FLEXI: Scale-resolving simulations of compressible turbulence on modern HPC systems

Join our own Anna Schwarz for the WHPC Europa invited session discussing our latest work preparing FLEXI for exascale and GPU-based systems.

Abstract

The application of scale-resolving simulations such as LES and DNS remains challenging in numerous engineering applications due to the substantial computational demands and the limited available computational resources. It was thoroughly demonstrated in the last decade that high-order numerical methods show significant advantages in terms of computational cost required to achieve a given accuracy for DNS and LES and that they can be implemented efficiently on traditional CPU-based system architectures. However, the improvements in performance for CPU-based systems has slowed down significantly over the last years. Consequently, more specialised architectures, such as GPUs, have become increasingly popular in the last decade. These architectures still promise considerable performance improvements per generation and thus allow more complex flow cases to be tackled, while also yielding better performance per invested amount of energy. However, CPU codes cannot be easily run on GPU hardware, since they exhibit a distinctly different architecture and require different programming languages. In this talk it is demonstrated how existing codebases can be translated efficiently to GPU hardware, both in terms of implementation effort and computational performance. For this, the unstructured DG code FLEXI~\cite{Krais2019,Kempf2024} is utilized as an example. Finally, it is shown that FLEXI is capable of tackling challenging large-scale problems with complex geometries on modern HPC systems.