An improved supercomputer sorting benchmark

Abstract

In this paper we propose that the process of sorting be more formally adopted as a performance benchmark for commercial supercomputer applications. To this end we have investigated the use of entropy as a measure of data distribution and propose that it, along with larger datasets, be added to existing sorting benchmarks (such as NAS). Some of the key points in adopting such a benchmark are presented and the results of applying such a benchmark to the CM-5 supercomputer are discussed. As a result of carefully examining this problem, we were able to sort 1 billion 32-bit keys in less than 17 seconds on a 1024 processor CM-5.