Implementation and Analysis of Performance Portable Programming Models for cloud resolving GPU-enabled climate modeling

We implemented and analysed three different performance portable programming models (OpenMP, OpenACC, and YAKL) in the cloud resolving model (CRM) used in the E3SM-MMF earth system model. The CRM solves a system of nonlinear inelastic momentum and scalar conservation and continuity equations using the finite difference method. Performance evaluation and analysis of these models are performed to understand which model allows portable performance by effectively utilizing hybrid many-core CPU and GPU heterogeneous architectures to be used in the future exascale supercomputers. Performance benchmark using Summit supercomputer on a range of parameters shows as most as 17x speedup for OpenMP, 13x speedup for OpenACC and 22x speedup for YAKL, with three-dimensional CRM model using 256 CRMs per node. These results demonstrate that OpenMP, OpenACC and YAKL are possible pathways to support the future heterogeneous exascale supercomputer systems. Detailed analysis on the performance data from these three programming models give us baselines for the legacy code refactoring and their data structures to these models.