Séminaire de Mécanique d'Orsay
Jeudi 26 mai à 14h au LIMSI
A new approach for Unstructured High Order Finite Volume codes on GPU
Matthieu Lefebvre
(et J.-M. Le Gouez [ONERA], A. Refloch'h [ONERA] and C. Basdevant [Univ. Paris 13])
ONERA, et Univ. Paris 13
 

For now a few years, an increasing number of studies are being conducted on how to implement CFD algorithms on GPUs, going from simplified ones on a single cartesian grid, up to representative grid systems for industrials applications (multi-blocks structured, arbitrary unstructured), only amenable on computer clusters. Willing to explore the limits of this approach, this paper dives into what is expected to be one of the worst case for GPU parallelization, the full turbulent Navier-Stokes equations on composite structured and unstructured grids. This type of grid system allows a good description of boundary layers flows, and an economy of memory in regions away from the aeronautic bodies. Following a short introduction to the numerical model and to the key points of the GPU, a discussion will be held about the implementation issues and the performance enhancements they bring.
Keywords: Computational Fluid Dynamics, Parallelization, GPU, Domain Partitioning, NXO, CUDA