Installation guide : direct solver

MUMPS - direct solver for real and complex valued unsymmetric matrices

Installing packages in Debian, Ubuntu or Redhat (MPI version)

For the Linux distributions Debian or Ubuntu install the package

libmumps-dev

using the packet manager. This package includes the library (installed to where the MPI libraries are installed, type locate libdmumps to find them) and the include files (copied to /usr/include).

Installing packages in Debian, Ubuntu or Redhat (sequential version)

For the Linux distributions Debian or Ubuntu install the package

libmumps-seq-dev

using the packet manager. This package includes the library (copied to /usr/lib or /usr/lib64) and the include files (copied to /usr/include/mumps_seq).

Known problems
  • MPI does not compile anymore. This is due to the fact that some partial MPI implementations are in /usr/include/mumps_seq. If you want to use the MPI version of MUMPS, install
    libmumps-dev
    instead.

Direct installation

If your Linux distribution is not able to install Debian or RPM packages follow this guide to install the library yourself.

BLACS

The BLACS (Basic Linear Algebra Communication Subprograms) project is an ongoing investigation whose purpose is to create a linear algebra oriented message passing interface that may be implemented efficiently and uniformly across a large range of distributed memory platforms. It's used by Scalapack.

  • Download both Blacs and Blacs patch and unpack them to $(CONCEPTS_DEV)/../BLACS. This can be done also by
    cd $(CONCEPTS_DEV)/.. 
    wget http://www.netlib.org/blacs/mpiblacs.tgz
    tar -xzf mpiblacs.tgz
    wget http://www.netlib.org/blacs/mpiblacs-patch03.tgz
    tar -zxf mpiblacs-patch03.tgz
    cd BLACS
  • Copy the file BMAKES/Bmake.MPI-LINUX into Bmake.inc and edit it. You need to change lines 50 ( MPIdir) to precise where your MPI installation is, and you need to replace line 53 ( MPILIB) by
   MPILIB = -L$(MPILIBdir) -lmpi -lmpi_f77 -lmpi_f90 -lmpi_cxx

if you're using OpenMPI.

You also need to replaces lines 198, 199 and 204 by

   F77            = gfortran
   F77NO_OPTFLAGS = -fPIC -I$(MPIINCdir)
   CCFLAGS        = -fPIC -O4 -I$(MPIINCdir)

  • Compile the library typing make mpi

ScaLAPACK

The ScaLAPACK (Scalable Linear Algebra PACKage) is a library of high-performance linear algebra routines for parallel distributed memory machines. ScaLAPACK solves dense and banded linear systems, least squares problems, eigenvalue problems, and singular value problems.

  • Download Scalapack and unpack it to $(CONCEPTS_DEV)/../scalapack-2.0.2. This can be done also by
    cd $(CONCEPTS_DEV)/.. 
    wget http://www.netlib.org/scalapack/scalapack-2.0.2.tgz
    tar -xzf scalapack-2.0.2.tgz
    cd scalapack-2.0.2

  • Run the configuration tool, typing ccmake .. Activate advanced editing (with key 't'), and edit both CMAKE_C_FLAGS and CMAKE_FORTRAN_FLAGS to add -fPIC., then press 'c' to run configuration script, and 'g' to generate and exit.

  • Compile it, typing make scalapack.

  • Modify the flag in $(CONCEPTS_DEV)/boost-build-modules.jam, to put
    SCALAPACK_PATH = $(CONCEPTS_DEV)/../scalapack-2.0.2/lib

MUMPS

http://graal.ens-lyon.fr/MUMPS (MUltifrontal Massively Parallel sparse direct Solver) is a massively parallel solver mainly developed at ENS Lyon.

  • Request http://graal.ens-lyon.fr/MUMPS/index.php?page=dwnld|here for downloading the library.
  • Get the file MUMPS_x.xx.x.tar.gz using the download link provided in your mail (in our example, MUMPS-5.1.1.tar.gz) and copy it to the same root folder as BLACS and ScaLAPACK, i.e. $(CONCEPTS_DEV)/..).
  • Unpack it using tar -zxf MUMPS-5.1.1.tar.gz, and go inside folder MUMPS-5.1.1
  • Copy the file Make.inc/Makefile.inc.generic = into =Makefile.inc and edit it. For example, here is the difference between the Make.inc/Makefile.inc.generic and the modified version Makefile.inc
94c94
< CC      = cc
---
> CC      = gcc
96c96
< FC      = f90
---
> FC      = gfortran
98c98
< FL      = f90
---
> FL      = gfortran
108c108
< SCALAP  = -lscalapack -lblacs
---
> SCALAP  = -L$(CONCEPTS_DEV)/../scalapack-2.0.2/lib -lscalapack $(CONCEPTS_DEV)/../BLACS/LIB/blacs_MPI-LINUX-0.a $(CONCEPTS_DEV)/../BLACS/LIB/blacsF77init_MPI-LINUX-0.a $(CONCEPTS_DEV)/../BLACS/LIB/blacsCinit_MPI-LINUX-0.a
111c111
< INCPAR  = -I/usr/include
---
> INCPAR  = -I/usr/mpi/gcc/openmpi-1.4.2/include
114c114
< LIBPAR  = $(SCALAP) -L/usr/lib -lmpi
---
> LIBPAR  = $(SCALAP) -L/usr/mpi/gcc/openmpi-1.4.2/lib64 -lmpi
143,145c143,145
< OPTF    = -O
< OPTC    = -O -I.
< OPTL    = -O
---
> OPTF    = -fPIC -O
> OPTC    = -fPIC -O -I.
> OPTL    = -fPIC -O
  • Compile using make alllib.
  • Modify the flag in $(CONCEPTS_DEV)/boost-build-modules.jam, to put
    MUMPS_PATH = $(CONCEPTS_DEV)/../MUMPS-5.0.2/lib
    and
    MUMPS_INCLPATH = $(CONCEPTS_DEV)/../MUMPS-5.0.2/include

Umfpack - direct solver for real valued unsymmetric matrices

Umfpack is a direct solver for real and complex matrices, however only the interface to routines for real matrices is implemented (version ≥ 4.1). Ubuntu or Debian:
  • Install the package libsuitesparse-dev. If only the package libsuitesparse is installed, the package will not be found creating an error like
    gcc.compile.c++ builds/operator/gcc-4/debug/umfpack.o
    operator/umfpack.cc:9:23: error: umfpack.h: No such file or directory 

Important: Make sure that the libraries libamd.so, libcamd.so, libcolamd.so, libccolamd.so, libcolord.so, libcholmod.so and libsuitesparseconfig.so are installed into /usr/lib or /usr/lib64. Otherwise specify the path UMFPACK_PATH in $(CONCEPTS_DEV)/boost-build-modules.jam. Furthermore, make sure that the include file umfpack.h is located in /usr/include, /usr/include/ufpack, or /usr/include/suitespare. Otherwise specify the include path UMFPACK_INCLPATH in $(CONCEPTS_DEV)/boost-build-modules.jam. Note that the libraries need to have exactly the name as given above. If they should have a suffix with the version number for example, you have to create symbolic links.

SuperLU - direct solver for real and complex valued unsymmetric matrices

Installing packages in Debian, Ubuntu or Redhat

For the Linux distributions Debian or Ubuntu install the package libsuperlu3-dev, or any higher version, using your packet manager. This package includes the library (installed to /usr/lib or /usr/lib64) and the include files (copied to /usr/include or /usr/include/superlu).

The SuperLU packages (RPM) for Redhat includes the library and includes files as well.

Direct installation

If your Linux distribution is not able to install Debian or RPM packages follow this guide to install the library yourself. The guide refers to installing SuperLU of version 3.1 but it can easily be adjusted for any other version.
  • Download SuperLU 3.1 and unpack it to /scratch/kersten/SuperLU (when Concepts is for example in /scratch/kersten/concepts-2). This can be done also by
    cd /scratch/kersten 
    wget http://crd.lbl.gov/~xiaoye/SuperLU/superlu_3.1.tar.gz
    tar -xzf superlu_3.1.tar.gz
    mv SuperLU_3.1 SuperLU
    cd SuperLU
  • Copy the file make.inc into /scratch/kersten/SuperLU
  • If your gcc is 4.xx or newer you have to edit make.inc and use something like
    SuperLUroot = /scratch/kersten/SuperLU_3.1 
    SUPERLULIB = $(SuperLUroot)/libsuperlu.a
    BLASLIB = /usr/lib/libblas.a
    TMGLIB = libtmglib.a
    LIBS = $(SUPERLULIB) $(BLASLIB)
    CC = gcc-3.4.2
    CFLAGS = -O3 -fPIC -fomit-frame-pointer -g
    FORTRAN = gfortran
    FFLAGS = -O -fPIC
    LOADER = gcc-3.4.2
    LOADOPTS = -O3 -fPIC
  • Change in the file /scratch/kersten/SuperLU/SRC/Makefile
    superlu_timer.o: superlu_timer.c ; $(CC) -c $(NOOPTS) $<
    into
    superlu_timer.o: superlu_timer.c ; $(CC) -c $(NOOPTS) -fPIC $< 
  • execute in /scratch/kersten/SuperLU
    make
  • When building Concepts (call of bjam) it indicates, that it finds SuperLU with the line
    * Activating support for direct solver SuperLU.
Known problems
  • Pre-installed BLAS in /usr/lib/libblas.a
    /scratch/kersten/SuperLU/lib/libsuperlu_3.1.a(xerbla.o):/scratch/kersten/SuperLU/SRC/xerbla.c:5: first defined here /usr/bin/ld: Warning: size of symbol `xerbla_' changed from 30 in /scratch/kersten/SuperLU/lib/libsuperlu_3.1.a(xerbla.o) to 38 in /usr/lib/libblas.a(xerbla.o)
    Solution: deactive xerbla.o in SRC/Makefile by changing
    LAAUX = lsame.o
    into
    LAAUX = lsame.o #xerbla.o

Page URL: http://wiki.math.ethz.ch/bin/view/Concepts/InstallationGuideDirectSolver
2020-10-26
© 2020 Eidgenössische Technische Hochschule Zürich