HP MLIB User's Guide Vol. 2 7th Ed.

Chapter 11 Introduction to Distributed SuperLU 719

Overview

11 Introduction to Distributed SuperLU

Overview

This chapter provides an overview of HP’s implementation of Distributed

SuperLU functionality.

This implementation provides the Distributed SuperLU library designed for

distributed memory parallel computers. It was based on the public-domain

SuperLU_DIST, which was developed at the Lawrence Berkeley National Lab

and the University of California at Berkeley.

HP’s Distributed SuperLU contains two libraries: SuperLU_DIST and

SuperLU_DIST8. The only difference between SuperLU_DIST and

SuperLU_DIST8 is that SuperLU_DIST uses 4-byte integers, but

SuperLU_DIST8 uses 8-byte integers. Each library contains a set of

subroutines to solve a sparse linear system. These libraries are implemented in

ANSI C, using HP Message Passing Interface (MPI) for communication. These

libraries include routines to handle both real and complex matrices using

double precision. The parallel routine names for the double-precision real

version start with the letters “pd” (e.g. pdgstrf). The parallel routine names for

the double-precision complex version start with letters “pz” (e.g. pzgstrf).

The routines can be called directly from C applications. They may also be called

from Fortran applications, however, “bridge” routines (in C) must be supplied.

For details about creating bridge routines, please refer to Section 2.9.2 in the

SuperLU User’s Guide available at http://www.nersc.gov/~xiaoye/SuperLU.

NOTE The SuperLU_DIST8 library assumes all integers are 8 bytes in length

but array lengths should not exceed 4-byte integer values. That is,

SuperLU_DIST8 integer arrays can store any 8-byte integer value, but

the length of the arrays must be smaller than 232. This reﬂects the

current HP MPI 1.8 functionality.