Workshops Archive

Analyzing User Logs as a First Diagnostic Step in Open OnDemand


Workshop: HUST-26: 13th International Workshop on HPC User Support Tools

Authors: Arsalan Akhter, Serena Caplins, Ganesan Murugan, and Ghanghoon Paik (Northeastern University)

Abstract: Open OnDemand (OOD) session failures generate developer-centric logs that may confuse researchers and burden HPC support staff with routine tickets. While large language models (LLMs) provide readable failure explanations, unconstrained models can hallucinate non-existent root causes. We present \textit{LogLight}, an asynchronous diagnostic panel running inside the OOD session card under the invoking user's Linux privileges without elevated service state. LogLight enforces a four-source trust hierarchy that prioritizes deterministic extraction from logs and scheduler accounting, incorporates an advisory site playbook, and restricts the LLM strictly to narrative synthesis. In a blind evaluation across 30 production sessions, binding deterministic findings into prompts significantly improved diagnostic accuracy and attribution over log-only inputs. Grounding enabled a lightweight Mistral-7B model to match the diagnostic performance of a GPT-OSS-20B model, offering an efficient, user-scoped architecture for failure analysis.


Back to HUST-26: 13th International Workshop on HPC User Support Tools Archive Listing

Back to Full Workshop Archive Listing