install-mpi.sh

#!/bin/bash

ucx_version=1.20.0
ompi_version=5.0.9
with_rocm=/opt/dtk

install_dir=/opt
ucx_prefix=${install_dir}/ucx
ompi_prefix=${install_dir}/mpi

exit_with_error( )
{
  if (( $1 == 2 )); then
    # Failure in some install step
    # Print some message about needed dependencies

    # dependencies needed for executable to build
    local library_dependencies_ubuntu=( "git" "make" "cmake" "libnuma-dev" "pkg-config" "autoconf" "libtool" "automake" "m4" "flex"  "libgomp1" "libibverbs-dev" "librdmacm-dev" )
    local library_dependencies_centos=( "git" "make" "cmake3" "gcc-c++" "rpm-build" "epel-release" "numactl-libs" "autoconf"         "libtool" "automake" "m4" "flex" "libgomp" "libibverbs-devel" "librdmacm-devel" )
    local library_dependencies_fedora=( "git" "make" "cmake" "gcc-c++" "libcxx-devel" "rpm-build" "numactl-libs"  "autoconf"         "libtool" "automake" "m4" "flex" "libgomp" "libibverbs-devel" "librdmacm-devel" )
    local library_dependencies_sles=(   "git" "make" "cmake" "gcc-c++" "libcxxtools9" "rpm-build" "libnuma-devel" "autoconf"         "libtool" "automake" "m4" "flex" "libgomp1" "libibverbs-devel" "librdmacm-devel" )

    printf "Installation failed. Some required packages may be missing.\n"
    printf "The following package manager install command may be needed:\n"
    case "${ID}" in
      ubuntu)
        printf "sudo apt install -y ${library_dependencies_ubuntu[*]}\n"
        ;;

      centos|rhel|tencentos|kylin)
        printf "sudo yum -y --nogpgcheck install ${library_dependencies_centos[*]}\n"
        ;;

      fedora|rocky)
        printf "sudo dnf install -y ${library_dependencies_fedora[*]}\n"
        ;;

      sles)
        printf "sudo zypper -n --no-gpg-checks install ${library_dependencies_sles[*]}\n"
        ;;
      *)  
        exit 2
        ;;
    esac
  fi  

  exit $1
}

check_exit_code( )
{
  if (( $? != 0 )); then
    err=$1
    msg=$2
    if [[ "$msg" == "" ]]; then
      msg="Unknown error"
    fi  
    echo "ERROR: $msg"
    exit $err
  fi  
}

# Clone and build OpenMPI+UCX in /opt
install_openmpi( )
{
  # Backup old installations if they exist (overwrite existing .bak if present)
  [ -d "${ompi_prefix}" ] && {
    rm -rf ${ompi_prefix}.bak; mv ${ompi_prefix} ${ompi_prefix}.bak;
    }
  [ -d "${ucx_prefix}" ] && {
    rm -rf ${ucx_prefix}.bak; mv ${ucx_prefix} ${ucx_prefix}.bak;
    }

  # OpenMPI and UCX install to one of these locations depending on OS
  local ucx_lib_folder=${ucx_prefix}/lib
  local ompi_lib_folder=${ompi_prefix}/lib
  local ucx_lib64_folder=${ucx_prefix}/lib64
  local ompi_lib64_folder=${ompi_prefix}/lib64

  local src_dir=${PWD}
  local ucx_src=${src_dir}/ucx-${ucx_version}
  local ucx_tarball=ucx-${ucx_version}.tar.gz
  local ompi_src=${src_dir}/openmpi-${ompi_version}
  local ompi_tarball=openmpi-${ompi_version}.tar.gz

  # Download UCX on demand
  if [ ! -d "${ucx_src}" ]; then
    if [ ! -f "${ucx_tarball}" ]; then
        wget https://github.com/openucx/ucx/releases/download/v${ucx_version}/${ucx_tarball}
    fi
    check_exit_code 2
    tar -zxf ${ucx_tarball}
    check_exit_code 2
  fi
  # Download OpenMPI on demand
  if [ ! -d "${ompi_src}" ]; then
    if [ ! -f "${ompi_tarball}" ]; then
        wget https://download.open-mpi.org/release/open-mpi/v${ompi_version%.*}/${ompi_tarball}
    fi
    check_exit_code 2
    tar -zxf ${ompi_tarball}
    check_exit_code 2
  fi

  # Build UCX
  cd ${ucx_src};
  ./contrib/configure-release --prefix=${ucx_prefix} \
    --enable-cma --enable-mt \
    --with-mlx5 --with-rc --with-ud --with-dc --with-dm --with-ib_hw_tm \
    --with-verbs=/usr/include --with-rdmacm=/usr \
    --with-rocm=${with_rocm} \
    --without-knem --without-cuda --without-java
  check_exit_code 2
  make -j$(nproc)
  check_exit_code 2
  make install
  check_exit_code 2
  cd ..

  # Check for successful build
  if ([ ! -f "${ucx_lib_folder}/libucm.so" ] || [ ! -f "${ucx_lib_folder}/libucp.so" ]  || \
      [ ! -f "${ucx_lib_folder}/libucs.so" ] || [ ! -f "${ucx_lib_folder}/libuct.so" ]) &&
     ([ ! -f "${ucx_lib64_folder}/libucm.so" ] || [ ! -f "${ucx_lib64_folder}/libucp.so" ]  || \
      [ ! -f "${ucx_lib64_folder}/libucs.so" ] || [ ! -f "${ucx_lib64_folder}/libuct.so" ]); then
    echo "Error: UCX install unsuccessful."
    exit 3
  fi

  export LD_LIBRARY_PATH="${ucx_prefix}/lib:${ucx_prefix}/lib64:${LD_LIBRARY_PATH}"
  export LIBRARY_PATH="${ucx_prefix}/lib:${ucx_prefix}/lib64:${LIBRARY_PATH}"
  export CPATH="${ucx_prefix}/include:${CPATH}"

  # Build OpenMPI
  cd ${ompi_src}
  ./configure --prefix=${ompi_prefix} \
    --with-ucx=${ucx_prefix} \
    --with-rocm=${with_rocm} \
    --enable-builtin-atomics \
    --enable-wrapper-rpath \
    --without-verbs --enable-mca-no-build=btl-uct
  check_exit_code 2
  make -j$(nproc)
  check_exit_code 2
  make install
  check_exit_code 2
  cd ..

  # Check for successful build
  if [ ! -f "${ompi_lib_folder}/libmpi.so" ] && [ ! -f "${ompi_lib64_folder}/libmpi.so" ]; then
    echo "Error: OpenMPI install unsuccessful."
    exit_with_error 2
  fi
}

install_openmpi