Instructor Notes

Timing

Episode	Teaching	Exercises
Embarrassingly parallel	15	0
Slurm arrays	15	25
QCG-PilotJob	25	20
Averages with a dependency	10	25
Idle time	10	20
Other tools (optional)	20	25

The last episode is optional. Keep FireWorks in it when you teach it. The first episode is a discussion. The measured table stays in the last episode.

Decks

QCG slides and examples 01 to 05: branch fix/slides-qcg-source, commit cc9675bb69b01d8b339808d6715d52cbdeb5afde.

Workflow slides and alternatives/: branch feat/workflow-alternatives, commit 6cada3eb5c7ae5e9fbed91d91ce27a34b0ae4b93.

PROVENANCE.md records both pairs. The hands-on tree is episodes/files/hands-on/. Leave the directory layout alone. alternatives/hyperqueue/average.sh extracts ../../example_04/input.tar.gz.

Prep

input.tar.gz is already in example_02 and example_04. prepare_input.sh builds input_mod/. Example 4 extracts with tar -xvf input.tar.gz.

Published scripts request the rome partition and, for example 3, 128 tasks per node. Leave those lines. Overrides go on the sbatch command. The hands-on README’s small-node forms are --array=0-99%10, QCG_MAX_INPUTS=100 with --ntasks=2, and TIMES_FILE=times_short.txt with --ntasks=2. Example 3’s README example is sbatch --nodes=3 --ntasks-per-node=2 qcg_job.sh. Check sinfo before you quote a core count. The course cluster is a SURF Research Cloud workspace. The course tab requests --partition=rome. The cpu-minimum tab is the two-core override when sinfo shows small nodes.

QCG in the published scripts is module load 2024 and module load QCG-PilotJob/0.14.1-gfbf-2024a. The 2025 stack has no build of it. EESSI 2025.06 does not include it. The qcg-modules tab is the load line for the room. episodes/files/exercises/qcg-three.py prints qcg-finished and then task-0. That prediction is the two print calls plus out_0.txt. It does not depend on a cluster hostname.

qcg_job.sh in example 4 calls qcg_job_v1.py. The other versions are comments. One active line at a time.

times_short.txt is the first 32 lines of times.txt (160 lines). Those 32 numbers sum to 338. The longest is 56 and the shortest is 2. Keep those lengths. The README says shorter sleeps make QCG’s per-task start-up dominate, and the comparison turns around.

The 25 October 2026 table is one run on two-core nodes, 2 slots, FireWorks on its one worker. Say so before you read a percentage. measurements/sacct.txt is Slurm’s Elapsed, AllocCPUS, state, and exit code for the nine job ids. The idle percentages use START and END inside the output files. Those two clocks differ on the profile jobs. Job 16235860 in the example 4 README is a different example (16 CPUs, CPUTimeRAW 7232, 2.0 SBU). It is not one of the nine lines.

example_01/array_job.sh has a non-breaking space in three echo strings, before $SLURM_ARRAY_TASK_ID, $SLURM_ARRAY_TASK_MAX, and $SLURM_ARRAY_TASK_MIN. The variables still expand. Leave the file.

The v0 comment says “CVS”. The inputs are CSV.

Common mistakes

One sbatch per short command. The first episode is about that cost. It is a discussion. There is no timing table in it.
Editing #SBATCH inside the published script. Pass --array, --ntasks, or --ntasks-per-node on the command line.
Submitting example 2 as --array=0-999 on a two-core node. Use --array=0-99%10.
Reading example 1’s hostnames as if they must all differ. TMPDIR differs because each array element has its own node-local directory. Hostnames differ when the nodes differ.
Running two versions at once in qcg_job.sh.
Treating v0’s result.csv as a finished aggregate. v0 has no after. v1 is the first file with after=job_names.
Replacing times_short.txt with smaller integers.
Starting a second rlaunch worker on the file LaunchPad. States get lost and the aggregate never runs. FireWorks in this tree is one worker and --ntasks=1.
Ranking the tools from the 25 October table as if it were general. The slide says it is one run.
Equating a span such as 267 s with the Elapsed field 00:04:35 on job 18158.