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

2025-03-12

AI Summary

To address the challenges of manually configuring OS kernels for diverse workloads and the lag between human-driven tuning and rapid kernel evolution, this paper proposes BYOS: an automated kernel configuration optimization framework that synergistically leverages large language models (LLMs) and a novel Operating-System-Directed dual-layer Knowledge Graph (OD-KG). OD-KG uniquely integrates kernel semantic structures with constraint logic, significantly enhancing LLMs’ domain-specific understanding and explainable reasoning capabilities. Coupled with configuration-space modeling and empirical validation, BYOS enables end-to-end high-performance configuration generation. Evaluated across multiple representative workloads, BYOS-generated configurations achieve 7.1%–155.4% performance improvements over vendor-default configurations, substantially reducing the time-to-deployment for customized OSes. This work establishes a new paradigm for LLM-driven intelligent configuration of systems software.

Technology Category

• Search and Optimization: Algorithm Configuration
• Machine Learning: Kernel Methods
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Kernel configurations play an important role in the performance of Operating System (OS). However, with the rapid iteration of OS, finding the proper configurations that meet specific requirements can be challenging, which can be primarily attributed to the default kernel provided by vendors does not take the requirements of specific workloads into account, and the heavyweight tuning process cannot catch up with the rapid evolving pace of the kernel. To address these challenges, we propose BYOS, a novel framework powered by Large Language Models (LLMs) to customize kernel configurations for diverse user requirements. By integrating OS-oriented Dual-layer Knowledge Graph (OD-KG) and corresponding reasoning strategy, BYOS enhanced the LLM's understanding of the characteristics and capabilities of OS, thus enabling customized, cost-effective, and convenient generation of kernel configurations. Experiments show that the kernels configured by BYOS outperform the default vendor-configured kernels by 7.1% to 155.4%, demonstrating the effectiveness and efficiency of BYOS in customizing kernel configurations. Our code is available at https://github.com/LHY-24/BYOS.

Problem

• Customizing kernel configurations for specific user requirements.
• Addressing challenges in OS performance due to rapid kernel evolution.
• Enhancing LLM understanding of OS characteristics for better configuration.

Innovation

• LLMs customize kernel configurations effectively
• OD-KG enhances OS understanding for LLMs
• BYOS outperforms default kernels significantly