Traffic modelling and design methodologies for broadband networks

Designers of modern high-speed complex networks, such as frame relay (FR) and ATM, face three principal areas of difficulty: the need to ensure accurate and efficient traffic modelling; the large number of combinations of control parameter settings that is available; and the rarity of some events which contribute to Quality of Service measures, resulting in a high cost of simulation. This paper shows how designers can evaluate and design such networks without making oversimplified assumptions. We review some of the relevant techniques and show how they can be used separately or in combination. We use the transform-expand-sample (TES) modelling technique to synthesize traffic based on measured autocorrelation and marginal probability distribution of the real traffic traces. For longer time-dependent models we use self-similar methods. For evaluating extremely low overflow probabilities for queues driven by bursty traffic, we show how we have applied importance sampling techniques to reduce simulation run times by several orders of magnitude. To cope with the large numbers of parameters, we use factorial experimental design to identify the most significant parameters and then to find near-optimal values for them by means of mean field annealing. Throughout the paper we give examples based on variable-bit-rate video traffic (MPEG-2 and H.261) in FR and ATM networks.