Parallelization of the AVL FIRE Benchmark with SVM-Fortran

This article outlines the parallelization of an irregular grid application with SVM-Fortran. It describes the different optimizations and their effectiveness. The parallelization was much simplified by the performance analysis tool OPAL, a source code based tool for requesting and analyzing runtime performance data. Although shared memory parallelization is easier than distributed memory parallelization, understanding and eliminating the overhead from page faults is impossible without such a tool. It relates the page faults to the arrays and to the location in the source code. An area which is not supported by OPAL but where supporting tools are highly desirable, is the performance degradation due to low utilization of the on-chip cache