High performance regional ocean modeling with GPU acceleration

The Regional Ocean Modeling System (ROMS) is an open-source, free-surface, primitive equation ocean model used by the scientific community for a diverse range of applications [1]. ROMS employs sophisticated numerical techniques, including a split-explicit time-stepping scheme that treats the fast barotropic (2D) and slow baroclinic (3D) modes separately for improved efficiency [2]. ROMS also contains a suite of data assimilation tools that allow the user to improve the accuracy of a simulation by incorporating observational data. These tools are based on four dimensional variational methods [3], which generate reliable results, but require more computational resources than without any assimilation of data. The implementation of ROMS supports two parallel computing models; a distributed memory model that utilizes Message Passing Interface (MPI), and a shared memory model that utilizes OpenMP. Prior research has shown that portions of ROMS can also be executed on a General Purpose Graphics Processing Unit (GPGPU) to take advantage of the massively parallel architecture available on those systems [4]. This paper presents a comparison between two forms of parallelism. NVIDIA Kepler K20X GPUs were used for performance measurement of GPU parallelism using CUDA while an Intel Xeon E5-2650 was used for shared memory parallelism using OpenMP. The implementation is benchmarked using idealistic marine conditions. Our experiments show that OpenMP was the fastest, followed closely by CUDA, while the normal serial version was considerably slower.