NVIDIA GPU feature discovery

Migrated from https://gitlab.com/nvidia/kubernetes/gpu-feature-discovery

Table of Contents

Overview
Beta Version
Prerequisites
Quick Start
The GFD Command line interface
Generated Labels
- MIG 'single' strategy
- MIG 'mixed' strategy
Deployment via helm
- Deploying via helm install with a direct URL to the helm package
Building and running locally on your native machine

Overview

NVIDIA GPU Feature Discovery for Kubernetes is a software component that allows you to automatically generate labels for the set of GPUs available on a node. It leverages the Node Feature Discovery to perform this labeling.

Beta Version

This tool should be considered beta until it reaches v1.0.0. As such, we may break the API before reaching v1.0.0, but we will setup a deprecation policy to ease the transition.

Prerequisites

The list of prerequisites for running the NVIDIA GPU Feature Discovery is described below:

nvidia-docker version > 2.0 (see how to install and its prerequisites)
docker configured with nvidia as the default runtime.
Kubernetes version >= 1.10
NVIDIA device plugin for Kubernetes (see how to setup)
NFD deployed on each node you want to label with the local source configured
- When deploying GPU feature discovery with helm (as described below) we provide a way to automatically deploy NFD for you
- To deploy NFD yourself, please see https://github.com/kubernetes-sigs/node-feature-discovery

Quick Start

The following assumes you have at least one node in your cluster with GPUs and the standard NVIDIA drivers have already been installed on it.

Node Feature Discovery (NFD)

The first step is to make sure that Node Feature Discovery is running on every node you want to label. NVIDIA GPU Feature Discovery use the local source so be sure to mount volumes. See https://github.com/kubernetes-sigs/node-feature-discovery for more details.

You also need to configure the Node Feature Discovery to only expose vendor IDs in the PCI source. To do so, please refer to the Node Feature Discovery documentation.

The following command will deploy NFD with the minimum required set of parameters to run gpu-feature-discovery.

kubectl apply -f https://raw.githubusercontent.com/NVIDIA/k8s-device-plugin/v0.16.2/deployments/static/nfd.yaml

Note: This is a simple static daemonset meant to demonstrate the basic features required of node-feature-discovery in order to successfully run gpu-feature-discovery. Please see the instructions below for Deployment via helm when deploying in a production setting.

Preparing your GPU Nodes

The following steps need to be executed on all your GPU nodes. This README assumes that the NVIDIA drivers and the nvidia-container-toolkit have been pre-installed. It also assumes that you have configured the nvidia-container-runtime as the default low-level runtime to use.

Please see: https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html

Deploy NVIDIA GPU Feature Discovery (GFD)

The next step is to run NVIDIA GPU Feature Discovery on each node as a Daemonset or as a Job.

Daemonset

kubectl apply -f https://raw.githubusercontent.com/NVIDIA/k8s-device-plugin/v0.16.2/deployments/static/gpu-feature-discovery-daemonset.yaml

Note: This is a simple static daemonset meant to demonstrate the basic features required of gpu-feature-discovery. Please see the instructions below for Deployment via helm when deploying in a production setting.

Job

You must change the NODE_NAME value in the template to match the name of the node you want to label:

export NODE_NAME=<your-node-name>
curl https://raw.githubusercontent.com/NVIDIA/k8s-device-plugin/v0.16.2/deployments/static/gpu-feature-discovery-job.yaml.template \
    | sed "s/NODE_NAME/${NODE_NAME}/" > gpu-feature-discovery-job.yaml
kubectl apply -f gpu-feature-discovery-job.yaml

Note: This method should only be used for testing and not deployed in a production setting.

Verifying Everything Works

With both NFD and GFD deployed and running, you should now be able to see GPU related labels appearing on any nodes that have GPUs installed on them.

$ kubectl get nodes -o yaml
apiVersion: v1
items:
- apiVersion: v1
  kind: Node
  metadata:
    ...

    labels:
      nvidia.com/cuda.driver.major: "455"
      nvidia.com/cuda.driver.minor: "06"
      nvidia.com/cuda.driver.rev: ""
      nvidia.com/cuda.runtime.major: "11"
      nvidia.com/cuda.runtime.minor: "1"
      nvidia.com/gpu.compute.major: "8"
      nvidia.com/gpu.compute.minor: "0"
      nvidia.com/gfd.timestamp: "1594644571"
      nvidia.com/gpu.count: "1"
      nvidia.com/gpu.family: ampere
      nvidia.com/gpu.machine: NVIDIA DGX-2H
      nvidia.com/gpu.memory: "39538"
      nvidia.com/gpu.product: A100-SXM4-40GB
      ...
...

The GFD Command line interface

Available options:

gpu-feature-discovery:
Usage:
  gpu-feature-discovery [--fail-on-init-error=<bool>] [--mig-strategy=<strategy>] [--oneshot | --sleep-interval=<seconds>] [--no-timestamp] [--output-file=<file> | -o <file>]
  gpu-feature-discovery -h | --help
  gpu-feature-discovery --version

Options:
  -h --help                       Show this help message and exit
  --version                       Display version and exit
  --oneshot                       Label once and exit
  --no-timestamp                  Do not add timestamp to the labels
  --fail-on-init-error=<bool>     Fail if there is an error during initialization of any label sources [Default: true]
  --sleep-interval=<seconds>      Time to sleep between labeling [Default: 60s]
  --mig-strategy=<strategy>       Strategy to use for MIG-related labels [Default: none]
  -o <file> --output-file=<file>  Path to output file
                                  [Default: /etc/kubernetes/node-feature-discovery/features.d/gfd]

Arguments:
  <strategy>: none | single | mixed

You can also use environment variables:

Env Variable	Option	Example
GFD_FAIL_ON_INIT_ERROR	--fail-on-init-error	true
GFD_MIG_STRATEGY	--mig-strategy	none
GFD_ONESHOT	--oneshot	TRUE
GFD_NO_TIMESTAMP	--no-timestamp	TRUE
GFD_OUTPUT_FILE	--output-file	output
GFD_SLEEP_INTERVAL	--sleep-interval	10s

Environment variables override the command line options if they conflict.

Generated Labels

Below is the list of the labels generated by NVIDIA GPU Feature Discovery and their meaning. For a similar list of labels generated or used by the device plugin, see here.

Note

Label values in Kubernetes are always of type string. The table's value type describes the type within string formatting.

Label Name	Value Type	Meaning	Example
nvidia.com/cuda.driver.major	Integer	(Deprecated) Major of the version of NVIDIA driver	550
nvidia.com/cuda.driver.minor	Integer	(Deprecated) Minor of the version of NVIDIA driver	107
nvidia.com/cuda.driver.rev	Integer	(Deprecated) Revision of the version of NVIDIA driver	02
nvidia.com/cuda.driver-version.major	Integer	Major of the version of NVIDIA driver	550
nvidia.com/cuda.driver-version.minor	Integer	Minor of the version of NVIDIA driver	107
nvidia.com/cuda.driver-version.revision	Integer	Revision of the version of NVIDIA driver	02
nvidia.com/cuda.driver-version.full	Integer	Full version number of NVIDIA driver	550.107.02
nvidia.com/cuda.runtime.major	Integer	(Deprecated) Major of the version of CUDA	12
nvidia.com/cuda.runtime.minor	Integer	(Deprecated) Minor of the version of CUDA	5
nvidia.com/cuda.runtime-version.major	Integer	Major of the version of CUDA	12
nvidia.com/cuda.runtime-version.minor	Integer	Minor of the version of CUDA	5
nvidia.com/cuda.runtime-version.full	Integer	Full version number of CUDA	12.5
nvidia.com/gfd.timestamp	Integer	Timestamp of the generated labels (optional)	1724632719
nvidia.com/gpu.compute.major	Integer	Major of the compute capabilities	7
nvidia.com/gpu.compute.minor	Integer	Minor of the compute capabilities	5
nvidia.com/gpu.count	Integer	Number of GPUs	2
nvidia.com/gpu.family	String	Architecture family of the GPU	turing
nvidia.com/gpu.machine	String	Machine type. If in a public cloud provider, value may be set to the instance type.	DGX-1
nvidia.com/gpu.memory	Integer	Memory of the GPU in mebibytes (MiB)	15360
nvidia.com/gpu.product	String	Model of the GPU. May be modified by the device plugin if a sharing strategy is employed depending on the config.	Tesla-T4
nvidia.com/gpu.replicas	String	Number of GPU replicas available. Will be equal to the number of physical GPUs unless some sharing strategy is employed in which case the GPU count will be multiplied by replicas.	4
nvidia.com/gpu.mode	String	Mode of the GPU. Can be either "compute" or "display". Details of the GPU modes can be found here	compute
nvidia.com/gpu.clique	String	GPUFabric ClusterUUID + CliqueID	7b968a6d-c8aa-45e1-9e07-e1e51be99c31.1

Depending on the MIG strategy used, the following set of labels may also be available (or override the default values for some of the labels listed above):

MIG 'single' strategy

With this strategy, the single nvidia.com/gpu label is overloaded to provide information about MIG devices on the node, rather than full GPUs. This assumes all GPUs on the node have been divided into identical partitions of the same size. The example below shows info for a system with 8 full GPUs, each of which is partitioned into 7 equal sized MIG devices (56 total).

Label Name	Value Type	Meaning	Example
nvidia.com/mig.strategy	String	MIG strategy in use	single
nvidia.com/gpu.product (overridden)	String	Model of the GPU (with MIG info added)	A100-SXM4-40GB-MIG-1g.5gb
nvidia.com/gpu.count (overridden)	Integer	Number of MIG devices	56
nvidia.com/gpu.memory (overridden)	Integer	Memory of each MIG device in megabytes (MB)	5120
nvidia.com/gpu.multiprocessors	Integer	Number of Multiprocessors for MIG device	14
nvidia.com/gpu.slices.gi	Integer	Number of GPU Instance slices	1
nvidia.com/gpu.slices.ci	Integer	Number of Compute Instance slices	1
nvidia.com/gpu.engines.copy	Integer	Number of DMA engines for MIG device	1
nvidia.com/gpu.engines.decoder	Integer	Number of decoders for MIG device	1
nvidia.com/gpu.engines.encoder	Integer	Number of encoders for MIG device	1
nvidia.com/gpu.engines.jpeg	Integer	Number of JPEG engines for MIG device	0
nvidia.com/gpu.engines.ofa	Integer	Number of OfA engines for MIG device	0

MIG 'mixed' strategy

With this strategy, a separate set of labels for each MIG device type is generated. The name of each MIG device type is defines as follows:

MIG_TYPE=mig-<slice_count>g.<memory_size>.gb
e.g.  MIG_TYPE=mig-3g.20gb

Label Name	Value Type	Meaning	Example
nvidia.com/mig.strategy	String	MIG strategy in use	mixed
nvidia.com/MIG_TYPE.count	Integer	Number of MIG devices of this type	2
nvidia.com/MIG_TYPE.memory	Integer	Memory of MIG device type in megabytes (MB)	10240
nvidia.com/MIG_TYPE.multiprocessors	Integer	Number of Multiprocessors for MIG device	14
nvidia.com/MIG_TYPE.slices.ci	Integer	Number of GPU Instance slices	1
nvidia.com/MIG_TYPE.slices.gi	Integer	Number of Compute Instance slices	1
nvidia.com/MIG_TYPE.engines.copy	Integer	Number of DMA engines for MIG device	1
nvidia.com/MIG_TYPE.engines.decoder	Integer	Number of decoders for MIG device	1
nvidia.com/MIG_TYPE.engines.encoder	Integer	Number of encoders for MIG device	1
nvidia.com/MIG_TYPE.engines.jpeg	Integer	Number of JPEG engines for MIG device	0
nvidia.com/MIG_TYPE.engines.ofa	Integer	Number of OfA engines for MIG device	0

Deployment via `helm`

The preferred method to deploy GFD is as a daemonset using helm. Instructions for installing helm can be found here.

As of v0.15.0, the device plugin's helm chart has integrated support to deploy GFD.

To deploy GFD standalone, begin by setting up the plugin's helm repository and updating it as follows:

helm repo add nvdp https://nvidia.github.io/k8s-device-plugin
helm repo update

Then verify that the latest release of the plugin is available (Note that this includes GFD ):

$ helm search repo nvdp --devel
NAME                     	  CHART VERSION  APP VERSION	DESCRIPTION
nvdp/nvidia-device-plugin	  0.16.2	 0.16.2		A Helm chart for ...

Once this repo is updated, you can begin installing packages from it to deploy GFD in standalone mode.

The most basic installation command without any options is then:

helm upgrade -i nvdp nvdp/nvidia-device-plugin \
  --version 0.16.2 \
  --namespace gpu-feature-discovery \
  --create-namespace \
  --set devicePlugin.enabled=false

Disabling auto-deployment of NFD and running with a MIG strategy of 'mixed' in the default namespace:

helm upgrade -i nvdp nvdp/nvidia-device-plugin \
  --version=0.16.2 \
  --set allowDefaultNamespace=true \
  --set nfd.enabled=false \
  --set migStrategy=mixed \
  --set devicePlugin.enabled=false

Note: You only need the to pass the --devel flag to helm search repo and the --version flag to helm upgrade -i if this is a pre-release version (e.g. <version>-rc.1). Full releases will be listed without this.

Deploying via `helm install` with a direct URL to the `helm` package

If you prefer not to install from the nvidia-device-plugin helm repo, you can run helm install directly against the tarball of the plugin's helm package. The example below installs the same chart as the method above, except that it uses a direct URL to the helm chart instead of via the helm repo.

Using the default values for the flags:

helm upgrade -i nvdp \
  --namespace gpu-feature-discovery \
  --set devicePlugin.enabled=false \
  --create-namespace \
  https://nvidia.github.io/k8s-device-plugin/stable/nvidia-device-plugin-0.16.2.tgz

Building and running locally on your native machine

Download the source code:

git clone https://github.com/NVIDIA/k8s-device-plugin

Get dependencies:

make vendor

Build it:

make build

Run it:

./gpu-feature-discovery --output=$(pwd)/gfd