Poster: Intent-Aware and LLM-Empowered Automatic Sampling Optimization for In-band Network Telemetry

In-band network telemetry (INT) enables real-time state measurement along the packet path. However, since INT inherently incurs transmission overhead due to its in-packet data insertion, sampling is commonly employed. To manage INT sampling in dynamic network environments without prior knowledge, large language models (LLMs) can be an effective solution due to their remarkable reasoning ability. In this work, we present Auto-INT, an automatic INT system that determines optimal sampling intervals by LLM agents. Auto-INT leverages in-context learning (ICL) of LLMs to infer the sampling interval that satisfies a target accuracy based on the action-reward history. Furthermore, it maintains a long-term memory of evaluated intervals, enabling faster and more accurate decision-making. Finally, Auto-INT interprets high-level operator intents to select an appropriate target accuracy from the trade-off between accuracy and overhead. Through our prototype, we demonstrate that Auto-INT converges to a suitable interval within only a few steps.