Mixed MPI. OpenMP Programming: A Study in Parallelisation of a CFD Multiblock

摘要

Two of the most common and best standardized methods for parallel coding of scientific and engineering applications are OpenMP and MPI. OpenMP is now widely supported on shared memory systems while MPI is available on almost all distributed and shared memory machines. With the advent of more machines that combine both shared and distributed memory architectures it is becoming more common to combine both techniques in a single application. In this report the authors look at a simple case study of the parallelization of a multiblock CFD code using both methods. The objective was to rapidly add some parallelism to an existing serial code to exploit a small Beowulf cluster of dual processor PC nodes. It is shown that combining both techniques can yield better performance than either method on its own, without much additional work in code and algorithm restructuring.