EXCELLERAT P2 D3.1 Report on Exa-Enabling Methodologies

The scalability of CODA was analysed on the largest available partition of DLR’s main production system. The performance of the code was tested on various upcoming CPU architectures and on Nvidia A100 GPUs. RWTH significantly improved the performance and parallel efficiency for large-scale multiphysics simulations with the code m-AIA. The code was scaled up to utilise the full HAWK HPC system of about 500,000 compute cores. BSC tested Alya on 3 different machines: MareNostrum IV and MN3 CTE-Power at BSC, and Polaris at Argonne Leadership Computing Facility. Moreover, it has recently received computational resources at the Karolina EuroHPC supercomputer and is currently starting tests on that machine too. Better memory management has allowed to significantly improve scalability on GPUs. CINECA and URMLS are using a multi-paradigm approach for their code FLEW which involves a two-phase code development line. Code development, for example with new algorithms, uses a backend “master paradigm” based on CUDA Fortran. This main code is then transformed to “secondary paradigms” using automatic translations in house developed tools. The UL team focused on the field line tracing optimisation. The code L2G has been optimised for better computational efficiency. The algorithmic improvements of the code are octree space partitioning and bounding box partitioning of large meshes to decrease time of line triangle intersection checks.