MueLu: Designing a New Multigrid Solver for the Trilinos Project
SESSION: Research Poster Reception
EVENT TYPE: Poster
TIME: 5:15PM - 7:00PM
AUTHOR(S):Jérémie Gaidamour, Jonathan J. Hu, Christopher M. Siefert, Ray S. Tuminaro
ABSTRACT: Smoothed Aggregation (SA) multigrid is a widely used algorithm
for the scalable solution of linear systems that are at the heart of many
DOE parallel scientific computations. However, SA's cost grows
unacceptably for problems involving multiple physics and large number of
near-kernel components. This issue becomes more pronounced for stretched meshes
large material variation, and higher dimensionality.
To address SA's shortcomings, we have developed a new algorithm
that allows explicit control of the grid transfer operators' sparsity patterns.
This ensures low cost while tailoring performance for anisotropic problems. A
key to the new algorithm is keeping good convergence by minimizing energies
of grid transfer operators while capturing important near-kernel components.
This algorithm is the basis of a new parallel multigrid solver in
the Trilinos project. The solver will leverage the modern Trilinos software
stack that abstracts the notion of a compute node in order to
harness multi-core CPUs and GPUs.
Jérémie Gaidamour - Sandia National Laboratories
Jonathan J. Hu - Sandia National Laboratories
Christopher M. Siefert - Sandia National Laboratories