The course has three sessions of four hours. Part I is the Introduction to HPC session: fundamentals, EESSI and first batch jobs. Use the running order below for that class. The full episode timings are reference estimates for teaching each complete lesson; do not add them together as the Part I timetable.
Times are elapsed from the start. Add the room's start time. This plan includes a 15-minute break and a final ten-minute recovery and exit-check block. Learners need a terminal, their own course account and the setup page.
Fundamentals slides pages 1-5; Why use a cluster?. Explain the goal: submit a correct job, retain its output and explain its resource request. Ask for one task from the room that needs more memory, arithmetic or independent runs.
Slides pages 22-26; Connecting. Spend ten minutes signing in and fifteen on whoami, pwd, hostname and a private working directory. Check that every pair can open a shell. Pair a learner with an access problem while the instructor handles the account privately.
Slides pages 6-17; Exploring the cluster. Explain login versus compute nodes, then ask pairs to locate their current host and read sinfo. A node allocation does not make a serial program parallel.
Slides pages 27-30 and 40-45; Shared stack and Finding modules. Use fifteen minutes to explain and fifteen to purge, load the course stack and record Python's version and module list. Keep the course-cluster tab selected. The client-container episode is a fallback reading, not an image download during class.
Slides pages 31-39 and 46-48; Scheduling jobs and Python and OpenMP. Explain the script for ten minutes and allow twenty for the Python submission, queue inspection and output. Submit python.sh from its hands-on directory. The output must contain sum: 499999500000. Save the job ID. Use the script's Python module version; it differs from the module discovery example.
Leave the return time visible. A pending job may finish during the break. Do not submit a second copy because the first is pending.
Slides pages 11-21 and 34-39; Accounting on Snellius and Resources. Give fifteen minutes of explanation and fifteen for pairs to calculate 16 cores for two hours on Snellius rome: 32 SBU. Explain the 16-core minimum there and the course cluster's smaller CPU request. SRC workspace credits and job SBU are different accounting systems.
Slides pages 46-48; Python and OpenMP. Explain processes versus threads for ten minutes; give twenty for compilation and the two-CPU job, then ten to compare the two results. Use sbatch --cpus-per-task=2 pi.sh. Expect two pi = 3.141592654 lines, with threads 1 and 2. Timing may differ; correctness must agree. The compiler step belongs in an allocation as shown below.
QCG deck pages 1-13; Job arrays. Explain an array for ten minutes and spend ten reading its index and concurrency limit. Submit only the small classroom example below. Do not run the 100-input comparison or install workflow tools in Part I.
Fundamentals slides pages 18-21; Transferring files. Ask each pair to choose a durable location for the job script and output. Distinguish home/project, shared scratch and node-local temporary files. The detailed filesystem and transfer lessons belong to Part II.
Each pair shows a job ID, the sum, the two pi lines, the loaded environment and the output location. Ask what it requested and where it ran. This block also absorbs short queue and access delays. Keep the break and the first successful job; shorten the array discussion if the room needs more help.
Keep the teacher workbook on a separate screen. The matching slide pack contains the concise controller lesson and the audited extended decks. Page references above refer to the PDF pages in that pack. The workbook provides extended explanations, worked answers, misconceptions and recovery guidance. Its storage, workflow and performance chapters are references for Parts II and III, not additional Part I requirements.
The learner website is a Carpentries Workbench lesson, not a remote desktop or a computing allocation. Open the course portal from the private welcome note in a separate tab. Keep the roster, MFA setup and individual enrollment URLs off the projector.
The website's exercise files are in a separate checkout from ~/introduction-to-supercomputing. Clone once; inspect an existing directory instead of cloning over it.
git clone https://github.com/HaoZeke/supercomputing_eessi_workbench.git "$HOME/supercomputing_eessi_workbench"
cd "$HOME/supercomputing_eessi_workbench/lessons/fundamentals/episodes/files/hands-on"
sbatch --partition=rome --cpus-per-task=1 python.sh
squeue -u "$USER"Read slurm-JOBID.out using the job ID that sbatch returns. The first line names the allocated host and the sum line is sum: 499999500000. These files match the published fundamentals hands-on examples. Keep #SBATCH --partition=rome in the scripts and use command-line overrides for the classroom request.
Compile within a small allocation, then submit the published OpenMP script:
srun --partition=rome --ntasks=1 --cpus-per-task=2 --time=00:05:00 bash compilepi
sbatch --partition=rome --cpus-per-task=2 pi.shThe published script requests 32 CPUs, so the two-CPU override matters. Read the output only when the job has finished. It must report the same nine decimal digits for one and two threads. Do not promise that this short loop scales well.
For the array discussion, from your own working directory:
sbatch --partition=rome --ntasks=1 --cpus-per-task=1 --time=00:02:00 \
--array=0-3%2 --output='intro-array-%A_%a.out' \
--wrap='printf "index=%s host=%s\n" "$SLURM_ARRAY_TASK_ID" "$(hostname)"'Expect four separate output files, with indices 0, 1, 2 and 3. At most two array elements run together. The hostname is assigned by Slurm and is not a fixed answer. This is an index demonstration; it does not measure parallel speedup.
sinfo; do not quote the diagram as the live inventory.GCC/14.2.0 are available. Do not substitute a different version in a teaching script without a rehearsal.Use this branch only when pairs have the Python result, the correct pi lines and a resource explanation by 03:20. It replaces the 03:20-03:50 array/storage block; it does not extend the class. Stop at 03:50 for the exit check. Choose individual previews when the group needs more practice with the core.
The five previews take 25 minutes and leave five minutes of transition margin. No preview requires a database, a new software installation, profiler setup or a full transfer experiment. The links retain the detailed exercises for independent use.
Part II covers file systems, moving data and many tasks through QCG-PilotJob. Part III covers hybrid MPI/OpenMP and performance. Use the individual lesson instructor notes for their complete episodes. Cut several-streams, other-tools and energy before sacrificing the core exercise for those sessions.
The hybrid lesson is CC-BY 4.0, adapted from Introduccion a la Supercomputacion by Jose Luis Bosque, Borja Perez, Ramon Beivide and Carmen Martinez, Universidad de Cantabria, by the EuroCC NL team at SURF. Credits are on the lesson index.