Workshops Archive

Tracking HPC Software Stack Drift and Performance with Benchpark


Workshop: Fourth International Workshop on HPC Testing and Evaluation of Systems, Tools, and Software (HPCTESTS 2026)

Authors: Matthew Chung (UC Riverside), Joshua Vining (Texas A&M University), Michael McKinsey (Lawrence Livermore National Laboratory), Stephanie Brink (Lawrence Livermore National Laboratory (LLNL)), and Olga Pearce (Lawrence Livermore National Laboratory)

Abstract: As HPC systems grow in architectural complexity, system software stacks undergo continuous, autonomous development separate from HPC applications. Existing benchmarking frameworks do not fully automate evaluation of the changing software stacks, and a binary pass/fail result provides insufficient information to root cause test failures or performance changes. We present a continuous integration (CI) methodology that tracks changes in the software stack, tests functionality and performance, and identifies software and configuration changes that may explain failures or performance regressions. We demonstrate our methodology on HPC systems in two different compute centers, and attribute performance regressions to updates in the ROCm runtime library. By associating experiment results with software changes, our methodology provides a basis for diagnosing and reviewing evolving HPC environments.


Back to Fourth International Workshop on HPC Testing and Evaluation of Systems, Tools, and Software (HPCTESTS 2026) Archive Listing

Back to Full Workshop Archive Listing