Demonstrate remote parallel rendering from HPC

[johnkit]

Background

In present-day HPC systems, remote visualization uses ParaView server running on the HPC system and connected by reverse SSH tunnel to a ParaView client on the user's desktop:

Current remote HPC visualization (no Jupyter support)

To provide remote visualization in Jupyter, we envision 2 main conceptual changes

1. Replace ParaView server with ParaViewWeb server (or whatever it's called now) to provide an http/websocket interface.
2. Replace the HPC login node with the JupyterHub server connected to the computing nodes. (Modern HPC system are now including JupyterHub in their infrastructure because of user demand.)

So the PVWeb server will connect via http/websocket to JupyterHub which will relay the connection to the end user's Jupyter notebook. Although the specifics of how this works is not well defined (maybe using jupyter http proxy?) the tech staff at NERSC has told us that this doable.

Envisioned remote HPC visualization for Jupyter

[jourdain]

Hum, I'm not sure about what you are suggesting.

I was thinking we would run mpi pvserver on the HPC and have your Jupyter connect to it using regular paraview with a simple.Connect call (can be a reverse connection if need be). Then leverage trame to do the viz part within Jupyter.
That's it.

[aashish24]

@jourdain I think the main work here is going to be integration with JupyterHub. As that would need to spin a new rendering server and Jupyter client together in order to enable visualization support for each user. @johnkit would you agree?

[jourdain]

Hum, I'm not sure what you mean. Maybe we can sync up on that.