Allow setting runtimeClassName

[murata-yu]

Hi, there.

In k3s containerd, set runtimeClass to pod without setting default-runtime to use GPU.
We can't set this in helm chart values.
Do you have any plans to change it so that we can set the runtimeClass?

[nikkon-dev]

@murata-yu,

If I got your question right, you should address this to the nvidia-container-toolkit. The dcgm-exporter does not have control over the runtime - it just needs to see GPUs inside the container, and the way those are visible does not matter.

[murata-yu]

OK, thanks.

[retsamedoc]

Er, pushing this off to the nvidia-container-toolkit would require changing the default runtime. Using runtimeClass is more correct and some CNCF kubernetes distros (like k3s) are auto detecting and exposing these 3rd party runtimes this way. Being able to run both runtimes is good from a security perspective (keep the attack vector small whenever possible) while also allowing me to utilize newer features in my other services.

The other tools being used as part of GPU Operator allow for runtimeClass to be set in the values.yaml (and DCGM Exporter will not start due to missing libraries if it is not using the right runtime).

[retsamedoc]

Er, how was this fixed?

[nvvfedorov]

Please confirm that the issue still is not resolved.

[nvvfedorov]

@retsamedoc , Please confirm. Do you want to be able to specify runtime in the helm chart?

[retsamedoc]

Yes. I would like to specify the runtime in the helm chart using the values.yaml similar to the other charts used for GPU operator.

[nvvfedorov]

@retsamedoc , As I see, today you can specify the runtimeClass when you install the dcgm-exporter with the helm:

dcgm-exporter/deployment/templates/daemonset.yaml

Line 48 in bb492ac

{{- if .Values.runtimeClassName }}

.

[nvvfedorov]

Done: #291