Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

install CUDA under host_injections #410

Merged
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 16 additions & 0 deletions EESSI-install-software.sh
Original file line number Diff line number Diff line change
Expand Up @@ -187,6 +187,22 @@ fi
# assume there's only one diff file that corresponds to the PR patch file
pr_diff=$(ls [0-9]*.diff | head -1)

# install any additional required scripts
# order is important: these are needed to install a full CUDA SDK in host_injections
install_scripts_changed=$(cat ${pr_diff} | grep '^+++' | cut -f2 -d' ' | sed 's@^[a-z]/@@g' | grep '^install_scripts.sh$' > /dev/null; echo $?)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@casparvl I messed up a little by merging this as now the file checked here will not be updated in the other PR, and as a result the scripts will not get installed. It doesn't matter too much as I think the use of $pr_diff is overkill here. I think we should always run the installer script and instead of using cp use cp -u, that could be implemented as part of #434

Also, this was a little wrong anyway, as it doesn't check if the scripts installed by the installer have changed

if [ ${install_scripts_changed} == '0' ]; then
# for now, this just reinstalls all scripts. Note the most elegant, but works
${TOPDIR}/install_scripts.sh --prefix ${EESSI_CVMFS_REPO}
fi

# Install full CUDA SDK in host_injections
# Hardcode this for now, see if it works
# TODO: We should make a nice yaml and loop over all CUDA versions in that yaml to figure out what to install
${EESSI_CVMFS_REPO}/gpu_support/nvidia/install_cuda_host_injections.sh 12.1.1
Copy link
Member

@ocaisa ocaisa Dec 19, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Wow, I really messed up here...the arguments are incorrect!


# Install drivers in host_injections
${EESSI_CVMFS_REPO}/gpu_support/nvidia/link_nvidia_host_libraries.sh

# use PR patch file to determine in which easystack files stuff was added
for easystack_file in $(cat ${pr_diff} | grep '^+++' | cut -f2 -d' ' | sed 's@^[a-z]/@@g' | grep '^easystacks/.*yml$' | egrep -v 'known-issues|missing'); do

Expand Down
71 changes: 65 additions & 6 deletions eessi_container.sh
Original file line number Diff line number Diff line change
Expand Up @@ -30,8 +30,8 @@
# -. initial settings & exit codes
TOPDIR=$(dirname $(realpath $0))

source ${TOPDIR}/scripts/utils.sh
source ${TOPDIR}/scripts/cfg_files.sh
source "${TOPDIR}"/scripts/utils.sh
source "${TOPDIR}"/scripts/cfg_files.sh

# exit codes: bitwise shift codes to allow for combination of exit codes
# ANY_ERROR_EXITCODE is sourced from ${TOPDIR}/scripts/utils.sh
Expand All @@ -46,6 +46,7 @@ SAVE_ERROR_EXITCODE=$((${ANY_ERROR_EXITCODE} << 8))
HTTP_PROXY_ERROR_EXITCODE=$((${ANY_ERROR_EXITCODE} << 9))
HTTPS_PROXY_ERROR_EXITCODE=$((${ANY_ERROR_EXITCODE} << 10))
RUN_SCRIPT_MISSING_EXITCODE=$((${ANY_ERROR_EXITCODE} << 11))
NVIDIA_MODE_UNKNOWN_EXITCODE=$((${ANY_ERROR_EXITCODE} << 12))

# CernVM-FS settings
CVMFS_VAR_LIB="var-lib-cvmfs"
Expand All @@ -72,12 +73,17 @@ display_help() {
echo " -a | --access {ro,rw} - ro (read-only), rw (read & write) [default: ro]"
echo " -c | --container IMG - image file or URL defining the container to use"
echo " [default: docker://ghcr.io/eessi/build-node:debian11]"
echo " -h | --help - display this usage information [default: false]"
echo " -g | --storage DIR - directory space on host machine (used for"
echo " temporary data) [default: 1. TMPDIR, 2. /tmp]"
echo " -h | --help - display this usage information [default: false]"
echo " -i | --host-injections - directory to link to for host_injections "
echo " [default: /..storage../opt-eessi]"
echo " -l | --list-repos - list available repository identifiers [default: false]"
echo " -m | --mode MODE - with MODE==shell (launch interactive shell) or"
echo " MODE==run (run a script or command) [default: shell]"
echo " -n | --nvidia MODE - configure the container to work with NVIDIA GPUs,"
echo " MODE==install for a CUDA installation, MODE==run to"
echo " attach a GPU, MODE==all for both [default: false]"
echo " -r | --repository CFG - configuration file or identifier defining the"
echo " repository to use [default: EESSI via"
echo " default container, see --container]"
Expand Down Expand Up @@ -111,6 +117,8 @@ VERBOSE=0
STORAGE=
LIST_REPOS=0
MODE="shell"
SETUP_NVIDIA=0
ADDITIONAL_SINGULARITY_FLAGS=
REPOSITORY="EESSI"
RESUME=
SAVE=
Expand Down Expand Up @@ -141,6 +149,10 @@ while [[ $# -gt 0 ]]; do
display_help
exit 0
;;
-i|--host-injections)
USER_HOST_INJECTIONS="$2"
shift 2
;;
-l|--list-repos)
LIST_REPOS=1
shift 1
Expand All @@ -149,6 +161,11 @@ while [[ $# -gt 0 ]]; do
MODE="$2"
shift 2
;;
-n|--nvidia)
SETUP_NVIDIA=1
NVIDIA_MODE="$2"
shift 2
;;
-r|--repository)
REPOSITORY="$2"
shift 2
Expand Down Expand Up @@ -224,6 +241,13 @@ if [[ "${MODE}" != "shell" && "${MODE}" != "run" ]]; then
fatal_error "unknown execution mode '${MODE}'" "${MODE_UNKNOWN_EXITCODE}"
fi

# Also validate the NVIDIA GPU mode (if present)
if [[ ${SETUP_NVIDIA} -eq 1 ]]; then
if [[ "${NVIDIA_MODE}" != "run" && "${NVIDIA_MODE}" != "install" && "${NVIDIA_MODE}" != "all" ]]; then
fatal_error "unknown NVIDIA mode '${NVIDIA_MODE}'" "${NVIDIA_MODE_UNKNOWN_EXITCODE}"
fi
fi

# TODO (arg -r|--repository) check if repository is known
# REPOSITORY_ERROR_EXITCODE
if [[ ! -z "${REPOSITORY}" && "${REPOSITORY}" != "EESSI" && ! -r ${EESSI_REPOS_CFG_FILE} ]]; then
Expand Down Expand Up @@ -310,12 +334,25 @@ fi
# |-overlay-work
# |-home
# |-repos_cfg
# |-opt-eessi (unless otherwise specificed for host_injections)

# tmp dir for EESSI
EESSI_TMPDIR=${EESSI_HOST_STORAGE}
mkdir -p ${EESSI_TMPDIR}
[[ ${VERBOSE} -eq 1 ]] && echo "EESSI_TMPDIR=${EESSI_TMPDIR}"

# Set host_injections directory and ensure it is a writable directory (if user provided)
if [ -z ${USER_HOST_INJECTIONS+x} ]; then
# Not set, so use our default
HOST_INJECTIONS=${EESSI_TMPDIR}/opt-eessi
mkdir -p $HOST_INJECTIONS
else
# Make sure the host_injections directory specified exists and is a folder
mkdir -p ${USER_HOST_INJECTIONS} || fatal_error "host_injections directory ${USER_HOST_INJECTIONS} is either not a directory or cannot be created"
HOST_INJECTIONS=${USER_HOST_INJECTIONS}
fi
[[ ${VERBOSE} -eq 1 ]] && echo "HOST_INJECTIONS=${HOST_INJECTIONS}"

# configure Singularity: if SINGULARITY_CACHEDIR is already defined, use that
# a global SINGULARITY_CACHEDIR would ensure that we don't consume
# storage space again and again for the container & also speed-up
Expand Down Expand Up @@ -394,12 +431,34 @@ fi
[[ ${VERBOSE} -eq 1 ]] && echo "SINGULARITY_HOME=${SINGULARITY_HOME}"

# define paths to add to SINGULARITY_BIND (added later when all BIND mounts are defined)
BIND_PATHS="${EESSI_CVMFS_VAR_LIB}:/var/lib/cvmfs,${EESSI_CVMFS_VAR_RUN}:/var/run/cvmfs"
BIND_PATHS="${EESSI_CVMFS_VAR_LIB}:/var/lib/cvmfs,${EESSI_CVMFS_VAR_RUN}:/var/run/cvmfs,${HOST_INJECTIONS}:/opt/eessi"
# provide a '/tmp' inside the container
BIND_PATHS="${BIND_PATHS},${EESSI_TMPDIR}:${TMP_IN_CONTAINER}"

[[ ${VERBOSE} -eq 1 ]] && echo "BIND_PATHS=${BIND_PATHS}"

# Configure anything we need for NVIDIA GPUs and CUDA installation
if [[ ${SETUP_NVIDIA} -eq 1 ]]; then
if [[ "${NVIDIA_MODE}" == "run" || "${NVIDIA_MODE}" == "all" ]]; then
# Give singularity the appropriate flag
ADDITIONAL_SINGULARITY_FLAGS="--nv ${ADDITIONAL_SINGULARITY_FLAGS}"
[[ ${VERBOSE} -eq 1 ]] && echo "ADDITIONAL_SINGULARITY_FLAGS=${ADDITIONAL_SINGULARITY_FLAGS}"
fi
if [[ "${NVIDIA_MODE}" == "install" || "${NVIDIA_MODE}" == "all" ]]; then
# Add additional bind mounts to allow CUDA to install within a container
# (Experience tells us that these are necessary, but we don't know _why_
# as the CUDA installer is a black box. The suspicion is that the CUDA
# installer gets confused by the permissions on these directories when
# inside a container)
EESSI_VAR_LOG=${EESSI_TMPDIR}/var-log
EESSI_USR_LOCAL_CUDA=${EESSI_TMPDIR}/usr-local-cuda
mkdir -p ${EESSI_VAR_LOG}
mkdir -p ${EESSI_USR_LOCAL_CUDA}
BIND_PATHS="${BIND_PATHS},${EESSI_VAR_LOG}:/var/log,${EESSI_USR_LOCAL_CUDA}:/usr/local/cuda"
[[ ${VERBOSE} -eq 1 ]] && echo "BIND_PATHS=${BIND_PATHS}"
fi
fi

# set up repository config (always create directory repos_cfg and populate it with info when
# arg -r|--repository is used)
mkdir -p ${EESSI_TMPDIR}/repos_cfg
Expand Down Expand Up @@ -562,8 +621,8 @@ if [ ! -z ${EESSI_SOFTWARE_SUBDIR_OVERRIDE} ]; then
fi

echo "Launching container with command (next line):"
echo "singularity ${RUN_QUIET} ${MODE} ${EESSI_FUSE_MOUNTS[@]} ${CONTAINER} $@"
singularity ${RUN_QUIET} ${MODE} "${EESSI_FUSE_MOUNTS[@]}" ${CONTAINER} "$@"
echo "singularity ${RUN_QUIET} ${MODE} ${ADDITIONAL_SINGULARITY_FLAGS} ${EESSI_FUSE_MOUNTS[@]} ${CONTAINER} $@"
singularity ${RUN_QUIET} ${MODE} ${ADDITIONAL_SINGULARITY_FLAGS} "${EESSI_FUSE_MOUNTS[@]}" ${CONTAINER} "$@"
exit_code=$?

# 6. save tmp if requested (arg -s|--save)
Expand Down
209 changes: 209 additions & 0 deletions gpu_support/nvidia/install_cuda_host_injections.sh
Original file line number Diff line number Diff line change
@@ -0,0 +1,209 @@
#!/usr/bin/env bash

# This script can be used to install CUDA under the `.../host_injections` directory.
# This provides the parts of the CUDA installation that cannot be redistributed as
# part of EESSI due to license limitations. While GPU-based software from EESSI will
# _run_ without these, installation of additional CUDA software requires the CUDA
# installation(s) under `host_injections` to be present.
#
# The `host_injections` directory is a variant symlink that by default points to
# `/opt/eessi`, unless otherwise defined in the local CVMFS configuration (see
# https://cvmfs.readthedocs.io/en/stable/cpt-repo.html#variant-symlinks). For the
# installation to be successful, this directory needs to be writeable by the user
# executing this script.

# Initialise our bash functions
TOPDIR=$(dirname $(realpath $BASH_SOURCE))
source "$TOPDIR"/../../scripts/utils.sh

# Function to display help message
show_help() {
echo "Usage: $0 [OPTIONS]"
echo "Options:"
echo " --help Display this help message"
echo " --accept-cuda-eula You _must_ accept the CUDA EULA to install"
echo " CUDA, see the EULA at"
echo " https://docs.nvidia.com/cuda/eula/index.html"
echo " -c, --cuda-version CUDA_VERSION Specify a version o CUDA to install (must"
echo " have a corresponding easyconfig in the"
echo " EasyBuild release)"
echo " -t, --temp-dir /path/to/tmpdir Specify a location to use for temporary"
echo " storage during the CUDA install"
echo " (must have >10GB available)"
}

# Initialize variables
install_cuda_version=""
eula_accepted=0

# Parse command-line options
while [[ $# -gt 0 ]]; do
case "$1" in
--help)
show_help
exit 0
;;
-c|--cuda-version)
if [ -n "$2" ]; then
install_cuda_version="$2"
shift 2
else
echo "Error: Argument required for $1"
show_help
exit 1
fi
;;
--accept-cuda-eula)
eula_accepted=1
shift 1
;;
-t|--temp-dir)
if [ -n "$2" ]; then
CUDA_TEMP_DIR="$2"
shift 2
else
echo "Error: Argument required for $1"
show_help
exit 1
fi
;;
*)
show_help
fatal_error "Error: Unknown option: $1"
;;
esac
done

# Make sure EESSI is initialised
check_eessi_initialised

# Make sure the CUDA version supplied is a semantic version
is_semantic_version() {
local version=$1
local regex='^[0-9]+\.[0-9]+\.[0-9]+$'

if [[ $version =~ $regex ]]; then
return 0 # Return success (0) if it's a semantic version
else
return 1 # Return failure (1) if it's not a semantic version
fi
}
if ! is_semantic_version "$install_cuda_version"; then
show_help
error="\nYou must provide a semantic version for CUDA (e.g., 12.1.1) via the appropriate\n"
error="${error}command line option. This script is intended for use with EESSI so the 'correct'\n"
error="${error}version to provide is probably one of those available under\n"
error="${error}$EESSI_SOFTWARE_PATH/software/CUDA\n"
fatal_error "${error}"
fi

# Make sure they have accepted the CUDA EULA
if [ "$eula_accepted" -ne 1 ]; then
show_help
error="\nYou _must_ accept the CUDA EULA via the appropriate command line option.\n"
fatal_error "${error}"
fi

# As an installation location just use $EESSI_SOFTWARE_PATH but replacing `versions` with `host_injections`
# (CUDA is a binary installation so no need to worry too much about the EasyBuild setup)
cuda_install_parent=${EESSI_SOFTWARE_PATH/versions/host_injections}

# Only install CUDA if specified version is not found.
# (existence of easybuild subdir implies a successful install)
if [ -d "${cuda_install_parent}"/software/CUDA/"${install_cuda_version}"/easybuild ]; then
echo_green "CUDA software found! No need to install CUDA again."
else
# We need to be able write to the installation space so let's make sure we can
if ! create_directory_structure "${cuda_install_parent}"/software/CUDA ; then
fatal_error "No write permissions to directory ${cuda_install_parent}/software/CUDA"
fi

# we need a directory we can use for temporary storage
if [[ -z "${CUDA_TEMP_DIR}" ]]; then
tmpdir=$(mktemp -d)
else
tmpdir="${CUDA_TEMP_DIR}"/temp
if ! mkdir "$tmpdir" ; then
fatal_error "Could not create directory ${tmpdir}"
fi
fi

required_space_in_tmpdir=50000
# Let's see if we have sources and build locations defined if not, we use the temporary space
if [[ -z "${EASYBUILD_BUILDPATH}" ]]; then
export EASYBUILD_BUILDPATH=${tmpdir}/build
required_space_in_tmpdir=$((required_space_in_tmpdir + 5000000))
fi
if [[ -z "${EASYBUILD_SOURCEPATH}" ]]; then
export EASYBUILD_SOURCEPATH=${tmpdir}/sources
required_space_in_tmpdir=$((required_space_in_tmpdir + 5000000))
fi

# The install is pretty fat, you need lots of space for download/unpack/install (~3*5GB),
# need to do a space check before we proceed
avail_space=$(df --output=avail "${cuda_install_parent}"/ | tail -n 1 | awk '{print $1}')
if (( avail_space < 5000000 )); then
fatal_error "Need at least 5GB disk space to install CUDA under ${cuda_install_parent}, exiting now..."
fi
avail_space=$(df --output=avail "${tmpdir}"/ | tail -n 1 | awk '{print $1}')
if (( avail_space < required_space_in_tmpdir )); then
error="Need at least ${required_space_in_tmpdir} disk space under ${tmpdir}.\n"
error="${error}Set the environment variable CUDA_TEMP_DIR to a location with adequate space to pass this check."
error="${error}You can alternatively set EASYBUILD_BUILDPATH and/or EASYBUILD_SOURCEPATH "
error="${error}to reduce this requirement. Exiting now..."
fatal_error "${error}"
fi

if ! command -v "eb" &>/dev/null; then
echo_yellow "Attempting to load an EasyBuild module to do actual install"
module load EasyBuild
# There are some scenarios where this may fail
if [ $? -ne 0 ]; then
error="'eb' command not found in your environment and\n"
error="${error} module load EasyBuild\n"
error="${error}failed for some reason.\n"
error="${error}Please re-run this script with the 'eb' command available."
fatal_error "${error}"
fi
fi

cuda_easyconfig="CUDA-${install_cuda_version}.eb"

# Check the easyconfig file is available in the release
# (eb search always returns 0, so we need a grep to ensure a usable exit code)
eb --search ^${cuda_easyconfig}|grep CUDA > /dev/null 2>&1
# Check the exit code
if [ $? -ne 0 ]; then
eb_version=$(eb --version)
available_cuda_easyconfigs=$(eb --search ^CUDA-*.eb|grep CUDA)

error="The easyconfig ${cuda_easyconfig} was not found in EasyBuild version:\n"
error="${error} ${eb_version}\n"
error="${error}You either need to give a different version of CUDA to install _or_ \n"
error="${error}use a different version of EasyBuild for the installation.\n"
error="${error}\nThe versions of available with the current eb command are:\n"
error="${error}${available_cuda_easyconfigs}"
fatal_error "${error}"
fi

# We need the --rebuild option, as the CUDA module may or may not be on the
# `MODULEPATH` yet. Even if it is, we still want to redo this installation
# since it will provide the symlinked targets for the parts of the CUDA
# installation in the `.../versions/...` prefix
# We install the module in our `tmpdir` since we do not need the modulefile,
# we only care about providing the targets for the symlinks.
extra_args="--rebuild --installpath-modules=${tmpdir}"

# We don't want hooks used in this install, we need a vanilla CUDA installation
touch "$tmpdir"/none.py
# shellcheck disable=SC2086 # Intended splitting of extra_args
eb --prefix="$tmpdir" ${extra_args} --accept-eula-for=CUDA --hooks="$tmpdir"/none.py --installpath="${cuda_install_parent}"/ "${cuda_easyconfig}"
ret=$?
if [ $ret -ne 0 ]; then
fatal_error "CUDA installation failed, please check EasyBuild logs..."
else
echo_green "CUDA installation at ${cuda_install_parent}/software/CUDA/${install_cuda_version} succeeded!"
fi
# clean up tmpdir
rm -rf "${tmpdir}"
fi
Loading