Adaptive multivariate regression for advanced memory system evaluation: application and experience

Recent advances in latency hiding techniques have made performance evaluation of memory hierarchies a more difficult task. Applications compiled for a particular architecture may be executed on vastly different memory hierarchy implementations. There is a need for performance analysis techniques that provide methods for understanding the interaction between applications and a given memory hierarchy. In this paper, we present a statistical approach to performance analysis of advanced memory hierarchy implementations. The method involves the utilization of previously available statistical analysis techniques coupled with scalability analysis. The result is a novel step-wise approach to understanding the hierarchical memory performance of scientific applications. We apply the method to several scientific applications of interest to the accelerated strategic computing initiative (ASCI) over the SGI machines PowerChallenge and Origin 2000. Results indicate some codes are statistically identical in memory performance, while others vary greatly. Furthermore, some codes do not take advantage of the performance enhancements to the memory system found in the Origin 2000.