Open Access

Peter Münch

• For numerical computations based on finite element methods (FEM), it is common practice to assemble the system matrix related to the discretized system and to pass this matrix to an iterative solver. However, the assembly step can be costly and the matrix might become locally dense, e.g., in the context of high-order, high-dimensional, or strongly coupled multicomponent FEM, leading to high costs when applying the matrix due to limited bandwidth on modern CPU- and GPU-based hardware. Matrix-free algorithms are a means of accelerating FEM computations on HPC systems, by applying the effect of the system matrix without assembling it. Despite convincing arguments for matrix-free computations as a means of improving performance, their usage still tends to be an exception at the time of writing of this thesis, not least because they have not yet proven their applicability in all areas of computational science, e.g., solid mechanics. In this thesis, we further develop a state-of-the-artFor numerical computations based on finite element methods (FEM), it is common practice to assemble the system matrix related to the discretized system and to pass this matrix to an iterative solver. However, the assembly step can be costly and the matrix might become locally dense, e.g., in the context of high-order, high-dimensional, or strongly coupled multicomponent FEM, leading to high costs when applying the matrix due to limited bandwidth on modern CPU- and GPU-based hardware. Matrix-free algorithms are a means of accelerating FEM computations on HPC systems, by applying the effect of the system matrix without assembling it. Despite convincing arguments for matrix-free computations as a means of improving performance, their usage still tends to be an exception at the time of writing of this thesis, not least because they have not yet proven their applicability in all areas of computational science, e.g., solid mechanics. In this thesis, we further develop a state-of-the-art matrix-free framework for high-order FEM computations with focus on the preconditioning and adopt it in novel application fields. In the context of high-order FEM, we develop means of improving cache efficiency by interleaving cell loops with vector updates, which we use to increase the throughput of preconditioned conjugate gradient methods and of block smoothers based on additive Schwarz methods; we also propose an algorithm for the fast application of hanging-node constraints in 3D for up to 137 refinement configurations. We develop efficient geometric and polynomial multigrid solvers with optimized transfer operators, whose performance is experimentally investigated in detail in the context of locally refined meshes, indicating the superiority of global-coarsening algorithms. We apply the developed solvers in the context of novel stage-parallel implicit Runge–Kutta methods and demonstrate the benefit of stage–parallel solvers in decreasing the time to solution at the scaling limit. Novel challenging application fields of matrix-free computations include high-dimensional computational plasma physics, solid-state-sintering simulations with a high and dynamically changing number of strongly coupled components, and coupled multiphysics problems with evaluation and integration at arbitrary points. In the context of these fields, we detail computational challenges, propose modified versions of the standard matrix-free algorithms for high-performance computing, and discuss preconditioning-related topics. The efficiency of the derived algorithms on the node level and at extreme scales is demonstrated experimentally on SuperMUC-NG, one of Germany’s leading supercomputers, with up to 150k processes and by solving systems of up to 5 × 1012 unknowns. Such problem sizes would not be conceivable for equivalent matrix-based algorithms. The major achievements of this thesis allow to run larger simulations faster and more efficiently, enabling progress and new possibilities for a range of application fields in computational science.…