Model-based approach to TCP-friendly congestion control.

摘要

Continuous media (CM) applications such as streaming audio and video are being rapidly deployed throughout the Internet. These applications must be able to co-exist with each other and with traditional TCP-based applications.; One requirement for such co-existence is the implementation of congestion control mechanism by CM applications. TCP-based applications reduce their sending rate in face of network congestion. Many CM applications run on top of UDP, as TCP's error control mechanisms are ill suited for CM applications. UDP-based applications must provide their own congestion control.; In this dissertation, we work towards development of a congestion control protocol for CM applications that will facilitate co-existence of CM applications with each other, and with TCP flows. We address this problem in two steps.; First, we develop an analytical model of TCP. Such a model is essential to understand TCP's congestion control behavior under a variety of network conditions. Our model provides a closed form equation that can be used to calculate the send rate of a TCP connection, given the packet loss rate, the round trip time and the base timeout value. We verify the accuracy of our model against measurements taken in the Internet.; The next step is to develop a congestion control protocol that is suitable for the CM applications and also shares the bandwidth fairly with TCP connections that travel over the same network path. We develop a protocol that uses our TCP model as the control function of a feedback loop. The protocol design is deliberately kept simple to allow us to evaluate the feasibility of the model-based approach to congestion control. The simple protocol is evaluated using simulations and Internet experiments. Our results show that the protocol performs well under a variety of network conditions.; We then refine the simple protocol, trading off design simplicity to achieve better performance. We introduce sophisticated mechanisms to track packet loss rate, control rate increase and handle startup behavior. We evaluate the protocol using simulations and Internet experiments. Our results show that the more complex protocol provides a smooth sending rate and is TCP-friendly under a wide variety of networking conditions.