Asynchronous transfer mode traffic modelling and dimensioning using artificial neural networks

ATM (asynchronous transfer mode) networks use statistical multiplexing to i ncrease the utilization of network resources at the cell level. The ATM sta ndard provides comprehensive controls at the cell level to ensure accurate allocation; to achieve high overall ATM utilization levels, however, accura te resource allocation is also required at the call level. Call-level traff ic needs to be continually monitored and dimensioned to ensure that resourc es are not mis-allocated. In this paper an overall architecture for ATM dim ensioning, called NN-DIM, is introduced. The key prediction component is im plemented using neural networks, and produces predictions at minute and hou rly levels. The NN-DIM system assimilates these into network-dimensioning r ecommendations. The performance of the neural-network-based predictors is e valuated on real data. When compared with simple predictors it shows no per formance improvement at the minute level, but shows a significant improveme nt at hourly level where variation is greater. The paper concludes with an evaluation of the overall NN-PIM system. It is shown to have no quality of service violations while using a reasonable number of dimensioning interven tions. (C) 1999 Elsevier Science Ltd. All rights reserved.