Architecture

The following figure shows the high-level architecture of JupyterHub on Jean Zay platform.

The main take-away of this architecture is that every JupyterLab instances are either launched by SSH'ing' into a dedicated frontal node or on a compute node via a SLURM job. Just before spawning a JupyterLab instance, the user will be presented with a form to choose between frontal node (SSH spawner) or compute node (SLURM spawner). In addition, if user chose to use a compute node, it is possible to configure the SLURM job details like number of nodes, number of GPUs, etc,. The main job step of this job will be JupyterLab instance and this it will be submitted to SLURM scheduler. Once JupyterLab instance starts running, the configurable HTTP proxy will add the JupyterLab routes and serve them back to the user.