Parallelization of some scientific codes using profiling tools and OpenMP

Abstract

The purpose of this report is to exchange our experience with parallelizing existing scientific codes by utilizing profiling tools and the OpenMP application programming interface (API) for multi-platform shared-memory parallel programming in C/C++ and Fortran. Profiling is a very good tool to get an indication on which parts of the program to concentrate on in order to parallelize scientific codes. In general there will be necessary to get more detailed information regarding the relative importance of different sub-blocks inside an interesting subroutine. This information might be obtained by the use of manually inserted timers inside the subroutines. In order to get more reliable results cpu-timers should be used instead of wall-clock-timers. The main goal was to parallelize the Simra CFD-code as much as possible. Some initial work on less complex and smaller programs did undoubtedly lead to better results for the Simra code. At the moment about 65% of the program has been parallelized. When using 16 cores on the Njord supercomputer, the global speedup resulting from this work is between 2.2 and 2.3 depending on the problem size.

Oppdragsgiver SINTEF ICT Applied Mathematics