CoreOS, Nvidia kernel module Dockerfile
Shell
Switch branches/tags
Nothing to show
Clone or download
Pull request Compare This branch is 30 commits ahead of hookenz:master.
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
devices @ aa6cc1b
.gitmodules
Dockerfile
README.md
install_nvidia.sh

README.md

CoreOS-Nvidia

This is a Dockerfile that builds a container with nvidia GPU drivers and CUDA.

"Installation"

The purpose of the steps in this section is to load the NVIDIA drivers.

  1. Build: docker build -t nvidia github.com/majidaldo/coreos-nvidia[#abranch] (can be built on any machine but not recommended. alternatively just download the image: docker pull majidaldoiongithub/coreos-nvidia:abranch)
  2. Run: docker run --rm --privileged nvidia (on the NVIDIA machine of course)
    You should see your NVIDIA harware printed.
  3. Confirm module is loaded: lsmod | grep -i nvidia
    You should see:
nvidia_uvm             77824  0
nvidia               8560640  1 nvidia_uvm
i2c_core               45056  3 i2c_i801,ipmi_ssif,nvidia

Usage

To make use of CUDA in a docker container, you must recreate the installation environment in the container. The Dockerfile contains instructions to recreate the environment if it is used as a base for other images. Just insert FROM nvidia at the beginning of your CUDA application Dockerfile (You could also just modify the 'installation' Dockerfile itself or enter the installation container.).

You can also use some containerized CUDA-enabled applications built by others. As a (tested) example, you can run CUDA-enabled Theano if the FROM instruction is changed to FROM nvidia.

After building the CUDA application image, run it passing NVIDIA devices. This script will generate a device list in a format suitable for use with docker run options. It also creates NVIDIA devices on the host so it has to run with escalated privileges (sudo).

docker build -t cuda-app ./cuda-app
NV_DEVICES=$(sudo ./coreos-nvidia/devices/nvidia_devices.sh)
docker run -ti $NV_DEVICES cuda-app

You can start over just by removing the modules sudo rmmod nvidia_uvm && sudo rmmod nvidia and building again (rmomod might not be necessary since the driver installation removes the older driver but I haven't checked).

Notes

There may be a better base image, and some post build cleanup that could be done to make the container smaller.

Currently the build uses uname -r to detect the running kernel version. There maybe be a better way to do this that offers more flexibility.

General Comments