CONJUGATE GRADIENTS PARALLELIZED ON THE HYPERCUBE

Abstract

For the solution of discretized ordinary or partial differential equations it is necessary to solve systems of equations with coefficient matrices of different sparsity pattern, depending on the discretization method; using the finite element method (FE) results in largely unstructured systems of equations. A frequently used iterative solver for systems of equations is the method of conjugate gradients (CG) with different preconditioners. On a multiprocessor system with distributed memory, in particular the data distribution and the communication scheme depending on the used data struture are of greatest importance for the efficient execution of this method. Here, a data distribution and a communication scheme are presented which are based on the analysis of the column indices of the non-zero matrix elements. The performance of the developed parallel CG-method was measured on the distributed-memory-system INTEL iPSC/860 of the Research Centre Jülich with systems of equations from FE-models. The parallel CG-algorithm has been shown to be well suited for both regular and irregular discretization meshes, i.e. for coefficient matrices of very different sparsity pattern.