Statistical approaches for congestion control in computer networks .

摘要

Active Queue Management (AQM) aims at minimizing queuing delay while maximizing the bottleneck link throughput. The methods described in this dissertation can be exploited to develop improved AQM mechanisms. Each method provides features that can be combined with other mechanisms to improve or modify the behavior of AQM schemes. Combined, these methods can minimize the variability of the queue, thus allowing small queuing delays and small buffer sizes, while maintaining high link utilization.;One of the methods exploits the statistical detection theory to develop detection mechanisms that can further enhance current AQM methods. The detection of congestion is performed using a Maximum Likelihood Ratio Test (MLRT), an asymptotically powerful unbiased test, which indicates that the likelihood of congestion grows exponentially with the queue occupancy level. The mathematical expression of the likelihood of congestion depends on parameters that can be easily measured or estimated.;Another principle indicates that the statistical characteristics of packet markings provide a performance bound of AQM in relation to the variance of the queue occupancy. The analysis thus shows a limitation of using traditional probabilistic marking. To overcome this limitation, a simple marking strategy based on the concept of error diffusion is proposed. Compared to probabilistic drops, it is found that error diffusion markings can significantly reduce the variance of the queue occupancy.;The utility of these fundamental principles is demonstrated in an AQM mechanism which, as shown in this dissertation, results in lower router queue occupancy, higher bottleneck link utilization, lower drop rate, and faster reaction to changes in the network state than other AQM schemes.