We present EPIC, an LLM-orchestrated multi-agent framework that enables dynamic, on-demand operational data analytics for HPC systems. EPIC addresses critical limitations of conventional ODA approaches such as long turnaround times due to batch-oriented analytics workflows, rigid analytical pipelines requiring manual reconfiguration for new use cases, results constrained to predefined static dashboards that cannot adapt to
novel queries, and high dependency on data science expertise to translate stakeholder questions into executable analyses. EPIC employs hierarchical architecture where a top-level LLM orchestrator coordinates three specialized LLM-interfaced agents: (1) a document analysis agent that retrieves and synthesizes
information from unstructured HPC documentation, tables, and images; (2) a telemetry descriptive analytics agent that translates natural language to analytical queries executed against structured telemetry logs; and (3) a predictive analytics agent that uses LLM-based feature extraction to parse queries and invoke regression models for jobs’ energy, power, and temperature predictions. Through extensive evaluation on the Frontier exascale system, we demonstrate that fine-tuned smaller LLMs outperform large foundation models on specialized tasks, with our telemetry analytics agent achieving 87.4% accuracy compared to 69.3% for GPT-4o. We demonstrate cross-system generalizability on Fugaku supercomputer datasets with comparable results, such as achieving 90% accuracy by descriptive analytics agent. Our hierarchical orchestration strategy using open-weight LLMs reduces operational costs by 19× compared to exclusive use of proprietary models while achieving higher accuracy on specialized tasks.
