Performance and implementation of a geometric multigrid solver with Trilinos

by Mr Matthias Frey (Paul Scherrer Institut)

Thursday 14 Jun 2018, 14:00 → 15:00 Europe/Zurich

OHSA/E13 (Paul Scherrer Institu)

Abstract: The Poisson problem arising from large-scale N-body problems cou- pled with Maxwell's equations in the electrostatic limit represents an accuracy and efficiency bottleneck in simulations of neighbouring bunch eects in high intensity cyclotrons. Standard particle-in-cell models are not able to capture the tiny eects between bunches without wasting memory in regions of void due to the uniformity of the ne mesh. Block-structured adaptive mesh refinement algorithms are a suitable method to overcome this issue. Their hierarchy of levels and grids is applied to solve Poisson's equation using an adaptive geometric multgrid algorithm. This talk presents a new implementation of Martin's and Cartwright's algorithm based on Trilinos. Furthermore, a benchmark study of various preconditioners and solvers is shown with a comparison to AMReX's multigrid solver. A scalability test up to 13'824 cores shows a parallel efficiency of around 60% on Piz Daint.