Parallel Hybrid Particle Simulations Using MPI and OpenMP

We present a library for the parallel computation of particle simulations called ParaSPH. It is portable and performs well on a variety of parallel architectures with shared and distributed memory. We give details of the parallelization for hybrid architectures (clustered SMPs) using MPI and OpenMP and discuss implementation issues, performance results and memory consumption of the code on two parallel architectures, a Linux Cluster and a Hitachi SR8000-F1. We show the advantage of hybrid parallelization over pure message-passing especially for large node numbers for which we gain a maximum speedup of about 350 for hybrid parallelization compared to 120 for message-passing.