BYOS: Knowledge-driven Large Language Models Bring Your Own Operating System More Excellent

TL;DR

This paper tackles the challenge of OS kernel tuning by introducing BYOS, a knowledge-driven framework that couples large language models with an OS-oriented dual-layer knowledge graph (OD-KG) to bridge abstract tuning goals and concrete kernel options. The framework comprises three innovations: structured knowledge construction and mapping via OD-KG, knowledge-driven configuration generation, and continuous knowledge maintenance to track kernel evolution. Empirical results show BYOS achieves substantial performance gains and improved stability across multiple Linux distributions, workloads, and evolving kernel versions, outperforming baselines and reducing LLM hallucinations. The work demonstrates that grounding LLM-driven optimization in structured, maintainable knowledge can yield scalable, real-world improvements in system software optimization.

Abstract

Operating System (OS) kernel tuning involves systematically adjusting kernel configurations to optimize system performance. Despite recent advancements in large language models (LLMs), kernel tuning remains a critical challenge due to: (1) the semantic gap between abstract tuning objective and concrete config options, (2) insufficient environmental interaction induces LLM hallucinations, and (3) the rapid evolution of kernel versions. To address these challenges, we propose BYOS, a LLM-powered framework that automates kernel tuning through three key innovations: structured knowledge construction and mapping, knowledge-driven configuration generation, and continuous knowledge maintenance. Extensive experiments show that BYOS achieves 7.1%-155.4% performance improvements over default configurations across standard OS benchmarks and real-world applications, demonstrating structured knowledge representation can overcome key limitations of pure LLM solutions for system optimization. Our code is available at https://github.com/LHY-24/BYOS.

Figures (15)

• Figure 1: An example of kernel tuning task.
• Figure 2: Challenges on OS kernel tuning using LLM. The first challenge is that it is difficult for LLM to map fuzzy objectives to precise configuration items. The second challenge is that LLM can hallucinate, resulting in giving non-existent configurations. The third challenge is the rapid iteration of the kernel configuration, which typically changes every few months.
• Figure 3: An overview of our BYOS framework. First, we construct the OD-KG by a pre-built corpus(Knowledge Construction and Mapping \ref{['Method_1']}). We then generate kernel configuration with the help of OD-KG(Knowledge-driven Configuration Generation \ref{['Method_2']}). Since kernel config might be changed frequently, we need to add new configs, delete deprecated configs and update changed configs to OD-KG(Continuous Knowledge Maintenance \ref{['Method_3']})
• Figure 4: Result of LEBench: the heatmap shows the relative latency changes in kernel operations for each method, compared to the default configuration. Green indicates reduced latency (better), while red denotes increased latency (worse).
• Figure 5: Performance evaluation across kernel versions (a) and real-world applications (b). The x-axis shows the performance normalized to the default configuration. Nginx and Apache are measured in Requests per Second (RPS), Redis in Operations per Second (OPS), and PostgreSQL in Queries per Second (QPS).