A novel method to estimate performance for a high performance computation workload

Abstract

Given the rapid change in processor architecture in the past years, there is a driving necessity to assess processor performance for a high performance computation workload. Assessing performance for a given workload is important to understand which architecture parameters the workload performance is sensitive to. A given workload can be categorized as memory bounded, compute bounded, or in between. In this paper we present performance sensitivity analysis for a high performance computation workload using LS-DYNA/car2car. We derive a sensitivity analysis for this workload with respect to different processor architecture parameters such as number for threads and memory. We also propose a performance estimation analytical model in which we can estimate performance for LS-DYNA workload by changing specific processor architecture parameters. The models is verified to estimate performance for different processor architectures with error margin <;5%.