Performance Engineering with a Coding Agent: A Real HPC Case Study

Abstract

"Vibe coding" — letting an AI agent produce software mostly on its own, calling tools, testing, and fixing its own mistakes in a loop, with little human interaction along the way — is everywhere right now. This tutorial looks at a different use of the same AI agents: hands-on practice in data analysis and exploration, and in performance engineering under real hardware constraints on an actual HPC system.

Partial differential equations (PDEs) are a common way to model reality across physics and other fields, and getting a PDE solver to run correctly and effectively on real hardware is one of the most demanding engineering tasks. This tutorial works through one such example: our own implementation of a solver — computing the recombination correction factor for a pulsed proton beam in an ionization chamber — with a physically validated simulation, an analytic correctness check, and a slow, naive implementation to start from.

Participants work on that naive implementation using opencode, an AI coding agent that can use several models hosted directly on Cyfronet's infrastructure — including GLM 5.2, whose coding ability is similar to Claude Sonnet's — at no cost to the user. All data and tasks stay within Cyfronet’s infrastructure. Three skills matter here: using the agent as a data-analysis tool (profile, plot, and explain a run instead of hand-typing matplotlib), directing it toward optimization that is both algorithmic and hardware-aware (using Athena's real core count, cache size, and NUMA layout as constraints, not guesses), and using it safely — permissions, review discipline, and version control, so it can't do real damage.

Level

Target audience

Training is intended for users who want to use the AI agents for their work on HPC machines.

Agenda

11:00–13:30 — Infrastructure, safety, and diagnosis

• Introduction to the domain problem, and a walkthrough of the naive implementation of the PDE solver. 

• Cyfronet-hosted LLMs: why running GLM 5.2 on-premises matters — free at this scale, data never leaves the cluster, coding ability similar to Claude Sonnet's

• Using an agent safely: permission modes, reviewing diffs before accepting them, git commits as checkpoints, never approving destructive commands or job submissions blindly — the rules for the rest of the day

• Setup: Athena account, opencode + GLM 5.2 token, a quick "hello agent" test — and the day's submission rule: the agent drafts job scripts, participants review and submit them by hand

• Orientation: the problem we're solving, how the starting code works, and the materials everyone begins with — the code itself, a way to measure performance, and a way to check the results are still correct

• Hands-on: treat the agent as a lab assistant, not an author — "run the sweep, plot wall time and speedup vs. thread count, explain the shape" — and write a short diagnosis based on the output

14:20–16:00 — Optimization: algorithmic and hardware-aware

• Reframing the question: instead of "make it faster," have the agent check the hardware it's running on — cores, cache, NUMA layout.

• Hands-on: algorithmic optimization — restructuring the hot loop

• Hands-on: hardware-aware optimization — deciding how, and whether, to parallelize

• Re-run the timing harness and the correctness check after every change

16:30–18:00 — Solution reveal, security debrief, and wrap-up

• Synthesis: compare each participant's findings against the optimized reference

• Discussion: where the agent's hardware reasoning held up under measurement, where it didn't, and how participants told the difference

• Security retrospective: LLMs executing risky commands, why the rules held, and what to take away for using coding agents on shared HPC accounts

• Wrap-up: agent as an analysis-and-optimization tool, not autocomplete; the case for on-prem LLMs; pointers to the reference repo

Requirements

No recommended training courses before this one.

No account required.

Working knowledge of Python and basic experience with AI agents will be benefitial.

Venue

The workshop will be conducted during PPAM 2026 conference.

Language

English

Duration

7 hours

Registration

The Registration via PPAM2026 conference site.

Organized by

Funded by

Acknowledgements

This event is partially funded by the EuroCC 3 project.

The project has received funding from the European High-Performance Computing Joint Undertaking (JU) under grant agreement No 101101903. The JU receives support from the Digital Europe Programme and Austria, Belgium, Bulgaria, Croatia, Cyprus, Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Iceland, Ireland, Italy, Latvia, Lithuania, Luxembourg, Montenegro, Netherlands, North Macedonia, Norway, Poland, Portugal, Romania, Serbia, Slovakia, Slovenia, Spain, Sweden, and Türkiye.