Node Feature Discovery

Install Node Feature Discovery (NFD) next to Mokka and verify that NFD turns Mokka's feature file into the same GPU-presence label that a real node would expose.

Mokka does not write Kubernetes node labels. Its node agent writes pci-10de.present=true to the NFD local-source directory, and NFD owns the resulting feature.node.kubernetes.io/pci-10de.present=true label. This keeps the simulated path aligned with a real deployment.

Prerequisites

• Docker and Kind
• Helm 3.8 or newer
• kubectl

This guide takes about 10 minutes and does not require a GPU.

Step 1 — Create a cluster and install Mokka

Kind's default cluster contains only a control-plane node. That node may be tainted NoSchedule, so the NFD worker would never run there. Create one worker explicitly for this guide:

$ cat > mokka-nfd-kind.yaml <<'EOF'
kind: Cluster
apiVersion: kind.x-k8s.io/v1alpha4
nodes:
- role: control-plane
- role: worker
EOF
$ kind create cluster --name mokka-nfd --config mokka-nfd-kind.yaml
$ helm install nvml-mock oci://ghcr.io/nvidia/k8s-test-infra/chart/nvml-mock \
    --namespace mokka --create-namespace \
    --wait --timeout 120s
$ kubectl label node --all mokka.nvidia.com/type=sgpu
$ rm mokka-nfd-kind.yaml

The chart's default nodeLabels.featuresDir is /etc/kubernetes/node-feature-discovery/features.d. It is mounted from each node so that NFD can read the file after the node agent writes it.

Step 2 — Install NFD

The repository's NFD end-to-end scenario is pinned to version 0.19.0. Use the same version here so the local-source behavior is reproducible:

$ helm repo add nfd https://kubernetes-sigs.github.io/node-feature-discovery/charts
$ helm repo update
$ helm upgrade --install nfd nfd/node-feature-discovery \
    --namespace node-feature-discovery --create-namespace \
    --version 0.19.0 \
    --wait --timeout 120s

Wait for NFD to become ready before checking the label:

$ kubectl -n node-feature-discovery wait \
    --for=condition=ready pod \
    -l app.kubernetes.io/name=node-feature-discovery \
    --timeout=120s

Step 3 — Verify the label and its provenance

NFD polls the local-source directory and adds the namespaced label to nodes where the feature file exists:

$ kubectl get nodes \
    -o custom-columns='NODE:.metadata.name,NFD_GPU:.metadata.labels.feature\.node\.kubernetes\.io/pci-10de\.present'

The simulated nodes should report true:

NODE                         NFD_GPU
mokka-nfd-control-plane      true
mokka-nfd-worker             true

The input is a feature file, not a label written by Mokka. You can inspect it from the node-agent container:

$ kubectl -n mokka exec daemonset/nvml-mock -c node-agent -- \
    cat /host/etc/kubernetes/node-feature-discovery/features.d/nvml-mock.features

Expected output:

pci-10de.present=true

NFD records labels that it owns in the nfd.node.kubernetes.io/feature-labels node annotation. Check that annotation when a workload or admission rule needs to distinguish NFD-owned labels from labels written by another component:

$ kubectl get node mokka-nfd-control-plane \
    -o jsonpath='{.metadata.annotations.nfd\.node\.kubernetes\.io/feature-labels}{"\n"}'

The annotation includes pci-10de.present without the feature.node.kubernetes.io/ prefix.

Use the label in a workload

Select simulated nodes with the NFD label just as you would select a real GPU node, then confirm that Kubernetes scheduled the pod on a labelled node:

$ kubectl apply -f - <<'EOF'
apiVersion: v1
kind: Pod
metadata:
  name: nfd-labelled-workload
spec:
  nodeSelector:
    feature.node.kubernetes.io/pci-10de.present: "true"
  containers:
    - name: shell
      image: busybox:1.36
      command: ["sleep", "300"]
EOF
$ kubectl get pod nfd-labelled-workload -o wide

Troubleshooting

The label is missing. Check that the Mokka DaemonSet is ready and that the feature file is present in the node-agent container. The NFD chart and Mokka must use the same featureFilesDir and nodeLabels.featuresDir paths.

The NFD pod is ready but no label appears. Inspect the NFD worker logs:

$ kubectl -n node-feature-discovery logs -l app.kubernetes.io/name=node-feature-discovery

The node has a label but the workload is pending. Confirm that the pod's nodeSelector uses the complete namespaced key and the string value "true".

Clean up

$ kind delete cluster --name mokka-nfd

Related

To read about	See
Mokka chart values	Installation
GPU allocation through the NVIDIA device plugin	NVIDIA Device Plugin
NFD label behavior in the chart	Helm chart reference