Metrics Exporter

Metrics Exporter#

Configure Metrics Exporter#

To enable the Metrics Exporter alongside the Network Operator, configure the fields under the spec.metricsExporter section in the NetworkConfig Custom Resource (CR):

apiVersion: amd.com/v1alpha1
kind: NetworkConfig
metadata:
    name: example-networkconfig
spec:
    ...
    metricsExporter:
        # Enable the Metrics Exporter component (default: false)
        enable: true

        # Specify the Metrics Exporter image
        image: "docker.io/rocm/k8s-network-metrics-exporter:nic-v1.0.0"

        # Image pull policy (default: IfNotPresent, or Always if tag is :latest)
        imagePullPolicy: "IfNotPresent"

        # Port for metrics endpoint (default: 5001)
        port: 5001

        # Service type for metrics access (default: ClusterIP)
        serviceType: "NodePort"

        # NodePort for external access (should be between 30000-32767)
        # Works with serviceType: "NodePort"
        nodePort: 32501

        # Use host networking (default: true)
        hostNetwork: true
        ...
    ...

Field Description#

Field Name

Description

Default Value

enable

Enable or disable the Metrics Exporter (true/false)

false

image

Container image to use for the Metrics Exporter

-

imagePullPolicy

Image pull policy: Always, Never, IfNotPresent

IfNotPresent

imageRegistrySecret

Secret for pulling images from private registries

-

port

Internal port for metrics endpoint

5001

serviceType

Service type for metrics access: ClusterIP, NodePort

ClusterIP

nodePort

External port for NodePort service (30000-32767)

hostNetwork

Enable host networking for the exporter pods

true

selector

Node selector for fine-grained pod placement

-

tolerations

Pod tolerations for scheduling

-

upgradePolicy

DaemonSet upgrade strategy configuration

-

config

Configmap containing exporter config.json

-

rbacConfig

Optional RBAC proxy configuration

-

Note:

  • The ImagePullPolicy field defaults to Always if the image tag is :latest, or to IfNotPresent for other tags. This follows the default Kubernetes behavior for ImagePullPolicy.

  • For the exporter pod to be able to fetch all metrics, we recommend running the pod with hostNetwork set to true, which is the default behavior.

The Metrics Exporter is deployed as a DaemonSet, which means one pod runs on each node that matches the specified selector.

Node Selection Behavior#

The NetworkConfig CR has a global spec.selector field that controls deployment of all operands (like device plugin, node labeller and metrics exporter) under the NetworkConfig. However, you can override this with component-specific selectors:

  • Global selector: Controls all operands when no component-specific selector is set

  • Component selector: When metricsExporter.selector is specified, it provides fine-grained control and overrides the global NetworkConfig selector for the Metrics Exporter component only

Deployment#

Metrics Exporter pods will start automatically after you update the NetworkConfig CR with the metrics exporter configuration:

kubectl get pods -n kube-amd-network
NAME                                                              READY   STATUS    RESTARTS   AGE
amd-network-operator-kmm-controller-8558dd8554-pnklg              1/1     Running   0          23s
amd-network-operator-kmm-webhook-server-6d54d5556-wn6dr           1/1     Running   0          23s
amd-network-operator-multus-multus-zm75t                          1/1     Running   0          23s
amd-network-operator-network-operator-charts-controller-ma64rjp   1/1     Running   0          23s
amd-network-operator-node-feature-discovery-gc-77d6d6449c-t85rz   1/1     Running   0          23s
amd-network-operator-node-feature-discovery-master-869f4bbprrhw   1/1     Running   0          23s
amd-network-operator-node-feature-discovery-worker-vbcxx          1/1     Running   0          23s
test-networkconfig-device-plugin-l89f9                            1/1     Running   0          8s
test-networkconfig-metrics-exporter-htdew                         1/1     Running   0          8s

Accessing Metrics#

The Metrics Exporter creates a Kubernetes service to expose metrics. For more information about Kubernetes services, see the official kubernetes documentation.

From Within the Cluster (ClusterIP)#

Access metrics from within the cluster using the ClusterIP:

# Access metrics via ClusterIP
curl http://<cluster-ip>:<port>/metrics

Note: The <cluster-ip> can be obtained by running kubectl get svc -n kube-amd-network and finding the metrics exporter service, the cluster IP is listed in the CLUSTER-IP column. The <port> is configured in your NetworkConfig (default: 5001).

From Outside the Cluster (NodePort)#

To access metrics from outside the cluster, you must enable NodePort by setting serviceType: "NodePort" in your NetworkConfig.

# Access metrics via any node's IP and NodePort
curl http://<node-ip>:<node-port>/metrics

Note: The <node-ip> can be the IP address of any node in your cluster. The <node-port> is either auto-assigned by Kubernetes or explicitly set via the nodePort field in your NetworkConfig.

Advanced Configuration#

To customize metrics fields, labels and other advanced setting, create a ConfigMap with the desired values and reference it in your NetworkConfig.

kubectl apply -f path/to/your/configmap.yaml

An example ConfigMap is available here: configmap.yaml

Note: When the Metrics Exporter is deployed through the Network Operator, GPU metrics are automatically disabled via the monitor-gpu=false argument (not user-configurable). This means:

  • Only NIC-related metrics are exported

  • Including GPU fields in your ConfigMap will not enable GPU metrics collection

  • The example ConfigMap is a generic configuration that works with both GPU and Network operators - each operator exports only its relevant metrics