Load Test: Benchmarking Throughput

Abstract

Acquisition of the best high performance computers (HPCs) for the High Performance Computing Modernization Program’s (HPCMP) annual technology insertion (TI) cycle is guided by estimating the performance of the new machines based upon the computational technical areas (CTAs) for which they will be used. For each CTA, one or more codes are used to estimate the aggregate load across all HPC systems in the HPCMP. The primary aim is to purchase equipment that is best suited for each of the CTAs. To this end, each code is benchmarked on every major system in the HPCMP using standard and large test cases that represent the anticipated computational load. The scalability and efficiency of the equipment are determined and help to guide the prospective purchases. The overall goal of each shared resource center is to provide quality service while executing as many HPC tasks as possible. The job mix varies from small jobs requiring, at most, one node and short execution times, to large jobs requiring significant portions of entire systems, including large amounts of memory, dedicated to their execution. Therefore, the load test was devised to benchmark the throughput of existing and proposed HPC systems using a mix of jobs that represents the anticipated usage by the CTAs. This paper will present a brief description of the load test, including the job mix by CTA and the representative codes of each. The number of times each job is executed at each processor count will be described. Data from a run of the load test conducted at the U.S. Army Engineer Research and Development Center Major Shared Resource Center will be used to demonstrate the benchmark.