lacek/Dockerfile

## README.md

      
    Raw
  

              README.md
            
          
    Nvidia Docker-in-Docker

Problem

Q: Can Nvidia GPU be used in Docker daemon inside another Docker daemon?
A: Untested according to nvidia-docker wiki.
Testbed


Ubuntu 14.04.1 LTS (Kernel: 3.13.0-170-generic)
Docker CE 18.06.1-ce
nvidia-docker2 2.0.3
Nvidia driver 430.34
i7-3930K CPU @ 3.20GHz
GTX 680

Test

Test Outer Docker

docker run --runtime=nvidia --rm nvidia/cuda:10.1-base nvidia-smi
Looks good:
Fri Jul 12 09:30:46 2019       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 430.34       Driver Version: 430.34       CUDA Version: 10.1     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 680     Off  | 00000000:01:00.0 N/A |                  N/A |
| 30%   37C    P0    N/A /  N/A |      0MiB /  4035MiB |     N/A      Default |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|    0                    Not Supported                                       |
+-----------------------------------------------------------------------------+

Build Image

docker build -t docker:cuda .

Test Inner Docker

docker save nvidia/cuda:10.1-base > cuda_10.1-base.tar
docker run -it --rm --runtime=nvidia --privileged -v $PWD:/workdir -w /workdir docker:cuda bash
docker load < cuda_10.1-base.tar
docker run --runtime=nvidia --rm nvidia/cuda:10.1-base nvidia-smi
Benchmark with Tensorflow

Building Tensorflow from source

GTX680 is pretty old and has compute capability of 3.0, which is not supported by current builds of Tensorflow. Following the instructions of official documentation, a wheel package is built with the image tensorflow/tensorflow:devel-gpu-py3 to include compute capability of 3.0.
Benchmark Outer Docker

docker run -it --runtime=nvidia --rm -v $PWD:/workdir -w /workdir tensorflow/tensorflow:1.14.0-gpu-py3 bash
pip install tensorflow-1.14.0-cp36-cp36m-linux_x86_64.whl
python matmul.py gpu 10000
# 1st: Time taken: 0:00:02.589672
# 2nd: Time taken: 0:00:02.506626
# 3rd: Time taken: 0:00:02.570411
# Average: 2.56s

python matmul.py cpu 10000
# 1st: Time taken: 0:00:12.053325
# 2nd: Time taken: 0:00:12.104460
# 3rd: Time taken: 0:00:12.023136
# Average: 12.06s

# CPU:GPU ~ 4.71
Benchmark Inner Docker

docker save tensorflow/tensorflow:1.14.0-gpu-py3 > tensorflow_1.14.0-gpu-py3.tar
docker run -it --rm --runtime=nvidia --privileged -v $PWD:/workdir -w /workdir docker:cuda bash
docker load < tensorflow_1.14.0-gpu-py3.tar
docker run -it --runtime=nvidia --rm -v $PWD:/workdir -w /workdir tensorflow/tensorflow:1.14.0-gpu-py3 bash
pip install tensorflow-1.14.0-cp36-cp36m-linux_x86_64.whl
python matmul.py gpu 10000
# 1st: Time taken: 0:00:02.594973
# 2nd: Time taken: 0:00:02.574000
# 3rd: Time taken: 0:00:02.540780
# Average: 2.57s
# 0.39% increase

python matmul.py cpu 10000
# 1st: Time taken: 0:00:12.861318
# 2nd: Time taken: 0:00:12.685900
# 3rd: Time taken: 0:00:12.373268
# Average: 12.64s
# 4.81% increase

# CPU:GPU ~ 4.92
# 4.46% increase

  
## Dockerfile
FROM nvidia/cuda:10.1-base-ubuntu18.04

RUN  sed -i -E -e 's@//(archive|security).ubuntu@//hk.archive.ubuntu@' /etc/apt/sources.list \
    && apt-get update \
    && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
        apt-transport-https \
        ca-certificates \
        curl \
        gnupg2 \
        iptables \
    && curl -fsSL https://download.docker.com/linux/ubuntu/gpg | apt-key add - \
    && echo "deb https://download.docker.com/linux/ubuntu bionic stable" | tee /etc/apt/sources.list.d/docker.list \
    && curl -fsSL https://nvidia.github.io/nvidia-docker/gpgkey | apt-key add - \
    && curl -fsSL https://nvidia.github.io/nvidia-docker/ubuntu18.04/nvidia-docker.list -o /etc/apt/sources.list.d/nvidia-docker.list \
    && apt-get update \
    && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
        docker-ce=5:18.09.7~3-0~ubuntu-bionic \
        nvidia-docker2=2.0.3+docker18.09.7-3 \
        nvidia-container-runtime=2.0.0+docker18.09.7-3 \
    && rm -rf /var/lib/apt/lists/*

COPY entrypoint.sh /sbin/entrypoint.sh

VOLUME /var/lib/docker
EXPOSE 2375

ENTRYPOINT ["/sbin/entrypoint.sh"]

## entrypoint.sh
#!/bin/sh
# adapted from https://github.com/moby/moby/blob/3b5fac462d21ca164b3778647420016315289034/hack/dind

# apparmor sucks and Docker needs to know that it's in a container (c) @tianon
export container=docker

# as of docker 1.8, cgroups will be mounted in the container
if ! mountpoint -q /sys/fs/cgroup; then

    # First, make sure that cgroups are mounted correctly.
    CGROUP=/cgroup

    mkdir -p "$CGROUP"

    if ! mountpoint -q "$CGROUP"; then
        mount -n -t tmpfs -o uid=0,gid=0,mode=0755 cgroup $CGROUP || {
            echo >&2 'Could not make a tmpfs mount. Did you use --privileged?'
            exit 1
        }
    fi

    # Mount the cgroup hierarchies exactly as they are in the parent system.
    for HIER in $(cut -d: -f2 /proc/1/cgroup); do

        # If cgroup hierarchy is named(mounted with "-o name=foo") we
        # need to mount it in $CGROUP/foo to create exect same
        # directoryes as on host. Else we need to mount it as is e.g.
        # "subsys1,subsys2" if it has two subsystems

        # Named, control-less cgroups are mounted with "-o name=foo"
        # (and appear as such under /proc/<pid>/cgroup) but are usually
        # mounted on a directory named "foo" (without the "name=" prefix).
        # Systemd and OpenRC (and possibly others) both create such a
        # cgroup. So just mount them on directory $CGROUP/foo.

        OHIER=$HIER
        HIER="${HIER#*name=}"

        mkdir -p "$CGROUP/$HIER"

        if ! mountpoint -q "$CGROUP/$HIER"; then
            mount -n -t cgroup -o "$OHIER" cgroup "$CGROUP/$HIER"
        fi
        # Likewise, on at least one system, it has been reported that
        # systemd would mount the CPU and CPU accounting controllers
        # (respectively "cpu" and "cpuacct") with "-o cpuacct,cpu"
        # but on a directory called "cpu,cpuacct" (note the inversion
        # in the order of the groups). This tries to work around it.

        if [ "$HIER" = 'cpuacct,cpu' ]; then
            ln -s "$HIER" "$CGROUP/cpu,cpuacct"
        fi

        # If hierarchy has multiple subsystems, in /proc/<pid>/cgroup
        # we will see ":subsys1,subsys2,subsys3,name=foo:" substring,
        # we need to mount it to "$CGROUP/foo" and if there were no
        # name to "$CGROUP/subsys1,subsys2,subsys3", so we must create
        # symlinks for docker daemon to find these subsystems:
        # ln -s $CGROUP/foo $CGROUP/subsys1
        # ln -s $CGROUP/subsys1,subsys2,subsys3 $CGROUP/subsys1

        OSUBSYSTEMS="${HIER%name=*}"
        SUBSYSTEMS="$(echo $OSUBSYSTEMS | tr , ' ')"
        if [ "$SUBSYSTEMS" != "$OSUBSYSTEMS" ]; then
            for SUBSYS in $SUBSYSTEMS
            do
                ln -s "$CGROUP/$HIER" "$CGROUP/$SUBSYS"
            done
        fi
    done
fi

if [ -d /sys/kernel/security ] && ! mountpoint -q /sys/kernel/security; then
    mount -t securityfs none /sys/kernel/security || {
        echo >&2 'Could not mount /sys/kernel/security.'
        echo >&2 'AppArmor detection and --privileged mode might break.'
    }
fi

# Mount /tmp (conditionally)
if ! mountpoint -q /tmp; then
    mount -t tmpfs none /tmp
fi

# Start docker service and poll
service docker start

TIMEOUT=$(( $(date +%s) + 30 ))
until docker info >/dev/null 2>&1; do
    if [ $(date +%s) -ge $TIMEOUT ]; then
        echo >&2 'Timed out trying to connect to internal docker host.'
        break
    fi
    sleep 1
done

if [ $# -gt 0 ]; then
    exec "$@"
fi

## matmul.py
# copied from https://databricks.com/tensorflow/using-a-gpu
import sys
import numpy as np
import tensorflow as tf
from datetime import datetime

device_name = sys.argv[1]  # Choose device from cmd line. Options: gpu or cpu
shape = (int(sys.argv[2]), int(sys.argv[2]))
if device_name == "gpu":
    device_name = "/gpu:0"
else:
    device_name = "/cpu:0"

with tf.device(device_name):
    random_matrix = tf.random_uniform(shape=shape, minval=0, maxval=1)
    dot_operation = tf.matmul(random_matrix, tf.transpose(random_matrix))
    sum_operation = tf.reduce_sum(dot_operation)


startTime = datetime.now()
with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as session:
        result = session.run(sum_operation)
        print(result)

# It can be hard to see the results on the terminal with lots of output -- add some newlines to improve readability.
print("\n" * 5)
print("Shape:", shape, "Device:", device_name)
print("Time taken:", datetime.now() - startTime)

print("\n" * 5)
	FROM nvidia/cuda:10.1-base-ubuntu18.04

	RUN sed -i -E -e 's@//(archive\|security).ubuntu@//hk.archive.ubuntu@' /etc/apt/sources.list \
	&& apt-get update \
	&& DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
	apt-transport-https \
	ca-certificates \
	curl \
	gnupg2 \
	iptables \
	&& curl -fsSL https://download.docker.com/linux/ubuntu/gpg \| apt-key add - \
	&& echo "deb https://download.docker.com/linux/ubuntu bionic stable" \| tee /etc/apt/sources.list.d/docker.list \
	&& curl -fsSL https://nvidia.github.io/nvidia-docker/gpgkey \| apt-key add - \
	&& curl -fsSL https://nvidia.github.io/nvidia-docker/ubuntu18.04/nvidia-docker.list -o /etc/apt/sources.list.d/nvidia-docker.list \
	&& apt-get update \
	&& DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
	docker-ce=5:18.09.7~3-0~ubuntu-bionic \
	nvidia-docker2=2.0.3+docker18.09.7-3 \
	nvidia-container-runtime=2.0.0+docker18.09.7-3 \
	&& rm -rf /var/lib/apt/lists/*

	COPY entrypoint.sh /sbin/entrypoint.sh

	VOLUME /var/lib/docker
	EXPOSE 2375

	ENTRYPOINT ["/sbin/entrypoint.sh"]
	#!/bin/sh
	# adapted from https://github.com/moby/moby/blob/3b5fac462d21ca164b3778647420016315289034/hack/dind

	# apparmor sucks and Docker needs to know that it's in a container (c) @tianon
	export container=docker

	# as of docker 1.8, cgroups will be mounted in the container
	if ! mountpoint -q /sys/fs/cgroup; then

	# First, make sure that cgroups are mounted correctly.
	CGROUP=/cgroup

	mkdir -p "$CGROUP"

	if ! mountpoint -q "$CGROUP"; then
	mount -n -t tmpfs -o uid=0,gid=0,mode=0755 cgroup $CGROUP \|\| {
	echo >&2 'Could not make a tmpfs mount. Did you use --privileged?'
	exit 1
	}
	fi

	# Mount the cgroup hierarchies exactly as they are in the parent system.
	for HIER in $(cut -d: -f2 /proc/1/cgroup); do

	# If cgroup hierarchy is named(mounted with "-o name=foo") we
	# need to mount it in $CGROUP/foo to create exect same
	# directoryes as on host. Else we need to mount it as is e.g.
	# "subsys1,subsys2" if it has two subsystems

	# Named, control-less cgroups are mounted with "-o name=foo"
	# (and appear as such under /proc/<pid>/cgroup) but are usually
	# mounted on a directory named "foo" (without the "name=" prefix).
	# Systemd and OpenRC (and possibly others) both create such a
	# cgroup. So just mount them on directory $CGROUP/foo.

	OHIER=$HIER
	HIER="${HIER#*name=}"

	mkdir -p "$CGROUP/$HIER"

	if ! mountpoint -q "$CGROUP/$HIER"; then
	mount -n -t cgroup -o "$OHIER" cgroup "$CGROUP/$HIER"
	fi
	# Likewise, on at least one system, it has been reported that
	# systemd would mount the CPU and CPU accounting controllers
	# (respectively "cpu" and "cpuacct") with "-o cpuacct,cpu"
	# but on a directory called "cpu,cpuacct" (note the inversion
	# in the order of the groups). This tries to work around it.

	if [ "$HIER" = 'cpuacct,cpu' ]; then
	ln -s "$HIER" "$CGROUP/cpu,cpuacct"
	fi

	# If hierarchy has multiple subsystems, in /proc/<pid>/cgroup
	# we will see ":subsys1,subsys2,subsys3,name=foo:" substring,
	# we need to mount it to "$CGROUP/foo" and if there were no
	# name to "$CGROUP/subsys1,subsys2,subsys3", so we must create
	# symlinks for docker daemon to find these subsystems:
	# ln -s $CGROUP/foo $CGROUP/subsys1
	# ln -s $CGROUP/subsys1,subsys2,subsys3 $CGROUP/subsys1

	OSUBSYSTEMS="${HIER%name=*}"
	SUBSYSTEMS="$(echo $OSUBSYSTEMS \| tr , ' ')"
	if [ "$SUBSYSTEMS" != "$OSUBSYSTEMS" ]; then
	for SUBSYS in $SUBSYSTEMS
	do
	ln -s "$CGROUP/$HIER" "$CGROUP/$SUBSYS"
	done
	fi
	done
	fi

	if [ -d /sys/kernel/security ] && ! mountpoint -q /sys/kernel/security; then
	mount -t securityfs none /sys/kernel/security \|\| {
	echo >&2 'Could not mount /sys/kernel/security.'
	echo >&2 'AppArmor detection and --privileged mode might break.'
	}
	fi

	# Mount /tmp (conditionally)
	if ! mountpoint -q /tmp; then
	mount -t tmpfs none /tmp
	fi

	# Start docker service and poll
	service docker start

	TIMEOUT=$(( $(date +%s) + 30 ))
	until docker info >/dev/null 2>&1; do
	if [ $(date +%s) -ge $TIMEOUT ]; then
	echo >&2 'Timed out trying to connect to internal docker host.'
	break
	fi
	sleep 1
	done

	if [ $# -gt 0 ]; then
	exec "$@"
	fi
	# copied from https://databricks.com/tensorflow/using-a-gpu
	import sys
	import numpy as np
	import tensorflow as tf
	from datetime import datetime

	device_name = sys.argv[1] # Choose device from cmd line. Options: gpu or cpu
	shape = (int(sys.argv[2]), int(sys.argv[2]))
	if device_name == "gpu":
	device_name = "/gpu:0"
	else:
	device_name = "/cpu:0"

	with tf.device(device_name):
	random_matrix = tf.random_uniform(shape=shape, minval=0, maxval=1)
	dot_operation = tf.matmul(random_matrix, tf.transpose(random_matrix))
	sum_operation = tf.reduce_sum(dot_operation)


	startTime = datetime.now()
	with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as session:
	result = session.run(sum_operation)
	print(result)

	# It can be hard to see the results on the terminal with lots of output -- add some newlines to improve readability.
	print("\n" * 5)
	print("Shape:", shape, "Device:", device_name)
	print("Time taken:", datetime.now() - startTime)

	print("\n" * 5)