Bimodal Traffic for Computer Interconnects

Message sizes are currently not modeled in performance analysis of interconnection network components (ICNC) of parallel machines such as switches and network interface cards (NIC). However, this becomes important as the gap, between processor and chip I/O performance widens. It is therefore necessary to include size modeling into commonly known traffic sources. Our study shows that message length distributions can be assumed to be bimodal, and presents four parameters that allow to mathematically describe such a source. Finally, we present a straight-forward, easy configurable bimodal traffic source realized by a four-state discrete Markov chain. It can be used as an orthogonal refinement to temporal and spatial traffic distributions in the analysis of ICNC for developing realistic performance models.