| A congestion avoidance system aims to maintain a stable operation point of high-throughput and low-delay, improving the performance of both throughput-sensitive and delay-sensitive applications. Two fundamental components are needed in the network to enable congestion avoidance: A congestion detection mechanism and a congestion notification mechanism. The successful operation of a congestion avoidance system depends on the amount of feedback it creates under given traffic conditions. If the packet marking is too aggressive, then the throughput performance will degrade, and if the packet marking is too conservative, then congestion will not be avoided. However, traffic conditions, in particular the number of active TCP flows, change dramatically over time. The main contribution of this thesis is an adaptive active queue management algorithm that can maintain a network state of high-throughput and low-delay under changing traffic conditions. The algorithm is able to adapt the marking behavior when the number of TCP connections change, and therefore outperforms nonadaptive schemes like RED. This thesis also shows that in addition to the number of TCP connections, the heterogeneity in the network also affects the operating state of a congestion avoidance system. Multiple bottleneck points with different buffer management schemes, and the mixture of ECN-capable and ECN-incapable flows create challenges for maintaining the desired high-throughput/low-delay state. The proposed adaptive method provides significant advantages under these network dynamics also. |
