A hybrid multigrid solver for high-order discontinuous Galerkin methods combining coarsening in the mesh size h and the polynomial degree p has been developed and is presented in this Master’s thesis. It can efficiently solve second-order partial differential equations for complex geometries on modern CPU hardware as well as on massively parallel and distributed systems. The given operator is rediscretized on every level, and the size of the coarse-grid problem is reduced as much as possible. Almost all multigrid components are evaluated in a highly efficient matrix-free way, based on sum factorization, and an algebraic multigrid solver is applied for solving the coarse-grid problem. A detailed investigation of a variety of multigrid design choices, including alternative p-coarsening strategies and the auxiliary space idea, is presented. The implementation efficiency of all multigrid components on their own is demonstrated using experimental measurements, conducted for standard finite element methods and discontinuous Galerkin methods for 2D and 3D, as well as for Cartesian and curved meshes. The results are compared to theoretical expectations. The overall efficiency of the developed hybrid multigrid solver is demonstrated by its application to the Poisson problem, to the convection–diffusion equation, and to the unsteady incompressible Navier–Stokes equations. Strong- and weak-scaling results, showing excellent parallel efficiency, as well as a novel strong-scaling model of the developed hybrid multigrid solver are presented.
«
A hybrid multigrid solver for high-order discontinuous Galerkin methods combining coarsening in the mesh size h and the polynomial degree p has been developed and is presented in this Master’s thesis. It can efficiently solve second-order partial differential equations for complex geometries on modern CPU hardware as well as on massively parallel and distributed systems. The given operator is rediscretized on every level, and the size of the coarse-grid problem is reduced as much as possible. Alm...
»