On end-to-end congestion avoidance for TCP/IP

A TCP/IP network utilizes several congestion control schemes: end-to-end flow control and congestion avoidance, gateway congestion control, and explicit closed-loop feedback (i.e., source quench). The evolution of TCP/IP includes enhanced gateway congestion control algorithms (i.e., Random Early Detect) and a variety of incremental improvements to TCP including selective acknowledgement and possibly end-to-end congestion avoidance (i.e., TCP/Vegas). We focus on end-to-end congestion avoidance algorithms for TCP, specifically those algorithms that use change in packet transit times as an indicator of network congestion. TCP/Vegas is the most well known algorithm based on this form of congestion control. We find that TCP/Vegas does increase throughput primarily by avoiding time-outs. However its assessment of congestion is prone to significant error which can lead to increased queue levels at the bottleneck link. By studying TCP/Vegas and other algorithms, our goal is to understand the issues associated with end-to-end congestion avoidance schemes that monitor change in packet delays. This paper is organized as follows. First we introduce end-to-end congestion avoidance. Next, using simulation, we explore the various issues and challenges associated with end-to-end congestion avoidance by demonstrating and analyzing several end-to-end congestion avoidance algorithms. We conclude with a discussion of key issues associated with end-to;end congestion avoidance and identify future work.