SuperLU_MT contains a set of subroutines to solve a sparse linear system
A*X=B. It uses Gaussian elimination with partial pivoting (GEPP).
The columns of A may be preordered before factorization; the
preordering for sparsity is completely separate from the factorization.
SuperLU_MT is a parallel extension to the serial SuperLU library.
SuperLU_MT is implemented in ANSI C, with multithreading extension,
for example, using POSIX threads or OpenMP. Currently, only the LU
factorization routine, which is the most time-consuming part of the
solution process, is parallelized on machines with a shared address space.
The other routines, such as column preordering and the forward and
back substitutions are performed sequentially.
The library provides functionality for both real and complex
matrices, in both single and double precision.
The distribution contains the following directory structure:
SuperLU_MT_2.2/README instructions on installation
SuperLU_MT_2.2/CBLAS/ BLAS routines in C, functional but not fast
SuperLU_MT_2.2/DOC/ Users’ Guide
SuperLU_MT_2.2/EXAMPLE/ example programs
SuperLU_MT_2.2/INSTALL/ test machine dependent parameters; the Users’ Guide.
SuperLU_MT_2.2/MAKE_INC/ sample machine-specific make.inc files
SuperLU_MT_2.2/SRC/ C source code, to be compiled into libsuperlu_mt.a
SuperLU_MT_2.2/TESTING/ driver routines to test correctness
SuperLU_MT_2.2/lib/ SuperLU_MT library archive libsuperlu_mt.a
SuperLU_MT_2.2/Makefile top level Makefile that does installation and testing
SuperLU_MT_2.2/make.inc compiler, compile flags, library definitions and C
preprocessor definitions, included in all Makefiles.
(You may need to edit it to suit your system
before compiling the whole package.)
Before installing the package, please examine the three things dependent
on your system setup:
Edit the make.inc include file.
This make include file is referenced inside each of the Makefiles
in the various subdirectories. As a result, there is no need to
edit the Makefiles in the subdirectories. All information that is
machine specific has been defined in this include file.
The following machine-specific make.inc files are provided in the
MAKE_INC/ directory:
make.cray -D__CRAY Cray C90/J90
make.ibm -D__PTHREAD IBM Power series
make.origin -D__ORIGIN SGI/Cray Origin2000
make.sun -D__SOLARIS Sun Ultra Enterprise servers
In each case, the CFLAGS should include a CPP definition to choose the
proper thread program interface. For example, in make.pthread, need:
CFLAGS = -D__PTHREAD ...
When you have selected the machine to which you wish to install SuperLU_MT,
copy the appropriate sample include file (if one is present) into
make.inc. For example, if you wish to run SuperLU_MT on a XE6 system,
you can do
cp MAKE_INC/make.xe6 make.inc
For the systems other than listed above, some porting effort is needed
for parallel factorization routines. Please refer to the Users’ Guide
for detailed instructions on porting.
The other CPP definitions can be set in CFLAGS as follows:
o -D_LONGINT
use 64-bit integers for indexing sparse matrices. (default is 32-bit)
o -DPRNTlevel=[0,1,2,…]
printing level to show solver's execution details. (default is 0)
o -DDEBUGlevel=[0,1,2,…]
diagnostic printing level for debugging purpose. (default is 0)
The BLAS library.
** NOTE: must link with a single-thread BLAS library. *
The parallel routines in SuperLU_MT uses some sequential BLAS routines
within each process (or thread). If there is BLAS library available on
your machine, you may define the following in the file make.inc:
BLASDEF = -DUSE_VENDOR_BLAS
BLASLIB = <BLAS library you wish to link with>
The CBLAS/ subdirectory contains the part of the C BLAS needed by
SuperLU_MT package. However, these codes are intended for use only if
there is no faster implementation of the BLAS already available on your
machine. In this case, you should go to the top-level SuperLU_MT/
directory and do the following:
In make.inc, undefine (comment out) BLASDEF, and define:
BLASLIB = ../lib/libblas$(PLAT).a
Type:
make blaslib
to make the BLAS library from the routines in the CBLAS/ subdirectory.
C preprocessor definition CDEFS.
In the header file SRC/Cnames.h, we use macros to determine how
C routines should be named so that they are callable by Fortran.
(Some vendor-supplied BLAS libraries do not have C interface. So the
re-naming is needed in order for the SuperLU BLAS calls (in C) to
interface with the Fortran-style BLAS.)
The possible options for CDEFS are:
o -DAdd_: Fortran expects a C routine to have an underscore
postfixed to the name;
o -DNoChange: Fortran expects a C routine name to be identical to
that compiled by C;
o -DUpCase: Fortran expects a C routine name to be all uppercase.
A Makefile is provided in each subdirectory. The installation can be done
completely automatically by simply typing “make” at the top level.
REFERENCES
[1] A Supernodal Approach to Sparse Partial Pivoting,
James W. Demmel, Stanley C. Eisenstat, John R. Gilbert, Xiaoye S. Li
and Joseph W.H. Liu,
SIAM J. on Matrix Anal. and Appl., vol 20(3), 720-755, 1999.
[2] An Asynchronous Parallel Supernodal Algorithm for Sparse Gaussian
Elimination, James W. Demmel, John R. Gilbert and Xiaoye S. Li,
SIAM J. Matrix Anal. Appl., vol. 20(4), 915-952, 1999.
[3] Sparse Gaussian Elimination on High Performance Computers,
Xiaoye S. Li, Tech report UCB//CSD-96-919, Computer Science Division,
U.C. Berkeley, September, 1996, Ph.D dissertation.
SuperLU_MT contains a set of subroutines to solve a sparse linear system A*X=B. It uses Gaussian elimination with partial pivoting (GEPP). The columns of A may be preordered before factorization; the preordering for sparsity is completely separate from the factorization.
SuperLU_MT is a parallel extension to the serial SuperLU library. SuperLU_MT is implemented in ANSI C, with multithreading extension, for example, using POSIX threads or OpenMP. Currently, only the LU factorization routine, which is the most time-consuming part of the solution process, is parallelized on machines with a shared address space. The other routines, such as column preordering and the forward and back substitutions are performed sequentially. The library provides functionality for both real and complex matrices, in both single and double precision.
The distribution contains the following directory structure:
SuperLU_MT_2.2/README instructions on installation SuperLU_MT_2.2/CBLAS/ BLAS routines in C, functional but not fast SuperLU_MT_2.2/DOC/ Users’ Guide SuperLU_MT_2.2/EXAMPLE/ example programs SuperLU_MT_2.2/INSTALL/ test machine dependent parameters; the Users’ Guide. SuperLU_MT_2.2/MAKE_INC/ sample machine-specific make.inc files SuperLU_MT_2.2/SRC/ C source code, to be compiled into libsuperlu_mt.a SuperLU_MT_2.2/TESTING/ driver routines to test correctness SuperLU_MT_2.2/lib/ SuperLU_MT library archive libsuperlu_mt.a SuperLU_MT_2.2/Makefile top level Makefile that does installation and testing SuperLU_MT_2.2/make.inc compiler, compile flags, library definitions and C preprocessor definitions, included in all Makefiles. (You may need to edit it to suit your system before compiling the whole package.)
Before installing the package, please examine the three things dependent on your system setup:
Edit the make.inc include file. This make include file is referenced inside each of the Makefiles in the various subdirectories. As a result, there is no need to edit the Makefiles in the subdirectories. All information that is machine specific has been defined in this include file.
The following machine-specific make.inc files are provided in the MAKE_INC/ directory:
make.inc CPP in CFLAGS Platforms
make.xe6 OpenMP or POSIX threads
In each case, the CFLAGS should include a CPP definition to choose the proper thread program interface. For example, in make.pthread, need:
When you have selected the machine to which you wish to install SuperLU_MT, copy the appropriate sample include file (if one is present) into make.inc. For example, if you wish to run SuperLU_MT on a XE6 system, you can do
For the systems other than listed above, some porting effort is needed for parallel factorization routines. Please refer to the Users’ Guide for detailed instructions on porting.
The other CPP definitions can be set in CFLAGS as follows: o -D_LONGINT
o -DPRNTlevel=[0,1,2,…]
o -DDEBUGlevel=[0,1,2,…]
The BLAS library.
** NOTE: must link with a single-thread BLAS library. *
The parallel routines in SuperLU_MT uses some sequential BLAS routines within each process (or thread). If there is BLAS library available on your machine, you may define the following in the file make.inc:
The CBLAS/ subdirectory contains the part of the C BLAS needed by SuperLU_MT package. However, these codes are intended for use only if there is no faster implementation of the BLAS already available on your machine. In this case, you should go to the top-level SuperLU_MT/ directory and do the following:
In make.inc, undefine (comment out) BLASDEF, and define: BLASLIB = ../lib/libblas$(PLAT).a
Type: make blaslib to make the BLAS library from the routines in the CBLAS/ subdirectory.
C preprocessor definition CDEFS. In the header file SRC/Cnames.h, we use macros to determine how C routines should be named so that they are callable by Fortran. (Some vendor-supplied BLAS libraries do not have C interface. So the re-naming is needed in order for the SuperLU BLAS calls (in C) to interface with the Fortran-style BLAS.) The possible options for CDEFS are:
A Makefile is provided in each subdirectory. The installation can be done completely automatically by simply typing “make” at the top level.
REFERENCES
[1] A Supernodal Approach to Sparse Partial Pivoting, James W. Demmel, Stanley C. Eisenstat, John R. Gilbert, Xiaoye S. Li and Joseph W.H. Liu, SIAM J. on Matrix Anal. and Appl., vol 20(3), 720-755, 1999.
[2] An Asynchronous Parallel Supernodal Algorithm for Sparse Gaussian Elimination, James W. Demmel, John R. Gilbert and Xiaoye S. Li, SIAM J. Matrix Anal. Appl., vol. 20(4), 915-952, 1999.
[3] Sparse Gaussian Elimination on High Performance Computers, Xiaoye S. Li, Tech report UCB//CSD-96-919, Computer Science Division, U.C. Berkeley, September, 1996, Ph.D dissertation.
Xiaoye S. Li, Lawrence Berkeley National Lab, xsli@lbl.gov James Demmel, UC Berkeley, demmel@cs.berkeley.edu John R. Gilbert, UC Santa Barbara, gilbert@cs.ucsb.edu