A Vector-space Model for Parallel Workload Characterization

AbstractSoftware parallelism is a key factor in performance of parallel systems. In this paper we discuss a parallel-instruction vector space model for workload representation and comparison. This model will be compared with the parallelism-matrix technique, which is based on the Frobenius matrix norm. The latter compares two workloads based on identical parallel instructions only, whereas the former compares two workloads based on all parallel instructions. It will be shown that the parallel-instruction vector space method outperforms the parallelism-matrix method in time and space, as well as in accuracy. Further, it will be shown that this model provides a useful framework for the design and analysis of benchmarks. This will be demonstrated by analyzing some of the NASA/NAS Parallel Benchmark workloads and their performance measurements in the context of the model