Restrict certain devices from being discovered and reported by the plugin

[piotrgrubicki]

Is your feature request related to a problem? Please describe.
I have a desktop with Iris Xe and ARC 770. clinfo reports:

  Device Name                                     Intel(R) Arc(TM) A770M Graphics
  Device Name                                     Intel(R) Iris(R) Xe Graphics

The device plugin discovers both cards and reports them in k8s as gpu.intel.com/i915
When running a AI training on k8s, I need to run it on an accelerator, in my case an ARC card. With the current solution, my pod gets the Iris Xe device and fails.

Describe the solution you'd like
A way to specify in the device plugin deamon set yaml a list of devices not to be discovered and reported in k8s.

[eero-t]

I vaguely remember some earlier discussion about possible options for supporting development cases (developer machines with non-server GPU configurations), e.g. to ignore integrated GPUs. I could not find in which ticket(s) that happened or what the conclusions were though.

(@tkatila is currently on vacation, he can hopefully give better answer when he's back.)

But in general, K8s device plugins are intended to be used with nodes where all the resources of given type have same capacity.

For heterogeneous device support (on same node), you need K8s DRA driver: https://github.com/intel/intel-resource-drivers-for-kubernetes/#readme

K8s DRA API graduated to v1beta1 in K8s v1.32: https://github.com/kubernetes/enhancements/blob/master/keps/sig-node/4381-dra-structured-parameters/README.md

[mythi]

I could not find in which ticket(s) that happened or what the conclusions were though.

#1815 (comment)

[tkatila]

There are a few options that could help here:

• White/black listing with device ids
  • Only allow certain devices to be introduced as resources
  • Could be tuned/modified with re-configuring the plugin
  • Requires knowledge of device ids
• Detection of integrated/discrete GPUs
  • Not possible from sysfs, requires LevelZero
• Resource naming with device type/name
  • Rename the resource type based on the device id or name: i915-arc, i915-flex, i915-max, i915-integrated etc.
  • Requires logic to detect devices via LevelZero or some sort of lookup-table for known devices.

First option is the easiest because it can be tuned based on the scenario and can support new devices also. If we combine it with overlays, we could hide some of the device id complexities.