Towards exafoam: Deployment of unstructured mesh implict CFD code suitable for exascale

OpenFOAM acts as a major player in the Open Source CFD arena, due to its flexibility, but on the other side due to its complexity it is hard to define correctly performance figure and scaling. From previous work, it is known that the scalability of the linear solvers restrict the parallel usage up to the order of few thousand of cores [1]. Meantime, the technological trends of exascale (computing systems capable of at least one exaFLOPs calculations per second) HPC (High Performance Computing) is moving towards the use of order of millions of cores [2, 3]. The actual trend of the HPC cluster is to use node with high-density of core (more than 50 cores per per CPU), with more cache level and with the order of O(100,000) of interconnected nodes and O(10,000,000) of available cores [7]. OpenFOAM is among the most used CFD code in HPC. To be able to strengthen his position in the competitive CFD market, it is mandatory to enable the code to efficiently use the current and the forthcoming HPC hardware available in the worldwide market. A recent survey, funded by the U.S. Department of Energy Office of Science on Economic Models linking HPC and ROI, has investigated how HPC can improve economic success and increase scientific innovation. The report indicates very substantial returns for investments in HPC, that is $38,7 on average of profits (or cost savings) per dollar of HPC invested. The long-term aim of this collaborative project is to enable the deployment of unstructured mesh implicit open source CFD code (i.e. OpenFOAM) more suitable for exascale computing. To get such ambitious goal, the actual bottlenecks have to be taken into account and specific objectives have to be achieved. Up to date, the well known bottlenecks for a full enabling of OpenFOAM for massively parallel clusters are ✓ the limit in the parallelism paradigm ✓ the I/O data storage system The objectives that have to be achieved to get the aforementioned aim are: ✓ Profiling OpenFOAM on HPC clusters to analyse the parallel performances ✓ Optimize the code on the selected and emerging HPC architectures (e.g. Intel Phi-KNL) ✓ Benchmarking the code on the actual petascale HPC architectures ✓ Optimize the actual version of Pstream and/or use a multi-threaded hybridization of the linear algebra libraries to get petascale performance ✓ Transition to exascale: algorithmic rewrite of the Pstream core library This is a co-design project, which should involve several partners in the Consortium: owner of the OpenFOAM® trademark, HPC centres and specialists, mathematical modelling and experts, application partners, hardware/software vendor, final end users which will act as drivers or observers. The platform to implement such project is the forthcoming call FET Proactive - High Performance Computing H2020-FETHPC-2016-2017. The specific challenge of the call is to take advantage of the full capabilities of exascale computing by addressing the topic of mathematics and algorithms for extreme scale HPC systems and applications working with extreme data.