ML/Jupyter pod

Do you need to run your own pod?

We provide the JupyterLab service running in our cluster, which is great if you need to quickly run your workflow and do not want to learn any kubernetes. Simply follow the above link and use your institutional credentials to login using CILogon. Choose the hardware specs to spawn your instance. Once authenticated you can run Jupyter notebooks as usual.
Your persistent home folder initially will be limited to 5GB. If you need more, you can request it to be extended. You can also request for cephFS storage that is mounted to a shared disk space. Please use this to store all the data, code and results that you would need for long experiments.

NOTE: your jupyter container will shut down 1hr after your browser disconnects from it. If you need your job to keep running, don’t close the browser window. You could either use a desktop with a persistent Internet connection or only use this for testing your code.

NOTE: Available images are described in the scientific images section.

If you need more: Step by Step Tensorflow with Jupyter

If you need to create your own jupyterlab instance, you can run it as a Jupyter container in Nautilus by following step by step instructions below. It will be automatically destroyed in 6 hours. Use the examples for Jobs once you’re done with interactive debugging and ready to run a bigger job.

Step 1: Set context

Let’s set the default namespace to avoid typing it for every command. Use your own namespace in the following command:

$ kubectl config set-context nautilus --namespace=<YOUR_NAMESPACE>

Check the default context with command and check your own namespace is listed under the NAMESPACE column

$ kubectl config get-contexts
CURRENT   NAME       CLUSTER    AUTHINFO                                  NAMESPACE
*         nautilus   nautilus   YOUR_NAMESPACE

Back to top

Step 2: Create Tensorflow pod
  1. Create a tensorflow pod
    You can copy and paste the following text in your terminal window:

    cat << EOF | kubectl create -f -
    apiVersion: v1
    kind: Pod
      name: gpu-pod-example
      - name: gpu-container
        command: ["sleep", "infinity"]

    Or you can create a tensorflow-pod.yaml file with the content:

    apiVersion: v1
    kind: Pod
      name: gpu-pod-example
      - name: gpu-container
        command: ["sleep", "infinity"]

    and then deploy it with

    kubectl create -f tensorflow-pod.yaml
  2. Check the pod is running
    $ kubectl get pods
    NAME                                 READY     STATUS    RESTARTS   AGE
    gpu-pod-example                      1/1       Running   0          1m

    The output should give you the list of pods running in your namespace, including gpu-pod-example.

  3. Login into your tensorflow pod
    Run the kubectl command listed below. In the command, the -i flag asks for bash to be executed interactively and -t flag allocates an access to a terminal. That is to say you want to have an interactive shell login on your pod.
    $ kubectl exec -it gpu-pod-example  bash

    A successful execution should result in the output similar to:

    To run a command as administrator (user "root"), use "sudo <command>".
    See "man sudo_root" for details.

At this point you can already run some code from command line. Keep reading if you want jupyter GUI interface.

(If you stop here, don’t forget to tear down from step 4 )

Back to top

Step 3: Run Jupyter notebook
  1. Start Jupyter notebook
    In the newly running bash prompt running on the nautilus kubernetes container:
    jovyan@gpu-pod-example:~$ jupyter notebook --ip=''

    You should see the output similar to:

    [I 20:41:07.016 NotebookApp] Writing notebook server cookie secret to /home/jovyan/.local/share/jupyter/runtime/notebook_cookie_secret
    [I 20:41:07.490 NotebookApp] [jupyter_nbextensions_configurator] enabled 0.4.1
    [I 20:41:07.552 NotebookApp] JupyterLab extension loaded from /opt/conda/lib/python3.6/site-packages/jupyterlab
    [I 20:41:07.553 NotebookApp] JupyterLab application directory is /opt/conda/share/jupyter/lab
    [I 20:41:07.559 NotebookApp] Serving notebooks from local directory: /home/jovyan
    [I 20:41:07.559 NotebookApp] The Jupyter Notebook is running at:
    [I 20:41:07.559 NotebookApp] http://(gpu-pod-example or
    [I 20:41:07.559 NotebookApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation).
    [C 20:41:07.567 NotebookApp]
        To access the notebook, open this file in a browser:
        Or copy and paste one of these URLs:
            http://(gpu-pod-example or

    Take note of the token at the end of the output. We will use it to login:

  2. Setup port forwarding
    In a separate terminal window on your local machine run port-forward to access the pod. Here we tell kubectl to forward any request to your local host port 8888 from a remote host (container) port 8888.
    kubectl port-forward gpu-pod-example 8888:8888

    A successful command results in the output

    Forwarding from -> 8888
    Forwarding from [::1]:8888 -> 8888
  3. Connect to Jupyter
    In a new web browser window connect to an address localhost:8888, you should see the following page:

    Enter the previously saved token on the Password or Token line at the top of the page. Optionally, to set a password follow direcitons on the page. Click appropriate Log in button.

    You get a Jupyter notebook with access to tensorflow:

    Once you are finished, we’ll tear down our setup.

Back to top

Step 4: Tear down

It is important to tear down everything once we are done with the Jupyter to free up resources for others.

  1. Close your web browser window with Jupyter notebook once you are finished.
  2. Shutdown the proxy
    In the terminal window where you executed port forwarding command, press Ctrl-C. This will stop local machine connection to the remote container.
  3. Shutdown the Jupyter Server
    Press Ctrl-C twice on the terminal window where you are logged in on a container and where jupyter server is running. You should see something along:
    Shutdown this notebook server (y/[n])? ^C
    I 17:44:16.657 NotebookApp] Shutting down 2 kernels
    [I 17:44:22.181 NotebookApp] Kernel shutdown: 651a1901-da9b-4465-9bbb-1b77775ee5ca
    [I 17:44:22.191 NotebookApp] Kernel shutdown: 93b5f0d5-b2ba-44a4-af7e-fb46eb19dc48

    This will stop the running jupyter notebook, but will not free up the resources taken by our Pod.

  4. Logout form your container
    Exit the current container shell and you should be back on your machine:
    root@gpu-pod-example:/notebooks# exit
  5. Delete the pod
    Check the pod is still running
    $ kubectl get pods
    NAME              READY     STATUS    RESTARTS   AGE
    gpu-pod-example   1/1       Running   0          36m

    Remove the pod

    $ kubectl delete pod gpu-pod-example

    Check the pod is terminating

    $ kubectl get pods
    NAME              READY     STATUS        RESTARTS   AGE
    gpu-pod-example   1/1       Terminating   0          37m

    You can see the pod is currently being terminated under the STATUS column. If you retry after a minute or so you should see no resources anymore.

    $ kubectl get pods
    No resources found.

    This will free up the resources taken by our Pod.

Back to top