On bandsolver using skyline storage

Abstract

A simple out-of-core bandsolver using skyline storage is presented. Without dividing the equations into blocks, the efficiency of the working vector can be greatly improved by a dynamic storage scheme to hold as many equations as possible at any one time. Compared to the blocked-skyline method, the total number of data transfers from core to disk and vice versa is reduced and the amount of active coefficients being processed in a single data transfer is maximized. The matrix decomposition without using direct-access files is first presented to explain how triangulation is done on a one-dimensional working array using skyline profile storage. Its extension to allow for out-of-core storage is then discussed in detail. Computer codes for both versions with and without using out-of-core storage are given in the Appendix. © 1992.