userdocker-diffproml

userdocker-diffproml is a fork of userdocker which adds the following features:

• support for arguments with values (as opposed to just flags) in the config files
• support for the docker stop and exec subcommand
• automatic container names based on the username

These features will be merged into userdocker if and when the original author approves of the changes. At the moment it is not possible to install userdocker-diffproml alongside userdocker, since they use the same paths and executables.

userdocker-diffproml has received partial funding from the German Federal Ministry for Education and Research under grant agreement n° 01|S17075.

Userdocker

Userdocker is a wrapper that allows admins to grant restricted docker commandline access to users.

Note

Userdocker is currently in BETA state. Despite our ongoing efforts to test on our local infrastructure, further testing, reviewing and feedback are very welcome. Use with caution and watch the GitHub repo for issues and new releases!

Userdocker is aimed towards scientific high performance computing and cluster setups, as they exist in most universities or research groups. Often, such scientific computations have peculiar dependencies that are difficult to satisfy across linux distributions (and drive admins crazy ;) ).

In theory such use-cases could largely benefit from docker, as it would allow users to easily define environments themselves and run them basically without negative performance impact, as they run directly on the host's kernel. In reality however granting docker commandline access to users effectively makes them root equivalent on the host (root in container, volume mount...), making this prohibitive for cluster computing.

Userdocker solves this problem by wrapping the docker command and just making the safe parts available to users. Admins can decide what they consider safe (with sane defaults). The userdocker command largely follows the docker commandline syntax, so users can use it as an in-place replacement for the docker command.

Feedback / bugreports / contributions welcome:

https://github.com/saviola777/userdocker (this fork)

https://github.com/joernhees/userdocker (original repository)

Sample Usage:

# command line help (including subcommands the user is allowed to execute)
sudo userdocker -h

# (docker images) list images (and useful tree visualization)
sudo userdocker images
sudo userdocker dockviz

# (docker run) run a debian image with user (read-only) mounted home
sudo userdocker run -it --rm -v $HOME:$HOME:ro debian bash

# (docker attach) re-attach to own container after connection loss
sudo userdocker attach 438c7648e76b

# (docker stop) stop own container
sudo userdocker stop 438c7648e76b

# (docker ps) list running containers
sudo userdocker ps

# (docker pull / load) pull or load
sudo userdocker pull debian
sudo userdocker load < image.tar.gz

# (nvidia-docker) extensions for nvidia GPU support, see nvidia-smi -L for GPU IDs
alias nvidia-userdocker='userdocker --executor=nvidia-docker'
NV_GPU="GPU-37f6b436-5153-16af-0d26-88b96cd391bf,MIG-af8113ae-fec4-52d4-9cd7-299e6db5f9c6" nvidia-userdocker run -it --rm nvcr.io/nvidia/tensorflow
userdocker ps --gpu-used
userdocker ps --gpu-free

Features:

• Similar commandline interface as docker ... called userdocker ...

• Support for several docker commands / plugins (docker, nvidia-docker)

• Fine granular configurability for admins in /etc/userdocker/ allows to:

  • restrict runnable images if desired (allows admin reviews)
  • restrict run to locally available images
  • restrict available mount points (or enforce them, or default mount)
  • probe mounts (to make sure nfs automounts don't make docker sad)
  • enforce non-root user in container (same uid:gid as on host)
  • enforce dropping caps
  • enforce environment vars
  • enforce docker args
  • restrict port publishing
  • explicitly white-list available args to user
  • restrict allowed GPU access / reservations via NV_GPU (using GPU IDs obtained from nvidia-smi -L)

• System wide config + overrides for individual groups, gids, users, uids.

• Easy extensibility for further subcommands and args.

Installation:

The installation of userdocker-diffproml works in three steps:

1. Install package:

First make sure that docker is installed:

sudo docker version

Afterwards, as userdocker-diffproml is written in python3 and not yet available as python package:

sudo pip3 install -U https://github.com/saviola777/userdocker/archive/diffproml-1.2.0.tar.gz

The above is the preferable way of installation of the latest stable release.

If you want to try the current master (stable dev):

sudo pip3 install -U https://github.com/saviola777/userdocker/archive/master.tar.gz

Alternatively (and to contribute), you can clone this repo and execute:

sudo python3 setup.py install

This will give you a userdocker command that you can test with:

userdocker -h

2. Configuration:

Copy the default config to /etc/userdocker/config.py, then edit the file. The config contains tons of comments and explanations to help you make the right decisions for your scenario.

sudo cp /etc/userdocker/default.py /etc/userdocker/config.py

3. Allowing users to run sudo userdocker:

You should now allow the users in question to run sudo userdocker. This is basically done by adding a /etc/sudoers.d/userdocker file. If you want to grant this permission to all users in group users, add the following two lines:

Defaults env_keep += "NV_GPU"
%users ALL=(root) NOPASSWD: /usr/local/bin/userdocker

The first is strongly recommended in case you want to allow users to use nvidia GPUs from within docker containers via nvidia-docker (see EXECUTORS in config). Without it they cannot pass the NV_GPU environment variable to the userdocker (and thereby nvidia-docker) command to select their desired GPU(s). Make sure to include the right path to the userdocker executable, if in doubt or the above does not work, run whereis userdocker to find the correct path.

FAQ:

Why sudo?

Because it supports logging and is in general a lot more configurable than the alternatives. For example if you only want to make userdocker available on some nodes in your cluster, you can use the Host_List field:

%users node1,node2,node4=(root) /usr/local/bin/userdocker