HPC Code Generation Framework for the Discontinuous Galerkin Method as a Technology Demonstrator for Digital Twins
DescriptionNowadays, automatic code generation and domain-specific languages are among the most promising methodologies for the future development of numerical weather prediction models and digital twins. As a technology demonstrator for new weather and climate models, we present the automatic code generation framework ExaStencils with its Python frontend GHODDESS, which is specialized towards discontinuous Galerkin discretizations for a shallow water model. Our implementation serves as a test platform for a range of new numerical, algorithmic, and computational technologies with the potential to be incorporated into digital twins after successful evaluation. First, we re-formulate the conservative shallow water equations in a way that avoids fraction-type nonlinearities and is suited for quadrature-free integration. Furthermore, we show an algorithm re-design that achieves improved hardware usage on a heterogeneous CPU-GPU system and significantly speeds up the computations. Lastly, we present masked block-structured grids for realistic ocean domains, which on the one hand, are capable of accurately meshing fine-scale geometric features and, on the other, offer performance benefits associated with structured grid models.
TimeWednesday, June 2815:00 - 15:30 CEST
Climate, Weather and Earth Sciences