A022-11
Progress towards accelerating Unified Model on hybrid multi-core systems: Porting CASIM and SOCRATES to GPU

Monday, 7 December 2020: 16:30
Virtual
Wei Zhang1, Min Xu2, Matthew R Norman1, Salil Mahajan3, Mario Morales-Hernandez2 and Katherine J Evans2, (1)Oak Ridge National Lab, Oak Ridge, TN, United States, (2)Oak Ridge National Laboratory, Oak Ridge, TN, United States, (3)Oak Ridge Nat'l Lab, Oak Ridge, TN, United States
Abstract:
The Cloud-AeroSol Interacting Microphysics (CASIM) and Suite of Community Radiation Codes based on Edwards & Slingo, 1996 (SOCRATES) schemes used in the Met Office (UKMO) Unified Model (UM) are ported to GPU on Oak Ridge Leadership Computing Facility’s (OLCF) Summit machine. Initial performance results are reported, and validation of output is described. These two schemes dominate the computational cost of the full UM. OpenACC directives are used to offload the most intensive computations from the CPU to GPUs. Both manual refactoring and automatic source-to-source compiler (CLAW) are used. Optimizing loops and increasing parallelism bring better performance of the tests with CASIM and SOCRATES running on GPU. These improvements are expected to enable UM to run at high resolution on hybrid multi-core systems.