which should be optimized. This note puts forward a proposal that effectively merges these separate schools of thought. It describes a self-tuning algorithm called SPRED (Short Packet Random Early Drop) for router queue management, which is demonstrated to be able to optimize the flow of short packets which require low latency, whilst maintaining bulk flows using long packets. The key advantage of this approach is that preference may be given to “real-time” voice and video flows. An attractive attribute is that the algorithm is adaptive, and is able to adjust to varying traffic conditions without manual intervention.