CESMTuner: An Auto-Tuning Framework for the Community Earth System Model

The growing scientific demands of climate predication and climate projection have promoted to manage the computational resources of climate model rationally. The Community Earth System Model (CESM) is one of the state-of-the-art and the most widely-used coupled models for simulating the earth system. Although considerable effort has been put to improve the scalability of single component, CESM is still struggling with the poor performance due to load balance across components. To solve this problem, an easy-used and easy-ported auto-tuning framework named CESMTuner is proposed in this paper. It targets to reduce the time consumed of CESM as much as possible by looking for the optimal process configuration. In which, a novel process layout searching algorithm is presented that can look for the optimal process count of each component as well as the best process layout across components simultaneously. Moreover, a lightweight and accurate performance model is built to reduce searching overhead effectively. With the evaluation over TianHe-1A, CESMTuner can achieve 58.49% performance improvement compared to the widely-used sequential process layout and achieve 38.23% performance improvement compared to the heuristic branch and bound algorithm based on the performance model of simply fitting each component's runtime.