Machine-Learning-Based Load Balancing for Community Ice Code Component in CESM

Load balancing scientific codes on massively parallel architectures is becoming an increasingly challenging task. In this paper, we focus on the Community Earth System Model, a widely used climate modeling code. It comprises six components each of which exhibits different scalability patterns. Previously, an analytical performance model has been used to find optimal load-balancing parameter configurations for each component. Nevertheless, for the Community Ice Code component, the analytical performance model is too restrictive to capture its scalability patterns. We therefore developed machine-learning-based load-balancing algorithm. It involves fitting a surrogate model to a small number of load-balancing configurations and their corresponding runtimes. This model is then used to find high-quality parameter configurations. Compared with the current practice of expert-knowledge-based enumeration over feasible configurations, the machine-learning-based load-balancing algorithm requires six times fewer evaluations to find the optimal configuration.