Skip to content

RBLN Metrics Exporter

With Metrics Exporter, Prometheus collects NPU metrics and Grafana visualizes them.

The default local mode collects metrics from its own node, and gateway mode collects metrics from multiple remote hosts.

Kubernetes Installation (Helm)

Deploy Metrics Exporter through the NPU Operator Helm chart. First, complete the prerequisites and driver configuration in NPU Operator Installation.

Use the following setting in Helm values.

1
2
metricsExporter:
  enabled: true

Check that the exporter pod on each NPU node is Ready, then forward the metrics endpoint to a local port.

1
2
3
4
kubectl get pods -n rbln-system -o wide \
    -l app.kubernetes.io/component=metrics-exporter
kubectl port-forward -n rbln-system \
    service/rbln-metrics-exporter-service 9090:9090

Query the metrics from another terminal.

1
curl --fail http://127.0.0.1:9090/metrics

See Collect NPU metrics with Prometheus for ServiceMonitor configuration and scrape verification.

Bare-Metal Installation (Docker)

Install Docker and the RBLN driver on the Linux host, and check that RSMD is running. The following example assumes its gRPC endpoint is 127.0.0.1:50051.

1
sudo systemctl status rbln-smd

Run the Metrics Exporter container on each NPU host. The container reaches the RSMD on the same host through the host network (--network host), and --kubernetes-mode=off is set because this is not a Kubernetes environment.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
export EXPORTER_VERSION=v0.4.0

docker run -d \
    --name rbln-metrics-exporter \
    --restart unless-stopped \
    --network host \
    -e PROMETHEUS_METRIC_NAMES=true \
    -e NODE_NAME="$(hostname)" \
    docker.io/rebellions/rbln-metrics-exporter:${EXPORTER_VERSION} \
    --mode=local \
    --kubernetes-mode=off \
    --rbln-daemon-url=127.0.0.1:50051 \
    --port=9090

Check that the container is running, then query the metrics.

1
2
docker ps --filter name=rbln-metrics-exporter
curl --fail http://127.0.0.1:9090/metrics

Metric Names and Collection Verification

Setting PROMETHEUS_METRIC_NAMES=true selects rbln_npu_* names instead of their legacy RBLN_DEVICE_STATUS:* aliases.

After the first collection cycle, check the /metrics response for rbln_up 1 and rbln_npu_* metrics. If rbln_up stays at 0, check the RSMD service and connection address. For metric and label definitions, see Metrics Reference.