Efficient Parallel Solvers for the FireStar3D Wildfire Numerical Simulation Model
Abstract
This paper presents efficient parallel methods for solving
ill-conditioned linear systems arising in fluid dynamics problems. The
first method is based on the Modified LU decomposition, applied as a
preconditioner to the Conjugate gradient algorithm. Parallelization of
this method is based on the use of nested twisted factorization. Another
method is based on a highly parallel Algebraic multigrid algorithm with a
new smoother developed for anisotropic grids. Performance comparisons
demonstrate superiority of new methods over commonly used variants of
the Conjugate gradient method.