Add defaults for GPU PCI passthrough

[Alex-Welsh]

This change is designed to drastically simplify pci passthrough GPU configuration.

The idea is that we have a data dictionary containing common GPU types, and templating to write that data to nova config files based on a simple group-to-gpu map.

Configuring passthrough is as simple as creating a dictionary like this:

gpu_group_map:
  compute_a100:
    - a100_80
  compute_v100:
    - v100_32
  compute_multi_gpu:
    - a100_80
    - v100_32

(and passing through the group names to Kolla-Ansible)

The pci-passthrough.yml playbook manages host configuration (and is pre-hooked to overcloud host configure)

Templates for nova-compute.conf, nova-api.conf, and nova-scheduler.conf have been added.

Changes have now been tested in a prod environment

[priteau]

This should be using the stackhpc.linux collection if possible: stackhpc/ansible-collection-linux#28

[Alex-Welsh]

This should be using the stackhpc.linux collection if possible: stackhpc/ansible-collection-linux#28

Agreed, though I'd rather get this merged so we can start using it, then update once the collection supports it

[Alex-Welsh]

I tried out the changes on a client deployment with three GPU types, worked very well

[MoteHue]

Just some docs changes now.
I've used this at a different customer site, worked a treat

Done

[Alex-Welsh]

Since we're not far from the 2025.1 release, should we target that instead?

[MoteHue]

Since we're not far from the 2025.1 release, should we target that instead?

As long as we backport, because we'll still need this for deployments before the next upgrade season

[Alex-Welsh]

Since we're not far from the 2025.1 release, should we target that instead?

As long as we backport, because we'll still need this for deployments before the next upgrade season

If we want it in 2024.1 anyway, we should just include it here and merge it up like normal

[MoteHue]

LGTM now, thanks @Alex-Welsh!

[bbezak]

bringing defaults nova-scheduling filters for everybody - not sure if that is best way. Plus empty pci stanzas in default nova configs. Can this be refactored as a opt-in feature ? - potentially as a gpu-passthrough mix-in environment, or at least opt-in variable.