Using Computational Similarity to Analyze the Performance Data of the NAS Parallel Benchmarks

One of the major problems in using complex application benchmarks is the di culty to relate the measured timing information to basic properties of the hardware and of the application used. In this paper we extend the concept of computational similarity invented by Roger Hockney to analyze complex benchmarks. We demonstrate this method for the published performance data of the NAS Parallel Benchmarks. We nd that it is only to a very limited amount possible to get information about the details of the timing relations. For a detailed study, access to the actual implementation from the di erent vendors would be necessary to get analytical timing relations.