PMPS(3): a performance model of parallel systems

In this paper, an open performance model framework PMPS(n) and a realization of this framework PMPS(3), including memory, I/O and network, are presented and used to predict runtime of NPB benchmarks on P4 cluster. The experimental results demonstrates that PMPS(3) can work much better than PERC for I/O intensive applications, and can do as well as PERC for memory-intensive applications. Through further analysis, it is indicated that the results of the performance model can be influenced by the data correlations, control correlations and operation overlaps and which must be considered in the models to improve the prediction precision. The experimental results also showed that PMPS(n) be of great scalability.