Multigrid preconditioners, coupled with Krylov iterative solvers, are widely used in the parallel solution of large and sparse linear systems, because of their optimality in the solution of linear systems arising from the discretization of scalar elliptic Partial Differential Equations (PDEs) on regular grids. Optimality, also known as algorithmic scalability, is the property of having a computational cost per iteration that depends linearly on the problem size, and a convergence rate that is independent of the problem size.
Multigrid preconditioners are based on a recursive application of a two-grid process consisting of smoother iterations and a coarse-space (or coarse-level) correction. The smoothers may be either basic iterative methods, such as the Jacobi and Gauss-Seidel ones, or more complex subspace-correction methods, such as the Schwarz ones. The coarse-space correction consists of solving, in an appropriately chosen coarse space, the residual equation associated with the approximate solution computed by the smoother, and of using the solution of this equation to correct the previous approximation. The transfer of information between the original (fine) space and the coarse one is performed by using suitable restriction and prolongation operators. The construction of the coarse space and the corresponding transfer operators is carried out by applying a so-called coarsening algorithm to the system matrix. Two main approaches can be used to perform coarsening: the geometric approach, which exploits the knowledge of some physical grid associated with the matrix and requires the user to define transfer operators from the fine to the coarse level and vice versa, and the algebraic approach, which builds the coarse-space correction and the associate transfer operators using only matrix information. The first approach may be difficult when the system comes from discretizations on complex geometries; furthermore, ad hoc one-level smoothers may be required to get an efficient interplay between fine and coarse levels, e.g., when matrices with highly varying coefficients are considered. The second approach performs a fully automatic coarsening and enforces the interplay between fine and coarse level by suitably choosing the coarse space and the coarse-to-fine interpolation (see, e.g., [3, 24, 22] for details.)
MLD2P4 uses a pure algebraic approach, based on the smoothed aggregation algorithm [2, 26], for building the sequence of coarse matrices and transfer operators, starting from the original one. A decoupled version of this algorithm is implemented, where the smoothed aggregation is applied locally to each submatrix [25]. A brief description of the AMG preconditioners implemented in MLD2P4 is given in Sections 4.1-4.3. For further details the reader is referred to [4, 5, 7, 8].
We note that optimal multigrid preconditioners do not necessarily correspond to minimum execution times in a parallel setting. Indeed, to obtain effective parallel multigrid preconditioners, a tradeoff between the optimality and the cost of building and applying the smoothers and the coarse-space corrections must be achieved. Effective parallel preconditioners require algorithmic scalability to be coupled with implementation scalability, i.e., a computational cost per iteration which remains (almost) constant as the number of parallel processors increases.