Parallelizing a highly vectorized multigrid code with zebra relaxation

After a brief introduction to multigrid methods, the author discusses some of the algorithmic choices in MGZEB, a parallelized highly vectorized multigrid code for the solution of linear systems resulting from the seven-point discretization of general linear second-order elliptic partial differential equations in two dimensions. He describes the minimization of the scalar operation count, the vector tuning on a vector-register machine, and the parallelization of the already existing highly vectorized MGZEB code. At present the same algorithm would be used with the same scalar operation count on a scalar uniprocessor. The overall parallel vector-performance using autotasking on a Cray Y-MP4/464 is discussed.<
>