Keywords: supercomputer, parallel and vector computing, computational fluid dynamics, computer architecture, domain decomposition, OpenMP, MPI, GPU.

A successful solver for fluid dynamics equations should be accurate, efficient and robust. Moreover, for large-scale applications, the solver should be parallelized and vectorized for running on the parallel and vector computer platforms. In this review paper after a review of the history of computer architecture from vector supercomputers to current multi-core CPU architectures, algorithms developed for high performance computing in computational fluid dynamics are discussed in relation to parallel architectures. The focus is on parallel as well as vectorized solvers for fluid dynamics equations. For the vector pipeline supercomputers, the parallel strategy is to make a longer array for the do-loops. However, for the current massively parallel computer, parallel strategy depends on the memory systems. After the desirable form of the algorithms for fluid dynamics equations is presented, the parallelization and vectorization are discussed. Domain decomposition with load balancing is one of the key techniques in high performance computing and is discussed from the viewpoint of the grid system, using both structured and unstructured grids. Results based on our numerical experiences with the parallel implementations of the unstructured and structured grid methods are introduced. In the laboratory computer environment, a personal computer with multi-cores is usually used for scientific computing. The numerical experiments of the implementation of the parallel algorithms for structured and unstructured grids were performed targeting a conventional personal computer with four cores and a standard workstation of four CPUs (total 24 cores), and the parallel performances are discussed.

In this paper, the use of the graphic processor unit (GPU) for scientific computing is also discussed from the point of view of computational fluid dynamics. The programming model of GPU computing is SPMD (single program multiple data), which is the standard programming model for computational fluid dynamics. Thus the GPU would be a powerful tool for computational fluid dynamics. The recent research activity using GPU for fluid dynamics simulations is reviewed.

purchase the full-text of this paper (price £20)

go to the previous paper
return to the table of contents
return to Computational Technology Reviews
purchase this volume (price £80 +P&P)