Mathematical Modelling of Torus Networks under Bursty Traffic

Many performance models for interconnection networks in parallel computing systems have been reported under the assumption of non-bursty Poisson arrival process. However, network traffic loads often exhibit bursty nature, which can markedly degrade the system performance. In order to develop a cost-efficient performance evaluation tool, this paper proposes a new analytical model for torus networks in the presence of bursty and batch arrival traffic modelled by the Compound Poisson Process (CPP) with geometrically distributed batch sizes. This study presents a novel approach to calculate the average degree of virtual channel multiplexing by virtue of a Markovian chain. The analytical results are validated through extensive simulation experiments. Index Terms Multicomputers, Interconnection Networks, Virtual Channels, Performance Evaluation