Tasks

Edit This Page

Schedule GPUs

FEATURE STATE: Kubernetes 1.10 beta
This feature is currently in a beta state, meaning:

  • The version names contain beta (e.g. v2beta3).
  • Code is well tested. Enabling the feature is considered safe. Enabled by default.
  • Support for the overall feature will not be dropped, though details may change.
  • The schema and/or semantics of objects may change in incompatible ways in a subsequent beta or stable release. When this happens, we will provide instructions for migrating to the next version. This may require deleting, editing, and re-creating API objects. The editing process may require some thought. This may require downtime for applications that rely on the feature.
  • Recommended for only non-business-critical uses because of potential for incompatible changes in subsequent releases. If you have multiple clusters that can be upgraded independently, you may be able to relax this restriction.
  • Please do try our beta features and give feedback on them! After they exit beta, it may not be practical for us to make more changes.

Kubernetes includes experimental support for managing AMD and NVIDIA GPUs (graphical processing units) across several nodes.

This page describes how users can consume GPUs across different Kubernetes versions and the current limitations.

Using device plugins

Kubernetes implements Device PluginsContainers running in Kubernetes that provide access to a vendor specific resource. to let Pods access specialized hardware features such as GPUs.

As an administrator, you have to install GPU drivers from the corresponding hardware vendor on the nodes and run the corresponding device plugin from the GPU vendor:

When the above conditions are true, Kubernetes will expose amd.com/gpu or nvidia.com/gpu as a schedulable resource.

You can consume these GPUs from your containers by requesting <vendor>.com/gpu just like you request cpu or memory. However, there are some limitations in how you specify the resource requirements when using GPUs:

  • GPUs are only supposed to be specified in the limits section, which means:
    • You can specify GPU limits without specifying requests because Kubernetes will use the limit as the request value by default.
    • You can specify GPU in both limits and requests but these two values must be equal.
    • You cannot specify GPU requests without specifying limits.
  • Containers (and Pods) do not share GPUs. There’s no overcommitting of GPUs.
  • Each container can request one or more GPUs. It is not possible to request a fraction of a GPU.

Here’s an example:

apiVersion: v1
kind: Pod
metadata:
  name: cuda-vector-add
spec:
  restartPolicy: OnFailure
  containers:
    - name: cuda-vector-add
      # https://github.com/kubernetes/kubernetes/blob/v1.7.11/test/images/nvidia-cuda/Dockerfile
      image: "k8s.gcr.io/cuda-vector-add:v0.1"
      resources:
        limits:
          nvidia.com/gpu: 1 # requesting 1 GPU

Deploying AMD GPU device plugin

The official AMD GPU device plugin has the following requirements:

  • Kubernetes nodes have to be pre-installed with AMD GPU Linux driver.

To deploy the AMD device plugin once your cluster is running and the above requirements are satisfied:

kubectl create -f https://raw.githubusercontent.com/RadeonOpenCompute/k8s-device-plugin/v1.10/k8s-ds-amdgpu-dp.yaml

You can report issues with this third-party device plugin by logging an issue in RadeonOpenCompute/k8s-device-plugin.

Deploying NVIDIA GPU device plugin

There are currently two device plugin implementations for NVIDIA GPUs:

Official NVIDIA GPU device plugin

The official NVIDIA GPU device plugin has the following requirements:

  • Kubernetes nodes have to be pre-installed with NVIDIA drivers.
  • Kubernetes nodes have to be pre-installed with nvidia-docker 2.0
  • Kubelet must use Docker as its container runtime
  • nvidia-container-runtime must be configured as the default runtime for Docker, instead of runc.
  • The version of the NVIDIA drivers must match the constraint ~= 361.93

To deploy the NVIDIA device plugin once your cluster is running and the above requirements are satisfied:

kubectl create -f https://raw.githubusercontent.com/NVIDIA/k8s-device-plugin/1.0.0-beta/nvidia-device-plugin.yml

You can report issues with this third-party device plugin by logging an issue in NVIDIA/k8s-device-plugin.

NVIDIA GPU device plugin used by GCE

The NVIDIA GPU device plugin used by GCE doesn’t require using nvidia-docker and should work with any container runtime that is compatible with the Kubernetes Container Runtime Interface (CRI). It’s tested on Container-Optimized OS and has experimental code for Ubuntu from 1.9 onwards.

You can use the following commands to install the NVIDIA drivers and device plugin:

# Install NVIDIA drivers on Container-Optimized OS:
kubectl create -f https://raw.githubusercontent.com/GoogleCloudPlatform/container-engine-accelerators/stable/daemonset.yaml

# Install NVIDIA drivers on Ubuntu (experimental):
kubectl create -f https://raw.githubusercontent.com/GoogleCloudPlatform/container-engine-accelerators/stable/nvidia-driver-installer/ubuntu/daemonset.yaml

# Install the device plugin:
kubectl create -f https://raw.githubusercontent.com/kubernetes/kubernetes/release-1.14/cluster/addons/device-plugins/nvidia-gpu/daemonset.yaml

You can report issues with using or deploying this third-party device plugin by logging an issue in GoogleCloudPlatform/container-engine-accelerators.

Google publishes its own instructions for using NVIDIA GPUs on GKE .

Clusters containing different types of GPUs

If different nodes in your cluster have different types of GPUs, then you can use Node Labels and Node Selectors to schedule pods to appropriate nodes.

For example:

# Label your nodes with the accelerator type they have.
kubectl label nodes <node-with-k80> accelerator=nvidia-tesla-k80
kubectl label nodes <node-with-p100> accelerator=nvidia-tesla-p100

Automatic node labelling

If you’re using AMD GPU devices, you can deploy Node Labeller. Node Labeller is a controllerA control loop that watches the shared state of the cluster through the apiserver and makes changes attempting to move the current state towards the desired state. that automatically labels your nodes with GPU device properties.

At the moment, that controller can add labels for:

  • Device ID (-device-id)
  • VRAM Size (-vram)
  • Number of SIMD (-simd-count)
  • Number of Compute Unit (-cu-count)
  • Firmware and Feature Versions (-firmware)
  • GPU Family, in two letters acronym (-family)

    • SI - Southern Islands
    • CI - Sea Islands
    • KV - Kaveri
    • VI - Volcanic Islands
    • CZ - Carrizo
    • AI - Arctic Islands
    • RV - Raven

      kubectl describe node cluster-node-23
      Name:               cluster-node-23
      Roles:              <none>
      Labels:             beta.amd.com/gpu.cu-count.64=1
                      beta.amd.com/gpu.device-id.6860=1
                      beta.amd.com/gpu.family.AI=1
                      beta.amd.com/gpu.simd-count.256=1
                      beta.amd.com/gpu.vram.16G=1
                      beta.kubernetes.io/arch=amd64
                      beta.kubernetes.io/os=linux
                      kubernetes.io/hostname=cluster-node-23
      Annotations:        kubeadm.alpha.kubernetes.io/cri-socket: /var/run/dockershim.sock
                      node.alpha.kubernetes.io/ttl: 0
      …
      

With the Node Labeller in use, you can specify the GPU type in the Pod spec:

apiVersion: v1
kind: Pod
metadata:
  name: cuda-vector-add
spec:
  restartPolicy: OnFailure
  containers:
    - name: cuda-vector-add
      # https://github.com/kubernetes/kubernetes/blob/v1.7.11/test/images/nvidia-cuda/Dockerfile
      image: "k8s.gcr.io/cuda-vector-add:v0.1"
      resources:
        limits:
          nvidia.com/gpu: 1
  nodeSelector:
    accelerator: nvidia-tesla-p100 # or nvidia-tesla-k80 etc.

This will ensure that the Pod will be scheduled to a node that has the GPU type you specified.

Feedback