High Productivity Computing Systems Analysis and Performance

Abstract

Abstract : Due to its focus on performance, the Defense High Performance Computing (HPC) community has always had a need for performance measurement, evaluation, and comparison. Nevertheless, there have been no widely accepted standards for measuring the performance of HPC systems, much less their productivity when applied to the unique computational challenges facing Defense scientists and engineers as well as operational users. In coordination with MITRE and Lincoln Laboratory, the University of Southern California's Information Sciences Institute (ISI) led an effort to begin addressing this issue in support of DARPA's High Productivity Computing Systems (HPCS) program. This final report describes how ISI formed a team of experts that helped enable the HPCS program to determine Defense high performance computing needs. It then describes how ISI and its subcontractors, in a supplementary effort, initiated the HPCS phase two development time and execution time productivity activities. The outcome of this project was a new understanding of the various components of the productivity of Defense. HPC systems together with a methodology for measuring productivity from both the perspective of application developers as well as users of HPC systems. In addition, a new set of HPCS benchmarks were developed, including a set of discrete mathematics kernels.