5:15 PM - 6:45 PM
[AAS02-P08] Performance evaluation of the GPU-enabled weather model SCALE
In this study, we evaluated the performance of the GPU-ported version of a meteorological model; Scalable Computing for Advanced Library and Environment (SCALE: Nishizawa et al. 2015; Sato et al. 2015).
The regional version of SCALE (SCALE-RM) has been developed as a next-generation community model in Japan to conduct numerical simulations for regional scale atmospheric phenomena on various computer architectures, and it has been used by various studies. SCALE consists of a dynamical core and a physics scheme for calculating various physical processes such as cloud microphysics, atmospheric radiation, oceanic and land surface processes, and so on. SCALE has been used on Central Processing Unit (CPU)-based computers. However, the next-generation supercomputers are expected to be constructed with many nodes centered on Graphics Processing Units (GPUs), and the meteorological model working on GPU-based supercomputers are required.
To address this problem, the numerical weather models using GPUs are developing in the world (e.g., Sauer and Muñoz-Esparza, 2020, Muñoz-Esparza et al. 2022), and it is also required to develop GPU-ported version of SCALE.
In this study, we ported SCALE into GPU and evaluated the performance of the GPU-ported version of SCALE. There are several methods to port against GPU such as rewriting the source code by CUDA, inserting directives, and using a library for GPUs. We adopted the second one, inserting directive of OpenACC. In this presentation, we will present the evaluation results of the performance of GPU ported version of SCALE. We will also show the results of the acceleration by the method proposed by Yamazaki (2023), who tried to accelerate SCALE by launching GPU kernels asynchronously.
reference
Muñoz-Esparza, D. et al. (2022), The FastEddy® Resident-GPU Accelerated Large-Eddy Simulation Framework: Moist Dynamics Extension, Validation and Sensitivities of Modeling Non-Precipitating Shallow Cumulus Clouds, J. Adv. Model. Earth Sys. doi: 10.1029/2021MS002904
Nishizawa, S. et al. (2015), Influence of grid aspect ratio on planetary boundary layer turbulence in large-eddy simulations. Geosci. Model Dev., 8, 3393-3419
Sato, Y., et al. (2015), Impacts of cloud microphysics on trade wind cumulus: which cloud microphysics processes contribute to the diversity in a large eddy simulation?. Prog. Earth Planet. Sci., 2(23), doi:10.1186/s40645-015-0053-6.
Sauer, J. A.,amd Munoz-Esparza, D. (2020). The FastEddy® resident-GPU accelerated large-eddy simulation framework: model formulation, dynamical-core validation and performance benchmarks. J. Adv. Model. Earth Sys., 12(11), e2020MS002100, doi:10.1029/2020MS002100
Yamazaki, K. (2023), Porting an Atmospheric Model to GPUs: Low-cost to High-performance Results, Open Accelerated Computing Summit 2023, October 4, 2023 (https://www.openacc.org/events/open-accelerated-computing-summit-2023 )
The regional version of SCALE (SCALE-RM) has been developed as a next-generation community model in Japan to conduct numerical simulations for regional scale atmospheric phenomena on various computer architectures, and it has been used by various studies. SCALE consists of a dynamical core and a physics scheme for calculating various physical processes such as cloud microphysics, atmospheric radiation, oceanic and land surface processes, and so on. SCALE has been used on Central Processing Unit (CPU)-based computers. However, the next-generation supercomputers are expected to be constructed with many nodes centered on Graphics Processing Units (GPUs), and the meteorological model working on GPU-based supercomputers are required.
To address this problem, the numerical weather models using GPUs are developing in the world (e.g., Sauer and Muñoz-Esparza, 2020, Muñoz-Esparza et al. 2022), and it is also required to develop GPU-ported version of SCALE.
In this study, we ported SCALE into GPU and evaluated the performance of the GPU-ported version of SCALE. There are several methods to port against GPU such as rewriting the source code by CUDA, inserting directives, and using a library for GPUs. We adopted the second one, inserting directive of OpenACC. In this presentation, we will present the evaluation results of the performance of GPU ported version of SCALE. We will also show the results of the acceleration by the method proposed by Yamazaki (2023), who tried to accelerate SCALE by launching GPU kernels asynchronously.
reference
Muñoz-Esparza, D. et al. (2022), The FastEddy® Resident-GPU Accelerated Large-Eddy Simulation Framework: Moist Dynamics Extension, Validation and Sensitivities of Modeling Non-Precipitating Shallow Cumulus Clouds, J. Adv. Model. Earth Sys. doi: 10.1029/2021MS002904
Nishizawa, S. et al. (2015), Influence of grid aspect ratio on planetary boundary layer turbulence in large-eddy simulations. Geosci. Model Dev., 8, 3393-3419
Sato, Y., et al. (2015), Impacts of cloud microphysics on trade wind cumulus: which cloud microphysics processes contribute to the diversity in a large eddy simulation?. Prog. Earth Planet. Sci., 2(23), doi:10.1186/s40645-015-0053-6.
Sauer, J. A.,amd Munoz-Esparza, D. (2020). The FastEddy® resident-GPU accelerated large-eddy simulation framework: model formulation, dynamical-core validation and performance benchmarks. J. Adv. Model. Earth Sys., 12(11), e2020MS002100, doi:10.1029/2020MS002100
Yamazaki, K. (2023), Porting an Atmospheric Model to GPUs: Low-cost to High-performance Results, Open Accelerated Computing Summit 2023, October 4, 2023 (https://www.openacc.org/events/open-accelerated-computing-summit-2023 )