Analysis and parallelizationstrategies for Ruge-Stüben AMGon many-core processors

The Ruge-Stuben algebraic multigrid method (AMG) is an optimal-complexity black-box approach to solve linear systems arising in discretizations of e.g. elliptic PDEs. Recently, there has been a growing interest in parallelizing this method on many-core hardware, especially graphics processing units (GPUs). This type of hardware delivers high performance for highly parallel algorithms. In this work, we analyse convergence properties of recent AMG developments for many-core processors and propose to use more classical choices of AMG components for higher robustness. Based on these choices, we introduce many-core parallelization strategies for a robust hybrid many-core AMG. The strategies can be understood and applied without deep knowledge of a given many-core architecture. We use them to propose a new hybrid GPU implementation. The implementation is tested in an in-depth performance analysis, which outlines its good convergence properties and high performance in the solve phase