Correlating Workload Characteristics to Performance Metrics for theCray X-MP/Y-MP

Abstract

Workload characterization is essential for performance evaluation studies. For multiprocessor supercomputers, this characterization usually consists of program measurements from uniprocessor execution (e.g., average vector length, percentage vectorization, etc.). There is no consistent, quantitative correlation between such characteristics and performance metrics across different programs. We present a methodology for defining and measuring characterization parameters for both single and multi-processor executions. Using several production codes, we show for the Cray X-MP/Y-MP that these characterization parameters correlate consistently to observed performance metrics across different programs. Moreover, the correlation allows the identification of bottlenecks in system architecture that limit performance. Presentation slides from the Fourth SIAM Conference on Parallel Processing for Scientific Computing, December, 1989.