Progress for Optimizing the UM Physics for Accelerated Computing Systems

Our research has been focused on enabling the UK Meteorological Office (UKMO)’s Unified Model (UM) to take advantage of accelerated multi-core systems. One of the optimization targets is the Cloud-AeroSol Interacting Microphysics (CASIM). With a more detailed representation of the interaction between clouds and aerosols, CASIM simulates cloud processes more realistically, but is computationally intensive. We optimized CASIM by offloading the expensive parts to GPUs using OpenACC. Through experiments running CASIM within a simple model, we achieved a reasonable speedup when comparing CASIM's performance on a single Summit node (6 NVIDIA V100 GPUs vs. 42 IBM Power9 CPUs; speedup depends on problem size). In this presentation, we will show how GPU-enabled CASIM is implemented into the UM framework and its impact on cloud process performance within UM regional and global simulations. In conclusion, the presentation will outline how the power of GPU acceleration can be leveraged to optimize the role of CASIM in UM and not harm performance when the code runs on CPUs.