Performance evaluation of hybrid parallel computing for WRF model with CUDA and OpenMP

The efficiency of weather forecasting model is determined by the speed of its execution process. However, for about 50 years, its increasing performance has come from the underlying processor improvements, not due to large scale parallelism. This work aimed to evaluate whether a low-cost heterogeneous CPU/GPU computing architecture using hybrid parallelization method of CUDA and OpenMP will increase the speed of Weather Research and Forecasting (WRF) model. The parallelization of many-core GPU using CUDA was implemented to run WSM5, a crucial but computationally intensive finegrained parallelism part of WRF. The results obtained indicate that the speedup of WSM5 CUDA on Tesla C2050 GPU has increased significantly until 403× compared to WSM5 Fortran being run on a single core Intel Xeon X5650 CPU. This hybrid method has also increased a total speed of WRF model up to 12× faster than its implementation on a single core CPU and 27% faster than its implementation using OpenMP on 12-core CPUs. It was also revealed that a fast-math option in WSM5 CUDA code has not significantly influence the accuracy of overall WRF forecasting results.