Add daemon set as a way to deploy device plugin. #77
Conversation
"without NFD" is already possible. we need a mechanism that covers all plugins and is low maintenance. can you explain what problems the existing setup has? |
afaik we need to install operator in order to have it working, but like in mentioned docs it is possible to just deploy daemonset. Note: I created this PR as installation of apps in our cluster is done via helm charts so instead of manual deployment of daemon set I wanted to use official helm chart to do the same thing. |
charts/gpu-device-plugin/values.yaml
Outdated
| @@ -21,3 +21,10 @@ nodeSelector: | |||
| tolerations: [] | |||
|
|
|||
| nodeFeatureRule: true | |||
|
|
|||
| # to preserve backward compatibility | |||
| operator: true | |||
There was a problem hiding this comment.
| operator: true | |
| deployWithoutOperator: false |
charts/gpu-device-plugin/values.yaml
Outdated
| # to deploy the device plugin as a DaemonSet | ||
| daemonSet: | ||
| enabled: false |
| @@ -0,0 +1,79 @@ | |||
| {{- if .Values.daemonSet.enabled }} | |||
| @@ -2,7 +2,7 @@ | |||
| based on | |||
| deployments/operator/samples/deviceplugin_v1_gpudeviceplugin.yaml | |||
| */}} | |||
|
|
|||
| {{- if .Values.operator }} | |||
There was a problem hiding this comment.
Use ! deployWithoutOperator
| path: /var/run/cdi | ||
| type: DirectoryOrCreate | ||
| nodeSelector: | ||
| kubernetes.io/arch: amd64 |
There was a problem hiding this comment.
This should use nodeSelector from values.
There was a problem hiding this comment.
okay, but default value is:
intel.feature.node.kubernetes.io/gpu: 'true'
and I could not find a way to replace this selector with different one.
| metadata: | ||
| labels: | ||
| app: intel-gpu-plugin | ||
| spec: |
There was a problem hiding this comment.
I'd like to have the tolerations defined here as well.
| - name: HOST_IP | ||
| valueFrom: | ||
| fieldRef: | ||
| fieldPath: status.hostIP |
There was a problem hiding this comment.
You can drop this. It's only used with resourceManager which is being EoL'd.
| spec: | ||
| containers: | ||
| - name: intel-gpu-plugin | ||
| env: |
There was a problem hiding this comment.
Please also add support to configure GPU plugin with its different modes: sharedDevNum, enableMonitoring, allocationPolicy and logLevel. No need for the "resourceManager" as it's being EoL'd.
I don't see there's enough justification to accept the maintenance burden especially in this repo that is decoupled from the original reference YAML we have and as long as its GPU-only. |
Based on docs:
https://github.com/intel/intel-device-plugins-for-kubernetes/blob/main/cmd/gpu_plugin/advanced-install.md#install-to-all-nodes
Added possibility to deploy device plugin as a daemon set without NFD and operator.