Hybrid Parallel Solver for the Atmospheric Dynamical Core of the INM RAS Earth System Model
摘要
This paper considers hybrid parallel approaches for solving Helmholtz-type equation arising from semi-implicit time integration scheme of the atmospheric dynamical core of the INM RAS Earth System Model (INMCM). The original FFT-based solver designed for distributed memory systems requires global data transpositions. We employ a transpose-free approach suitable both for MPI+OpenMP and for heterogeneous CPU+GPU computations. We evaluated the parallel performance and scalability of the MPI+OpenMP implementation compared to the original solver using the same number of CPU cores for different spatial resolutions of INMCM. We also measured the speedup of the GPU implementation relative to the fastest CPU implementation.