A Parallel Row Distributed Linear Algebra System (PRDLA Users' Guide)

Abstract

The purpose of this system is to provide an easy-to-use et of basic parallel matrix manipulation subroutines, in C, for use on the Intel iPSC/860 hypercube. Most of the subroutines exhibit good performance, but not necessarily optimal. Users are welcome to use these subroutines as building blocks in their own codes, to modify if the situation warrants this, or just to study the implementation to help guide their own programming. The subroutines all operate under the assumption that matrices are distributed by row in a natural wrap mapping. That is, if we assume p processors, row 1 is mapped to processor 0, row 2 is mapped to processor 1, row p is mapped to processor p-1, row p+1 is mapped to processor 0, etc. If this is deemed to restrictive, the user is welcome to use these codes as a starting point and develop their own codes (perhaps under less restrictive assumptions). The codes are not directly portable to other computer systems. However, the codes do use PICL (the portable instrumented communica- tion library, developed at Oak Ridge National Laboratory); therefore, provided PICL is available the codes are easily ported.