TY - CONF AU - Huber, Markus AU - Kohl, Nils AU - Leleux, Philippe AU - Rüde, Ulrich AU - Thönnes, Dominik AU - Wohlmuth, Barbara TI - Massively Parallel Multigrid with Direct Coarse Grid Solvers VL - 50 CY - Jülich PB - Forschungszentrum Jülich GmbH Zentralbibliothek, Verlag M1 - FZJ-2020-01436 T2 - Publication Series of the John von Neumann Institute for Computing (NIC) NIC Series SP - 335 - 344 PY - 2020 AB - Multigrid methods play an important role in the numerical approximation of partial differential equations. As long as only a moderate number of processors is used, many alternatives can be used as solver for the coarsest grid. However, when the number of processors increases, then standard coarsening will stop while the problem is still large and the communication overhead for solving the corresponding coarsest grid problem may dominate. In this case, the coarsest grid must be agglomerated to only a subset of the processors. This article studies the use of sparse direct methods for solving the coarsest grid problem as it arises in a multigrid hierarchy. We use as test case a Stokes-type model and solve algebraic saddle point systems with up to O(10$^{11}$) degrees of freedom on a current peta-scale supercomputer. We compare the sparse direct solver with a preconditioned minimal residual iteration and show that the sparse direct method can exhibit better parallel efficiency. T2 - NIC Symposium 2020 CY - 27 Feb 2020 - 28 Feb 2020, Jülich (Germany) Y2 - 27 Feb 2020 - 28 Feb 2020 M2 - Jülich, Germany LB - PUB:(DE-HGF)8 ; PUB:(DE-HGF)7 UR - https://juser.fz-juelich.de/record/874426 ER -