Performance analysis of an ATM cell multiplexer with MMPP input and a neural connection admission control approach

ATM will be used to implement the BISDN. Because the cells of individual calls are combined in a statistical fashion, congestion may occur at multiplexers and switches in ATM networks. Understanding the queueing behaviour and the cell loss performance of ATM networks is very important for switch design, buffer dimensioning, and developing congestion control strategies. We study the performance of a finite-buffered statistical multiplexer with MMPP input. An exact analysis of the MMPP/D/1/K queue is carried out, yielding the loss probability. A supervised neural network is used for QOS estimation in admission control schemes so as to comply with ATM real time requirements. The inputs to the neural network are link statistics that are computationally easy to obtain and yield a good description of the aggregate link traffic. The neural network target QOS is derived by an accurate MMPP/D/1/K model to allow for a high utilization of network resources.