A highly-efficient automated optimization approach for kilometer-level resolution Earth system models on heterogeneous many-core supercomputers

Abstract. As coupled Earth system models advance, it becomes increasingly feasible to attain higher spatial resolutions, thereby enabling more precise simulations and predictions of the evolution of the Earth system. Consequently, there is an urgent demand of highly-efficient optimization for extensive scientific programs on more power-efficient heterogeneous many-core systems. This study introduces a highly-efficient optimization approach tailored for kilometer-level resolution Earth System Models (ESMs) operating on heterogeneous many-core supercomputers. Leveraging scalable model configurations and innovative tripolar ocean/sea-ice grids that bolster spatial accuracy and computational efficiency, we initially establish a series of high resolutions (HRs) within a solitary component (either the atmosphere or ocean) while maintaining a fixed resolution for the other, resulting in notable enhancements in both model performance and efficacy. Furthermore, we have devised an OpenMP tool specifically optimized for the new Sunway supercomputer, facilitating automated code optimization. Our approach is designed to be non-intrusive, minimizing the need for manual code alterations while ensuring both performance gain and code consistency. We adopt a hybrid parallelization strategy combining Athread and OpenMP, achieving full parallel coverage for code segments with a runtime proportion exceeding 1 %. After optimization, the atmosphere, ocean, and sea-ice models achieve speedups of 4.43×, 1.86×, and 2.43×, respectively. Consequently, the overall simulation performance of the 5-km/3-km coupled model reaches 222 SDPD. This achievement renders multiple decadal scientific numerical simulations utilizing such HR coupled simulations feasible. Our work signifies a pivotal advancement in Earth system modeling, providing a robust framework for high-resolution climate simulations on more ubiquitous (next-generation) heterogeneous supercomputing platforms, such as GPUs, with minimal additional effort.