Re: [atomic-devel] Add Nvidia GPU to Centos Atomic Node

[Dusty Mabe <dusty dustymabe com>]

On 11/01/2017 03:39 PM, hskarlupka wrote:
> Hello,
> 

Hi Heath!

> I'm running a 4 node Kubernetes bare metal cluster using Centos Atomic 
> as the base OS.  My team is interested in installing a Nvidia GPU in one 
> of the nodes.  One concern I have is on the installation of the Nvidia 
> drivers.  The first thing that comes to mind is that it won't like the 
> /usr RO filesystem.  The only guides I've seen are for CoreOS Nvidia 
> installs that use a set of Docker containers to install drivers and use 
> the device:
> 
> https://github.com/src-d/coreos-nvidia
> 
> https://github.com/ryanolson/CoreOS-GPU
> 
> Has anyone run into this before for CentOS Atomic?
> 

Unfortunately I don't have good answers for you on this one. I was going to suggest
package layering via `rpm-ostree install foo.rpm` but there are some compatibility 
issues with the nvidia rpms and rpm-ostree that don't quite work [1] [2].

Really you have two options.

1. unlock your ostree and install the rpms
2. build the kernel module for the kernel you are targeting and deliver via a system
   container like in https://github.com/giuseppe/hellomod


The `unlock your ostree` approach basically converts the atomic host back into a system
that is completely writable:

- ostree admin unlock --hotfix
- alias yum='/usr/share/yum-cli/yummain.py'
- yum install -y epel-release kernel-{devel,headers}-$(uname -r) https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-repo-rhel7-9.0.176-1.x86_64.rpm
- yum install -y nvidia-kmod

should do it.

Sorry I don't have better answers on this right now.

Dusty 


[1] https://github.com/projectatomic/rpm-ostree/issues/1091
[2] https://github.com/projectatomic/rpm-ostree/issues/233#issuecomment-342643731