|
Abstract : |
TCP congestion control [9] is designed for network stability, robustness and opportunistic use of network buffer and bandwidth resources on an end-to-end per-connection basis. Upon detecting packet loss, TCP infers congestion and trades off per-user goodput for network stability. Specifically, TCP throughput can be approximated by a function which is inversely proportional to the round trip time, the timeout delays and the square root of loss probability [16]. While the use of packet loss as an indicator of congestion is a robust technique, packet loss itself has a profound effect on performance-- especially in terms of the variance in goodput seen by individual connections. This "fairness " problem also results in what is commonly known as the "World Wide Wait " experienced by a majority of interactive Internet applications such as WWW or ftp. Another auxiliary problem in TCP congestion control is the lack of control over bottleneck queueing delay due to the end-to-end nature of control. In this paper, we evaluate three proposed solutions for these problems- an improved drop scheme (RED), a bit-based explicit congestion notification scheme (ECN) and a scheme which explicitly and transparently controls TCP rate (Packeteer TCP rate control). Our studies indicate marked improvements in fairness as we move from RED through ECN to TCP rate control. All schemes control bottleneck queueing delay, but trade off other measures such as drop rate, utilization and fairness, with TCP rate control exhibiting the best performance in terms of all metrics. In terms of deployment flexibility, TCP rate control and RED allow widespread and immediate deployment because they are transparent to hosts (ECN is not because it requires TCP protocol modifications). The minimal state requirements and protocol transparency of RED allows it a large deployment space. 1, |
