Workshop: Workshop on Operational Data Analytics in HPC (HPC-ODA)
Authors: Siqi Deng (Lawrence Berkeley National Laboratory (LBNL))
Abstract: HPC incident investigation requires system administrators to correlate metrics, logs, and ticket history across heterogeneous data sources. We present Perlmutter-Agent, a locally deployed, tool-grounded large language model (LLM) assistant for incident investigation on NERSC’s Perlmutter supercomputer. An orchestrator converts Slack messages into structured requests and dispatches them to specialized metrics, logs, and ticket sub-agents. Guided by site-specific agent skills, the sub-agents query operational data through MCP servers and return evidence that the orchestrator summarizes into likely causes. All agents use an open-weight Qwen3.8-27B model served locally by vLLM, keeping operational data from external model providers. In a restricted beta, 339 of 380 user requests (89.2\%) completed successfully; diagnostic accuracy and operator usefulness remain to be evaluated.
Back to Workshop on Operational Data Analytics in HPC (HPC-ODA) Archive Listing
Back to Full Workshop Archive Listing