Adapting Large Language Models for Improving TCP Fairness over WiFi

TL;DR

This work tackles the challenge of optimizing TCP performance in heterogeneous networks by leveraging large language models (LLMs) to reduce manual engineering and improve generalization. The proposed TCP-LLM framework integrates an encoder, a task-specific LLM head, and Low-Rank Adaptation (LoRA) to enable efficient, real-time CCA selection, flow fairness, and starvation prevention. Key innovations include bridging modality gaps with an Integrated Encoder, direct CCA prediction with a non-autoregressive head, and parameter-efficient adaptation via $W = W_0 + AB$. Empirical evaluations show TCP-LLM outperforming DRL and traditional CCAs in fairness, stability, and adaptability across dynamic Wi-Fi scenarios, highlighting the potential of LLMs to transform TCP optimization and deployment in real-world networks.

Abstract

The new transmission control protocol (TCP) relies on Deep Learning (DL) for prediction and optimization, but requires significant manual effort to design deep neural networks (DNNs) and struggles with generalization in dynamic environments. Inspired by the success of large language models (LLMs), this study proposes TCP-LLM, a novel framework leveraging LLMs for TCP applications. TCP-LLM utilizes pre-trained knowledge to reduce engineering effort, enhance generalization, and deliver superior performance across diverse TCP tasks. Applied to reducing flow unfairness, adapting congestion control, and preventing starvation, TCP-LLM demonstrates significant improvements over TCP with minimal fine-tuning.

Figures (15)

• Figure 1: Visualization of performance efficiency and reliability of LLMs in comparison to TCP-LLM in CCA selection TCP task
• Figure 2: Comparision of the trainable parameters and resource-intensive adaptation cost of a whole parameter fine-tuning. TCP-LLM's low-rank TCP adaptation approach minimizes the cost of implementing LLM in TCP.
• Figure 3: TCP-LLM features an encoder for data processing, a task-specific head, and Low-Rank Adaptation for efficient learning, with extensible principles demonstrated on three key TCP tasks.
• Figure 4: Graphical presentation of the integrated encoder of TCP-LLM for TCP data encoding.
• Figure 5: The detailed design of TCP-LLM Head and implementation in solving TCP-specific problems.