Reproduce “Distilling Large Language Models for Network Active Queue Management” on OpenResearch

About this paper

Shiva Raj Pokhrel, Deol Satish, Jonathan Kua, Anwar Walid

Researchers at Deakin University and Columbia University developed AQM-LLM, a framework that distills Large Language Models (LLMs) for Active Queue Management (AQM) within the L4S architecture. It achieved 97.56% accuracy in predicting AQM actions, while significantly reducing median queue delay and increasing bandwidth utilization compared to traditional L4S AQM, with inference latency reduced to 0.0391 seconds.

Key insights

• LLMs can be effectively adapted for latency-sensitive network control tasks by designing specialized input encoders and modifying the output head for single-inference decisions.
• Parameter-efficient fine-tuning methods like LoRA are crucial for making large LLMs computationally feasible for real-time network management, reducing training overheads significantly.
• Reinforcement learning, combined with real-world network data collected from kernel logs, enables LLMs to learn and execute adaptive, proactive congestion control policies that surpass static rule-based systems.

Summary by alphaXiv.