Parallelization of a block tridiagonal solver in HPF on an IBM SP2
The aim of this article is to describe how an existing FORTRAN-code for the solution of block tridiagonal systems of linear equations can be parallelized by using High Performance Fortran (HPF). The algorithm we consider in this paper, consists of a complete LU decomposition. In order to obtain a well-parallelizable algorithm, a simultaneous reordering of both rows and columns of the coefficient matrix is performed before the LU decomposition is constructed. Numerical results obtained on an IBM SP2 using the x1hpf-compiler will be compared with numerical results obtained on a Cray T3D.
Keywordsparallel linear solver distributed memory HPF IBM SP2 Cray T3D
Unable to display preview. Download preview PDF.
- 1.D. Heller. Some aspects of the cyclic reduction algorithm for block tridiagonal linear systems. SIAM J. Numer. Anal., 13:484–496, 1978.Google Scholar
- 2.W. Kerner, S. Poedts, J.P. Goedbloed, G.T.A. Huysmans, B. Keegan, and E. Schwartz.-. In P. Bachman and D.C. Robinson, editors, Proceedings of 18th Conference on Controlled Fusion and Plasma Physics. EPS: Berlin, 1991. IV.89–IV.92.Google Scholar
- 3.M. Nool and A. van der Ploeg. A Parallel Jacobi-Davidson Method for solving Generalized Eigenvalue Problems in linear Magnetohydrodynamics. Technical report, CWI, 1997. to appear.Google Scholar
- 4.G.L.G. Sleijpen and H.A. van der Vorst. A Jacobi-Davidson iteration method for linear eigenvalue problems. SIAM Journal on Matrix Analysis and Applications, 17(2), April 1996.Google Scholar
- 5.A. van der Ploeg. Reordering Strategies and LU-decomposition of Block Tridiagonal Matrices for Parallel Processing. Technical Report NM-R9618, CWI, October 1996.Google Scholar