Challenges and advances towards exascale computing in Earth System Modelling
Convener:
Daniel Caviedes-Voullième
|
Co-conveners:
Denise DegenECSECS,
Hui TangECSECS,
Sabine Grießbach
The challenges are manyfold: leveraging accelerators, coupling models, adding processes, parametrisations, and feedbacks relevant at higher resolutions, ensuring accurate and stable algorithms, dealing with increased I/O, big-data ready workflows, compressing data, in situ processing, real-time applications, etc. Solutions to these challenges must also guarantee future-proofing codes, performance-portability across current (and future) architectures, domain-specific languages, code sustainability (including readability, maintenance, etc), improved energy use efficiency and minimising the carbon footprint of ESM. These expectations require an intense cross-disciplinary effort between geoscientists of all Earth system compartments, computer scientists, applied mathematicians, and software engineers.
Harnessing exascale resources requires a range of solutions ranging from improving parallel efficiency and scalability of numerical methods and coupling strategies, to porting code to accelerators, leveraging heterogeneous and modular computing, novel I/O solutions and data compression. Exascale ESM will also rely on Machine Learning (ML) techniques for probabilistic and inverse modelling, data assimilation, uncertainty quantification, as well as ML-based model surrogates and ML-driven model operation.
This session is a forum to discuss these challenges and solutions across all Earth system compartments, involving domain scientists, applied mathematicians, computer scientists, and high-performance computing experts. The session welcomes contributions addressing both open challenges and current advances to achieve exascale-readiness for Earth System Models across all compartments of the Earth system and across different methodologies and technologies. We especially encourage contributions discussing challenges and advances potentially applicable to exascale modelling efforts across Earth system compartments and digital twins.