Trace Reduction for Performance Improvement Assessment of Message Passing Parallel Programs

摘要

This paper proposes a trace reduction method for assessing the improvability of the performance of message passing parallel programs. This assessment is based on a what-if prediction approach that forecasts future program performance, for example, the execution time if the target program is modified according to typical tuning techniques. Our method reduces the size of trace files by aggregating records of communications that do not change the predicted execution time. In order to avoid recording such useless information, our method automatically identifies them during program execution by comparing the occurrence time of sends and receives. In case studies, our method reduces the analysis time for what-if predictions as well as the size of trace files roughly by half. We also discuss the usability of our method.