Instructor Notes

Teach this in one day next to the slide decks in https://github.com/sara-nl/introduction-to-supercomputing. Each episode below names the deck to have open. The episode slide deck in slides/ is the backup if a course deck will not build.

Timing

Page Teaching Exercises
what-is-a-supercomputer.md 30 15
software-with-eessi.md 30 20
slurm-jobs.md 35 25
filesystems-and-data.md 30 20
many-tasks.md 30 20
performance-first-steps.md 25 20

Teaching is 180 minutes and exercises are 120 minutes. That is a full day with breaks, not six hours of talking. If you are short on time, keep the prediction in each episode and skip the second challenge.

Deck map

  • Episodes 1 and 2 follow course-module-fundamentals-of-hpc/slides/fundamentals-of-hpc.pdf. Stop the software episode at module load. Do not open EasyBuild. Point at https://hpceb.rgoswami.me/ and move on.
  • Episode 3 uses the scheduler section of that same fundamentals deck. The batch files in episodes/files/ are the ones to project.
  • Episode 4 follows course-module-snellius-fs/slides/snellius-fs.pdf. Draw /home, /project, and /scratch for the course cluster beside the Snellius table.
  • Episode 5 follows course-module-qcg-pilotjob/slides/qcg-pilotjob-slides.pdf for the pilot, then course-module-qcg-pilotjob/slides/workflow-alternatives.pdf for HyperQueue, Snakemake, and FireWorks. Ten minutes on the alternatives is enough.
  • Episode 6 follows course-module-performance-analysis/performance-analysis.pdf. Stop at the roofline picture. Do not start a profiler.

Before the room arrives

Confirm which machine the room will actually use. Snellius and the course cluster do not share a host name. Put the course address on the board as MANAGER and leave snellius.surf.nl written next to it so the two ssh lines are not mixed up.

On the course cluster, check module load 2024 and then module load QCG-PilotJob from a clean shell. If the site module is missing, teach the QCG section from the script and the figure, and do not ask the room to install a private copy.

If nobody has a cluster login, the EESSI client container covers episodes 2 and 6 and the pi program. It does not cover sbatch. Say that at the start.

Common mistakes

  • SSH to Snellius from a course-cluster shell, or the other way around.
  • A long Python run on the login node. SAMPLE_COUNT=1000 is the only size that belongs there.
  • module load on top of another stack without module purge.
  • Asking for 1 core on Snellius rome and then being surprised by a 16-core charge.
  • A second sbatch while the first job is still PD.
  • Copying a directory of tiny files with rsync and then blaming the network.
  • module spider QCG-PilotJob inside the EESSI container. It is not there.
  • Timing a larger SAMPLE_COUNT and calling the program slower.

What to collect

One real job id from the room, one pi-*.out line, and the eight-core efficiency 0.625 from the table. Those three show that login, the queue, and the timing episode all happened.

Building the lesson site

From a checkout, with pixi:

BASH 

pixi run build
pixi run gate
pixi run check
pixi run slides
pixi run site

pixi run check reruns the export, the offline exercises, the EESSI output blocks, and the slide build. The EESSI check needs the client image named by EESSI_SIF, or /tmp/eessi-client.sif if that variable is unset. A paragraph note from the prose gate does not fail the gate. A forbidden word, a dash, or a voice warning does.

What a supercomputer is

Software from a shared stack

Submitting a job

Files and moving data

Many small tasks

Timing and scaling