Reference

Short meanings for the words this lesson uses. The command you type is in the episode. This page is the reminder.

Machines

---

• Login node. The shared SSH host. Edit and submit here. Do not run the long job here. On Snellius the name starts with int.
• Compute node. The host that runs the job. Slurm starts your script there.
• Snellius. The national system. ssh YOUR_LOGIN@snellius.surf.nl after the SURF account exists. https://www.surf.nl/en/services/compute/snellius-the-national-supercomputer
• Course cluster. The Magic Castle workspace for this course. The address is in the welcome note. https://github.com/ComputeCanada/magic_castle
• Scheduler. The program that waits for resources and then starts the job. Here it is Slurm.

Software

---

• CVMFS. The file system that shows /cvmfs/software.eessi.io and fetches a file when you use it.
• EESSI. The shared software stack. Start 2025.06 with source /cvmfs/software.eessi.io/versions/2025.06/init/lmod/bash. https://www.eessi.io/isc26-tutorial/usage/
• Module. A name that puts one build on your PATH. module spider searches. module load selects. module purge clears.
• CPU target. The line the init script prints, such as a Sapphire Rapids path. Do not copy someone else's path into your script.
• EasyBuild. The tool that builds a module from a recipe. It is not taught here. The book is https://hpceb.rgoswami.me/ and the sources are https://github.com/HaoZeke/what-easybuild-is-doing.

Slurm

---

• sbatch submits a script. https://slurm.schedmd.com/sbatch.html
• squeue shows waiting and running jobs. https://slurm.schedmd.com/squeue.html
• scancel removes a job id.
• #SBATCH is a request in the script header. The class table is episodes/files/slurm-directives.tsv.
• Job array. One script, many tasks, one index each. https://slurm.schedmd.com/job_array.html
• SLURM_ARRAY_TASK_ID is that index inside the task.
• %j in an output name becomes the job id. %A is the array id and %a is the task id.

Files

---

• /home on Snellius is small, backed up, and capped at 200 GiB.
• /scratch-shared keeps an unused file for 14 days.
• /scratch-local keeps a file for 6 days.
• /scratch-node and $TMPDIR disappear when the job ends.
• /projects is shared project storage. /archive is the long-term store.
• On the course cluster, /project is the large shared disk and /scratch is local to one machine.
• tar packs a tree into one file. rsync -avP copies it. sha256sum -c checks it.
• dcp, dsync, and dwalk are the parallel tools in mpiFileUtils. dcp is documented at https://mpifileutils.readthedocs.io/en/latest/dcp.1.html. Pack tiny files first.

Many tasks

---

• Pilot job. One Slurm allocation, and a manager inside it that starts the short tasks.
• QCG-PilotJob. The pilot manager on the course cluster. Load 2024, then QCG-PilotJob. It is not an EESSI module. https://qcg-pilotjob.readthedocs.io/en/latest/
• HyperQueue. A server that fills Slurm workers with tasks. https://github.com/It4innovations/hyperqueue
• Snakemake. Rules that rebuild files. https://snakemake.readthedocs.io/en/stable/
• FireWorks. Tasks stored in a database. https://materialsproject.github.io/fireworks/

Timing

---

• time -p prints real, user, and sys. real is the time you waited.
• Strong scaling. The problem stays the same size and you add cores.
• Weak scaling. The problem grows so each core keeps about the same work.
• Speedup. T(1) / T(N).
• Efficiency. T(1) / (N * T(N)). The worked eight-core value is 0.625.
• Bandwidth roof. The limit set by how fast bytes move.
• Peak roof. The limit set by how fast the cores do arithmetic.
• Arithmetic intensity. Arithmetic per byte moved. It decides which roof you hit.