Technical Papers Archive

How Much Can Large Language Models Optimize Job Management in HPC Systems?


Authors: Yiqin Dai, Huize Zhang, Yifu Gao, Yong Dong, Juan Chen, Ruibo Wang, Xinghua Cheng, Yuhan Cai, and Zekai Li (National University of Defense Technology (NUDT)); Chun Fan (Peking University); and Kai Lu (National University of Defense Technology (NUDT))

Abstract: Large language models are increasingly capable, yet their integration into HPC job management remains limited. This paper identifies three optimization opportunities across the job lifecycle: resource demand prediction at submission, adaptive backfill parameter reconfiguration during scheduling, and anomaly diagnosis during execution. We build a unified framework combining domain-specific supervised fine-tuning, retrieval-augmented generation, and agentic workflows, and further incorporate rigorous hallucination mitigation mechanisms to ensure safe and reliable deployment. Extensive experiments on multiple open-source and proprietary LLMs using several historical datasets demonstrate that our domain-specific models can consistently improve job resource demand prediction accuracy, enhance scheduling efficiency via online reconfiguration of backfill parameters, and provide interpretable root-cause triage through a dual-agent workflow. Ultimately, this work establishes a foundational step toward intelligent HPC environments.


Presentation: file

Back to Technical Papers Archive Listing