Congestion Control for Self‐Similar Network Traffic

Abstract

Analytical and empirical studies have shown that self-similar network traffic can have a detrimental impact on network performance including amplified queueing delay and packet loss rate. Given the ubiquity of scale-invariant hurstincss observed across diverse networking contexts, finding effective traffic control algorithms capable of delecting and managing self-similar traffic ha'> become an important problem. In this paper, we study congeslion control algorithms for improving throughput under self-similar traffic conditions. Although scale-invariant burstiness imposes a limit on the ability to conjointly achieve high quality of service (QoS) and utilization, the long-range dependence associated with self-similar traffic leaves open the possibility that the correlation structure may be exploited for performance enhancement purposes. We use a form of predictive congestion control called Selective Aggressiveness Control (SAC) to show that long-range dependence can be on-line detected and exploited to improve throughput. e show that the correlation structure present in long-range dependent traffic can be used to predict future traffic levels over time scales that are relevant to feedback congestion control. This on-line detected correlation structure is then used to drive a variable aggressiveness control which throttles the data rate upward if the future contention level is predicted to be low, being more aggressive the lower the predicted contention level. The predictive congestion control module runs on top of a generic rate-based linear increase/exponential decrease feedback control and is easily portable to other contexts. \Ve show that a significant improvement in throughput can be obtained if SAC is active and we show that the relative performance improvement increases as long-range dependence is increased. Lastly, we show that as the number of connections engaging in SAC increases both fairness and efficiency are preserved. 'Supported in part by NSF grant NCR-9714707. tConlact author; leI.: (765) 494-7821, fax.: (765) 494-0739.