A serial data-parallel multi-block method for compressible flow computations

Abstract

Block structured meshes not only provide the possibility to compute flows in complex geometries but also lend themselves in a natural way to coarse-grain parallel processing via the distribution of different blocks to different processors. Nevertheless, for some flow computations, a fine-grain data parallel implementation may be more appropriate. This chapter presents a study of such an implementation, which utilizes the simplicity of the data parallel approach. Particular attention is placed on a dynamic block management strategy that allows computations to be undertaken only in blocks where useful work is to be performed. The question of code portability among four different parallel computer systems is addressed in the chapter. This chapter concludes that the serial data-parallel multi-block method provides a number of advantages: (1) it retains the simplicity of the above-mentioned data parallel methods, because each block is treated individually in the same manner as for a single block computation; (2) it does not impose any parallelization constraints on the mesh generation procedure, in principle, any number of blocks of unequal size can be employed; the transfer of data between two blocks (block connectivity) is performed in a transparent manner via globally addressable memory contrasting with the explicit data transfer required by message passing implementations; (3) because individual blocks are treated sequentially, a simple dynamic block management algorithm can be applied to avoid performing unnecessary operations; and (4) the use of standard Fortran 90 facilitates code portability among different platforms supporting the data parallel programming method.