Abstract

We outline a methodology for the performance analysis of distributed admission control for adaptive applications, or long-lived flows, where each flow can transmit at one of a discrete set of rates, and can switch between these rates. Specific examples are certain voice over IP or conferencing applications. In the Internet, UDP based protocols such as RTP have the potential to allow sources to alter their sending rates using periodic feedback, where again the rate can take values from a limited set. In distributed admission control [1, 4, 3], the end-system or edge-device plays a key role in the admission decision: it probes the network with a small number of packets and decides to enter or not depending on the fate of these probe packets (where the network drops or marks packets to signal congestion). These packets are marked at resources (routers) if these resources are nearing congestion, and the probe packets are fed back to the source. The source then decides whether to enter, and what rate to choose, according to the fraction of probing packets that have been marked. In addition, the application can occasionally reprobe the network and alter its rate.