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.
- Configure a PodMonitoring resource for Managed Service for Prometheus
to collect the exported metrics.
These instructions apply only if you are using
managed collection
with Managed Service for Prometheus.
If you are using self-deployed collection, then see the
source repository
for DCGM Exporter
for installation information.
These instructions are provided as an example and are expected to work in most Kubernetes
environments. If you are having trouble installing an application or exporter due to restrictive
security or organizational policies, we recommend you consult open-source documentation for support.
For information about DCGM, see NVIDIA 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
1.21.4-gke.300 or later.
- You must be running Managed Service for Prometheus
with managed collection enabled. For more information, see
Get started with managed collection.
Verify that you have
sufficient quota for NVIDIA GPUs.
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-gpu -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 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
Install the DCGM exporter
We recommend that you install the DCGM exporter,
DCGM-Exporter, by using the following config:
# 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
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
- 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:
Set up port-forwarding with the following command:
kubectl -n gmp-public port-forward POD_NAME 9400
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
to emit.
Alternatively, consider using the official Helm chart
to install DCGM Exporter.
To apply configuration changes from a local file, run the following command:
kubectl apply -n NAMESPACE_NAME -f FILE_NAME
You can also
use Terraform
to manage your configurations.
Define a PodMonitoring resource
For target discovery, the Managed Service for Prometheus Operator
requires a PodMonitoring resource that corresponds to the
DCGM exporter in the same namespace.
You can use the following PodMonitoring configuration:
To apply configuration changes from a local file, run the following command:
kubectl apply -n NAMESPACE_NAME -f FILE_NAME
You can also
use Terraform
to manage your configurations.
Verify the configuration
You can use Metrics Explorer to verify that you correctly configured the
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:
-
In the navigation panel of the Google Cloud console,
select Monitoring, and then select
leaderboard Metrics explorer:
Go to Metrics explorer
- In the toolbar of the
query-builder pane, select the button whose name is either
code MQL or code PromQL.
- Verify that PromQL is selected
in the Language toggle. The language toggle is in the same toolbar that
lets you format your query.
- Enter and run the following query:
DCGM_FI_DEV_GPU_UTIL{cluster="CLUSTER_NAME", namespace="gmp-public"}
Troubleshooting
For information about troubleshooting metric ingestion problems, see
Problems with collection from exporters in
Troubleshooting ingestion-side problems.
