TRAFFIC PREDICTION AND DYNAMIC BANDWIDTH ALLOCATION OVER ATM - A NEURAL-NETWORK APPROACH

ATM (Asynchronous Transfer Mode) has been recommended as the transport vehicle for Broadband Integrated Service Digital Networks (BISDN). Th e ATM technology offers a great flexibility of transmission bandwidth allocation to accommodate diverse demands of multimedia connections, w hich include data, voice, video, graphics and images. One major applic ation in ATM networks is to provide real-time, low-loss and minimum-de lay transmission of variable bit rate (VBR) traffic which is highly bu rsty, non-stationary and correlated. In this work we adopt neural netw ork methodology to predict VER traffic represented by a continuous aut oregressive (AR) Markov model. We have found that a simple 1-5-1 backp ropagation neural network can accurately predict the VER traffic. Base d on prediction results obtained from neural networks, we propose a dy namic bandwidth allocation scheme for ATM. The proposed scheme is simu lated under various traffic loads and buffer sizes. Its performance in terms of cell loss, cell delay and link utilization is examined and c ompared with two other bandwidth allocation schemes: the static averag e bandwidth allocation scheme, and the optimal (ideal) bandwidth alloc ation scheme. Experiments show that for most of the time, performance of the proposed dynamic bandwidth allocation is much better than that of the static average bandwidth allocation, and in many cases it is ve ry close to that of the ideal bandwidth allocation.