PERFORMANCE PREDICTION OF PARALLEL APPLICATIONS

Abstract

Nowadays, the need for parallel computing keeps increasing with the advancement of technology towards multi-core processor. However, the performance of parallel computing is di±cult to predict due to many factors such as communication time, latency, etc. Increasing the parallel system's CPU speed by two does not necessarily reduce the total execution time by half. Before deciding to add more resources into a system or change the computational power, we must consider how much is the performance gain against the cost. In this dissertation, we try to develop a simple solution for parallel application performance prediction. We propose two methods: analytical and simulation. The analytical method deals with equations governing the parallel application execution. For analytical method, we use two different methods for two categories of parallel applications: balanced-load and imbalanced-load. We propose a novel approach on performance prediction of parallel application with imbalanced load. For simulation method, we use a simulation software developed at the University of Barcelona CEPBA-UPC. The simulation software is able to predict the performance of the par- allel applications on various parallel systems by setting the appropriate simulator configuration. We carried out experiments to validate the performance prediction result obtained using various methods that we have studied in this project. We carried out experiments to validate the performance prediction result obtained using various methods that we have studied in this project.