Scalable configuration of RED queue parameters

Abstract

Active queue management policies and in particular random early drop (RED) are being pushed as additional mechanisms in Internet routers to control congestion and keep network utilization high. RED's performance is highly dependent on the settings of its control parameters. However, no firm guidelines exist on configuring RED parameters and the current suggestions fail to provide the desired performance scalability. We propose a mechanism to configure RED parameters based on evaluating the expected instantaneous length of a M/sup x//M/1/K queue. We show that by setting the RED parameters min/sub th/ and max/sub th/ to lie on either side of this expected queue length, we can ensure that the queue is not underutilized and flows cut their rates before the onset of congestion. This setting also allows the operating point to perform satisfactorily over a wide range of flow counts thereby allowing for a higher degree of scalability. We also show that our proposed mechanism increases the queue goodput, reduces losses and timeouts and increases the fairness when compared to existing guidelines. Our proposals have been verified using extensive simulations.