The International Conference for High Performance Computing, Networking, Storage and Analysis
Creating a Framework for Systematic Benchmarking of High Performance Computing Systems.
Student: Mike Pozulp (College of William & Mary)
Supervisor: Tom Crockett (College of William & Mary)
Abstract: The Applications Performance & Productivity (APP) Group of the High-End Computing Capability (HECC) Project is tasked with ensuring maximal performance of the supercomputing systems operated at NASA Ames. As past of this task, one focus is to benchmark current and future systems so as to guide users to the most suitable resources for their applications. The challenges of benchmarking include: intractable combinations of parameters and system configurations, issues in organization and persistence of benchmarking results, and problems of determining benchmark performance or computational success. The benchmarking workflow has three essential steps: parameter selection, code execution, and result interpretation. A novel implementation of a semi-automated test framework has been utilized to analyze performance variation of the NAS Parallel Benchmark (NPB) applications running on the resident supercomputer, Pleiades. The framework has also been used to test the hypothesis that reductions in performance are correlated with the corresponding variation in network communication distance.