Skip to content
Snippets Groups Projects
Select Git revision
  • 6030fadef6e95542824adc9abb6906f717bc16cc
  • deep-sea default
  • jsc-ci-add-mt-deployment
  • deepsea-sh
  • master protected
  • 5.9.x
  • 5.9.2-1
7 results

psmpi

  • Clone with SSH
  • Clone with HTTPS
  • ParaStation MPI

    Introduction

    ParaStation MPI is an implementation of the Message-Passing Interface (MPI) Standard. Being an MPICH-derivate it bases on the MPICH-4.1.1 implementation.

    Overview diagram of psmpi

    As shown in the figure above, ParaStation MPI relies on the low-level communication layer pscom for the implementation of the PSP Device on the ADI-Layer of the MPICH architecture. ParaStation MPI provides full MPI-4 compliance including support for the MPI Sessions model. Furthermore, the Process Management Interface for Exascale (PMIx) is supported as interface to the process manager.

    Requirements

    Mandatory

    • C compiler with C99 support
    • ParaStation Communication library (pscom)

    Optional

    • ParaStation Management (psmgmt) process manager if you do not want to use the Hydra process manager that comes with MPICH
    • C++ compiler for applications written in C++
    • Fortran compiler for applications written in Fortran
    • PMIx library if you want to use PMIx as interface to the process manager (use of PMIx and Hydra is mutual exclusive)
    • hcoll library for hcoll support
    • hwloc library for hwloc support in MPICH/hydra
    • CUDA runtime library for CUDA awareness

    Installation

    Configure ParaStation MPI

    Download the source code from GitHub:

    $ git clone https://github.com/ParaStation/psmpi.git psmpi

    Run Autotools:

    $ cd /path/to/psmpi
    $ ./autogen.sh

    We encourage you not to build ParaStation MPI directly within the main source folder:

    $ mkdir build && cd build

    ParaStation MPI relies on Autotools as the build system requiring a configure step. The build system already provides so-called "confsets" that pass the necessary configuration arguments for a particular installation type to the underlying MPICH build system. It is strongly recommended to rely on these confsets for a proper installation! Currently, the following confsets are provided:

    default    : Like 'gcc'
    none       : Do not configure mpich. Prepare only for tar, rpm and srpm build
    
    gcc        : Use Gnu compiler (gcc)
    intel      : Use Intel compiler (icx)
    icx        : Like 'intel'
    pgi        : Portland group compiler (pgcc)
    nvhpc      : Nvidia HPC compiler (nvc)
    llvm       : llvm-based toolchains (e.g., AOCC)
    
    devel      : With error checking and debug info (gcc)
    debug      : Like 'devel' but without any optimizations
    
    user       : No predefined options
    ch3        : original mpich ch3 device (no parastation)
    ch4        : original mpich ch4 device (no parastation)

    The following example configures ParaStation MPI for using the gcc compiler:

    $ ../configure --prefix=/path/to/installation/dir --with-confset=default

    We recommend developers to use the devel confset for more error checking and debug infos at compile time.

    Optional configure arguments

    Argument Description
    --with-pscom[=PATH]  Use pscom as communication transport [PATH to pscom installation]
    --with-cuda  Enable CUDA awareness
    --with-hydra  Use MPICH's process manager Hydra
    --with-threading  Enable multi-thread support
    --with-msa-awareness  Enable MSA awareness like hierarchical collectives
    --with-session-statistics Enable the collection of statistical information
    --with-hcoll[=PATH] Enable hcoll support [PATH to hcoll installation]
    --with-hwloc[=PATH] Enable hwloc in MPICH/Hydra [built-in or PATH]
    --with-pmix[=PATH] Use PMIx as process manager interface [PATH to PMIx installation]

    Build ParaStation MPI

    For a successful build, you have to provide a path to your pscom installation via the --with-pcom[=PATH] configure option (defaults to /opt/parastation) or alternatively by setting the LIBRARY_PATH and CPATH environment variables:

    $ export LIBRARY_PATH=/path/to/pscom/installation/lib[64]:${LIBRARY_PATH}
    $ export CPATH=/path/to/pscom/installation/include:${CPATH}

    Now, ParaStation MPI can be built and installed in accordance with the configuration arguments:

    $ make -j8 && make install

    Prepare the environment

    To use ParaStation MPI for building and running MPI applications, it is advisable to adjust the environment properly. This can be done, e.g., by using the following bash script:

    #!/bin/bash
    
    if [ $# -eq 0 ]; then
        echo "ERROR: Please provide the path to ParaStation MPI. Abort!"
        exit -1
    fi
    
    PARASTATION_MPI=`realpath ${1}`
    
    export PATH="${PARASTATION_MPI}/bin${PATH:+:}${PATH}"
    export CPATH="${PARASTATION_MPI}/include${CPATH:+:}${CPATH}"
    export LD_LIBRARY_PATH="${PARASTATION_MPI}/lib${LD_LIBRARY_PATH:+:}${LD_LIBRARY_PATH}"
    export LIBRARY_PATH="${PARASTATION_MPI}/lib${LIBRARY_PATH:+:}${LIBRARY_PATH}"

    Alternative Installation

    Instead of relying on the pscom as a shared-library, ParaStation MPI can be optionally compiled as a single shared-object by directly using the pscom sources. For doing so, the pscom source files are required:

    $ git clone https://github.com/ParaStation/pscom.git pscom

    After downloading the psmpi sources, the same configuration parameters apply as discussed above (see Configure). Additionally, you will need to add the --with-pscom-allin flag, e.g.:

    $ ../configure --prefix=/path/to/installation/dir --with-confset=default --with-pscom-allin=/path/to/pscom/sources

    By default, the pscom4ucp as well as the pscom4psm plugins are included firmly into ParaStation MPI if --with-pscom-allin is set and the related low-level drivers are found. For specifying the plugins to be built-in explicitly, the --with-pscom-builtin[=list] option can be used.

    Runtime Parameters

    Debugging

    Environment Variable Description
    PSP_DEBUG=0 only fatal conditions (like detected bugs)
    PSP_DEBUG=1 fatal conditions + errors (default)
    PSP_DEBUG=2 + warnings
    PSP_DEBUG=3 + information
    PSP_DEBUG=4 + debug
    PSP_DEBUG=5 + verbose debug
    PSP_DEBUG=6 + tracing calls
    PSP_DEBUG_VERSION=1 Show always the pscom version (info)
    PSP_DEBUG_CONTYPE=1 Show connection types (info)
    PSP_HARD_ABORT=0 Process termination in MPI_Abort() via exit() (default)
    PSP_HARD_ABORT=1 Process termination in MPI_Abort() via PMI_Abort()
    PSP_HARD_ABORT=2 Process termination in MPI_Abort() via _exit()
    PSP_HARD_ABORT=3 Process termination in MPI_Abort() via abort()
    PSP_CUDA_ENFORCE_STAGING=1 Enforce staging of CUDA buffers via host memory (with a significant performance penalty!)

    Feature Activation

    Environment Variable Description Required build config options
    PSP_CUDA=1 Enable/Disable CUDA awareness (default = 0) --with-cuda
    PSP_HCOLL=1 Enable/Disable HCOLL support (default = 0) --with-hcoll[=PATH]
    PSP_SMP_AWARENESS=1 Take locality information into account (default = 1)
    PSP_SMP_AWARE_COLLOPS=1 Enable/Disable SMP-aware collectives (default = 0) --with-msa-awareness
    PSP_MSA_AWARENESS=1 Take topology information into account (default = 0) --with-msa-awareness
    PSP_MSA_AWARE_COLLOPS=1 Enable/Disable MSA-aware collectives (default = 0) --with-msa-awareness

    Statistical Analysis

    Environment Variable Description Required build config options
    PSP_HISTOGRAM=1 Enable the collection of statistical data --with-session-statistics
    PSP_HISTOGRAM_MIN=x Set the lower message size limit for the histogram --with-session-statistics
    PSP_HISTOGRAM_MAX=y Set the upper message size limit for the histogram --with-session-statistics
    PSP_HISTOGRAM_SHIFT=z Bit shift for the number of bins of the histogram --with-session-statistics
    PSP_HISTOGRAM_CONTYPE=con Limit the histogram to a particular connection type --with-session-statistics
    PSP_HCOLL_STATS=1 Enable the collection of HCOLL usage statistics --with-session-statistics --with-hcoll[=PATH]

    Test Suite

    MPICH has a test suite that can also be used and is even extended by ParaStation MPI.

    $ make test

    ...will run the complete test suite comprising tests in the following subfolders in mpich2/test/mpi/:

    parastation
    cuda
    attr
    ckpoint
    coll
    comm
    cxx
    datatype
    errhan
    errors
    f08
    f77
    f90
    ft
    group
    impls
    info
    io
    init
    mpi_t
    perf
    pt2pt
    rma
    spawn
    threads
    topo
    part
    session

    Selecting Test Directories and Test Sets

    However, you can also have only tests of a certain subdirectory to be executed by specifying a TESTDIR, for example:

    $ make test TESTDIR=pt2pt

    In addition, a TESTSET can be specified to further restrict the number of tests. The tests that belong to a certain test set are stated as a list in a file with the same name within the above listed subdirectories. Currently, only ps-test-minimal (minimal list of tests) and testlist (containing all tests) are valid test sets. So, for example, the following invocation runs all tests belonging to ps-test-minimal within all subdirectories:

    $ make test TESTSET=ps-test-minimal

    And if, for example, the test set ps-test-minimal should only be executed for tests within the subfolder pt2pt, then the following invocation is the means of choice:

    $ make test TESTSET=ps-test-minimal TESTDIR=pt2pt

    In the case that several directories and/or several test sets should be used for a run, one can do this by using comma-separated lists and the variables TESTDIRS and TESTSETS. (Please note the additional "S" at the end of the variable names!) However, in this case no explicit check for validity of the passed test sets or folder names is performed. Example:

    $ make test TESTDIRS=pt2pt,coll,info

    Furthermore, there is a "meta" set of tests that are specifically intended for large numbers of processes and/or large messages and data sets. This additional (and somehow orthogonal) set can be chosen via largetest as the make target:

    $ make largetest

    ...which can also be used in combination with the TESTSET/TESTSETS and the TESTDIR/TESTDIRS options.

    Using Test Configurations

    When running the tests, some settings like the mpiexec command to be used are pre-configured and depend on the configuration chosen for psmpi during its top-level configure call. However, there is still the possibility to overwrite some of such configurations by explicitly setting the following environment variables:

    • MPIEXEC: choose the path and the executable to be used for process start
    • MPIEXECARG: additional arguments to be passed to the mpiexec command

    For example:

    $ make test MPIEXEC=/opt/parastation/bin/mpiexec MPIEXECARG="-e MPIEXEC_UNIVERSE_SIZE=128"

    Furthermore, there is a small set of such overriding configurations that can also be chosen via the TESTCONF option. The following configurations are available:

    • hydra: Force the use of MPICH's Hydra mpiexec
    • psmgmt: Force the use of psmgmt's mpiexec in /opt/parastation/bin
    • verbose: Switch on verbose mode for runtests script
    • psmgmt/memcheck: Call psmgmt's mpiexec with --memcheck
    • psmgmt/valgrind: Call psmgmt's mpiexec with --valgrind
    • psmgmt/pmix: Call psmgmt's mpiexec with --pmix

    If multiple of such configurations are to be selected at the same time, then the TESTCONFS (mind the additional S at the end!) can be used. Example:

    $ make test TESTCONFS=hydra,verbose

    Contributing

    To ensure that all commits conform to the coding style, the pre-commit hook should be activated. Therefore, you have to link this from the top-level source directory:

    $ ln -s ../../scripts/hooks/pre-commit .git/hooks/pre-commit