We investigate the performance of the EFCI-based (explicit forward congestion indication) and ER-based (explicit rate; EPRCA in particular) algorithms for the rate-based flow control of the ABR (available bit rate) traffic in an ATM network. We consider the case of multiple switches in tandem. We present several definitions of a bottleneck, and provide conditions that determine which queue is the bottleneck. We show that it is not necessarily the queue with the slowest transmission rate that is the first to notify congestion, and that queues with faster transmission rates can increase considerably the congestion. Moreover, a queue may build up in a faster node before it builds in a slower one. We derive analytic formulas for the maximum queue length. We compare our results to those obtained by approximating a network by a simpler one, containing only the bottleneck switch. We show that the maximum queue lengths under the approximating approach may largely underestimate the ones obtained in the real network.

The best known congestion control mechanism for the TCP/IP protocols, is that proposed by Van Jacobson (1988) dubbed Tahoe. Revised two years after (Van Jacobson, 1990), by including fast recovery and fast retransmit (Stevens, 1994), the new version called Reno achieves more bandwidth utilization and retransmits less data than its predecessor. However its method of predicting the available bandwidth by provoking losses, is the main source of large averages and variances of rtt (round trip time) and also of aggressiveness (network saturation). Brakmo and Peterson (1995) proposed a new congestion avoidance algorithm dubbed Vegas based on some modifications to the source behavior of Reno. In this paper we use an analytic fluid approach in order to analyze the different features of both Vegas and Reno. We then use simulations to confirm our analytic results. When the available bandwidth is high, our results confirm the claims of Brakmo and Peterson (1995) and An et al. (1995); indeed Vegas can retransmit less than one-fifth as much data as Reno does, so that the higher the available bandwidth is, the more efficient Vegas is. However, under heavy congestion Vegas behaves like Reno and does not manage to make efficient use of its new mechanism for detection of congestion. The analytic results that we obtain are the evolution of the window size, round trip times and their averages, and the average throughput.