Introduction

PLASMA installer is a set of python scripts developed to ease the installation of PLASMA (Parallel Linear Algebra Software for Multicore Architectures) and all software packages required by PLASMA: BLAS, CBLAS, LAPACK & libtmg, LAPACKE and hwloc. (Section PLASMA Software Stack provides a brief description for each package.)

None of these packages has to be present on the system prior to running the PLASMA installer in order for the installation to succeed. If any component is missing, the implementation provided by Netlib will be automatically downloaded and installed prior to PLASMA installation. The only exception is the hwloc package, which is optional and will not be automatically installed if not already present at the time of running the installer.

The user should not download the PLASMA tarball prior to running the installer. The installer automatically downloads the PLASMA tarball. This behavior can be overwritten if desired. (Section Installation Options provides more details.)

Installing BLAS

It is crucial that the user provides an optimized implementation of BLAS. If appropriate BLAS implementation is not installed prior to PLASMA installation, the PLASMA installer will download and install the reference FORTRAN implementation of BLAS from Netlib. This implementation should be considered the definition of BLAS rather than an implementation. It is completely unoptimized for modern processors (what no compiler flag can remedy). The use of Netlib BLAS will produce an order of magnitude lower performance than optimized BLAS.

A common mistake when attempting to run with optimized BLAS, is to forget to notify the OS about the location of the dynamic library. This it done by including the location of the dynamic library in the LD_LIBRARY_PATH environment variable, e.g.:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/path_to_blas/lib/

Since version 2.5.1, PLASMA can exploits multithreaded BLAS libraries to improve performance of eigenvalue solvers. However, every other algorithm still use only sequential library. It is then recommended to linked with a sequential BLAS library if eigenvalue/singular value solver is not of interest. If PLASMA is linked against a multithreaded BLAS library, the multithreading can be disabled by setting the appropriate environment variable from the command prompt, for instance:

export OMP_NUM_THREADS=1
export MKL_NUM_THREADS=1
export GOTO_NUM_THREADS=1
export VECLIB_MAXIMUM_THREADS=1

If eigenvalue/singular value solvers are needed, it is recommended to link PLASMA with multithreaded Intel MKL or AMD ACML BLAS library. In order to let PLASMA switch automatically the number of threads in the proper section, it is required to specify which BLAS library is used in the make.inc file by either adding -DPLASMA_WITH_MKL or -DPLASMA_WITH_ACML to the compilation flags (CFLAGS). Then, PLASMA will change the number of threads by calling respectively, omp_num_threads() or mkl_num_threads(). It is important to note that with MKL, PLASMA has to disable the affinity of the Intel OpenMP scheduler by calling: kmp_set_defaults("KMP_AFFINITY=disabled")

Installing hwloc

The hwloc package is an optional component and PLASMA can be build without it. However, if present, hwloc will help with mapping of the software to the hardware resources. hwloc can be downloaded and installed from source code (http://www.open-mpi.org/projects/hwloc/). However, hwloc is included in many Linux distributions and can be easily installed using standard package management tools, e.g. the apt-get command line utility (requires root access):

apt-get install hwloc

Also, PLASMA installer uses the pkg-config utility to detect and use hwloc. If not already present, pkg-config can be installed in the same way, e.g.:

apt-get install pkg-config

Also, the location of the “hwloc.pc” file (commonly /usr/lib/pkgconfig) has to be included in the PKG_CONFIG_PATH environment variable, e.g.:

export PKG_CONFIG_PATH=$PKG_CONFIG_PATH:/usr/lib/pkgconfig

Installation Options

The installer provides the following set of command line options (default settings are given in square brackets):

-h or --help        : Display this help and exit.
--prefix=[DIR]      : Install libraries in the DIR directory. [./install]
--build=[DIR]       : Build libraries in the DIR directory. [./build]
--cc=[CMD]          : Use CMD to compile C files. [cc]
--fc=[CMD]          : Use CMD to compile Fortran files. [gfortran]
--cflags=[FLAGS]    : Use FLAGS for the C compiler. [-02]
--fflags=[FLAGS]    : Use FLAGS for the Fortran compiler. [-O2]
--ldflags_c=[FLAGS] : Use FLAGS for linking using C compiler.
--ldflags_fc=[flags]: Use FLAGS for linking using Fortran compiler.
                      (By default ldflags_fc = ldflags_c.)
--make=[CMD]        : Use CMD as the ``make'' command. [make]
--blaslib=[LIB]     : Use LIB as the BLAS library.
                      (The path should be absolute if --prefix is used.)
--cblaslib=[LIB]    : Use LIB as the CBLAS library.
                      (The path should be absolute if --prefix is used.)
--lapacklib=[LIB]   : Use LIB as the LAPACK library.
                      (The path should be absolute if --prefix is used.)
--lapclib=[LIB]     : Use LIB as the LAPACK C API.
                      (The path should be absolute if --prefix is used.)
--downblas          : Download and install BLAS from Netlib.
--downcblas         : Download and install CBLAS from Netlib.
--downlapack        : Download and install LAPACK from Netlib.
--downlapc          : Download and install LAPACK C API from Netlib.
--downall           : Download and install all missing components from Netlib.
--[no]testing       : Enables/disables testings.
                      If enabled, all external libraries are required and tested.
                      (Enabled by default.)
--nbcores=[CORES]   : Use CORES number of cores for testing. [half of all the cores]
--clean             : Clean up the installer directory.
--src               : Generate the make.inc file for PLASMA based on the set of provided options.
                      Automatically download and install all the missing components.
                      (Testing is deactivated.)
                      This option is mostly meant for developers, who check out PLASMA sources from its repository.

The installer can also be used without the wget command or access to a network connection. Prior to running the installer the tarball of each missing component has to be placed in the build/download directory. The components can be found in the following locations:

Installer Usage Examples

For an installation with gcc, gfortran and Netlib BLAS type:

./setup.py --cc=gcc --fc=gfortran --downblas

For an installation with gcc, gfortran and vecLib (Mac OS X) type:

./setup.py --cc=gcc --fc=gfortran --blaslib="-framework veclib"

For an installation with gcc, gfortran and ATLAS type:

./setup.py --cc=gcc --fc=gfortran --blaslib="-lf77blas -lcblas -latlas"

For an installation with gcc, gfortran and Goto BLAS type:

./setup.py --cc=gcc --fc=gfortran --blaslib="-lgoto"

For an installation with xlc, xlf and ESSL type:

./setup.py --cc=xlc --fc=xlf --blaslib="-lessl"

A common mistake, when attempting to build with optimized BLAS, is to forget to notify the compiler about the location of the library. This it done by including the location of the library in the LD_LIBRARY_PATH environment variable, e.g.:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/path_to_blas/lib

Alternatively, the -L compiler flag can be included in the --blaslib option (as show in further examples).

Build and install in the current directory using Intel ICC and IFORT for compilation.

Use sequential MKL for the EM64T architecture for BLAS. Use LAPACK and CBLAS from Netlib and not from MKL:

./setup.py --prefix="/opt" \
           --cc=icc        \
           --fc=ifort      \
           --blaslib="-L/intel/mkl/lib/em64t -lmkl_intel_lp64 -lmkl_sequential -lmkl_core" \
           --downlapack \
           --downcblas

Do the same, but instead of downloading PLASMA, generate a make.inc to use with PLASMA sources:

./setup.py --prefix="/opt" \
           --cc=icc        \
           --fc=ifort      \
           --blaslib="-L/intel/mkl/lib/em64t -lmkl_intel_lp64 -lmkl_sequential -lmkl_core" \
           --downlapack \
           --downcblas  \
           --src

Finally, an empty template for cutting and pasting:

./setup.py --prefix="" \
           --build="" \
           --cc= \
           --fc= \
           --cflags="" \
           --fflags="" \
           --ldflags_c="" \
           --ldflags_fc="" \
           --make= \
           --blaslib="" \
           --cblaslib="" \
           --lapacklib="" \
           --lapclib="" \
           --downblas \
           --downcblas \
           --downlapack \
           --downlapc \
           --downall \
           --[no]testing \
           --nbcores= \
           --src

Intel Math Kernel Library Link Line Advisor is an excellent tool for selecting the right link line when using Intel MKL library.

PLASMA Software Stack

BLAS

BLAS (Basic Linear Algebra Subprograms), are a de facto standard for basic linear algebra operations such as vector and matrix multiplication. FORTRAN implementation of BLAS is available from Netlib (Netlib BLAS). Also, C implementation of BLAS is included in GSL (GNU Scientific Library). Both these implementations are reference implementation of BLAS, are not optimized for modern processor architectures and provide an order of magnitude lower performance than optimized implementations. Highly optimized implementations of BLAS are available from many hardware vendors, such as Intel and AMD. Fast implementations are also available as academic packages, such as ATLAS Goto BLAS. The standard interface to BLAS is the FORTRAN interface.

CBLAS

CBLAS is a C language interface to BLAS. Most commercial and academic implementations of BLAS also provide CBLAS. Netlib provides a reference implementation of CBLAS on top of FORTRAN BLAS (Netlib CBLAS). Since GSL is implemented in C, it naturally provides CBLAS.

LAPACK

LAPACK (Linear Algebra PACKage) is a software library for numerical linear algebra, a successor of LINPACK and EISPACK and a predecessor of PLASMA. LAPACK provides routines for solving linear systems of equations, linear least square problems, eigenvalue problems and singular value problems. Most commercial and academic BLAS packages also provide some LAPACK routines.

CLAPACK

CLAPACK is a C implementation of LAPACK provided by Netlib (Netlib CLAPACK). CLAPACK is produced by automatic translation of FORTRAN LAPACK to C using the F2C utility (Netlib F2C). The purpose of CLAPACK is to provide LAPACK implementation in situations when FORTRAN compiler is not available. Despite being implemented in C, CLAPACK provides the “original” FORTRAN interface, which for the most part conforms to the GNU g77 ABI (Application Binary Interface).

LAPACKE

LAPACKE is a C language interface to LAPACK (or CLAPACK). It is produced by Intel in coordination with the LAPACK team and is available in source code from Netlib in its original version (Netlib LAPACKE) and from PLASMA website in an extended version (LAPACKE for PLASMA). In addition to implementing the C interface, LAPACKE also provides routines which automatically handle workspace allocation, making the use of LAPACK much more convenient.

libtmg

libtmg is a component of the LAPACK library, containing routines for generation of input matrices for testing and timing of LAPACK. The testing and timing suites of LAPACK require libtmg, but not the library itself. (The LAPACK library can be built and used without libtmg.)

hwloc

hwloc (Portable Hardware Locality) is a software package for accessing the topology of a multicore system including components like: cores, sockets, caches and NUMA nodes. hwloc (http://www.open-mpi.org/projects/hwloc/) is a sub-project of the OpenMPI project (http://www.open-mpi.org/).

pthread

POSIX threads library is required to run PLASMA on Unix-like systems. It is a standard component of any such system. Windows threads are used on Microsoft Windows systems.

Getting Help

Please use the PLASMA Users' Forum (http://icl.cs.utk.edu/plasma/forum/) to get help with PLASMA installation.