S-Check: a tool for tuning parallel programs

Abstract

We present a novel tool, called S-Check, for identifying performance bottlenecks in parallel and networked programs. S-Check is a highly-automated sensitivity analysis tool for programs that extends benchmarking and conventional profiling. It predicts how refinements in parts of a program are going to affect performance by making focal changes in code efficiencies and correlating these against overall program performance. This analysis is a sophisticated comparison that catches interactions arising from shared resources or communication links. S-Check's performance assessment ranks code segments "bottleneck" according to their sensitivity to the code efficiency changes. This rank-ordered list serves as a guide for tuning applications. In practice, S-Check code analysis yields faster parallel programs. A case study compares and contrasts sensitivity analyses of the same program on different architectures and offers solutions for performance improvement. An initial implementation of S-Check runs on Silicon Graphics multiprocessors and IBM SP machines. Particulars of the underlying methodology are only sketched with main emphasis given to details of the tool S-Check and its use.