Report summarising the adaptation of the proposed DSLs and the evaluation of the benchmarks and models proposed in this project - Deliverable D2.1

Autor: Ford, Rupert, Porter, Andrew, Roethlin, Matthias, Osuna, Carlos, Kühnlein, Christian, Kavcic, Iva, Nobre, Nuno
Rok vydání: 2023
Předmět:
DOI: 10.5281/zenodo.8046383
Popis: This deliverable discusses the extensions and developments performed to two DSLs, PSyclone and dusk/dawn, required to support the models of ESiWACE2. PSyclone was further developed to improve the performance and scalability of the Met Office’s LFRic atmosphere model and to improve the performance of the generated GPU version of the full NEMO model. PSyclone is already used in LFRic and as a result of these improvements the NEMO community are also considering adopting PSyclone for use in their model. The dusk/dawn DSL was also extended to support specific patterns of ICON and to provide integration of the DSL code into the Fortran codebase of ICON. Dusk/dawn provides a pythonic high-level language to express the ICON algorithms. The optimised backend provides significant performance improvement on NVIDIA GPUs over the existing operational OpenACC code. Lastly, the potential for interoperability between the two DSLs is demonstrated in a NEMO benchmark code and it is shown that by combining the two tools, greater performance and improved performance portability can be obtained.
ESiWACE2 stands for Centre of Excellence in Simulation of Weather and Climate in Europe Phase 2. ESiWACE2 is funded by the European Union's Horizon 2020 research and innovation programme (H2020-INFRAEDI-2018-1 call) under grant agreement 823988.
Databáze: OpenAIRE