TuxBot: Semantic-Aware Online OS Tuning with LLMs

Abstract

Online OS tuning can improve long-running services, but existing tuners are not well suited for live hosts. They treat scheduler, power, memory, and I/O controls as black-box variables and optimize a scalar reward. This approach ignores cross-knob policy structure, breaks down when application metrics are unavailable, and can push a running service into degraded regions that persist after a bad setting is removed. We present TuxBot, a bounded LLM-based tuner for steady-state online OS tuning. TuxBot constructs a structured decision context from telemetry, knob schemas, recent trajectories, and prior sessions; pairs low-latency and deeper reasoning models; and applies only typed, validated knob updates. TuxBot is able to reason about knob semantics, subsystem interactions, and indirect performance signals without granting the model unconstrained host control. Across 13 live workloads from five benchmark suites, tuning up to 41 Linux parameters, TuxBot consistently outperforms prior tuners while avoiding catastrophic configurations. On the 11 workloads where baselines do not fail catastrophically, TuxBot improves stable-phase performance by 87.2% over defaults, compared with 50.5% for MLOS, the strongest non-LLM baseline; on the remaining two workloads, baselines enter severe-degradation regions that TuxBot avoids altogether. Remarkably, using only machine-level metrics, TuxBot still achieves a 64.0% improvement, outperforming MLOS with access to direct application-level metrics.