Nov 2026 · upcoming · SC ’26 · Chicago · conferenceUnderstanding Reliability Bottlenecks in LLM Agents for HPC I/O Telemetry Analysis
C. Egersdoerfer, O. Yildiz, K. Harms, R. Latham, R. Ross, D. Dai
This work investigates what makes LLM agents reliable at HPC I/O telemetry analysis using CIODI-Bench, 49 expert-curated tasks over job-centric I/O telemetry data collected on the Polaris system at Argonne National Lab. It shows how reliability bottlenecks change with task structure, identifying when domain knowledge and targeted review help—and where more complex agent workflows add cost without improving results.
Nov 2025 · SC ’25 · St. Louis · conferenceC. Egersdoerfer, P. Carns, S. Snyder, R. Ross, D. Dai
STELLAR is an autonomous tuning method for HPC storage systems that uses LLM agents to reduce reliance on manual expertise and extensive trial-and-error search. It reuses lessons across workloads and achieved performance comparable to expert tuning within five attempts in Lustre experiments.
Jul 2025 · SSDBM ’25 · Columbus · posterC. Egersdoerfer, A. Sareen, J. L. Bez, S. Byna, D. D. Xu, D. Dai
This poster presents an approach to making HPC I/O diagnosis more accessible through structured trace analysis and research-informed LLM guidance. It demonstrates interactive optimization advice and evaluates both proprietary and open-source models, exploring alternatives to dependence on frontier-model services.
Jun 2025 · IPDPS ’25 · Milano · conferenceC. Egersdoerfer, A. Sareen, J. L. Bez, S. Byna, D. D. Xu, D. Dai
IOAgent is an LLM-based method for diagnosing HPC I/O performance problems that combines application traces with knowledge retrieved from research papers. It reduces reliance on specialist interpretation by producing explanations with supporting references and allowing users to explore diagnoses through follow-up questions.
Nov 2024 · SC-W ’24 · Atlanta · workshopC. Egersdoerfer, Md. H. Rashid, D. Dai, B. Fang, N. Tallent
This work combines application I/O traces with storage-server measurements to detect interference and classify its severity for individual applications. It accounts for how workloads respond differently to shared-storage contention, providing a more application-specific view than aggregate system activity alone.
Jul 2024 · HotStorage ’24 · Santa Clara · workshopC. Egersdoerfer, A. Sareen, J. L. Bez, S. Byna, D. Dai
ION is an LLM-based assistant for HPC I/O performance diagnosis that combines domain knowledge with generated analysis code. It contextualizes diagnostic warnings and supports follow-up questions, helping users better identify potentially inefficient access patterns in data-intensive scientific applications.
Jun 2023 · HPDC ’23 · Orlando · posterC. Egersdoerfer, D. Zhang, D. Dai
This work explores detecting file-system anomalies with ChatGPT without task-specific training or expert-labeled examples. A running summary carries context across log windows, enabling explanations alongside predictions and outperforming comparable approaches out of the box.
May 2023 · IPDPS ’23 · St. Petersburg · conferenceD. Zhang, C. Egersdoerfer, T. Mahmud, M. Zheng, D. Dai
Drill detects storage-system anomalies by combining log-message sentiment with the source-code context surrounding logging statements. This captures warning signs that message text alone can miss and improves detection of unfamiliar log events in HDFS and Lustre experiments.
Nov 2022 · FTXS ’22 · Dallas · workshopC. Egersdoerfer, D. Zhang, D. Dai
ClusterLog groups log messages by meaning and sentiment to reduce the complexity of learning patterns from noisy storage-system logs. It improves anomaly detection on Lustre and HDFS datasets and maintains performance with limited training data.
