Implementation and performance evaluation of SPAM particle code with OpenMP-MPI hybrid programming

In this paper, we implement a SPAM (Smooth Particle Applied Mechanics) code in both pure MPI and MPI-OpenMP hybrid manner, then compare and analize the performance of them on an SMPPC cluster. Our SPAM code is described to handle any mapping of spatial cells on to parallel MPI processes to exploit well load-balancing even with a relatively high communication cost. First we implement a parallelized version of the code with MPI, then parallelize each MPI process with OpenMP parallel-do feature. In the overall performance results, the performance of MPI-only code exceeds that of MPIOpenMP hybrid code except with a single node execution. We analized the detail of execution time and cache hit ratio on each processor, and found that the advantage of low communication cost between threads overcomes the disadvantage of low cache miss-hit ratio caused by the data localization problem in OpenMP. Another issue of this code is load imbalancing caused with the setting in our target problem mode. We consider a better cell mapping suitable for hybrid code and discuss the di culty to exploit high performance on this kind of hybrid programming.