[A-AS02] 高性能計算が拓く気象・気候・環境科学

[AAS02-11] Efforts toward optimization of global non-hydrostatic atmospheric model on GPU supercomputer

*八代 尚¹、山崎 一哉²、荒川 隆³、松岸 修平²、Intyisyar Shereo¹、中島 研吾² (1.国立環境研究所、2.東京大学、3.株式会社ClimTech)

キーワード：GPU、モデルシミュレーション、高性能計算、気候予測、機械学習、Python

To secure more computational resources for future weather/climate simulations, it is necessary to utilize supercomputers equipped with many energy-efficient accelerators. We present the results of several strategies to port the non-hydrostatic icosahedral atmospheric model (NICAM, Satoh et al., 2014) to GPU supercomputers. 1) We conducted GPU porting of the full application using OpenACC directives and evaluated computational performance. We identified and avoided several patterns that led to deteriorated computational speed throughout the optimization process. 2) Several kernels extracted from the dycore were rewritten from Fortran to Python. Leveraging the JAX library, we conducted offloading the computations to GPU. 3) Algorithm transformation: We created data-driven model components capable of utilizing dense matrix operations by learning the simulation results based on the physics-based process model.