mirror of https://github.com/kubernetes/kops.git
Updated Kops GPU Setup Hook
* Changed Dockerfile base image to debian for systemctl and bash. * Added autodetect of AWS ec2 instanceclass p2, p3, g3. * For each detected instance class, added the installation of the proper driver according to the specific NVIDIA hardware. - G3 instance types require Nvidia Grid Series/Grid K520 drivers - P2 instance types require Nvidia Tesla K-Series drivers - P3 instance types require Nvidia Tesla V-Series drivers * Set custom nvidia-smi configurations according to nvidia hardware per ec2 instanceclass, according to the AWS GPU optimization document. * Added the installation and patches of the latest cuda 9.1 libraries. * Added restart of kubelet on kube node at end of successful hook run, thereby fixing a race condition where kubelet would start before the Nvidia drivers were loaded, thus not allowing kubernetes to detect GPUS on the kube node. * Ensured build of nvidia drivers used same gcc version as that which built default kops kernel. * Fixed issue where *every* run of this container would download all the NVIDIA drivers + cuda libs (1GB+), by caching the files on the kube node. * Fixed issue where after reboot, subsequent runs of this script would fail because mknod would try to create a previously-created device node and fail. This previously caused download loop as systemd perpetually restarted the unit upon failure. * Tested with p2.xlarge, p3.2xlarge, and g3.4xlarge
This commit is contained in:
David C Wang
parent
3781d63ca3
commit
69ab306eac
|
|
@ -12,9 +12,12 @@
|
||||||
# See the License for the specific language governing permissions and
|
# See the License for the specific language governing permissions and
|
||||||
# limitations under the License.
|
# limitations under the License.
|
||||||
|
|
||||||
FROM alpine:3.6
|
FROM debian:jessie
|
||||||
|
# ^ Cannot be Alpine since it does not support systemctl
|
||||||
|
# ^ Systemctl is used to restart kubelet upon successful run of run.sh
|
||||||
|
|
||||||
|
RUN apt-get update && apt-get -yq install curl jq
|
||||||
|
|
||||||
RUN apk --no-cache add ca-certificates wget && update-ca-certificates
|
|
||||||
ADD run.sh /run.sh
|
ADD run.sh /run.sh
|
||||||
|
|
||||||
CMD /run.sh
|
CMD [ "/bin/bash", "/run.sh" ]
|
||||||
|
|
|
||||||
|
|
@ -12,50 +12,219 @@
|
||||||
# See the License for the specific language governing permissions and
|
# See the License for the specific language governing permissions and
|
||||||
# limitations under the License.
|
# limitations under the License.
|
||||||
|
|
||||||
#!/bin/sh
|
#!/bin/bash
|
||||||
|
set -euo pipefail
|
||||||
|
|
||||||
# Simple early detection of nvidia card
|
#################################################
|
||||||
grep 10de102d /proc/bus/pci/devices || exit 0
|
# Settings
|
||||||
|
|
||||||
# p2.xlarge
|
# A place on the host machine to cache these huge 1.6GB+ downloads in between reboots.
|
||||||
# 00f0 10de102d 4b 84000000 100000000c 0 8200000c 0 0 0 1000000 400000000 0 2000000 0 0 0 nvidia
|
ROOTFS_DIR=/rootfs
|
||||||
|
CACHE_DIR_HOST=/nvidia-bootstrap-cache
|
||||||
|
CACHE_DIR_CONTAINER="${ROOTFS_DIR}${CACHE_DIR_HOST}"
|
||||||
|
|
||||||
|
# AWS Instance Types to Nvidia Card Mapping (cut and pasted from AWS docs)
|
||||||
|
# Load the correct driver for the correct instance type
|
||||||
|
# Instances Product Type Product Series Product
|
||||||
|
# G2 GRID GRID Series GRID K520 <-- I think they meant G3
|
||||||
|
# P2 Tesla K-Series K-80
|
||||||
|
# P3 Tesla V-Series V100
|
||||||
|
# Both P2 and P3 are set for Cuda Toolkit 9.1
|
||||||
|
# http://www.nvidia.com/Download/index.aspx
|
||||||
|
declare -A class_to_driver_file
|
||||||
|
class_to_driver_file=( \
|
||||||
|
["g3"]="http://us.download.nvidia.com/XFree86/Linux-x86_64/367.124/NVIDIA-Linux-x86_64-367.124.run" \
|
||||||
|
["p2"]="http://us.download.nvidia.com/tesla/390.46/NVIDIA-Linux-x86_64-390.46.run" \
|
||||||
|
["p3"]="http://us.download.nvidia.com/tesla/390.46/NVIDIA-Linux-x86_64-390.46.run" \
|
||||||
|
)
|
||||||
|
declare -A class_to_driver_checksum
|
||||||
|
class_to_driver_checksum=( \
|
||||||
|
["g3"]="77f37939efeea4b6505842bed50445971992e303" \
|
||||||
|
["p2"]="57569ecb6f6d839ecc77fa10a2c573cc069990cc" \
|
||||||
|
["p3"]="57569ecb6f6d839ecc77fa10a2c573cc069990cc" \
|
||||||
|
)
|
||||||
|
|
||||||
|
# CUDA Files that need to be installed ~1.4GB
|
||||||
|
# First one is main installation
|
||||||
|
# Subsequent files are patches which need to be applied in order
|
||||||
|
# Order in the arrays below matters
|
||||||
|
# https://developer.nvidia.com/cuda-downloads
|
||||||
|
cuda_files=( \
|
||||||
|
"https://developer.nvidia.com/compute/cuda/9.1/Prod/local_installers/cuda_9.1.85_387.26_linux" \
|
||||||
|
"https://developer.nvidia.com/compute/cuda/9.1/Prod/patches/1/cuda_9.1.85.1_linux" \
|
||||||
|
"https://developer.nvidia.com/compute/cuda/9.1/Prod/patches/2/cuda_9.1.85.2_linux" \
|
||||||
|
"https://developer.nvidia.com/compute/cuda/9.1/Prod/patches/3/cuda_9.1.85.3_linux" \
|
||||||
|
)
|
||||||
|
cuda_files_checksums=( \
|
||||||
|
"1540658f4fe657dddd8b0899555b7468727d4aa8" \
|
||||||
|
"7ec6970ecd81163b0d02ef30d35599e7fd6e97d8" \
|
||||||
|
"cfa3b029b58fc117d8ce510a70efc848924dd565" \
|
||||||
|
"6269a2c5784b08997edb97ea0020fb4e6c8769ed" \
|
||||||
|
)
|
||||||
|
|
||||||
|
containsElement () { for e in "${@:2}"; do [[ "$e" = "$1" ]] && return 0; done; return 1; }
|
||||||
|
|
||||||
|
#################################################
|
||||||
|
# Ensure that we are on a proper AWS GPU Instance
|
||||||
|
|
||||||
|
AWS_INSTANCE_TYPE=$(curl -m 2 -fsSL http://169.254.169.254/latest/dynamic/instance-identity/document | jq -r ".instanceType" || true) # eg: p2.micro
|
||||||
|
AWS_INSTANCE_CLASS=$(echo $AWS_INSTANCE_TYPE | cut -d . -f 1 || true) # eg: p2
|
||||||
|
|
||||||
|
if [[ -z $AWS_INSTANCE_TYPE ]] || [[ -z $AWS_INSTANCE_CLASS ]]; then
|
||||||
|
echo "This machine is not an AWS instance"
|
||||||
|
echo " Exiting without installing GPU drivers"
|
||||||
|
exit 0
|
||||||
|
fi
|
||||||
|
|
||||||
|
classnames=${!class_to_driver_file[@]} # e.g. [ "g3", "p2", "p3" ]
|
||||||
|
if ! containsElement $AWS_INSTANCE_CLASS $classnames; then
|
||||||
|
echo "This machine is an AWS instance, but not a GPU instance"
|
||||||
|
echo " Exiting without installing GPU drivers"
|
||||||
|
exit 0
|
||||||
|
fi
|
||||||
|
|
||||||
|
echo "Identified machine as AWS_INSTANCE_TYPE[$AWS_INSTANCE_TYPE] AWS_INSTANCE_CLASS[$AWS_INSTANCE_CLASS]"
|
||||||
|
|
||||||
|
#################################################
|
||||||
|
# Install dependencies
|
||||||
|
|
||||||
|
# Install GCC and linux headers on the host machine
|
||||||
|
# This is unfortunate but necessary. That NVIDIA driver build must be
|
||||||
|
# compiled with the same version of GCC as the kernel. In addition,
|
||||||
|
# linux-headers are machine image specific.
|
||||||
|
|
||||||
|
if [[ ! -f ${ROOTFS_DIR}/usr/bin/gcc ]]; then
|
||||||
|
# Cuda requires regular stock gcc and host headers
|
||||||
|
chroot ${ROOTFS_DIR} apt-get update
|
||||||
|
# use --no-upgrade so that the c-libs are not upgraded, possible breaking programs and requiring restart
|
||||||
|
chroot ${ROOTFS_DIR} /bin/bash -c 'apt-get --no-upgrade -y install gcc linux-headers-$(uname -r)'
|
||||||
|
fi
|
||||||
|
|
||||||
|
if [[ ! -f ${ROOTFS_DIR}/usr/bin/gcc-7 ]]; then
|
||||||
|
echo "Installing gcc-7 on host machine"
|
||||||
|
|
||||||
|
# Temporarily add the debian "buster" repo where gcc-7 lives
|
||||||
|
# But first clear it out first if it already exists
|
||||||
|
sed -n '/buster/q;p' -i ${ROOTFS_DIR}/etc/apt/sources.list
|
||||||
|
echo "deb http://deb.debian.org/debian buster main" >> ${ROOTFS_DIR}/etc/apt/sources.list
|
||||||
|
|
||||||
|
# Install gcc-7
|
||||||
|
chroot ${ROOTFS_DIR} apt-get update
|
||||||
|
chroot ${ROOTFS_DIR} /bin/bash -c 'apt-get -y install linux-headers-$(uname -r)'
|
||||||
|
chroot ${ROOTFS_DIR} /bin/bash -c 'DEBIAN_FRONTEND=noninteractive apt-get -t buster --no-upgrade -y install gcc-7'
|
||||||
|
|
||||||
|
# Remove the debian "buster" repo line that was added above
|
||||||
|
sed -n '/buster/q;p' -i ${ROOTFS_DIR}/etc/apt/sources.list
|
||||||
|
chroot ${ROOTFS_DIR} apt-get update
|
||||||
|
fi
|
||||||
|
|
||||||
|
# Unload open-source nouveau driver if it exists
|
||||||
|
# The nvidia drivers won't install otherwise
|
||||||
|
# "g3" instances in particular have this module auto-loaded
|
||||||
|
chroot ${ROOTFS_DIR} modprobe -r nouveau || true
|
||||||
|
|
||||||
|
|
||||||
# This is pretty annoying.... note this is installed onto the host
|
#################################################
|
||||||
chroot /rootfs apt-get update
|
# Download and install the Nvidia drivers and cuda libraries
|
||||||
chroot /rootfs apt-get install --yes gcc
|
|
||||||
|
|
||||||
mkdir -p /rootfs/tmp
|
# Create list of URLs and Checksums by merging driver item with array of cuda files
|
||||||
cd /rootfs/tmp
|
downloads=(${class_to_driver_file[$AWS_INSTANCE_CLASS]} ${cuda_files[@]})
|
||||||
# TODO: We can't download over SSL - presents an akamai cert
|
checksums=(${class_to_driver_checksum[$AWS_INSTANCE_CLASS]} ${cuda_files_checksums[@]})
|
||||||
wget http://us.download.nvidia.com/XFree86/Linux-x86_64/375.39/NVIDIA-Linux-x86_64-375.39.run
|
|
||||||
echo '5e5b9fbf12f4f926ed70c1fe39f71d9d9f154abea0268b1cf035982b34bd7c94baef7667e4f647cc19a62702b46f63b3c3df9f1589261f7138ed2ff151af63cc NVIDIA-Linux-x86_64-375.39.run' | sha3sum -c - || exit 1
|
|
||||||
chmod +x NVIDIA-Linux-x86_64-375.39.run
|
|
||||||
chroot /rootfs /tmp/NVIDIA-Linux-x86_64-375.39.run --accept-license --ui=none
|
|
||||||
|
|
||||||
cd /rootfs/tmp
|
# Ensure that the cache directory exists
|
||||||
wget https://developer.nvidia.com/compute/cuda/8.0/Prod2/local_installers/cuda_8.0.61_375.26_linux-run
|
mkdir -p $CACHE_DIR_CONTAINER
|
||||||
chmod +x cuda_8.0.61_375.26_linux-run
|
|
||||||
# If we want to install samples as well, add: --samples
|
|
||||||
chroot /rootfs /tmp/cuda_8.0.61_375.26_linux-run --toolkit --silent
|
|
||||||
|
|
||||||
chroot /rootfs nvidia-smi -pm 1
|
# Download, verify, and execute each file
|
||||||
chroot /rootfs nvidia-smi -acp 0
|
length=${#downloads[@]}
|
||||||
chroot /rootfs nvidia-smi --auto-boost-default=0
|
for (( i=0; i<${length}; i++ )); do
|
||||||
chroot /rootfs nvidia-smi --auto-boost-permission=0
|
download=${downloads[$i]}
|
||||||
chroot /rootfs nvidia-smi -ac 2505,875
|
checksum=${checksums[$i]}
|
||||||
|
filename=$(basename $download)
|
||||||
|
filepath_host="${CACHE_DIR_HOST}/${filename}"
|
||||||
|
filepath_container="${CACHE_DIR_CONTAINER}/${filename}"
|
||||||
|
filepath_installed="${CACHE_DIR_CONTAINER}/${filename}.installed"
|
||||||
|
|
||||||
|
echo "Checking for file at $filepath_container"
|
||||||
|
if [[ ! -f $filepath_container ]] || ! (echo "$checksum $filepath_container" | sha1sum -c - 2>&1 >/dev/null); then
|
||||||
|
echo "Downloading $download"
|
||||||
|
curl -L $download > $filepath_container
|
||||||
|
chmod a+x $filepath_container
|
||||||
|
fi
|
||||||
|
|
||||||
# TODO: Problem ... why is this needed - why didn't this happen when we installed nvidia-uvm?
|
echo "Verifying sha1sum of file at $filepath_container"
|
||||||
# TODO: Problem ... we need to restart kubelet
|
if ! (echo "$checksum $filepath_container" | sha1sum -c -); then
|
||||||
|
echo "Failed to verify sha1sum for file at $filepath_container"
|
||||||
|
exit 1
|
||||||
|
fi
|
||||||
|
|
||||||
chroot /rootfs /sbin/modprobe nvidia-uvm
|
# Install the Nvidia driver and cuda libs
|
||||||
|
if [[ -f $filepath_installed ]]; then
|
||||||
|
echo "Detected prior install of file $filename on host"
|
||||||
|
else
|
||||||
|
echo "Installing file $filename on host"
|
||||||
|
if [[ $download =~ .*NVIDIA.* ]]; then
|
||||||
|
# Install the nvidia package (using gcc-7)
|
||||||
|
chroot ${ROOTFS_DIR} /bin/bash -c "CC=/usr/bin/gcc-7 $filepath_host --accept-license --silent"
|
||||||
|
touch $filepath_installed # Mark successful installation
|
||||||
|
elif [[ $download =~ .*local_installers.*cuda.* ]]; then
|
||||||
|
# Install the primary cuda library (using gcc)
|
||||||
|
chroot ${ROOTFS_DIR} $filepath_host --toolkit --silent --verbose
|
||||||
|
touch $filepath_installed # Mark successful installation
|
||||||
|
elif [[ $download =~ .*patches.*cuda.* ]]; then
|
||||||
|
# Install an update to the primary cuda library (using gcc)
|
||||||
|
chroot ${ROOTFS_DIR} $filepath_host --accept-eula --silent
|
||||||
|
touch $filepath_installed # Mark successful installation
|
||||||
|
else
|
||||||
|
echo "Unable to handle file $filepath_host"
|
||||||
|
fi
|
||||||
|
fi
|
||||||
|
done
|
||||||
|
|
||||||
if [ "$?" -eq 0 ]; then
|
#################################################
|
||||||
|
# Now that things are installed, let's output GPU info for debugging
|
||||||
|
chroot ${ROOTFS_DIR} nvidia-smi --list-gpus
|
||||||
|
|
||||||
|
# Configure and Optimize Nvidia cards now that things are installed
|
||||||
|
# AWS Optimizization Doc
|
||||||
|
# https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/optimize_gpu.html
|
||||||
|
# Nvidia Doc
|
||||||
|
# http://developer.download.nvidia.com/compute/DCGM/docs/nvidia-smi-367.38.pdf
|
||||||
|
|
||||||
|
# Common configurations
|
||||||
|
chroot ${ROOTFS_DIR} nvidia-smi -pm 1
|
||||||
|
chroot ${ROOTFS_DIR} nvidia-smi --auto-boost-default=0
|
||||||
|
chroot ${ROOTFS_DIR} nvidia-smi --auto-boost-permission=0
|
||||||
|
|
||||||
|
# Custom configurations per class of nvidia video card
|
||||||
|
case "$AWS_INSTANCE_CLASS" in
|
||||||
|
"g2" | "g3")
|
||||||
|
chroot ${ROOTFS_DIR} nvidia-smi -ac 2505,1177
|
||||||
|
;;
|
||||||
|
"p2")
|
||||||
|
chroot ${ROOTFS_DIR} nvidia-smi -ac 2505,875
|
||||||
|
chroot ${ROOTFS_DIR} nvidia-smi -acp 0
|
||||||
|
;;
|
||||||
|
"p3")
|
||||||
|
chroot ${ROOTFS_DIR} nvidia-smi -ac 877,1530
|
||||||
|
chroot ${ROOTFS_DIR} nvidia-smi -acp 0
|
||||||
|
;;
|
||||||
|
*)
|
||||||
|
;;
|
||||||
|
esac
|
||||||
|
|
||||||
|
# Load the Kernel Module
|
||||||
|
if ! chroot ${ROOTFS_DIR} /sbin/modprobe nvidia-uvm; then
|
||||||
|
echo "Unable to modprobe nvidia-uvm"
|
||||||
|
exit 1
|
||||||
|
fi
|
||||||
|
|
||||||
|
# Ensure that the device node exists
|
||||||
|
if ! chroot ${ROOTFS_DIR} test -e /dev/nvidia-uvm; then
|
||||||
# Find out the major device number used by the nvidia-uvm driver
|
# Find out the major device number used by the nvidia-uvm driver
|
||||||
D=`grep nvidia-uvm /proc/devices | awk '{print $1}'`
|
D=`grep nvidia-uvm /proc/devices | awk '{print $1}'`
|
||||||
|
chroot ${ROOTFS_DIR} mknod -m 666 /dev/nvidia-uvm c $D 0
|
||||||
chroot /rootfs mknod -m 666 /dev/nvidia-uvm c $D 0
|
|
||||||
else
|
|
||||||
echo "Unable to modprobe nvidia-uvm"
|
|
||||||
fi
|
fi
|
||||||
|
|
||||||
|
# Restart Kubelet
|
||||||
|
echo "Restarting Kubelet"
|
||||||
|
systemctl restart kubelet.service
|
||||||
|
|
|
||||||
Loading…
Reference in New Issue