FACTOR-ANALYSIS OF APPLICATIONS PERFORMANCE DATA FOR THE CRAY Y-MP

Factor analysis was applied to supercomputer applications performance data recorded by the hpm hardware performance monitor on the Gray Y-MP for a sample of 146 user codes. The analysis reduced the voluminous h pm data to five dominant factors which were used as performance metric s for applications code execution on the Y-MP. Case studies of Fortran kernels demonstrated that the proposed metrics accurately reflect cha nges in performance due to source code modifications. Therefore, the f actor-analysis model solution provides a supercomputer performance met ric that is both parsimonious and simple to interpret as an alternativ e to the complexities of hpm-derived performance attributes. Metrics s uch as these should prove to be of value in code optimization, user tr aining, and characterization of site-specific workloads.