hammady/docker-nvidia-ansible

An Ansible playbook to install docker and nvidia drivers on Ubuntu

docker-nvidia-ansible

An Ansible playbook to install docker and nvidia drivers on Ubuntu Linux x86_64 machines.

Included software

• Docker Engine with bash completion.
• NVIDIA driver and Container Toolkit.

System requirements

1. Ansible on your controller machine (e.g. laptop). This should be different from the machines you are installing the packages to. The easiest way to install ansible (assuming python is installed) is to use pip:

   pip install --user ansible

   Verify that ansible is installed by running:

   ansible --version

   You also need a couple of additional Ansible collections, install using the command:

   ansible-galaxy install nvidia.nvidia_driver nvidia.nvidia_docker

2. One or more SSH connections to the target machine(s) with public key authentication configured. You can do that by editing your .ssh/config. For example:

   # ~/.ssh/config
   Host awsgpu1
       HostName 100.100.100.100
       User ubuntu
       IdentityFile ~/.ssh/id_rsa
       IdentitiesOnly yes

3. Edit inventory.ini to include the target machines:

   # inventory.ini
   [all]
   awsgpu1 ansible_ssh_user=ubuntu
   awsgpu2 ansible_ssh_user=ubuntu
   ...

Installation

To launch the installation process that will configure all target machines at once, run the following in the root directory of this repository:

ansible-playbook docker-nvidia.yaml --inventory ./inventory.ini

After a couple of minutes, you should get a report similar to the below if everything goes well:

...
PLAY RECAP ****************************************************************************************************************************************************************************************************************************************
awsgpu1                    : ok=34   changed=14   unreachable=0    failed=0    skipped=6    rescued=0    ignored=0   

Testing

To test that the installation was successful, restart then ssh to the machine(s) and run the following:

docker run --rm --gpus all nvidia/cuda:11.6.2-base-ubuntu20.04 nvidia-smi

You should get a report similar to the below:

Thu Feb  2 01:25:42 2023
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 515.86.01    Driver Version: 515.86.01    CUDA Version: 11.7     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla T4            On   | 00000000:00:1E.0 Off |                    0 |
| N/A   29C    P8    14W /  70W |      2MiB / 15360MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+