<br />

This document describes how to configure your Google Kubernetes Engine deployment
so that you can use Google Cloud Managed Service for Prometheus to collect metrics from

NVIDIA Data Center GPU Manager. This document shows you how to do the following:

- Set up the exporter for DCGM to report metrics.

<br />

These instructions apply only if you are using [managed collection](https://docs.cloud.google.com/stackdriver/docs/managed-prometheus/setup-managed.md.txt)
with Managed Service for Prometheus.
If you are using self-deployed collection, then see the

[source repository](https://github.com/NVIDIA/dcgm-exporter)
for DCGM Exporter

for installation information.

These instructions are provided as an example and are expected to work in
most Kubernetes environments.
For information about
a managed DCGM offering, see
[Collect and view DCGM metrics](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/dcgm-metrics.md.txt).


If you are having trouble installing an
application or exporter due to restrictive security or organizational policies,
then we recommend you consult open-source documentation for support.

For information about NVIDIA Data Center GPU Manager, see [NVIDIA DCGM](https://developer.nvidia.com/dcgm).

## Prerequisites

To collect metrics from

DCGM

by using
Managed Service for Prometheus and managed collection, your deployment must
meet the following requirements:

- Your cluster must be running Google Kubernetes Engine version 1.28.15-gke.2475000 or later.
- You must be running Managed Service for Prometheus with managed collection enabled. For more information, see [Get started with managed collection](https://docs.cloud.google.com/stackdriver/docs/managed-prometheus/setup-managed.md.txt).

  <br />

<br />

- Verify that you have [sufficient quota for NVIDIA GPUs](https://docs.cloud.google.com/compute/resource-usage.md.txt#gpu_quota).

- To enumerate GPU nodes in your GKE cluster and their GPU types in the
  relevant cluster, run the following command:

  ```
  kubectl get nodes -l cloud.google.com/gke-accelerator -o jsonpath='{range .items[*]}{@.metadata.name}{" "}{@.metadata.labels.cloud\.google\.com/gke-accelerator}{"\n"}{end}'
  ```
- Note that you might have to [install a compatible NVIDIA GPU driver](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/gpus.md.txt#installing_drivers) on the nodes if automatic
  installation was disabled or not supported for your GKE version. To verify that
  the NVIDIA GPU device plugin is running, run the following command:

  ```
  kubectl get pods -n kube-system | grep nvidia-gpu-device-plugin
  ```

> [!NOTE]
> **Note:** these instructions only apply for GKE Standard clusters. Refer to [Run GPUs
> in GKE Standard node pools](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/gpus#create) to create a Standard node pool with NVIDIA GPUs.

## Install the DCGM exporter

We recommend that you install the DCGM exporter,
[`DCGM-Exporter`](https://github.com/NVIDIA/dcgm-exporter), by using the following config:

```yaml
# Copyright 2023 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: nvidia-dcgm
  namespace: gmp-public
  labels:
    app: nvidia-dcgm
spec:
  selector:
    matchLabels:
      app: nvidia-dcgm
  updateStrategy:
    type: RollingUpdate
  template:
    metadata:
      labels:
        name: nvidia-dcgm
        app: nvidia-dcgm
    spec:
      affinity:
        nodeAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
            nodeSelectorTerms:
            - matchExpressions:
              - key: cloud.google.com/gke-accelerator
                operator: Exists
      tolerations:
      - operator: "Exists"
      volumes:
      - name: nvidia-install-dir-host
        hostPath:
          path: /home/kubernetes/bin/nvidia
          type: Directory
      containers:
      - image: "nvcr.io/nvidia/cloud-native/dcgm:3.3.0-1-ubuntu22.04"
        command: ["nv-hostengine", "-n", "-b", "ALL"]
        ports:
        - containerPort: 5555
          hostPort: 5555
        name: nvidia-dcgm
        securityContext:
          privileged: true
        volumeMounts:
        - name: nvidia-install-dir-host
          mountPath: /usr/local/nvidia
---
apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: nvidia-dcgm-exporter
  namespace: gmp-public
  labels:
    app.kubernetes.io/name: nvidia-dcgm-exporter
spec:
  selector:
    matchLabels:
      app.kubernetes.io/name: nvidia-dcgm-exporter
  updateStrategy:
    type: RollingUpdate
  template:
    metadata:
      labels:
        app.kubernetes.io/name: nvidia-dcgm-exporter
    spec:
      affinity:
        nodeAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
            nodeSelectorTerms:
            - matchExpressions:
              - key: cloud.google.com/gke-accelerator
                operator: Exists
      tolerations:
      - operator: "Exists"
      volumes:
      - name: nvidia-dcgm-exporter-metrics
        configMap:
          name: nvidia-dcgm-exporter-metrics
      - name: nvidia-install-dir-host
        hostPath:
          path: /home/kubernetes/bin/nvidia
          type: Directory
      - name: pod-resources
        hostPath:
          path: /var/lib/kubelet/pod-resources
      containers:
      - name: nvidia-dcgm-exporter
        image: nvcr.io/nvidia/k8s/dcgm-exporter:3.3.0-3.2.0-ubuntu22.04
        command: ["/bin/bash", "-c"]
        args:
        - hostname $NODE_NAME; dcgm-exporter --remote-hostengine-info $(NODE_IP) --collectors /etc/dcgm-exporter/counters.csv
        ports:
        - name: metrics
          containerPort: 9400
        securityContext:
          privileged: true
        env:
        - name: NODE_NAME
          valueFrom:
            fieldRef:
              fieldPath: spec.nodeName
        - name: "DCGM_EXPORTER_KUBERNETES_GPU_ID_TYPE"
          value: "device-name"
        - name: LD_LIBRARY_PATH
          value: /usr/local/nvidia/lib64
        - name: NODE_IP
          valueFrom:
            fieldRef:
              fieldPath: status.hostIP
        - name: DCGM_EXPORTER_KUBERNETES
          value: 'true'
        - name: DCGM_EXPORTER_LISTEN
          value: ':9400'
        volumeMounts:
        - name: nvidia-dcgm-exporter-metrics
          mountPath: "/etc/dcgm-exporter"
          readOnly: true
        - name: nvidia-install-dir-host
          mountPath: /usr/local/nvidia
        - name: pod-resources
          mountPath: /var/lib/kubelet/pod-resources
---
apiVersion: v1
kind: ConfigMap
metadata:
  name: nvidia-dcgm-exporter-metrics
  namespace: gmp-public
data:
  counters.csv: |
    # Utilization (the sample period varies depending on the product),,
    DCGM_FI_DEV_GPU_UTIL, gauge, GPU utilization (in %).
    DCGM_FI_DEV_MEM_COPY_UTIL, gauge, Memory utilization (in %).

    # Temperature and power usage,,
    DCGM_FI_DEV_GPU_TEMP, gauge, Current temperature readings for the device in degrees C.
    DCGM_FI_DEV_MEMORY_TEMP, gauge, Memory temperature for the device.
    DCGM_FI_DEV_POWER_USAGE, gauge, Power usage for the device in Watts.

    # Utilization of IP blocks,,
    DCGM_FI_PROF_SM_ACTIVE, gauge, The ratio of cycles an SM has at least 1 warp assigned
    DCGM_FI_PROF_SM_OCCUPANCY, gauge, The fraction of resident warps on a multiprocessor
    DCGM_FI_PROF_PIPE_TENSOR_ACTIVE, gauge, The ratio of cycles the tensor (HMMA) pipe is active (off the peak sustained elapsed cycles)
    DCGM_FI_PROF_PIPE_FP64_ACTIVE, gauge, The fraction of cycles the FP64 (double precision) pipe was active.
    DCGM_FI_PROF_PIPE_FP32_ACTIVE, gauge, The fraction of cycles the FP32 (single precision) pipe was active.
    DCGM_FI_PROF_PIPE_FP16_ACTIVE, gauge, The fraction of cycles the FP16 (half precision) pipe was active.

    # Memory usage,,
    DCGM_FI_DEV_FB_FREE, gauge, Framebuffer memory free (in MiB).
    DCGM_FI_DEV_FB_USED, gauge, Framebuffer memory used (in MiB).
    DCGM_FI_DEV_FB_TOTAL, gauge, Total Frame Buffer of the GPU in MB.

    # PCIE,,
    DCGM_FI_PROF_PCIE_TX_BYTES, gauge, Total number of bytes transmitted through PCIe TX
    DCGM_FI_PROF_PCIE_RX_BYTES, gauge, Total number of bytes received through PCIe RX

    # NVLink,,
    DCGM_FI_PROF_NVLINK_TX_BYTES, gauge, The number of bytes of active NvLink tx (transmit) data including both header and payload.
    DCGM_FI_PROF_NVLINK_RX_BYTES, gauge, The number of bytes of active NvLink rx (read) data including both header and payload.
```
To verify that DCGM Exporter is emitting metrics on the expected endpoints, do the following:

<br />

1. Set up port-forwarding with the following command:

   ```
   kubectl -n gmp-public port-forward POD_NAME 9400
   ```
2. Access the endpoint `localhost:9400/metrics` by using the
   browser or the `curl` utility in another terminal session.

You can customize the ConfigMap section to select [which GPU metrics](https://docs.nvidia.com/datacenter/dcgm/latest/user-guide/feature-overview.html#profiling-metrics)
to emit.

Alternatively, consider using [the official Helm chart](https://github.com/NVIDIA/dcgm-exporter#quickstart-on-kubernetes)
to install DCGM Exporter.


To apply configuration changes from a local file, run the following command:

```
kubectl apply -n NAMESPACE_NAME -f FILE_NAME
```

<br />

You can also
[use Terraform](https://docs.cloud.google.com/stackdriver/docs/managed-prometheus/setup-managed.md.txt#terraform-scrape)
to manage your configurations.

## Define a PodMonitoring resource

For target discovery, the Managed Service for Prometheus Operator
requires a PodMonitoring resource that corresponds to
DCGM Exporter in the same namespace.

You can use the following PodMonitoring configuration:

```yaml
# Copyright 2023 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

apiVersion: monitoring.googleapis.com/v1
kind: ClusterPodMonitoring
metadata:
  name: nvidia-dcgm-exporter
  labels:
    app.kubernetes.io/name: nvidia-dcgm-exporter
    app.kubernetes.io/part-of: google-cloud-managed-prometheus
spec:
  selector:
    matchLabels:
      app.kubernetes.io/name: nvidia-dcgm-exporter
  endpoints:
  - port: metrics
    interval: 30s
  targetLabels:
    metadata: []
```

To apply configuration changes from a local file, run the following command:

```
kubectl apply -n NAMESPACE_NAME -f FILE_NAME
```

<br />

You can also
[use Terraform](https://docs.cloud.google.com/stackdriver/docs/managed-prometheus/setup-managed.md.txt#terraform-scrape)
to manage your configurations.

## Verify the configuration

You can use Metrics Explorer to verify that you correctly configured
DCGM Exporter. It might take one or two minutes for
Cloud Monitoring to ingest your metrics.

To verify the metrics are ingested, do the following:

1. In the Google Cloud console, go to the
   **Metrics explorer** page:

   [Go to **Metrics explorer**](https://console.cloud.google.com/monitoring/metrics-explorer)

   <br />

   If you use the search bar to find this page, then select the result whose subheading is
   **Monitoring**.
2. In the toolbar of the query-builder pane, select the button whose name is **PromQL**.
3. Enter and run the following query:  

   ```
   DCGM_FI_DEV_GPU_UTIL{cluster="CLUSTER_NAME", namespace="gmp-public"}
   ```

<br />

## Troubleshooting

For information about troubleshooting metric ingestion problems, see
[Problems with collection from exporters](https://docs.cloud.google.com/stackdriver/docs/managed-prometheus/troubleshooting.md.txt#exporter-problems) in [Troubleshooting ingestion-side problems](https://docs.cloud.google.com/stackdriver/docs/managed-prometheus/troubleshooting.md.txt#ingest-problems).