Performance of Parallel Eigensolvers on Electronic Structure Calculations II
Robert C. Ward
Many models employed to solve problems in quantum mechanics, such as electronic structure calculations, result in both linear and nonlinear eigenproblems. Solutions to the nonlinear problems, such as the Self-Consistent Field method, typically involve iterative schemes requiring the solution of a large symmetric linear eigenproblem during each iteration. This paper evaluates the performance of various popular and new parallel symmetric linear eigensolvers applied to such eigenproblems in electronic structure calculations on the IBM distributed memory supercomputer at the Oak Ridge National Laboratory. Results using established routines from ScaLAPACK and vendor optimized packages are presented, as well as from three recently developed parallel eigensolvers, two implementations of the Multiple Relatively Robust Representations algorithm and the block divide-andconquer algorithm. This paper updates an earlier version of this work reported in University of Tennessee Technical Report UT-CS-05-560 by Ward, Bai & Pratt in 2005.
Published 2006-09-01 04:00:00 as ut-cs-06-572 (ID:130)