MGMPI: A Parallel Multigrid Software Library

J. Bordner
University of California, San Diego
9500 Gilman Drive Dept. 0424
La Jolla, CA 92093-0424


MGMPI ( is a parallel multigrid software library under development at the Laboratory for Computational Astrophysics, currently located at the University of California, San Diego. MGMPI is written using a combination of C++ and Fortran-77, and is designed to solve 3D linear elliptic PDEs discretized on logically Cartesian grids. Parallelism is obtained by partitioning grids into subregions, and using MPI for communication between these subregions.

Current multigrid components include V-, W-, and F-cycle methods, operator-based prolongation and restriction, coarse-grid problem generation using harmonic averaging or Galerkin coarsening, and weighted point Jacobi and point red-black Gauss Seidel smoothers. Coarse-grid problems can be solved using one of a variety of Krylov subspace solvers, including CG and BiCG-STAB, or by applying any of the available smoothers.

Future components we hope to implement soon include semi-coarsening, and alternating line- and plane-smoothers. We also intend to include fast computational kernels that exploit properties of multi-level memory hierarchies, and to optimize parallel performance.