Title: TuneAgent: Reinforcement Learning-Driven Agentic Kernel Tuning for Operating Systems

Abstract:

Optimizing operating system (OS) performance through Linux kernel tuning is a critical yet difficult task, primarily due to the intricate nature of the kernel space, the scarcity of performance feedback, and the high sensitivity of workloads. This paper introduces TuneAgent, a novel agentic framework for Linux kernel tuning that leverages rule-based reinforcement learning (RL). In this approach, the kernel environment is modeled as a constrained RL setting, allowing large language models (LLMs) to independently navigate the kernel space while strictly adhering to rules that guarantee the validity and precision of configuration changes.

To overcome the challenge of sparse performance signals, we have developed structured reward functions. These rewards simultaneously encourage standardized reasoning, ensure configuration accuracy, and enhance performance sensitivity. Additionally, we introduce a two-phase training methodology. The first phase focuses on establishing format and semantic correctness, after which the model shifts to performance-oriented exploration. This strategy significantly accelerates convergence and minimizes computational overhead.

Our experimental evaluations indicate that TuneAgent consistently surpasses current baseline methods, delivering an overall performance gain of up to 5.6% relative to existing approaches, all while preserving a high degree of configuration validity. We also validate the framework’s robustness across a variety of real-world applications, underscoring its practical utility and adaptability in diverse deployment scenarios.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC