nekRS and Ginkgo - Towards Exascale CFD
Research topic and goals
JSC will be the hosting side of the first European Exascale supercomputer (JUPITER) that is going to become available in 2024. While both CPU-heavy and GPU-heavy Petascale supercomputers exist, JUPITER is targeted for 1.3 Exaflops in its GPU partition. Therefore, code frameworks and workflows need to be adopted. In the area of CFD – ranging from earth system modeling to combustion applications, nekRS is currently the most promising candidate for Exascale simulations at JSC. It is part of the Exascale benchmark suite and extended in the direction of chemistry flows. This effort is led by Mathis Bode (JSC). Hartwig Anzt (UTK) and his team are developing the linear algebra library Ginkgo for many years. It is a high-performance numerical library that supports NVIDIA GPUs and AMD GPUs. Ginkgo has been successfully used with other CFD codes and preliminary work with nekRS has been conducted. This JLESC project aims to bring nekRS and Ginkgo together. In a first step, the performance of Ginkgo for cold flows at scale is evaluated for multiple application cases. These cases range from mesoscale convection to complex cases with moving geometries. Throughput and scaling behavior will be studied on JUWELS Booster. Secondly, the potential of Ginkgo in the context of nekRS’s currently developed chemistry module is accessed. This includes the evaluation of batched solvers for solving coupled scalar equations. Finally, the combined behavior of nekRS and Ginkgo at Exascale size are tested on the new supercomputer JUPITER. Bottlenecks will be identified and improvements suggested. Ultimately, the goal is to provide a high-scaling, high-throughput version of nekRS for Exascale runs to the nekRS community.
Results for 2022/2023
None yet.
Visits and meetings
None yet.