Coarse-grain parallelizations of interval algorithms decomposing dense matrices and solving triangular systems on multicomputers (Q1340868)

Computing a scalar product in the usual way by adding up its summands can result in substantial losses in accuracy due to round-off. Using a `long accumulator' (in hard- or software), scalar products can be evaluated to maximum accuracy, thus ranking the optimal scalar product at the level of an elementary operation just as addition or multiplication. The paper develops methods for the parallel LU-factorization of a matrix and for solving triangular systems in a distributed memory, message- passing environment. All these methods make efficient use of the optimal scalar product. They therefore increase the numerical stability significantly. The methods require the matrix to be distributed by rows rather than by columns. Both, real and interval matrix entries are considered. The authors also present numerical experiments on a T805 transputer system under OCCAM-2.