Cause-Effect Analysis of Parallel Program Performance

This paper describes a general framework and several specific techniques for cause-effect analysis: an automated inference process that presents explanations for dynamic phenomena of parallel program executions in terms of underlying causes and the related source code. We illustrate the framework by describing the implementation of three analysis techniques: waiting time analysis identifies the cause of synchronization overhead as the differences in execution paths taken by synchronizing processors; protocol analysis identifies the sharing patterns to pages that produce invalidations in a DSM protocol; and transaction analysis identifies conflicts between transactions in a parallel file system that cause aborts. We present examples of how each technique can be used to understand observed performance effects, and how insights derived from these techniques suggest program modifications that lead to improved performance. 1 Introduction Many performance phenomena arise during execution ..