TOWARD HIGH-PERFORMANCE COMPUTATIONAL CHEMISTRY .1. SCALABLE FOCK MATRIX CONSTRUCTION ALGORITHMS

Several parallel algorithms for Fock matrix construction are described . The algorithms calculate only the unique integrals, distribute the F ock and density matrices over the processors of a massively parallel c omputer, use blocking techniques to construct the distributed data str uctures, and use clustering techniques on each processor to maximize d ata reuse. Algorithms based on both square and row-blocked distributio ns of the Fock and density matrices are described and evaluated. Varia nts of the algorithms are discussed that use either triple-sort or can onical ordering of integrals, and dynamic or static task clustering sc hemes. The algorithms are shown to adapt to screening, with communicat ion volume scaling down with computation costs. Modeling techniques ar e used to characterize algorithm performance. Given the characteristic s of existing massively parallel computers, all the algorithms are sho wn to be highly efficient for problems of moderate size. The algorithm s using the row-blocked data distribution are the most efficient. (C) 1996 by John Wiley & Sons, Inc.