An improved LBM approach for heterogeneous GPU-CPU clusters

Graphics Processing Units, originally developed for computer games, now provide computational power for scientific applications. And the usage of the hybrid parallel computing programming environment with MPI and CUDA technology to build the simple CPU+GPU-based stream processor cluster system become popular. The Lattice-Boltzmann method is a distribution-function based approach to numerical fluid mechanics. However, in most implementations for heterogeneous clusters, they didn´t take into account the computing power of CPU and didn´t optimize the communication well. In this paper, we analyze these methods adapted to the heterogeneous clusters, and propose an improved LBM approach for heterogeneous GPU-CPU clusters which is suitable for load balancing and heterogeneous computations on CPUs and GPUs.