TuxBot: Semantic-Aware Online OS Tuning with LLMs

Modern operating systems expose many runtime knobs for scheduling, power, memory, and I/O, but most production services still run near defaults because online tuning is hard. A host-side tuner must adjust knobs while the service keeps running, often without direct access to the application metric it wants to optimize. Existing Bayesian and RL tuners search over numeric configurations and scalar rewards, but a live OS executes those values as coupled policies. As a result, a semantically bad step can push the system into a long-lived degraded region, proxy rewards can mislead tuning when application metrics are unavailable, and safe exploration becomes harder as the control surface grows. Our key insight is that LLMs can address these failures by reasoning over the meaning of knobs and telemetry. They can reject semantically inconsistent configurations and infer likely application progress from joint machine-level signals. Recent LLM-based tuning systems use this reasoning mostly offline, outside the control loop, because online use must also manage model latency and cost, preserve workload-specific experience, and keep actuation safe. We present TuxBot, a bounded LLM-based controller for steady-state online OS tuning. TuxBot builds a structured decision context from telemetry, knob schemas, recent trajectory, and prior summaries; combines low-latency and deeper reasoning models; and applies changes only through typed, validated actuation. This design lets TuxBot exploit LLM strengths—reasoning about knob semantics, subsystem interactions, and indirect performance signals—without putting a slow, unconstrained model in charge of the host. Across live workloads including Wikipedia, Xapian, TPC-C, Silo, and Sysbench, TuxBot improves performance by an average of 145.1% over state-of-the-art baselines, including MLOS, while incurring only a few cents in LLM API costs. More strikingly, TuxBot still outperforms them by 113.3% when it is limited to machine-level telemetry and the baselines are given direct access to application-level metrics.