MPTD: A Scalable and Flexible Performance Prediction Framework for Parallel Systems

Abstract

The increasing complexities of today’s parallel systems pose new challenges for performance prediction. Effective performance prediction can provide insight, deepen understanding and further identify potential performance bottlenecks of system/application combinations. In this paper, we present and evaluate a multi-phase trace-driven (MPTD) performance prediction framework for parallel systems. In the trace generation phase, based on a relatively simple performance model, MPTD performs parallel performance simulation to generate primary prediction results and traces rapidly. In the trace adjustment phase, traces are transformed or re-simulated based on performance models of new component architecture or more detailed performance models. This phase is self-repeatable (it can be performed more than once and need not go back to the former phase) to enable more flexible reuse of traces. We implemented an instantiation of MPTD to predict the performance of popular multi-core cluster systems. Analysis and tests show that MPTD is scalable, flexible, and can help researchers for better balancing accuracy and efficiency of performance prediction.KeywordsPerformance PredictionMessage Passing InterfaceParallel SystemTarget ProcessPerformance ToolThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.