Statistical performance modeling: Case study of the NPB 2.1 results

Abstract

With the results of version 2.1 a consistent set of performance measurements of the NAS Parallel Benchmarks (NPB) are available. Unchanged portable MPI code was used for this set of 269 single measurements. In this study we investigate how this amount of information can be condensed. We present a methodology for analyzing performance data not requiring detailed knowledge of the codes. For this we study several different generic timing models and fit the reported data. We show that with a joint timing model for all codes and all systems the data can be fitted reasonable well. The timing model also contains only a minimal set of free parameters. This method is usable in all cases where the analysis of results from complex application code benchmarks is necessary.