It’s far to unclear from the documentation, first, how to limit labels you want to appear, but also, that when adding other Node Feature Discovery stuff (such as NVIDIA’s GPU feature discovery) that the namespaces are actually combined.
First make sure you have your .master.extraLabelNs set to -
nvidia.com. This effectively causes feature.node.kubernetes.io to be
fused (and ignored) with nvidia.com so that the
worker.config.core.labelWhiteList regular expression starts after the
/ in either of them.
You might struggle initially with what the regular expression actually
is. To me intuitively I thought nvidia.com needed to be included, but
as I just explained, that’s wrong. Instead, you just add the prefixes
from the nvidia.com namespace is if they were all combined with
everything from feature.node.kubernetes.io. Adding cuda|gpu|gfd
currently grabs all the stuff in nvidia.com.
worker:
config: |
core:
labelWhiteList: '^(custom|system|pci|cuda|gpu|gfd)'
Also do not forget to set your nodeSelector in the GPU helm chart
(DaemonSet) and that you include that label to link to in your white
list (ours is pci-0 something).
#k8s #nfd #gfd #gpu #featurediscovery #kubernetes #tips