| The congestion control function, as that performed by the Internet Transport Protocol (TCP), is an integral function of any packet network. In this dissertation, we develop models of TCP congestion control algorithm which enable an analytical evaluation of its performance. Three models are developed, motivated by three different scenarios commonplace in today's packet networks. The first model considers a scenario where a (possibly mobile) wireless user accesses an infrastructure network (e.g. the Internet) through a last hop lossy link with memory. A realistic channel-driven model for TCP operation is derived for both types of links; memoryless and with memory. The second model considers the end-to-end performance of TCP flows with different round-trip delays over a bottleneck link. Two queue management schemes are considered; traditional Tail-Drop and recently proposed Active Queue Management (AQM) schemes like Random Early Detection (RED). The analysis is extended to a network of AQM routers with TCP and non-TCP compliant (e.g. UDP) flows. Finally, we propose a new router that provides preferential packet treatment according to the packet marking at the edge of the network, called DRED, and extend the analytical results to the case of a differentiated services network. A globally convergent numerical technique is proposed for the performance evaluation of all of these cases, which renders an average error of less than 5% as compared to those obtained using simulations. Finally, the third model is aimed at the end-to-end performance of TCP over an infrastructure-less network (e.g. an Ad-hoe network), characterized by random abrupt changes in the end-to-end path. Each of the proposed models provides a set of new analytical results and insights into the transport-layer congestion control. While this dissertation considers specific congestion control algorithms (e.g. Old-Tahoe, Tahoe, Reno and New-Reno versions of TCP and RED and DRED implementations of AQM) the methodology is applicable in general to the evaluation of congestion control, which is essential in enabling making informed choices for the protocol design in current and future packet networks. |
