Accelerating cloud and radiation schemes within the Unified Model for CPU-GPU based high-performance computing systems
Abstract
Earth System Models (ESMs) simulate states of the coupled climate containing the atmosphere, ocean, cryosphere, land surface, and biogeochemistry using various mathematical equations and physical schemes. ESMs' spatial and temporal resolutions have been increasing with advances in HPC computational power. However, improvements in scalability and efficiency of dynamic and physical components of the models remain a challenge for the next generation of ESMs, especially the physics - due to parameterization complexity and heavy legacy of code structures. Cloud and radiation schemes are two of the most time-consuming/computationally intensive parameterizations in ESMs and their parallelism is generally implemented by assigning columns to MPI processors and distributing columns within a processor using OpenMP threads for physics computations. Generally, their performance could be improved if the most intensive computations are offloaded to GPUs from the host CPU. Using a single-column abstraction method, we have refactored portions of the cloud (CASIM) and radiation (SOCRATES) modules within the Met Office (UKMO) Unified Model (UM) to enable them to utilize GPUs on Oak Ridge Leadership Computing Facility's (OLCF) Summit machine. We implemented OpenACC directives using a combination of manual refactoring and an automatic source-to-source compiler (CLAW), where possible. Initial tests of CASIM and SOCRATES show enhanced performance. The CLAW compiler was found to be useful for automatic GPU porting providing some performance portability. Future work will focus on refining the OpenACC implementation as executed on Summit to further improve performance, especially within the larger coupled UM model.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.A13H3039X
- Keywords:
-
- 3336 Numerical approximations and analyses;
- ATMOSPHERIC PROCESSES;
- 3337 Global climate models;
- ATMOSPHERIC PROCESSES