Anciaux-Sedrakian, A. (IFP Energies nouvelles) | Eaton, J. (NVIDIA) | Gratien, J. (IFP Energies nouvelles) | Guignon, T. (IFP Energies nouvelles) | Havé, P. (IFP Energies nouvelles) | Preux, C. (IFP Energies nouvelles) | Ricois, O. (IFP Energies nouvelles)
Heterogeneous supercomputers combining "general-purpose graphic processor", "many integrated cores" and general-purpose CPUs promise to be the future major architecture, because they deliver excellent performance with limited power consumption and space occupancy. However, developing applications that achieve "real" scalability remains a subject of several research fields. This paper studies the possibility of exploiting the power of heterogeneous architecture for Geoscience dynamic simulations for real large-scale models. Geoscience dynamic simulations need scalable linear solvers to realize the potential of the latest high-performance-computing architectures, since it represents the most expensive part of the whole simulation. Several studies have already been done to accelerate sparse linear solver's preconditioners and they illustrated comparable or even improved performance for the academic test cases. However, for a real industrial test case with highly heterogeneous data, these preconditioners do not offer a robust solution, as CPR-AMG does. We study several implementations of AMG and CPR-AMG for heterogeneous architectures in this paper. We illustrate the obtained gain using one real field thermal model and two academic test cases all using a fully implicit scheme. For these models, the total simulation time is reduced by a significant factor when running on n CPU + n GPU, compared to n CPU only, without any accuracy loss.