Enable access to GPUs#

GPUs are heavily used in machine learning workflows, and we support GPUs on all major cloud providers.

Setting up GPU nodes#


Requesting Quota Increase#

On AWS, GPUs are provisioned by using P series nodes. Before they can be accessed, you need to ask AWS for increased quota of P series nodes.

  1. Login to the AWS management console of the account the cluster i in.

  2. Make sure you are in same region the cluster is in, by checking the region selector on the top right.

  3. Open the EC2 Service Quotas page

  4. Select ‘Running On-Demand P Instances’ quota

  5. Select ‘Request Quota Increase’.

  6. Input the number of vCPUs needed. This translates to a total number of GPU nodes based on how many CPUs the nodes we want have. For example, if we are using P2 nodes with NVIDIA K80 GPUs, each p2.xlarge node gives us 1 GPU and 4 vCPUs, so a quota of 8 vCPUs will allow us to spawn 2 GPU nodes. We should fine tune this calculation for later, but for now, the recommendation is to give users a p2.xlarge each, so the number of vCPUs requested should be 4 * max number of GPU nodes.

  7. Ask for the increase, and wait. This can take several working days.

Setup GPU nodegroup on eksctl#

We use eksctl with jsonnet to provision our kubernetes clusters on AWS, and we can configure a node group there to provide us GPUs.

  1. In the notebookNodes definition in the appropriate .jsonnet file, add a node definition for the appropriate GPU node type:

         instanceType: "p2.xlarge",
         tags+: {
             "k8s.io/cluster-autoscaler/node-template/resources/nvidia.com/gpu": "1"

    p2.xlarge gives us 1 K80 GPU and ~4 CPUs. The tags definition is necessary to let the autoscaler know that this nodegroup has 1 GPU per node. If you’re using a different machine type with more GPUs, adjust this definition accordingly.

  2. Render the .jsonnet file into a .yaml file that eksctl can use

    jsonnet <your-cluster>.jsonnet > <your-cluster>.eksctl.yaml
  3. Create the nodegroup

    eksctl create nodegroup -f <your-cluster>.eksctl.yaml --install-nvidia-plugin=false

    The --install-nvidia-plugin=false is required until this bug is fixed.

    This should create the nodegroup with 0 nodes in it, and the autoscaler should recognize this!

Setting up a GPU user profile#

Finally, we need to give users the option of using the GPU via a profile. This should be placed in the hub configuration:

        - display_name: "Large + GPU: p2.xlarge"
          description: "~4CPUs, 60G RAM, 1 NVIDIA K80 GPU"
            mem_limit: null
            mem_guarantee: 55G
            image: "pangeo/ml-notebook:<tag>"
              NVIDIA_DRIVER_CAPABILITIES: compute,utility
              nvidia.com/gpu: "1"
              node.kubernetes.io/instance-type: p2.xlarge
  1. If using a daskhub, place this under the basehub key.

  2. The image used should have ML tools (pytorch, cuda, etc) installed. The recommendation is to use Pangeo’s ml-notebook for tensorflow and pytorch-notebook for pytorch. Do not use the latest or master tags - find a specific tag listed for the image you want, and use that.

  3. The NVIDIA_DRIVER_CAPABILITIES environment variable tells the GPU driver what kind of libraries and tools to inject into the container. Without setting this, GPUs can not be accessed.

  4. The node_selector makes sure that these user pods end up on the appropriate nodegroup we created earlier. Change the selector and the mem_guarantee if you are using a different kind of node

Do a deployment with this config, and then we can test to make sure this works!


  1. Login to the hub, and start a server with the GPU profile you just set up.

  2. Open a terminal, and try running nvidia-smi. This should provide you output indicating that a GPU is present.

  3. Open a notebook, and run the following python code to see if tensorflow can access the GPUs:

    import tensorflow as tf

    This should output something like:

    [PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]
  4. Remember to explicitly shut down your server after testing, as GPU instances can get expensive!

If either of those tests fail, something is wrong and off you go debugging :)