Atmospheric Model Cluster Performance Evaluation on hybrid MPI/OpenMP/Cuda Programming Model Platform

This work discuss the parallel performance of a global numerical simulation model, Ocean-Land-Atmosphere Model (OLAM), on a hybrid multicore/GPU cluster environment, under the following programming models: 1) OLAM MPI implementation, on the multicore system, 2) OLAM hybrid MPI/OpenMP, which starts one MPI process on each node of the platform and one OpenMP thread on each core of the node, 3) OLAM hybrid MPI/OpenMP/Cuda implementation, which starts one MPI process on each node of the platform, one OpenMP threads on each core of the node and Cuda kernels on the GPUs. The results shows that the adopted programming model impacts significantly the performance of the application. We show that as we increase the number of cores, the OLAM MPI parallel implementation running one process on each cluster core executes faster than the other implementations.