Nvidia DCGM Exporter

Docs > Integrations > Nvidia DCGM Exporter

Supported OS Linux Windows Mac OS

Integration version3.3.0

Overview

This check submits metrics exposed by the NVIDIA DCGM Exporter in Datadog Agent format. For more information on NVIDIA Data Center GPU Manager (DCGM), see NVIDIA DCGM.

Starting from Agent release 7.47.0, the DCGM check is included in the Datadog Agent package. However, you need to spin up the DCGM Exporter container to expose the GPU metrics in order for the Agent to collect this data. As the default counters are not sufficient, Datadog recommends using the following DCGM configuration to cover the same ground as the NVML integration in addition to having useful metrics. This integration is fully supported by Agent 7.59+. Some telemetry may not be available for older agent versions.

# Format
# If line starts with a '#' it is considered a comment
# DCGM FIELD                                                      ,Prometheus metric type ,help message

# Clocks
DCGM_FI_DEV_SM_CLOCK                                              ,gauge                  ,SM clock frequency (in MHz).
DCGM_FI_DEV_MEM_CLOCK                                             ,gauge                  ,Memory clock frequency (in MHz).

# Temperature
DCGM_FI_DEV_MEMORY_TEMP                                           ,gauge                  ,Memory temperature (in C).
DCGM_FI_DEV_GPU_TEMP                                              ,gauge                  ,GPU temperature (in C).

# Power
DCGM_FI_DEV_POWER_USAGE                                           ,gauge                  ,Power draw (in W).
DCGM_FI_DEV_TOTAL_ENERGY_CONSUMPTION                              ,counter                ,Total energy consumption since boot (in mJ).

# PCIE
DCGM_FI_DEV_PCIE_REPLAY_COUNTER                                   ,counter                ,Total number of PCIe retries.

# 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 %).
DCGM_FI_DEV_ENC_UTIL                                              ,gauge                  ,Encoder utilization (in %).
DCGM_FI_DEV_DEC_UTIL                                              ,gauge                  ,Decoder utilization (in %).

# Errors and violations
DCGM_FI_DEV_XID_ERRORS                                            ,gauge                  ,Value of the last XID error encountered.

# Memory usage
DCGM_FI_DEV_FB_FREE                                               ,gauge                  ,Framebuffer memory free (in MiB).
DCGM_FI_DEV_FB_USED                                               ,gauge                  ,Framebuffer memory used (in MiB).

# NVLink
DCGM_FI_DEV_NVLINK_BANDWIDTH_TOTAL                                ,counter                ,Total number of NVLink bandwidth counters for all lanes.

# VGPU License status
DCGM_FI_DEV_VGPU_LICENSE_STATUS                                   ,gauge                  ,vGPU License status

# Remapped rows
DCGM_FI_DEV_UNCORRECTABLE_REMAPPED_ROWS                           ,counter                ,Number of remapped rows for uncorrectable errors
DCGM_FI_DEV_CORRECTABLE_REMAPPED_ROWS                             ,counter                ,Number of remapped rows for correctable errors
DCGM_FI_DEV_ROW_REMAP_FAILURE                                     ,gauge                  ,Whether remapping of rows has failed

# DCP metrics
DCGM_FI_PROF_PCIE_TX_BYTES                                        ,counter                ,The number of bytes of active pcie tx data including both header and payload.
DCGM_FI_PROF_PCIE_RX_BYTES                                        ,counter                ,The number of bytes of active pcie rx data including both header and payload.
DCGM_FI_PROF_GR_ENGINE_ACTIVE                                     ,gauge                  ,Ratio of time the graphics engine is active (in %).
DCGM_FI_PROF_SM_ACTIVE                                            ,gauge                  ,The ratio of cycles an SM has at least 1 warp assigned (in %).
DCGM_FI_PROF_SM_OCCUPANCY                                         ,gauge                  ,The ratio of number of warps resident on an SM (in %).
DCGM_FI_PROF_PIPE_TENSOR_ACTIVE                                   ,gauge                  ,Ratio of cycles the tensor (HMMA) pipe is active (in %).
DCGM_FI_PROF_DRAM_ACTIVE                                          ,gauge                  ,Ratio of cycles the device memory interface is active sending or receiving data (in %).
DCGM_FI_PROF_PIPE_FP64_ACTIVE                                     ,gauge                  ,Ratio of cycles the fp64 pipes are active (in %).
DCGM_FI_PROF_PIPE_FP32_ACTIVE                                     ,gauge                  ,Ratio of cycles the fp32 pipes are active (in %).
DCGM_FI_PROF_PIPE_FP16_ACTIVE                                     ,gauge                  ,Ratio of cycles the fp16 pipes are active (in %).

# Datadog additional recommended fields
DCGM_FI_DEV_COUNT                                                 ,counter                ,Number of Devices on the node.
DCGM_FI_DEV_FAN_SPEED                                             ,gauge                  ,Fan speed for the device in percent 0-100.
DCGM_FI_DEV_SLOWDOWN_TEMP                                         ,gauge                  ,Slowdown temperature for the device.
DCGM_FI_DEV_POWER_MGMT_LIMIT                                      ,gauge                  ,Current power limit for the device.
DCGM_FI_DEV_PSTATE                                                ,gauge                  ,Performance state (P-State) 0-15. 0=highest
DCGM_FI_DEV_FB_TOTAL                                              ,gauge                  ,
DCGM_FI_DEV_FB_RESERVED                                           ,gauge                  ,
DCGM_FI_DEV_FB_USED_PERCENT                                       ,gauge                  ,
DCGM_FI_DEV_CLOCK_THROTTLE_REASONS                                ,gauge                  ,Current clock throttle reasons (bitmask of DCGM_CLOCKS_THROTTLE_REASON_*)

DCGM_FI_PROCESS_NAME                                              ,label                  ,The Process Name.
DCGM_FI_CUDA_DRIVER_VERSION                                       ,label                  ,
DCGM_FI_DEV_NAME                                                  ,label                  ,
DCGM_FI_DEV_MINOR_NUMBER                                          ,label                  ,
DCGM_FI_DRIVER_VERSION                                            ,label                  ,
DCGM_FI_DEV_BRAND                                                 ,label                  ,
DCGM_FI_DEV_SERIAL                                                ,label                  ,

Docker

To configure the exporter in a Docker environment:

Create the file $PWD/default-counters.csv which contains the default fields from NVIDIA etc/default-counters.csv as well as additional Datadog-recommended fields. To add more fields for collection, follow these instructions. For the complete list of fields, see the DCGM API reference manual.

Run the Docker container using the following command:

sudo docker run --pid=host --privileged -e DCGM_EXPORTER_INTERVAL=5000 --gpus all -d -v /proc:/proc -v $PWD/default-counters.csv:/etc/dcgm-exporter/default-counters.csv -p 9400:9400 --name dcgm-exporter nvcr.io/nvidia/k8s/dcgm-exporter:3.1.7-3.1.4-ubuntu20.04

Kubernetes (DCGM exporter Helm chart)

The DCGM exporter can quickly be installed in a Kubernetes environment using the NVIDIA DCGM Exporter Helm chart. The instructions below are derived from the template provided by NVIDIA here.

Add the NVIDIA DCGM Exporter Helm repository and ensure it is up-to-date :

helm repo add gpu-helm-charts https://nvidia.github.io/dcgm-exporter/helm-charts && helm repo update

Create a ConfigMap containing the Datadog-recommended metrics from Installation, as well as the RoleBinding and Role used by the DCGM pods to retrieve the ConfigMap using the manifest below :

apiVersion: rbac.authorization.k8s.io/v1
kind: Role
metadata:
  name: dcgm-exporter-read-datadog-cm
  namespace: default
rules:
- apiGroups: [""]
  resources: ["configmaps"]
  resourceNames: ["datadog-dcgm-exporter-configmap"]
  verbs: ["get"]
---
apiVersion: rbac.authorization.k8s.io/v1
kind: RoleBinding
metadata:
  name: dcgm-exporter-datadog
  namespace: default
subjects:
- kind: ServiceAccount
  name: dcgm-datadog-dcgm-exporter
  namespace: default
roleRef:
  kind: Role 
  name: dcgm-exporter-read-datadog-cm
  apiGroup: rbac.authorization.k8s.io
---
apiVersion: v1
kind: ConfigMap
metadata:
  name: datadog-dcgm-exporter-configmap
  namespace: default
data:
  metrics: |
      # Copy the content from the Installation section.

Create your DCGM Exporter Helm chart dcgm-values.yaml with the following content :

# Exposing more metrics than the default for additional monitoring - this requires the use of a dedicated ConfigMap for which the Kubernetes ServiceAccount used by the exporter has access thanks to step 1.
# Ref: https://github.com/NVIDIA/dcgm-exporter/blob/e55ec750def325f9f1fdbd0a6f98c932672002e4/deployment/values.yaml#L38
arguments: ["-m", "default:datadog-dcgm-exporter-configmap"]

# Datadog Autodiscovery V2 annotations
podAnnotations:
  ad.datadoghq.com/exporter.checks: |-
    {
      "dcgm": {
        "instances": [
          {
            "openmetrics_endpoint": "http://%%host%%:9400/metrics"
          }
        ]
      }
    }    
# Optional - Disabling the ServiceMonitor which requires Prometheus CRD - can be re-enabled if Prometheus CRDs are installed in your cluster
serviceMonitor:
  enabled: false

Install the DCGM Exporter Helm chart in the default namespace with the following command, while being in the directory with your dcgm-values.yaml :
```
helm install dcgm-datadog gpu-helm-charts/dcgm-exporter -n default -f dcgm-values.yaml
```

Note: You can modify the release name dcgm-datadog as well as the namespace, but you must modify accordingly the manifest from step 1.

Kubernetes (NVIDIA GPU Operator)

The DCGM exporter can be installed in a Kubernetes environment by using NVIDIA GPU Operator. The instructions below are derived from the template provided by NVIDIA here.

Add the NVIDIA GPU Operator Helm repository and ensure it is up-to-date :

helm repo add nvidia https://helm.ngc.nvidia.com/nvidia && helm repo update

Follow the Custom Metrics Config instructions with the CSV from Installation :
- Fetch the metrics file and save as dcgm-metrics.csv: curl https://raw.githubusercontent.com/NVIDIA/dcgm-exporter/main/etc/dcp-metrics-included.csv > dcgm-metrics.csv
- Edit the metrics file by replacing its content with the Datadog-provided mapping.
- Create a namespace gpu-operator if one is not already present: kubectl create namespace gpu-operator.
- Create a ConfigMap using the file edited above: kubectl create configmap metrics-config -n gpu-operator --from-file=dcgm-metrics.csv

Create your GPU Operator Helm chart dcgm-values.yaml with the following content:

# Refer to NVIDIA documentation for the driver and toolkit for your GPU-enabled nodes - example below for Amazon Linux 2 g5.xlarge
driver:
  enabled: true
toolkit:
  version: v1.13.5-centos7
# Using custom metrics configuration to collect recommended Datadog additional metrics - requires the creation of the metrics-config ConfigMap from the previous step
# Ref: https://docs.nvidia.com/datacenter/cloud-native/gpu-operator/latest/getting-started.html#custom-metrics-config
dcgmExporter:
  config:
    name: metrics-config
  env:
  - name: DCGM_EXPORTER_COLLECTORS
    value: /etc/dcgm-exporter/dcgm-metrics.csv
# Adding Datadog autodiscovery V2 annotations
daemonsets:
  annotations:
    ad.datadoghq.com/nvidia-dcgm-exporter.checks: |-
      {
        "dcgm": {
          "instances": [
            {
              "openmetrics_endpoint": "http://%%host%%:9400/metrics"
            }
          ]
        }
      }

Install the DCGM Exporter Helm chart in the default namespace with the following command, while being in the directory with your dcgm-values.yaml:
```
helm install datadog-dcgm-gpu-operator -n gpu-operator nvidia/gpu-operator -f dcgm-values.yaml
```

Configuration

Host

Metric collection

Edit the dcgm.d/conf.yaml file (located in the conf.d/ folder at the root of your Agent’s configuration directory) to start collecting your GPU Metrics. See the sample dcgm.d/conf.yaml for all available configuration options.

instances:

   ## @param openmetrics_endpoint - string - required
   ## The URL exposing metrics in the OpenMetrics format.
   ##
   ## Set this to <listenAddress>/<handlerPath> as configured in your DCGM Server
   #
   - openmetrics_endpoint: http://localhost:9400/metrics

Use the extra_metrics configuration field to add metrics that go beyond the ones Datadog supports out of the box. See the NVIDIA docs for the full list of metrics that dcgm-exporter can collect. Make sure to enable these fields in the dcgm-exporter configuration as well.

Docker

Metric collection

Set Autodiscovery Integrations Templates as Docker labels on your DCGM exporter container:

LABEL "com.datadoghq.ad.check_names"='["dcgm"]'
LABEL "com.datadoghq.ad.init_configs"='[{}]'
LABEL "com.datadoghq.ad.instances"='[{"openmetrics_endpoint": "http://%%host%%:9400/metrics"}]'

Kubernetes

Note: If you followed the instructions for the DCGM Exporter Helm chart or GPU Operator, the annotations are already applied to the pods and the instructions below can be ignored.

To configure this check for an Agent running on Kubernetes:

Metric collection

Set Autodiscovery Integrations Templates as pod annotations on your application container. Aside from this, templates can also be configured with a file, a configmap, or a key-value store.

Annotations v2 (for Datadog Agent v7.47+)

apiVersion: v1
kind: Pod
metadata:
  name: '<POD_NAME>'
  annotations:
    ad.datadoghq.com/dcgm.checks: |
      {
        "dcgm": {
          "init_config": {},
          "instances": [
            {
              "openmetrics_endpoint": "http://%%host%%:9400/metrics"
            }
          ]
        }
      }      
spec:
  containers:
    - name: dcgm

When you’re finished making configuration changes, restart the Agent.

Validation

Run the Agent’s status subcommand and look for dcgm under the Checks section.

Adjusting Monitors

The out-of-the-box monitors that come with this integration have some default values based on their alert thresholds. For example, the GPU temperature is determined based on an acceptable range for industrial devices. However, Datadog recommends that you check to make sure these values suit your particular needs.

Data Collected

Metrics

dcgm.clock_throttle_reasons (gauge)	Current clock throttle reasons (bitmask of DCGMCLOCKSTHROTTLEREASON*)
dcgm.correctable_remapped_rows.count (count)	Number of remapped rows for correctable errors. Shown as row
dcgm.dec_utilization (gauge)	Decoder utilization (in %). Shown as percent
dcgm.device.count (count)	Change in number Devices on the node. Shown as device
dcgm.device.total (gauge)	Number of Devices on the node.
dcgm.dram.active (gauge)	Ratio of cycles the device memory interface is active sending or receiving data (in %). Shown as fraction
dcgm.enc_utilization (gauge)	Encoder utilization (in %). Shown as percent
dcgm.fan_speed (gauge)	Fan speed for the device in percent 0-100. Shown as percent
dcgm.frame_buffer.free (gauge)	Free Frame Buffer in MB. Shown as megabyte
dcgm.frame_buffer.reserved (gauge)	Reserved Frame Buffer in MB. Shown as megabyte
dcgm.frame_buffer.total (gauge)	Total Frame Buffer of the GPU in MB. Shown as megabyte
dcgm.frame_buffer.used (gauge)	Used Frame Buffer in MB. Shown as megabyte
dcgm.frame_buffer.used_percent (gauge)	Percentage used of Frame Buffer: Used/(Total - Reserved). Range 0.0-1.0 Shown as fraction
dcgm.gpu_utilization (gauge)	GPU utilization (in %). Shown as percent
dcgm.gr_engine_active (gauge)	Ratio of time the graphics engine is active (in %). Shown as fraction
dcgm.mem.clock (gauge)	Memory clock frequency (in MHz). Shown as megahertz
dcgm.mem.copy_utilization (gauge)	Memory utilization (in %). Shown as percent
dcgm.mem.temperature (gauge)	Memory temperature (in C). Shown as degree celsius
dcgm.nvlink_bandwidth.count (count)	Change in number of NVLink bandwidth counters for all lanes
dcgm.nvlink_bandwidth.total (gauge)	Total number of NVLink bandwidth counters for all lanes
dcgm.pcie_replay.count (count)	Change in number of PCIe retries.
dcgm.pcie_replay.total (gauge)	Total number of PCIe retries.
dcgm.pcie_rx_throughput.count (count)	Change in PCIe Rx utilization information.
dcgm.pcie_rx_throughput.total (gauge)	PCIe Rx utilization information.
dcgm.pcie_tx_throughput.count (count)	Change PCIe Tx utilization information.
dcgm.pcie_tx_throughput.total (gauge)	PCIe Tx utilization information
dcgm.pipe.fp16_active (gauge)	Ratio of cycles the fp16 pipes are active (in %). Shown as fraction
dcgm.pipe.fp32_active (gauge)	Ratio of cycles the fp32 pipes are active (in %). Shown as fraction
dcgm.pipe.fp64_active (gauge)	Ratio of cycles the fp64 pipes are active (in %). Shown as fraction
dcgm.pipe.tensor_active (gauge)	Ratio of cycles the tensor (HMMA) pipe is active (in %). Shown as fraction
dcgm.power_management_limit (gauge)	Current power limit for the device. Shown as watt
dcgm.power_usage (gauge)	Power draw (in W). Shown as watt
dcgm.pstate (gauge)	Performance state (P-State) 0-15. 0=highest
dcgm.row_remap_failure (gauge)	Whether remapping of rows has failed.
dcgm.slowdown_temperature (gauge)	Slowdown temperature for the device. Shown as degree celsius
dcgm.sm_active (gauge)	The ratio of cycles an SM has at least 1 warp assigned (in %). Shown as fraction
dcgm.sm_clock (gauge)	SM clock frequency (in MHz). Shown as megahertz
dcgm.sm_occupancy (gauge)	The ratio of number of warps resident on an SM (in %). Shown as fraction
dcgm.temperature (gauge)	GPU temperature (in C). Shown as degree celsius
dcgm.total_energy_consumption.count (count)	Change in energy consumption (in mJ). Shown as millijoule
dcgm.total_energy_consumption.total (gauge)	Total energy consumption since boot (in mJ)
dcgm.uncorrectable_remapped_rows.count (count)	Change in number of remapped rows for uncorrectable errors. Shown as row
dcgm.uncorrectable_remapped_rows.total (gauge)	Total number of remapped rows for uncorrectable errors.
dcgm.vgpu_license_status (gauge)	vGPU License status
dcgm.xid_errors (gauge)	Value of the last XID error encountered.

Events

The DCGM integration does not include any events.

Service Checks

See service_checks.json for a list of service checks that this integration provides.

Troubleshooting

Metric Mapping

If you have added some metrics that don’t appear in the metadata.csv above but appear in your account with the format DCGM_FI_DEV_NEW_METRIC, remap these metrics in the dcgm.d/conf.yaml configuration file:

    ## @param extra_metrics - (list of string or mapping) - optional
    ## This list defines metrics to collect from the `openmetrics_endpoint`, in addition to
    ## what the check collects by default. If the check already collects a metric, then
    ## metric definitions here take precedence. Metrics may be defined in 3 ways:
    ...

The example below appends the part in NEW_METRIC to the namespace (dcgm.), giving dcgm.new_metric:

    extra_metrics:
    - DCGM_FI_DEV_NEW_METRIC: new_metric

DCGM field is enabled but not being submitted?

If a field is not being collected even after enabling it in default-counters.csv and performing a curl request to host:9400/metrics, the dcgm-exporter developers recommend looking at the log file at var/log/nv-hostengine.log.

Note: The dcgm-exporter is a thin wrapper around lower-level libraries and drivers which do the actual reporting.

Increased Resource Consumption

In some cases, the DCGM_FI_DEV_GPU_UTIL metric can cause heavier resource consumption. If you’re experiencing this issue:

Disable DCGM_FI_DEV_GPU_UTIL in default-counters.csv.
Make sure the following fields are enabled in default-counters.csv:
- DCGM_FI_PROF_DRAM_ACTIVE
- DCGM_FI_PROF_GR_ENGINE_ACTIVE
- DCGM_FI_PROF_PCIE_RX_BYTES
- DCGM_FI_PROF_PCIE_TX_BYTES
- DCGM_FI_PROF_PIPE_FP16_ACTIVE
- DCGM_FI_PROF_PIPE_FP32_ACTIVE
- DCGM_FI_PROF_PIPE_FP64_ACTIVE
- DCGM_FI_PROF_PIPE_TENSOR_ACTIVE
- DCGM_FI_PROF_SM_ACTIVE
- DCGM_FI_PROF_SM_OCCUPANCY
Restart both dcgm-exporter and the Datadog Agent.

Need help?

Contact Datadog support.

Nvidia DCGM Exporter

Overview

Setup

Installation

Docker

Kubernetes (DCGM exporter Helm chart)

Kubernetes (NVIDIA GPU Operator)

Configuration

Host

Metric collection

Docker

Metric collection

Kubernetes

Metric collection

Validation

Adjusting Monitors

Data Collected

Metrics

Events

Service Checks

Troubleshooting

Metric Mapping

DCGM field is enabled but not being submitted?

Increased Resource Consumption

Need help?

Further Reading