Algebraic Multigrid (AMG) for Higher-Order Finite Elements

Luke Olson
Division of Applied Mathematics, Brown University

Jeff Heys, Arizona State University
Tom Manteuffel, University of Colorado at Boulder
Steve McCormick, University of Colorado at Boulder


Abstract

In this talk we consider two related approaches for solving linear systems that arise from a higher-order finite element discretization of elliptic partial differential equations. The first approach explores direct application of an algebraic-based multigrid method (AMG) to iteratively solve the linear systems that result from higher-order discretizations. While the choice of basis used on the discretization has a significant impact on the performance of the solver, results indicate that AMG is capable of solving operators from both Poisson's equation and a first-order system least-squares (FOSLS) formulation of Stoke's equation in a scalable manner, nearly independent of basis order, p. The second approach incorporates preconditioning based on a bilinear finite element mesh overlaying the entire set of degrees of freedom in the higher-order scheme. AMG is applied to the operator based on the bilinear finite element discretization and is used as a preconditioner in a conjugate gradient (CG) iteration to solve the algebraic system derived from the high-order discretization. This approach is also nearly independent of p. We present several numerical examples that support each method and discuss the computational advantages of the preconditioning implementation.