KnowOS: Knowledge-driven Large Language Models for Operating System Kernel Tuning

ICLR 2026 Conference Submission12795 Authors

Keywords: Large Language Model, Operating System, Knowledge Graph, Linux

Abstract: Operating System (OS) kernel tuning involves systematically optimizing kernel configurations to enhance system performance. Despite recent advancements in large language models (LLMs), kernel tuning remains a significant challenge due to: (1) the semantic gap between abstract tuning objectives and the specific config options, (2) the limited environmental interaction leading to LLM hallucinations, and (3) the rapid evolution of kernel versions. To address these challenges, we introduce KnowOS, a framework powered by knowledge-driven LLMs for automating kernel tuning. KnowOS leverages three key innovations: structured knowledge construction and mapping, knowledge-driven configuration generation, and continuous knowledge maintenance. Extensive experiments demonstrate that KnowOS achieves performance improvements ranging from 7.1\% to 155.4\% over default configurations across standard OS benchmarks and real-world applications. These results highlight the potential of structured knowledge representations in overcoming the limitations of pure LLM-based approaches for system optimization. Our code is available at https://anonymous.4open.science/r/KnowOS-B274.

Primary Area: infrastructure, software libraries, hardware, systems, etc.

Submission Number: 12795