Verifying HPC applications actually run¶
Systems: Aurora, Crux, Frontier, Odo, Perlmutter, Polaris, Sirius, Sophia, Sunspot · Outcome: Ongoing coverage sweep
What this covers¶
This isn't one scientific investigation — it's a systematic "does the app actually run and produce a physically sensible result" check across roughly 38 HPC applications and 9 DOE systems. Each check goes beyond "it compiled": Trinity runs a tiny real physics or science problem per application and confirms the output is physically sensible, not just that the job exited cleanly.
Coverage so far¶
The playbook covers 38 applications across 9 systems (181 possible application/system pairs). As of this writing, 27 pairs have a recorded pass/partial/fail outcome, with more being filled in as agents run — this is ongoing infrastructure verification work, not a finished dataset.
Highlights¶
- BerkeleyGW on Aurora: dielectric screening parameter (epsilon-infinity) of 16.96, matching expected physics for silicon.
- block2 (DMRG quantum chemistry) matched exact full configuration interaction reference energies on Aurora, Frontier, and Polaris — a difference as small as 2.25×10⁻⁹ Hartree on Polaris.
- OpenMC (Monte Carlo particle transport) on Polaris: computed a criticality eigenvalue k-eff = 1.020 ± 0.016 for a benchmark reactor problem.
- PySCF (quantum chemistry) on Aurora and Crux: total energy of -74.963 Hartree for a water molecule, matching the expected reference value.
- WarpX (plasma/particle-in-cell) and Nyx (cosmological hydrodynamics) both completed basic verification runs on Polaris.