A Parallelization and Performance Optimization Framework for Mesh-Based Simulations Using Target Execution Models

Abstract

Mesh-based applications usually use very big mesh structures for simulation accuracy, and they are coded in different programming languages using different parallel runtime support systems. They run in various parallel computation systems. It is a big problem that in such a situation how can domain expert build a mesh-based application which well fit the different performance requirements for different target computation systems. Another problem is that it becomes harder to create well performed parallel programs for modern computing systems for domain experts even for computing specialists because modern computing systems become more complicated. In this paper an automatic parallelizing and performance optimizing framework is proposed and implemented for mesh-based application domains through a translation from domain description into internal programming model. Using internal programming models and their performance models with target system execution models, domain descriptions can be effectively parallelized and optimized into parallel code for target systems. Several CFD codes are tested in the framework as case-studies and experiments shown that more efficient codes, which are more scalable and of higher performance, can be generated for different target computation systems.