A class of alternate strip-based domain decomposition methods for elliptic partial differential Equations

Abstract

The domain decomposition strategies proposed in this thesis are efficient preconditioning techniques with good parallelism properties for the discrete systems which arise from the finite element approximation of symmetric elliptic boundary value problems in two and three-dimensional Euclidean spaces. For two-dimensional problems, two new domain decomposition preconditioners are introduced, such that the condition number of the preconditioned system is bounded independently of the size of the subdomains and the finite element mesh size. First, the alternate strip-based (ASB2) preconditioner is based on the partitioning of the domain into a finite number of nonoverlapping strips without interior vertices. This preconditioner is obtained from direct solvers inside the strips and a direct fast Poisson solver on the edges between strips, and contains two stages. At each stage the strips change such that the edges between strips at one stage are perpendicular on the edges between strips at the other stage. Next, the alternate strip-based substructuring (ASBS2) preconditioner is a Schur complement solver for the case of a decomposition with multiple nonoverlapping subdomains and interior vertices. The subdomains are assembled into nonoverlapping strips such that the vertices of the strips are on the boundary of the given domain, the edges between strips align with the edges of the subdomains and their union contains all of the interior vertices of the initial decomposition. This preconditioner is produced from direct fast Poisson solvers on the edges between strips and the edges between subdo- mains inside strips, and also contains two stages such that the edges between strips at one stage are perpendicular on the edges between strips at the other stage. The extension to three-dimensional problems is via solvers on slices of the domain.