This repository contains Golang bindings and DCGM-Exporter for gathering GPU telemetry in Kubernetes.
Golang bindings are provided for the following two libraries:
- NVIDIA Management Library (NVML) is a C-based API for monitoring and managing NVIDIA GPU devices.
- NVIDIA Data Center GPU Manager (DCGM) is a set of tools for managing and monitoring NVIDIA GPUs in cluster environments. It's a low overhead tool suite that performs a variety of functions on each host system including active health monitoring, diagnostics, system validation, policies, power and clock management, group configuration and accounting.
You will also find samples for both of these bindings in this repository.
This is based on NVIDIA/gpu-monitoring-tools. The original monitoring tools can only monitor specific NVIDIA GPU card, with the Kubernetes pods name listed which built by NVIDIA/k8s-device-plugin.
This repository allows to monitor the utilization such as sm, dec, env and memory of Kubernetes pods and containers, which may be built by any third party of gpu device plugin. For example:
DCGM_FI_K8S_MEM_COPY_UTIL{gpu="0",UUID="GPU-de4b1bb0-3ec3-67ed-b3e2-c32d8546e818",device="nvidia0",container="benchmark-0",namespace="default",pod="benchmark-0"} 30
DCGM_FI_K8S_ENC_UTIL{gpu="0",UUID="GPU-de4b1bb0-3ec3-67ed-b3e2-c32d8546e818",device="nvidia0",container="benchmark-0",namespace="default",pod="benchmark-0"} 0
DCGM_FI_K8S_DEC_UTIL{gpu="0",UUID="GPU-de4b1bb0-3ec3-67ed-b3e2-c32d8546e818",device="nvidia0",container="benchmark-0",namespace="default",pod="benchmark-0"} 0
DCGM_FI_K8S_GPU_UTIL{gpu="0",UUID="GPU-de4b1bb0-3ec3-67ed-b3e2-c32d8546e818",device="nvidia0",container="benchmark-0",namespace="default",pod="benchmark-0"} 47
DCGM_FI_K8S_GPU_UTIL{gpu="0",UUID="GPU-de4b1bb0-3ec3-67ed-b3e2-c32d8546e818",device="nvidia0",container="benchmark-1",namespace="default",pod="benchmark-1"} 17
DCGM_FI_K8S_MEM_COPY_UTIL{gpu="0",UUID="GPU-de4b1bb0-3ec3-67ed-b3e2-c32d8546e818",device="nvidia0",container="benchmark-1",namespace="default",pod="benchmark-1"} 10
DCGM_FI_K8S_ENC_UTIL{gpu="0",UUID="GPU-de4b1bb0-3ec3-67ed-b3e2-c32d8546e818",device="nvidia0",container="benchmark-1",namespace="default",pod="benchmark-1"} 0
DCGM_FI_K8S_DEC_UTIL{gpu="0",UUID="GPU-de4b1bb0-3ec3-67ed-b3e2-c32d8546e818",device="nvidia0",container="benchmark-1",namespace="default",pod="benchmark-1"} 0
Above outputs the utilization of all Kubernetes pods and containers run by some other K8S device plugin for GPU sharing.
The repository also contains DCGM-Exporter. It exposes GPU metrics exporter for Prometheus leveraging NVIDIA DCGM.
To gather metrics on a GPU node, simply start the dcgm-exporter container:
$ docker run -d --gpus all --rm -p 9400:9400 nvidia/dcgm-exporter:2.0.13-2.1.1-ubuntu18.04
$ curl localhost:9400/metrics
# HELP DCGM_FI_DEV_SM_CLOCK SM clock frequency (in MHz).
# TYPE DCGM_FI_DEV_SM_CLOCK gauge
# HELP DCGM_FI_DEV_MEM_CLOCK Memory clock frequency (in MHz).
# TYPE DCGM_FI_DEV_MEM_CLOCK gauge
# HELP DCGM_FI_DEV_MEMORY_TEMP Memory temperature (in C).
# TYPE DCGM_FI_DEV_MEMORY_TEMP gauge
...
DCGM_FI_DEV_SM_CLOCK{gpu="0", UUID="GPU-604ac76c-d9cf-fef3-62e9-d92044ab6e52"} 139
DCGM_FI_DEV_MEM_CLOCK{gpu="0", UUID="GPU-604ac76c-d9cf-fef3-62e9-d92044ab6e52"} 405
DCGM_FI_DEV_MEMORY_TEMP{gpu="0", UUID="GPU-604ac76c-d9cf-fef3-62e9-d92044ab6e52"} 9223372036854775794
...
Note: Consider using the NVIDIA GPU Operator rather than DCGM-Exporter directly.
Ensure you have already setup your cluster with the default runtime as NVIDIA.
The recommended way to install DCGM-Exporter is to use the Helm chart:
$ helm repo add gpu-helm-charts \
https://nvidia.github.io/gpu-monitoring-tools/helm-charts
Update the repo:
$ helm repo update
And install the chart:
$ helm install \
--generate-name \
gpu-helm-charts/dcgm-exporter
Once the dcgm-exporter pod is deployed, you can use port forwarding to obtain metrics quickly:
$ kubectl create -f https://raw.githubusercontent.com/NVIDIA/gpu-monitoring-tools/master/dcgm-exporter.yaml
# Let's get the output of a random pod:
$ NAME=$(kubectl get pods -l "app.kubernetes.io/name=dcgm-exporter" \
-o "jsonpath={ .items[0].metadata.name}")
$ kubectl port-forward $NAME 8080:9400 &
$ curl -sL http://127.0.01:8080/metrics
# HELP DCGM_FI_DEV_SM_CLOCK SM clock frequency (in MHz).
# TYPE DCGM_FI_DEV_SM_CLOCK gauge
# HELP DCGM_FI_DEV_MEM_CLOCK Memory clock frequency (in MHz).
# TYPE DCGM_FI_DEV_MEM_CLOCK gauge
# HELP DCGM_FI_DEV_MEMORY_TEMP Memory temperature (in C).
# TYPE DCGM_FI_DEV_MEMORY_TEMP gauge
...
DCGM_FI_DEV_SM_CLOCK{gpu="0", UUID="GPU-604ac76c-d9cf-fef3-62e9-d92044ab6e52",container="",namespace="",pod=""} 139
DCGM_FI_DEV_MEM_CLOCK{gpu="0", UUID="GPU-604ac76c-d9cf-fef3-62e9-d92044ab6e52",container="",namespace="",pod=""} 405
DCGM_FI_DEV_MEMORY_TEMP{gpu="0", UUID="GPU-604ac76c-d9cf-fef3-62e9-d92044ab6e52",container="",namespace="",pod=""} 9223372036854775794
...
To integrate DCGM-Exporter with Prometheus and Grafana, see the full instructions in the user guide.
dcgm-exporter is deployed as part of the GPU Operator. To get started with integrating with Prometheus, check the Operator user guide.
dcgm-exporter is actually fairly straightforward to build and use.
Ensure you have the following:
$ git clone https://github.com/NVIDIA/gpu-monitoring-tools.git
$ cd gpu-monitoring-tools
$ make binary
$ sudo make install
...
$ dcgm-exporter &
$ curl localhost:9400/metrics
# HELP DCGM_FI_DEV_SM_CLOCK SM clock frequency (in MHz).
# TYPE DCGM_FI_DEV_SM_CLOCK gauge
# HELP DCGM_FI_DEV_MEM_CLOCK Memory clock frequency (in MHz).
# TYPE DCGM_FI_DEV_MEM_CLOCK gauge
# HELP DCGM_FI_DEV_MEMORY_TEMP Memory temperature (in C).
# TYPE DCGM_FI_DEV_MEMORY_TEMP gauge
...
DCGM_FI_DEV_SM_CLOCK{gpu="0", UUID="GPU-604ac76c-d9cf-fef3-62e9-d92044ab6e52"} 139
DCGM_FI_DEV_MEM_CLOCK{gpu="0", UUID="GPU-604ac76c-d9cf-fef3-62e9-d92044ab6e52"} 405
DCGM_FI_DEV_MEMORY_TEMP{gpu="0", UUID="GPU-604ac76c-d9cf-fef3-62e9-d92044ab6e52"} 9223372036854775794
...
With dcgm-exporter you can configure which fields are collected by specifying a custom CSV file.
You will find the default CSV file under etc/dcgm-exporter/default-counters.csv in the repository, which is copied on your system or container at
/etc/dcgm-exporter/default-counters.csv
The format of this file is pretty straightforward:
# 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).
A custom csv file can be specified using the -f option or --collectors as follows:
$ dcgm-exporter -f /tmp/custom-collectors.csv
Notes:
- Always make sure your entries have 3 commas (',')
- The complete list of counters that can be collected can be found on the DCGM API reference manual: https://docs.nvidia.com/datacenter/dcgm/latest/dcgm-api/group__dcgmFieldIdentifiers.html
You can find the official NVIDIA DCGM-Exporter dashboard here: https://grafana.com/grafana/dashboards/12239
You will also find the json file on this repo under grafana/dcgm-exporter-dashboard.json
Pull requests are accepted!
Checkout the Contributing document!
- Please let us know by filing a new issue
- You can contribute by opening a pull request