Skip to content

Book a consultation

  • Loading…
↑ ↓ to move · Enter to open · Esc to close Ctrl K

RFC 0002

Open Concept

A helper tool for every researcher

Why P1 offers a small, optional tool for interactive GPU sessions, so every researcher gets the same safe way in.

Author
Benjamin Starostka Jakobsen
Created
14 November 2026
Updated
19 November 2026
Comments
Any channel, such as the compute needs survey
Address
hpc.aicentre.dk/rfc/0002/

1 Scope

When researchers start on P1 compute, one of the first things they ask is how to get an interactive session on a GPU, ideally in the editor or notebook they already use. Getting there takes several steps, and they differ from one cluster to the next. This RFC is about why P1 offers a small command line tool that takes care of them.

Using it is optional. If you're happy with SSH, the scheduler and your own setup, keep it.

2 Why

So what do researchers do today? Often they ask an AI agent, which will happily write a script that runs an SSH server inside the job, opens a tunnel, or routes traffic through a relay. Every researcher ends up with a slightly different version, and some are a lot more invasive than others. It's easy to end up with a port that anyone on the cluster can connect to, a token in a log that other users can read, or a GPU left allocated for hours after the work ended. What a shared system allows isn't always easy to tell, yet the researcher is the one carrying the risk.

And that's only for those who know to ask. HPC and IT experience varies a lot between researchers. Some have run jobs on clusters for years, some are just starting out, and many are somewhere in between. Should how much you get out of the compute depend on where you are on that scale, or on whether someone happened to point you to the right setup?

We think everyone with access to P1 compute should be able to use it productively, whatever their background. A shared tool like this could help with that:

  • One command starts a GPU session, opens your editor or notebook in it, and ends the job when you're done.
  • Everyone uses the same approach, reviewed in the open, instead of a different script each.
  • It checks your setup and tells you what to fix.
  • The commands are the same on every provider.

3 Proposal

P1 offers one small command line tool for interactive sessions and the other everyday steps on its compute. It sits on top of what the providers offer and doesn't replace any of it. If you prefer your own workflow, nothing changes for you. Where a provider has its own recommended way to reach a job, the tool uses it.

4 Requirements

IDRequirement
CL1The tool must be optional. Everything it does must also be possible without it, using the access methods the provider documents.
CL2The tool must work for someone with no custom setup. It sets up what it needs, or explains what's missing and how to fix it.
CL3The tool must only use the access the researcher already has, and must only run work through the provider's scheduler.
CL4Where a provider recommends a way to reach a job, the tool must use it. Otherwise, it must use the least invasive method that works, and tell the researcher which one it's using.
CL5The tool must not let other users reach a session or read its tokens.
CL6The tool must end the job when the session ends, and must end an idle GPU job after a set time.
CL7The tool must not collect data about the researcher or their work.
CL8The tool's source code must be public.
CL9The tool should offer the same commands on every provider.

5 Example (informative)

The aicentre CLI is a concept of such a tool. Running aicentre code --gpu starts a GPU job and opens an editor in your browser. Press Ctrl-C and the job is cancelled. Leave the browser closed for 10 minutes and the job ends on its own. The source is on GitHub.

6 Open questions

  1. Which methods should the tool be allowed to use? Should the providers or the HPC governance board sign off on each one?
  2. Should the tool grow into a web interface, for researchers who don't use a terminal?