Assessing the HPC Performance of CODA for the NASA Common Research Model

Abstract

In this paper we study the HPC performance of the new CFD software by ONERA, DLR and Airbus (CODA) on a range of CPU hardware types for the NASA Common Research Model, using meshes up to 192M cells. We explore both the weak and strong scaling, as well as the influence of the number of OMP threads and MPI ranks on the overall runtime. A number of different generations of AMD-based CPU architectures are assessed and compared against two generations of Arm-based CPUs. It is found that the code demonstrates excellent weak and strong scaling, with greater than linear scaling down to 2000 cells per CPU core. As expected the newer generation of AMD CPUs with greater memory bandwidth and higher cores per node, result in better over runtime performance. We see as good or better parallel efficiency using ethernet-based Elastic Fabric Adapter (EFA) interconnect versus Infiniband. The single-socket Arm64-based processors offer similar performance to the x86-based alternatives which suggests those more power-efficient processors can help to reach both performance and sustainability targets. Future work will be to extend this comparison to GPUs and Arm64 based Graviton4, which offer even greater potential for performance improvements.