getting started

from zero to a finished job in about five minutes

1. log in

Once I've made your account, SSH into the head node:

ssh [email protected]

That drops you on the login node. Don't run real work there; it's for editing files and submitting jobs. Anything heavy gets killed so the scheduler stays responsive.

2. check that the cluster sees you

sinfo          # what nodes exist and what state they're in
squeue         # what's queued or running right now
sacct          # your own job history

3. write a job script

A job script is a normal shell script with #SBATCH lines at the top telling the scheduler what you need. Save this as hello.sh: 

#!/bin/bash
#SBATCH --job-name=hello
#SBATCH --partition=main
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=1
#SBATCH --mem=512M
#SBATCH --time=00:05:00
#SBATCH --output=hello-%j.out

echo "running on $(hostname)"
sleep 30
echo "done"

4. submit it

sbatch hello.sh        # queue the job, prints a job id
squeue -u $USER        # watch it
cat hello-<jobid>.out   # read the output when it finishes
scancel <jobid>         # kill it if you need to

house rules

  • Always set --time. Jobs without a sane limit block everyone else.
  • Ask for what you'll use. Reserving all 12 cores for a single-threaded script is the one thing that'll get your account paused.
  • Nothing illegal, no crypto mining, no scraping other people's sites.
  • Your home directory isn't backed up. Keep anything you care about elsewhere too.
  • The cluster lives in my house. It goes down sometimes. That's part of the charm.