Lecture 2 Exercises — SSH & Slurm Practice (1-week workload)

Objective: practice secure access patterns and Slurm job engineering inside the Docker cluster. Deliver a short lab notebook (Markdown) with command transcripts, screenshots, and timing tables. Keep code under codes/lab02/ unless otherwise noted.

Exercise 1 — Hardened SSH workflow¶

1. Generate two SSH key pairs: one Ed25519 for personal logins, one RSA limited to Git interactions. Store them under ~/.ssh/keys/lecture02/ with descriptive names.

2. Create a multi-host SSH config:

   • sspa-controller (localhost:2222) using the Ed25519 key.

   • sspa-worker1 reachable via ProxyJump sspa-controller.

   • A fake campus cluster entry that demonstrates ProxyJump chaining plus a local port forward example.

3. Set up local and remote SSH agents (ssh-agent + ssh-add). Document how you forwarded the agent through the controller to worker1 (ssh -A).

4. Provide a troubleshooting section: how to rotate host keys, what to do when permissions are wrong, and how to test tunnels with nc or curl.

Exercise 2 — Job arrays & dependencies¶

1. Extend the provided vec_add_array.sbatch to cover ntasks {1,2,4} and capture results into $WORK/results/vec_add_ntasks-<n>.txt.

2. Modify montecarlo_array.sbatch so that each array task writes metadata into $SCRATCH/montecarlo/meta/<jobid>_<taskid>.json alongside the .npz file.

3. Write a new cleanup job cleanup.sh that deletes scratch artifacts older than 2 days (use find -mtime +2). Chain it with --dependency=afterany:<aggregate_jobid>.

4. Keep an experiment log summarizing queue wait time, run time, and resource requests. Include at least one scontrol show job output annotated to explain key fields.

Exercise 3 — Interactive debugging workflow¶

1. Request an interactive allocation (salloc --nodes=1 --time=00:20:00) and within it:

   • Launch htop or top to observe CPU usage.

   • Run srun --pty bash on worker1 and verify $SLURM_JOB_ID remains the same.

2. Demonstrate live code editing: modify montecarlo.py to print progress every 5 seconds, run it interactively, then revert the change.

3. Capture the commands required to forward a Jupyter notebook from worker1 back to your laptop using SSH tunnels (ssh -L). Explain how you would adapt the ports when bridging through the controller.

Submission checklist¶

• Provide SLURM job IDs, command transcripts, and relevant log snippets.

• Include any helper scripts you wrote (cleanup, metadata writer, etc.).

• Note open issues or unanswered questions you encountered; these feed future lectures.