Fast Performance Prediction of Master-Slave Programs by Partial Task Execution

In this paper, it is proposed to rapidly and accurately predict performance of master-slave (MS) parallel programs. To provide rapid prediction with high accuracy, our method reduces direct execution of the target MS program and estimates execution time of tasks of the program from only some directly executed tasks. In this estimation, we use a linear interpolation in order to reproduce the original order of task assignment, which affects the prediction accuracy of a performance saturation point. The experimental result shows that our proposed method predicts the performance of MS programs 1.7 times faster, at least, than the measured execution time which corresponds to the minimum time taken to predict the performance by prediction methods based on direct execution. Furthermore, our method predicts the performance with 7% error being as good as that of existing prediction method. Key-Words: Master-slave, Performance prediction, Parallel computational model, Simulation