lab-hpc: clusters, Slurm, and safety¶
The foundation skill. It loads before any ssh or Slurm work and hands the agent the lab's cluster map, the workflow, and the hard safety rules. The other cluster skills build on it.
The two clusters¶
| arc (chimera) | ircbc | |
|---|---|---|
| Hardware | GPU cluster, H100 nodes | CPU cluster |
| Software | Recent OS; pixi runs natively | CentOS 7, so modern tools need Singularity containers |
| Access | ssh alias arc or chimera-login |
ssh alias ircbc, behind a VPN you start by hand |
| Lab perks | Free zhoulab_gpu_priority partition, plus free preemptible tiers |
Shared store of prebuilt container images |
Which cluster fits a task is your call. The agent asks, or reads the answer from your request. It never picks one in silence.
What the skill enforces¶
Preflight first. The agent checks that your ~/.ssh/config defines the lab
aliases, and reports a line per cluster: arc_hpc is arc, ircbc_hpc is ircbc.
If the one you need is missing, it refuses the request and points you at the setup
steps. It never asks you for a raw address or a password.
Work belongs on a compute node. A login node is a doorway, not a workspace: one
shared machine per cluster, and a crowd of agents each firing off "small" commands
is what brings it down. So ssh arc '<command>' is the wrong move even for a
one-liner, and it is rarely needed, because squeue, sbatch and scancel all
run from the compute node anyway.
Two reasons to touch a login node, and no others. Either you hold no job yet, and
only a login node can submit the first one. Or the compute node truly lacks
something: on ircbc that means internet access and git, and on arc it means
nothing at all. The arc compute nodes have their own internet and a full
toolchain, so once you land on one, stay.
Reuse the idle job. The lab parks a long-lived idle GPU job on arc, so a node is
usually waiting for you. The agent finds it with squeue and works there. If no
job is running anywhere, it offers to start one first.
Show before submit. You see the exact sbatch or srun before it runs.
Stop at a down VPN. If ircbc does not answer, the agent asks you to check the VPN rather than retrying in a loop.
The workflow it follows¶
Code is written on your own machine and travels through GitHub: edit, git push,
then git pull on the cluster, then run on a compute node. The agent knows the
usual code directory on each cluster, and reads your own settings from
~/.claude/lab/personal.md.
When it triggers¶
Any mention of the clusters or of remote compute: ssh to a lab server, Slurm
commands (sbatch, srun, squeue), GPU or compute nodes, moving files to a
cluster, testing code remotely.