The parallel properties of three fast direct solution methods for linear systems with separable block tridiagonal matrices and a related C/MPI code are studied. Fast algorithm for separation of variables and two new variants of the generalized marching algorithm are first summarized. The results from numerical tests performed on two coarse-grained parallel architectures are then reported. The obtained speed-up and efficiency coefficients are compared. The presented results confirm that not always the best sequential solver has the best parallel performance.
展开▼
