| The TCP congestion control role in the Internet at large is intimately associated with understanding the underlying end-to-end path characteristics context (bottleneck rate, buffer sizes, end-to-end delay, loss rate) on which the algorithm perform. This dissertation focus on the study of new path characteristics estimation algorithms that can be easily used inside TCP itself in a passive and pervasive manner. These algorithms provide further transparency to user applications about the path context, and it make possible the enhancement of network throughput performance without badly impacting standard TCP NewReno. In addition, they can be used by network administrators to formulate a big picture of existing network constraints. The dissertation proposes and studies several techniques to estimate end-to-end capacity, maximum storage buffer size and discriminate the cause of loss, from within every TCP socket. In particular, data mining can be achieved from a TCP sender side (TCP Probe), a TCP receiver side (TCP Receiver Probe) or under NIC interrupt moderation (TCP IC-Aware Probe). Our studies are complemented with mathematical analysis, extensive simulation using state-of-the-art tools and actual Internet measurements that show a good match between theory and real-world. As the consistency of the estimators is shown in practice, we apply them on the classical congestion control problem using two distinct schemes. The first one called CCP, a delay-based approach, allows a stable throughput regardless of session characteristics due to its innovative nature of ratet = f(storage), instead of rate t = f(ratet -1). The second scheme, TCP-Adaptive Westwood is a hybrid delay and rate-based congestion control algorithm, that expands on TCP Westwood research, and obtains the best tradeoff of higher utilization with minimal harm to standard TCP Newreno, when compared to other leading proposals. Finally, we delve into other applications, like PathCrawler, that makes heavy use of estimator techniques to support network management tasks, e.g. discovering the underlying infra-structure capacity to a set of web-servers. |
