Porting the Princeton Ocean Model to GPUs

摘要

While GPU is becoming a compelling acceleration solution for a series of scientific applications, most existing work on climate models only achieved limited speedup. It is due to partial porting of the huge code and the memory bound inherence of these models. In this work, we design and implement a customized GPU-based acceleration of the Princeton Ocean Model (gpuPOM). Based on Nvidia's state-of-the-art GPU architectures (K20X and K40m), we rewrite the original model from the Fortran into the CUDA-C completely. Several accelerating methods, including optimizing memory access in a single GPU, overlapping communication and boundary operations among multiple GPUs, are presented. The experimental results show that the gpuPOM on one K40m GPU achieves 6.9-fold to 17.8-fold speedup and 5.8-fold to 15.5-fold speedup on one K20X GPU comparing with different Intel CPUs. Further experiments on multiple GPUs indicate that the performance of the gpuPOM on a super-workstation containing 4 GPUs is equivalent to a powerful cluster consisting of 34 pure CPU nodes with over 400 CPU cores.