cc-hpc-itwm / gpispace

GPI-Space: Memory Driven Computing and Big Data

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

GPI-Space

GPI-Space is a task-based workflow management system for parallel applications. It allows the developers to build domain-specific workflows using their own parallelization patterns, data management policies and I/O routines, while relying on the runtime system for the workflow management. The GPI-Space ecosystem "auto-manages" the application runs with dynamic scheduling, in-built distributed memory transfer and distributed task execution.

Installation

GPI-Space targets x86-64 Linux systems. Other architectures are not supported at this point. It is built and tested daily on

  • Centos 7
  • Oracle Linux 8
  • Ubuntu 18.04 LTS
  • Ubuntu 20.04 LTS

The virtual memory layer can be backed with either Ethernet, Infiniband or BeeOND.

GPI-Space is usually installed on cluster systems where a shared filesystem is available on all nodes. It is suggested to use a previously empty directory on such a shared filesystem for the GPI-Space installation and all manually installed dependencies.

The source and build directories are not required to be shared and for best performance it is recommended to place them in a fast (local) file system.

Spack Package

The recommended way of installing GPI-Space is with the Spack package manager:

The most recent version of GPI-Space can be installed using the following command:

spack install gpi-space

NOTE:

There might be a time delay between a GPI-Space release and its availability in the Spack package.


Manual Installation

Alternatively, GPI-Space can also be installed manually. GPI-Space requires some third-party dependencies. Most of the following are provided out of the box as (-devel/-dev) packages or other binary distributions on most supported operating systems.

Boost and GPI-2 require a manual installation on all systems. libssh2 is also required to be built manually on some systems, due to build version compatibility issues with other dependencies.

Note that some GPI-Space components can be disabled, which may remove some dependencies needed. Also see the "Building GPI-Space" section and subsection "Optional Components" below.

  • GCC (>= 5.5.0), or compatible compiler
  • CMake (>= 3.16)
    • Some distributions name the binary cmake3 while others use cmake. Snippets below assume cmake.
  • pkg-config (>= 0.23) , or an equivalent pkgconf. Both offer the pkg-config program.
  • OpenSSL (>= 0.9)
    • When using OpenSSL >= 1.1, Boost >= 1.62 and libssh2 (>= 1.8) are required for compatibility.
  • hwloc (>= 1.10)
  • Qt5 (>= 5.9)
    • Only required if GPI-Space is built with -D GSPC_WITH_MONITOR_APP=ON.
  • chrpath (>= 0.13)
  • Boost (>= 1.61, <= 1.63)
    • When using OpenSSL >= 1.1, Boost >= 1.62 is required.
  • libssh2 (>= 1.7, built with OpenSSL backend)
    • When SSH keys are generated with OpenSSH >= 7.8, only libssh2 >= 1.9 is compatible.
    • Among others, Ubuntu 20.04 ships with libgcrypt as backend as well as OpenSSH 8.2, thus needs a custom installation of libssh2.
  • GPI-2 (>= 1.5.0)
    • Only required if GPI-Space is built with -D GSPC_WITH_IML=ON.

WARNING:

It is important to keep the versions of all dependencies used in GPI-Space as well as the applications using GPI-Space synchronized to avoid hard to debug conflicts at run-time. Using a single environment to build the entire stack is strongly encouraged.



WARNING:

Boost 1.61 is not compatible with OpenSSL >= 1.1.



WARNING:

  • libssh2 1.7 is not compatible with OpenSSL >= 1.1.
  • libssh2 <= 1.8 is incompatible with the new default SSH-key format in OpenSSH >= 7.8.

Building GPI-Space

GPI-Space can be built as described below. The dependencies are found via various environment variables in addition to being searched in system directories. Some dependencies may have multiple ways to be found. The dependency-install snippets above already set the corresponding variables. It may be required to set additional environment or CMake variables if packages are not installed system-wide.

  • Boost: BOOST_ROOT. Newer versions use CMake's new config format and will be automatically preferred if existing. Using config format installations can be disabled with the CMake option -DBoost_NO_BOOST_CMAKE=on.
  • GPI-2: PKG_CONFIG_PATH
  • libssh2: Libssh2_ROOT. By default CMake's new config format will be automatically preferred if existing. Using config format installations can be disabled with the CMake option -DLibssh2_NO_CMAKE=on or the environment variable Libssh2_NO_CMAKE=on.
  • OpenSSL: OPENSSL_ROOT_DIR
  • hwloc: PKG_CONFIG_PATH
  • Qt5: PATH (the qmake/qmake-qt5 binary), Qt5_ROOT, Qt5_DIR or CMAKE_PREFIX_PATH
  • chrpath: PATH

By default, GPI-Space does not build unit- and system tests. To enable them, the options -DBUILD_TESTING=ON and -DSHARED_DIRECTORY_FOR_TESTS=<shared-directory> need to be given to CMake by setting build_tests in the snippet below. SHARED_DIRECTORY_FOR_TESTS shall point to an existing but empty directory on a shared filesystem. The tests can then be run with GSPC_NODEFILE_FOR_TESTS=<path to nodefile> ctest within the build directory.

# to build tests, remove leading # and choose a shared directory to be
# used to store temporary data.
#build_tests="-DBUILD_TESTING=on -DSHARED_DIRECTORY_FOR_TESTS=<shared-directory>"
gpispace_version=main

wget "https://github.com/cc-hpc-itwm/gpispace/archive/${gpispace_version}.tar.gz" \
  -O "gpispace-${gpispace_version}.tar.gz"
tar xf "gpispace-${gpispace_version}.tar.gz"

cmake -D CMAKE_INSTALL_PREFIX="<gpispace-install-prefix>"  \
      -B "gpispace-${gpispace_version}/build" \
      -S "gpispace-${gpispace_version}"       \
      ${build_tests:-}
cmake --build "gpispace-${gpispace_version}/build" \
      -j $(nproc)
cmake --install "gpispace-${gpispace_version}/build"

After installing, the installation can be verified by running a simple self-test example as follows:


NOTE:

GPI-Space requires a working SSH environment with a password-less SSH-key when using the SSH RIF strategy, the default for most applications.

By default, ${HOME}/.ssh/id_rsa is used for authentication. If no such key exists,

ssh-keygen -t rsa -b 4096 -N '' -f "${HOME}/.ssh/id_rsa"
ssh-copy-id -f -i "${HOME}/.ssh/id_rsa" "${HOSTNAME}"

can be used to create and register one.


# to test with multiple nodes, set GSPC_NODEFILE_FOR_TESTS
#   Slurm: export GSPC_NODEFILE_FOR_TESTS="$(generate_pbs_nodefile)"
#   PBS/Torque: export GSPC_NODEFILE_FOR_TESTS="${PBS_NODEFILE}"
# and SWH_INSTALL_DIR:
#   export SWH_INSTALL_DIR=<a-shared-directory-visible-on-all-nodes>

"<gpispace-install-prefix>/share/GPISpace/doc/example/stochastic_with_heureka/selftest"

If GPI-Space has been built with testing enabled, then ctest can be used to execute the unit- and system tests:

# it is always required to set GSPC_NODEFILE_FOR_TESTS for ctest:
#   Slurm: export GSPC_NODEFILE_FOR_TESTS="$(generate_pbs_nodefile)"
#   PBS/Torque: export GSPC_NODEFILE_FOR_TESTS="${PBS_NODEFILE}"
#   single-host: export GSPC_NODEFILE_FOR_TESTS=/etc/hostname
#     or hostname > nodefile; export ...="${PWD}/nodefile"
export GSPC_NODEFILE_FOR_TESTS=<path-to-a-nodefile>

gpispace_version=main

cd "gpispace-${gpispace_version}/build"
ctest --output-on-failure \
      -j $(nproc)

Optional Components

GPI-Space supports disabling some components at configuration time to reduce the dependencies needed as well as build time.

GSPC Monitor

The gspc-monitor (also known as "gantt") application for execution monitoring requires Qt5. The following options can be used to enable or disable this feature (enabled by default):

Spack CMake
[+|~]monitor -D GSPC_WITH_MONITOR_APP=[ON|OFF]

IML

The Independent Memory Layer (IML) is an abstraction layer for easy distributed memory management, allowing applications to cache data, or to prepare and extract data independent of GPI-Space across runs. The following options can be used to enable or disable this feature (enabled by default):

Spack CMake
[+|~]iml -D GSPC_WITH_IML=[ON|OFF]

Next steps

After a successful GPI-Space installation, the next step is to make use of it and create the first application. For that, a step-by-step guide on how to get started and create a first application is available here.

About

GPI-Space: Memory Driven Computing and Big Data

License:GNU General Public License v3.0


Languages

Language:C++ 89.3%Language:CMake 8.4%Language:Dockerfile 1.6%Language:Shell 0.4%Language:Roff 0.1%Language:Makefile 0.1%Language:Haskell 0.0%Language:C 0.0%