I updated the Python for HPC session I taught at the SDSC Summer Institute 2026 (August 7, 2026). The material was heavily rewritten for this year - a substantial rework, not a refresh of the 2025 version.
What changed:
- Slides improved - a rebuilt deck with per-section time checks, an explicit action/stop marker on each slide, and a “check your mental model” and debrief prompt per block, so the class holds to its 170-minute clock.
- Notebook-first, timed flow - slides drive the action, notebooks carry the executable explanations: setup and environment staging -> Numba serial -> Numba threads -> Dask -> multi-node Dask -> recap.
- Environment staging reworked - the
pythonhpcconda env is built once and conda-packed to node-local scratch, cached in~/.galyleo/, then staged from that cache by Galyleo and the Dask workers, so nothing rebuilds per job or per node. - Multi-node Dask added - the same Dask expression run on two Expanse nodes (a Slurm scheduler + worker job).
- Real-world example - a PySM dust-emission kernel (
@njit(parallel=True)) shows Numba in a production codebase. - Optional AI-coding-assistant block - OpenCode and GitHub Copilot, plus a safety-boundary section, now offered as an optional slide set.
The session follows one practical path: time a slow Python loop, speed it up with Numba, choose threads vs. processes by the type of work, spread the work across chunks with Dask, then run it on two Expanse nodes.
Repository: https://github.com/sdsc/sdsc-summer-institute-2026 (session 6.1_python_for_HPC).
License: All teaching material is licensed under CC BY-NC-SA 4.0.
If you reuse or adapt any of it, please cite the DOI: 10.5281/zenodo.21909401 - and if you have any problems or feedback, open an issue.