A case study of the partitioning patterns for domain decomposition method on VPP 700 E

The most common parallelization strategy for many Computational Mechanics (typified by Computational Fluid Dynamics(CFD) applications) which use structured grids, involves the one directional partition based upon slads of grids. For parallelised versions of CFD codes to scale well we must employ two (or more) dimensional partitions. However, FORTRAN code implementations by multi-directional partitions are not easy and the computational performances by the partitioning patterns are significantly different on many parallel machines.1) This paper describes a case study approach to the multi-dimensional structured grid partitioning in code parallelisation, its implementation using FORTRAN and MPI (Message Passing Interface) on the VPP700E. The results of the tests on FORTRAN codes demonstrate clear improvements in parallel performance (over the one directional partition). This is matched by a reduction in the time required to implement the parallel CFD code when MPI program.