A Look at Performance and Scalability of the GPU Accelerated Sparse Linear System Solver Spliss.

Abstract A significant part in computational fluid dynamics (CFD) simulations is the solving of large sparse systems of linear equations resulting from implicit time integration of the Reynolds-averaged Navier-Stokes (RANS) equations. The sparse linear system solver Spliss aims to provide a linear solver library that, on the one hand, is tailored to these requirements of CFD applications but, on the other hand, independent of the particular CFD solver. Spliss allows leveraging a range of available HPC technologies such as hybrid CPU parallelization and the possibility to offload the computationally intensive linear solver to GPU accelerators, while at the same time hiding this complexity from the CFD solver. This work highlights the steps taken to establish multi-GPU capabilities for the Spliss solver allowing for efficient and scalable usage of large GPU systems. In addition, this work evaluates performance and scalability on CPU and GPU systems using a representative CODA test case as an example. CODA is the CFD software being developed as part of a collaboration between the French Aerospace Lab ONERA, the German Aerospace Center (DLR), Airbus, and their European research partners. CODA is jointly owned by ONERA, DLR and Airbus. The evaluation examines and compares performance and scalability in a strong scaling approach on Nvidia A100 GPUs and the AMD Rome architecture.