Requesting Intel ARC GPU Resource

[DizzieNight]

Describe the bug I am trying to add my arc gpu to my jellyfin pod. I have the NFD installed and it correctly labelling my node with the Intel Arc A310 with the following labels: nfd.node.kubernetes.io/feature-labels=gpu.intel.com/device-id.0300-56a6.count gpu.intel.com/device-id.0300-56a6.present gpu.intel.com/device-id.0380-1912.count gpu.intel.com/device-id.0380-1912.present gpu.intel.com/family,intel.feature.node.kubernetes.io/gpu

But I put the following into my jellyfin deployment: resources: requests: gpu.intel.com/i915: "1" limits: gpu.intel.com/i915: "1"

but it still won't find a node with a gpu. It keep coming up with the following error: 0/8 nodes are available: 1 node(s) were unschedulable, 3 node(s) had untolerated taint {node-role.kubernetes.io/control-plane: }, 4 Insufficient gpu.intel.com/i915. preemption: 0/8 nodes are available: 4 No preemption victims found for incoming pod, 4 Preemption is not helpful for scheduling.

To Reproduce Install Arc GPU, install NFD and request i915.

Expected behavior Jellyfin pod should attach to worker 4 which has the Arc A310 GPU

Screenshots If applicable, add screenshots to help explain your problem.

System (please complete the following information):

• OS version: Talos v1.8.2
• Kernel version: 6.6.58-talos
• Device plugins version: v0.31.0
• Hardware info: Baremetal

Additional context Add any other context about the problem here.

[tkatila]

Hi @DizzieNight, did you also deploy the Intel GPU device plugin to the cluster? NFD only doesn't yet suffice.

[DizzieNight]

I just checked and it doesn't seem so actually. I get this error when trying to install using the helm chart

Helm install failed for release node-feature-discovery/intel-gpu-plugin with chart intel-device-plugins-gpu@0.31.1: unable to build kubernetes objects from release manifest: resource mapping not found for name: "gpudeviceplugin" namespace: "" from "": no matches for kind "GpuDevicePlugin" in version "deviceplugin.intel.com/v1" ensure CRDs are installed first

I couldn't find where to install the CRDs though. Any thoughts?

[tkatila]

Helm install builds on the operator. Please see the steps here: https://github.com/intel/intel-device-plugins-for-kubernetes/blob/main/INSTALL.md#install-with-helm-charts

You can also install gpu-plugin via kubectl: https://github.com/intel/intel-device-plugins-for-kubernetes/blob/main/cmd/gpu_plugin/README.md#install-with-nfd

[DizzieNight]

I have installed everything and my node is getting the labels but nfd is coming up with a warning when installing:

W1115 12:36:57.853014 1148310 warnings.go:70] would violate PodSecurity "restricted:latest": restricted volume types (volumes "host-boot", "host-os-release", "host-sys", "host-usr-lib", "host-lib", "host-proc-swaps", "source-d", "features-d" use restricted volume type "hostPath")

And I am not sure how to fix it

[DizzieNight]

Nevermind, set the namespace to privileged using the following commands:

kubectl label namespace node-feature-discovery pod-security.kubernetes.io/enforce=privileged kubectl label namespace node-feature-discovery pod-security.kubernetes.io/audit=privileged kubectl label namespace node-feature-discovery pod-security.kubernetes.io/warn=privileged.

Although jellyfin still won't pick a node. The node I want jellyfin to install to has these labels, I don't see i915 here anywhere though. Does Arc use i915 as well or do I have to set it to something else?

[tkatila]

NFD seems to be working fine.

Can you check a few things: 1) Is the GPU device plugin running on the node? Check pods for that specific node. 1) Describe the target node and see if it has "gpu.intel.com/i915" resource?

[DizzieNight]

I did just notice the plugin isn't being created and this is what I get:

Not really sure how to fix it though

[mythi]

Not really sure how to fix it though

this looks to be the same pod security admission error you saw with NFD. The fix should be to label the operator/plugin namespace with the privileged PSA settings.

[DizzieNight]

Yep that worked, it's attaching now. Thank you for your help

[mythi]

Yep that worked, it's attaching now. Thank you for your help

I created #1909 to make the experience a bit smoother.