Preliminary support for NVIDIA docker has been implemented for 8.0, deeper integration with Bright Cluster Manager is still a work in progress.
It is recommended to have a separate software image for the nodes containing GPU's, for example /cm/images/gpu-image so we can install the package in there once and provision all the GPU nodes at once.
Step 1: Make sure docker is installed first
Docker may have been previously installed in a different category, it is possible to execute the setup again for the gpu category
Example run of cm-docker-setup -e -i:
root@mbrt-ubuntu-trunk:~# cm-docker-setup -e -i
Run docker on head node (default: "no"):
no,
yes
> no
Node categories (default: "default", set to "none" if you don't want to deploy against any categories):
default,
gpu,
kube,
kube-gpu,
none
> gpu
Additional docker registries (default: none):
>
Storage backend (default: devicemapper):
default,
devicemapper
>
Use block device (default: no):
no,
yes
>
Loopback file size in GB for containers data (default: 100)
>
Loopback file size in GB for metadata (default: 2):
>
Setting up docker engine ...
Docker Engine has been setup successfully.
The compute nodes where docker will run has finished imageupdate.
Step 2: Install the cm-nvidia-docker and cuda drivers inside the GPU image
Example for Ubuntu:
chroot /cm/images/gpu-image
apt install cm-nvidia-docker cuda-driver
systemctl enable nvidia-docker
Example for CentOS/RHEL:
yum --installroot=/cm/images/gpu-image install cm-nvidia-docker cuda-driver
chroot /cm/images/gpu-image
systemctl enable nvidia-docker
Full output example from an Ubuntu system:
root@mbrt-ubuntu-trunk:~# chroot /cm/images/gpu-image
root@mbrt-ubuntu-trunk:/# apt install cm-nvidia-docker cuda-driver
Reading package lists... Done
Building dependency tree
Reading state information... Done
The following NEW packages will be installed:
cm-nvidia-docker
0 upgraded, 1 newly installed, 0 to remove and 177 not upgraded.
Need to get 2,261 kB of archives.
After this operation, 14.0 MB of additional disk space will be used.
Get:1 mirror://updates.brightcomputing.com/deb/cm/trunk/ubuntu/mirrors.txt xenial/main amd64 cm-nvidia-docker amd64 1.0.1-100015-cm-7f66df1237 [2,261 kB]
Fetched 2,261 kB in 0s (2,544 kB/s)
E: Can not write log (Is /dev/pts mounted?) - posix_openpt (2: No such file or directory)
Selecting previously unselected package cm-nvidia-docker.
(Reading database ... 146472 files and directories currently installed.)
Preparing to unpack .../cm-nvidia-docker_1.0.1-100015-cm-7f66df1237_amd64.deb ...
Unpacking cm-nvidia-docker (1.0.1-100015-cm-7f66df1237) ...
Setting up cm-nvidia-docker (1.0.1-100015-cm-7f66df1237) ...
Configuring user
Setting up permissions
setcap cap_fowner+pe /cm/local/apps/nvidia-docker/1.0.1/bin/nvidia-docker-plugin
Running ldconfig
Processing triggers for libc-bin (2.23-0ubuntu7) ...
root@mbrt-ubuntu-trunk:/# systemctl enable nvidia-docker
Created symlink /etc/systemd/system/multi-user.target.wants/nvidia-docker.service, pointing to /lib/systemd/system/nvidia-docker.service.
Step 3: Reboot GPU nodes
Recommended way of provisioning is issuing a reboot, as an imageupdate is not sufficient.
root@mbrt-ubuntu-trunk:~# pdsh -g category=gpu reboot
Final step: Test if all components are working correctly
The nvidia-docker service should be running, module files should be loaded and running the "Hello world" which is running nvidia-smi inside a container to output the info about the available GPU(s).
root@node003:~# systemctl status nvidia-docker
● nvidia-docker.service - NVIDIA Docker plugin
Loaded: loaded (/lib/systemd/system/nvidia-docker.service; enabled; vendor preset: enabled)
Active: active (running) since Wed 2017-08-23 17:16:56 CEST; 37s ago
Docs: https://github.com/NVIDIA/nvidia-docker/wiki
Process: 1741 ExecStartPost=/bin/sh -c /bin/echo unix:///var/lib/nvidia-docker/nvidia-docker.sock > /etc/docker/plugins/nvidia-docker.spec (code=exited, status=0/SUCCESS)
Process: 1733 ExecStartPost=/bin/sh -c /bin/mkdir -p /etc/docker/plugins (code=exited, status=0/SUCCESS)
Main PID: 1730 (nvidia-docker-p)
Tasks: 5
Memory: 8.2M
CPU: 179ms
CGroup: /system.slice/nvidia-docker.service
└─1730 /cm/local/apps/nvidia-docker/current/bin/nvidia-docker-plugin -s /var/lib/nvidia-docker -d /usr/local/nvidia-docker
Aug 23 17:16:56 node003 systemd[1]: Starting NVIDIA Docker plugin...
Aug 23 17:16:56 node003 nvidia-docker-plugin[1730]: /cm/local/apps/nvidia-docker/current/bin/nvidia-docker-plugin | 2017/08/23 17:16:56 Loading NVIDIA unified memory
Aug 23 17:16:56 node003 systemd[1]: Started NVIDIA Docker plugin.
Aug 23 17:16:56 node003 nvidia-docker-plugin[1730]: /cm/local/apps/nvidia-docker/current/bin/nvidia-docker-plugin | 2017/08/23 17:16:56 Loading NVIDIA management library
Aug 23 17:16:56 node003 nvidia-docker-plugin[1730]: /cm/local/apps/nvidia-docker/current/bin/nvidia-docker-plugin | 2017/08/23 17:16:56 Discovering GPU devices
Aug 23 17:16:57 node003 nvidia-docker-plugin[1730]: /cm/local/apps/nvidia-docker/current/bin/nvidia-docker-plugin | 2017/08/23 17:16:57 Provisioning volumes at /usr/local/nvidia-docker
Aug 23 17:16:57 node003 nvidia-docker-plugin[1730]: /cm/local/apps/nvidia-docker/current/bin/nvidia-docker-plugin | 2017/08/23 17:16:57 Serving plugin API at /var/lib/nvidia-docker
Aug 23 17:16:57 node003 nvidia-docker-plugin[1730]: /cm/local/apps/nvidia-docker/current/bin/nvidia-docker-plugin | 2017/08/23 17:16:57 Serving remote API at localhost:3476
root@node003:~# module load docker/engine/1.12.6
root@node003:~# docker ps
CONTAINER ID IMAGE COMMAND CREATED STATUS PORTS NAMES
root@node003:~# module load nvidia-docker/1.0.1
root@node003:~# nvidia-docker ps
CONTAINER ID IMAGE COMMAND CREATED STATUS PORTS NAMES
root@node003:~# nvidia-docker run --rm nvidia/cuda nvidia-smi
Wed Aug 23 15:19:00 2017
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 375.66 Driver Version: 375.66 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 Tesla K40c Off | 0000:00:08.0 Off | 0 |
| 23% 32C P8 22W / 235W | 2MiB / 11439MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
root@node003:~#
NVIDIA docker will need to pull an image the first time, the image it will use to run above example.
The service will also copy and mount needed NVIDIA libraries inside the container, this may also take a short while in case it's the first run.