Solving almost block diagonal systems on parallel computers

Abstract Finite difference methods for nonlinear boundary value problems (BVP's) in ordinary differential equations involve solving systems of linear equations at an inner iteration. In the typical case of separated boundary conditions, these systems are almost block diagonal (ABD). A new ‘tearing’ algorithm for the parallel solution of those ABD's is presented here. It is an extension of one proposed by Dongarra and Johnsson [7] for positive definite or strictly diagonally dominant banded systems and it is similar to one proposed by Wright [15] for banded systems. We compare the cost of the new algorithm with costs of other algorithms which might be applied to ABD's. We simulate the use of the algorithm in solving BVP's. The new algorithm is not designed for a specific computer architecture and it is believed that the analysis is sufficiently general to be indicative of performance for most current parallel architectures.