GPU implementation of a SPH-ALE fluid dynamics solver

Marongiu, J.-C.; Leduc, J.; Schaller, M.

Abstract

Application of SPH methods in an industrial context is closely linked to the capacity of exploiting the very last and best computational resources available. Indeed SPH involves a computational load that can be significant compared to others and well established numerical methods. Two years ago appeared in the SPHERIC community the first full GPU implementation of a SPH flow solver, based on the high level API CUDA from Nvidia. This pioneering work revealed the high potential of the GPU technology and further works confirmed that hybrid implementations can challenge the classical MPI- or OpenMP-based approaches. However the SPH-ALE method exhibits important peculiarities that require a dedicated approach. The paper thus reports the implementation strategy that has been applied for the different parts of the algorithm, leading to the full GPU implementation of a second order SPH-ALE flow solver, including the management of all types of boundary conditions (solid walls, periodic, symmetry). The impact of numerical models on performance is analyzed. The respective assigned roles to CPU and GPU are also discussed, together with the choice of floating point accuracy. A detailed validation of the GPU implementation versus its CPU counterpart is also shown. The resulting flow solver can achieve speed-ups up to 40 on industrial applications.

Citation

Marongiu, J., Leduc, J., & Schaller, M. (2011). GPU implementation of a SPH-ALE fluid dynamics solver.

• Conference Name: 6th International SPHERIC Workshop
• Conference Location: Hamburg, Germany
• Publication Date: 2011-06
• Deposit Date: Feb 26, 2012
• Publisher URL: http://icc.dur.ac.uk/~jlvc76/Files/Papers/SPHERIC_2011.pdf