High performance Jupyter Notebooks
The aim of this project is to connect JupyterHub to a high-performance computer (HPC). By automatically offloading the computations in a Jupyter notebook to the HPC system, even complex calculations are possible. While JupyterHub is deployed on a regular server, the notebooks themselves are spawned and run on the remote HPC system using a workload manager, such as Slurm.
Motivation
The technical core of this project is the transparent integration of digital worksheets (Jupyter notebooks), in which learning content and programs can be displayed, edited and executed on the students' own laptops, with current cloud and high-performance computing (HPC) technologies. This provides the conditions for innovative, digital teaching that encourages independent and interactive development of, for example, data science applications, without imposing the complexity of using a high-performance computer system on the students. Instead, particularly computationally and data-intensive calculations are automatically offloaded to a high-performance computer, enabling even sophisticated analyses to be performed that would otherwise not be feasible on students' laptops.
Features and use cases
- Starting a jupyter notebook server on a remote HPC system in a pre-defined singularity container
- Quick config setup when using the Slurm configuration wizard
- Automatically create a singularity overlay so that user changes are persistent
- Great for managing courses with external participants
- Possibility to include files in the notebook directory using WebDAV
- Suitable for HPC users who have their own JupyterHub instance running and want to use HPC resources
- JupyterHub + High-Performance Computing
This section describes the required installations and configurations on the JupyterHub server.
The first thing you should do is install JupyterHub and BatchSpawner. For this purpose we provide an Ansible playbook which can be found in /jupyterhub-deployment/
. See the README for details. Alternatively, you can follow the official installation instructions.
If you decide to do the installations yourself, please proceed as follows:
- install JupyterHub
- install BatchSpawner
- install WrapSpawner (make sure to install it in the right environment:
/opt/jupyterhub/bin/pip3 install git+https://github.com/jupyterhub/wrapspawner
) - copy the JupyterHub configuration file
/jupyterhub-deployment/config_files/jupyterhub_config.py
to/opt/jupyterhub/etc/jupyterhub/
(you will most likely have to edit this file afterwards to make it fit your needs) - restart the JupyterHub service
A user called tunnelbot
is needed on the JupyterHub server. This user is responsible for starting an SSH tunnel between the compute node and the JupyterHub server. An SSH key pair for the above mentioned purpose must be generated. See /examples/jupyterhub_config.py
for more information.
JupyterHub extracts the execution host name of the HPC system (e.g. node01-002
). When a notebook server is started, an SSH tunnel is established using the notebook port.
In order for JupyterHub to be able to resolve the compute nodes host name, the /etc/hosts
file must be edited. An example entry might look like the following:
127.0.0.1 node01-001
127.0.0.1 node01-002
127.0.0.1 node01-003
...
127.0.0.1 node12-048
The actual node names depend on your HPC system of course.
This section describes the required installations and configurations of the HPC system to enable the interaction with the JuypterHub server.
- You need a user who is allowed to allocate resources on the HPC system
- With a SSH key pair. The public part must be deposited on the JupyterHub serer (
tunnelbot
user) - The public key part of the
tunnelbot
-user created on the JupyterHub (-> ~/.ssh/authorized_keys)
- With a SSH key pair. The public part must be deposited on the JupyterHub serer (
- Singularity (> v.3.7.0)
- mkfs/e2fsprogs with following option:
You can download and install the required files with pip.
You may want to build a small Python environment, or install the tools with --user
.
python3 -m pip install --user jh-hpc-interface
Singularity recipe examples are in the directory singularity/.
If you do not want to use singularity, then change the value of use_singularity
in jh_config.ini to false.
To build the container with the recipe files in singularity/ you have to clone this repository.
The following commands replace USER_ID in the recipes to the output of id -u
, create a new hidden file and build the singularity container with the new created file.
USER_ID=$(id -u) && sed "s/USER_ID/$USER_ID/" < singularity/Singularity > singularity/.recipefile_compute && singularity build --remote singularity/compute_jupyter.sif singularity/.recipefile_compute
USER_ID=$(id -u) && sed "s/USER_ID/$USER_ID/" < singularity/Singularity_Tensorflow > singularity/.recipefile_gpu && singularity build --remote singularity/gpu_jupyter.sif singularity/.recipefile_gpu
singularity build help section:
-r, --remote build image remotely (does not require root)
Please refer to the official docs on how to use the remote build feature: https://sylabs.io/docs/
In the directory bin/ is a script, which is deposited after the installation on the system.
With the following call you can display the location of the configuration file:
$ jh_wrapper getconfig
To learn more about the configuration file, see docs/jh_config.ini.md
With the configuration wizard you can prepare your HPC environment.
The script interactively goes through the configuration file and creates a temporary file which can be copied with a simple cp
.
To start the wizard type the following:
$ jh_slurm_wizard
You will find examples for the configuration files jh_config.ini and jupyterhub_config.py in the directory examples/.
By default the logs contain only information such as warnings or error messages. It is also possible to switch on the debug mode, which writes extended information into the log files.
Just set log_level
in the configuration file to 'DEBUG'.
Shibboleth authentication was set up for a JupyterHub server in a test environment. See ./shibboleth/
for an example configuration.
Installation instructions: https://nbgrader.readthedocs.io/en/latest/configuration/jupyterhub_config.html
To create an exchange directory for every user, just create an empty directory in $scratch_dir
and mount it into the container with $singularity_bind_extra
.
Since the containers run as user jovyan
, the value from the $JUPYTERHUB_USER
variable is automatically used.
See here for more information: https://jupyter.readthedocs.io/en/latest/community/content-community.html#what-is-a-jovyan
To make nbgrader_config.py available in the container, just append the file in $singularity_bind_extra
.
In case you are using Singularity, the host file system may be automatically mounted into the container when you start a Singularity Container.
A possible cause is the option mount hostfs
in singularity.conf
See here: https://sylabs.io/guides/3.5/admin-guide/configfiles.html#singularity-conf
See here for more information: https://jupyterhub.readthedocs.io/en/stable/reference/websecurity.html
You can increase the security by deactivating shell access for this user.
Just type:
usermod -s /bin/false tunnelbot
When problems occur with the JupyterHub, some information can be obtained from the logs when debug mode is enabled:
https://github.com/jupyterhub/jupyterhub/wiki/Debug-Jupyterhub