A parallel preconditioned conjugate gradient package for solving sparse linear systems on a Cray Y-MP (Q1181505): Difference between revisions

A package of preconditioned conjugate gradient iterative methods (Cray PCG) was developed to solve linear systems \(Ax=b\) (\(A\) is a general large sparse matrix) on shared-memory, vector/parallel computers such as a Cray Y-MP. The iterative methods implemented in Cray PCG have three types of operations: sparse matrix times a vector, setup and solution of a sparse triangular system, vector updates and dot products. Two basic types of preconditioning are available in Cray PCG: explicit scaling of the linear system by a diagonal matrix and usual implicit preconditioning. Reverse communication is discussed to make the iterative routines independent of the data structures needed to perform the preconditioning and matrix-vector product. Parallel matrix-vector product and matrix- transpose-vector product algorithms, implementation of preconditioners (incomplete Cholesky, \(LU\) factorizations, and polynomial preconditioners) are dealt with. Computational results are shown in tables (solution times, number of iterations, errors) for 4 methods.