Speaker
Description
We present a tutorial on PSCToolkit, a set of libraries for high-performance linear solvers on parallel computers. The toolkit provides many advanced features besides the usual Krylov solvers, such as support for Algebraic Multigrid (AMG) preconditioning techniques based on the aggregation of unknowns, employing both standard strength-of-connection measures and graph-matching strategies. Moreover, it includes a variety of highly parallel smoothers, such as polynomial accelerators for weighted Jacobi methods, Additive Schwarz schemes, and block-Jacobi variants of Gauss–Seidel iterations and approximate inverse approaches. The toolkit provides easy mechanisms to run on GPUs supporting the CUDA programming environment; it is currently being integrated into dealii.X framework for use in the project; it has been tested for scaling on the Leonardo platform up to 8192 GPUs with linear systems of size up to 10^11.
