Scholarlink

In-band Network Telemetry (INT) enables fine-grained network state measurement by collecting flow and device status information on a per-hop basis. Leveraging source routing technology, INT can designate the transmission path of probe packets, thereby achieving comprehensive coverage of network topology measurements. While efficient path planning algorithms lie at the core of INT-based full-network telemetry, existing approaches exhibit limitations in flexibility and robustness when confronted with diverse requirements. In this paper, we propose INT-LLM, an intent-driven INT path planning algorithm based on large language models (LLMs), to enhance the flexibility and adaptability of path planning. INT-LLM employs an LLM to interpret user intent and translate it into path planning parameters, combining the efficiency of traditional algorithms with the flexibility of LLMs to rapidly generate optimal path planning schemes. This is the first work to introduce LLMs into INT path planning. We construct a dataset comprising data center network topologies and user intents specifications for experimental validation, followed by an empirical evaluation of INT-LLM. The results demonstrate that INT-LLM achieves a 12.72% higher intent-matching score compared to the best-performing baseline, underscoring its superior performance in diverse telemetry tasks.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

INT-LLM: Adaptive Path Planner for In-Band Network Telemetry via Large Language Models

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Anyi Li,
  • Weijia Shi,
  • Tingting Zhang,
  • Baokang Zhao

摘要

In-band Network Telemetry (INT) enables fine-grained network state measurement by collecting flow and device status information on a per-hop basis. Leveraging source routing technology, INT can designate the transmission path of probe packets, thereby achieving comprehensive coverage of network topology measurements. While efficient path planning algorithms lie at the core of INT-based full-network telemetry, existing approaches exhibit limitations in flexibility and robustness when confronted with diverse requirements. In this paper, we propose INT-LLM, an intent-driven INT path planning algorithm based on large language models (LLMs), to enhance the flexibility and adaptability of path planning. INT-LLM employs an LLM to interpret user intent and translate it into path planning parameters, combining the efficiency of traditional algorithms with the flexibility of LLMs to rapidly generate optimal path planning schemes. This is the first work to introduce LLMs into INT path planning. We construct a dataset comprising data center network topologies and user intents specifications for experimental validation, followed by an empirical evaluation of INT-LLM. The results demonstrate that INT-LLM achieves a 12.72% higher intent-matching score compared to the best-performing baseline, underscoring its superior performance in diverse telemetry tasks.