A note on the vectorization of algebraic multigrid algorithms (Q911699)

The authors discuss the vectorization of the algebraic multigrid method (AMG) for solving large sparse linear systems. The implementation is based on a combination of Gauss-Seidel and Jacobi relaxations that depends on work vector length. Experiments deal with the discretization of a self-adjoint elliptic equation on a square and with a random problem; they show that vectorized AMG performs considerably better than the sequential version, with respect to total time employed. The codes have been run on a Cray XMP machine and a Fujitsu vector processor.