Evaluation of two acceleration techniques in a multithreaded 2D Poisson equation solver

Two acceleration techniques, based on additive corrections are evaluated with a multithreaded 2D Poisson equation solver. The popular multigrid algorithm with 2-level grid is compared with the traditional block-correction strategy. In both, single-processor and distributed architectures, block correction is faster than the multigrid due mainly to the smaller cost that the solution of a 1D linear system has over one 2D linear system. Results in both cluster tested show that block correction can reduce significantly the computing time in the solution of very large linear systems. These calculations confirm that the Red/Black ordering is effective only if data fit entirely in cache memory.