Modernizing the ICON Codebase through Language-Interoperable Memory Management and a Consistent Test
DescriptionOver the past decades, processor performance has grown dramatically. Memory bandwidth and latency have become critical factors in solving HPC problems. Most HPC machines now use accelerators such as GPUs with independent memory. In addition, CPUs now feature many cores and special fast memory, like the Fujistu A64FX and Intel Sapphire Rapids. This work presents a memory-manager concept for the ICON climate and weather prediction model. The memory manager written in C++ allows supporting vendor-native parallel programming frameworks, which are also C++-based, like CUDA, HIP, and SYCL, or portability layers like Kokkos. At the same time, the language interoperability enables backward compatibility with the legacy code in Fortran while introducing concepts facilitating better ICON component interfaces and implementations for better efficiency. This work also introduces a consistent test workflow for memory management. The consistent test ensures the portability of the memory manager in different front-ends and can be extended to future efforts of ICON modernization.
TimeWednesday, June 2812:00 - 12:30 CEST
Climate, Weather and Earth Sciences