Support CloudBank Classroom NAIRR showcase [Community Success]

[aprilmj]

Context

Eric is presenting at NAIRR on March 10, and wants to showcase a hub running GPUs there (by showcase we mean have ~40 people use it live).

This card captures the work we'll do to support that set up - it's still being defined as we work & is a little bit of working ahead on our roadmap plus a little bit of 1:1 community success/support.

What we need to do

• Make sure there's enough cloud quota for 40-50 users
• Make sure there is enough disk space for that many users
• Evaluate if we need to increase disk performance for that time
• Actually spin up 40-50 simultaneous servers to check if resource exhaustion is a problem
• Setup unlisted choice so users can bring their own images if needed cloudbank: Allow unlisted choice for GPU demo hub #7757
• Setup dynamic image building so users can build their own images with git repos
• Upgrade version of JupyterHub fancy profiles so users can use 'permalinks'
• Switch to 'workshop hub auth' (with single shared password for everyone)
• Set max runtime of a server (to prevent cost overruns)
• Investigate improving GPU diagnostics

Definition of Done

• We're clear on what infrastructure we want to set up
• We set that up 😆
• We've tested the hub(s)
• Eric's comfortable running a demo/showcase

[yuvipanda]

#7757 is there!

[yuvipanda]

We also identified GPU quota as a big potential problem, and it was - the quota limit was only 16. I have made two requests, and increased it to 32, and then 48 T4 GPUs. They may still be exhausted, so we should test.

[yuvipanda]

#7759 does what we can to work around the potential GPU exhaustion issues.

[yuvipanda]

I spent a bunch of time with @sean-morris on the phone today planning as well

[yuvipanda]

Just a quick summary of how this came to be, and timeline.

1. Feb 5, 2026: We (me, @aprilmj and others) have a delivery meeting with @ericvd-ucb and @sean-morris. We identify an upcoming NAIRR classroom workshop that @ericvd-ucb is leading as a great opportunity to deliver value for our cloudbank classroom community, and help set up a GPU hub for the demo. We decide Support providing cost effective % of GPUs for students on a hub initiatives#10 is important but not necessary for the workshop, as we expect maybe 15-20 users at most and that can be supported with dedicated GPUs. We decide we'll find a good community to trial Support providing cost effective % of GPUs for students on a hub initiatives#10 with and try it out, rougly over the summer
2. Feb 19, 2026: I ping Eric on slack to firm up plans for the GPU splitting initiative with a timeline, so we can schedule it in time. I propose a specific plan that would have us working on that in a few months
3. Feb 25, 2026: I ping eric again to look for a response
4. Feb 26, 2026: Eric meets me at the JupyterHub Roadmapping workshop, and we discuss the workshop. I learn that a few things have changed, but the thing that primarily concerns me is that we will have 40-50 users! I know we haven't done that with GPUs before, and from prior experience following along work that @agoose77 did with our LEAP community, I know that GPU instance exhaustion is a common problem. We identify a bunch of things to do to make the workshop a success (all in the body of this issue now), but I recognize that we need to test that GPUs can be accessible for that many people, and if not, explicitly plan for that.
5. Feb 27, 2026: I try to provision 50 GPUs, and GCP runs out of resources after 32. And that feels like a good day (just vague intuition). I recognize that we must offer GPU splitting if we are to support 40-50 users. While Support providing cost effective % of GPUs for students on a hub initiatives#10 was originally written for cost control, we learn there's another really important component - just resource availability. The world is such that getting 40 low powered GPUs (T4 GPUs are several generations old by now) from Google Cloud is pretty hard right now, so we have to offer splitting to support that many users. I try to do the minimal work to make this possible.