A Machine Learning-Based AQM to Synergize Heterogeneous Congestion Control Algorithms

The coexistence of heterogeneous congestion control algorithms causes network unfairness and performance degradation. However, existing solutions suffer from the following issues: poor isolation reduces the overall performance, while sensitivity to tuning complicates deployment. In this work, we propose Warbler, a machine learning-driven active queue management (AQM) framework. Warbler classifies flows based on traffic characteristics and utilizes machine learning to adaptively control the bandwidth allocation to improve fairness. We implemented and evaluated the Warbler prototype on a programmable switch. The experimental results show that Warbler significantly improves the network performance, achieving a near-optimal Jain’s fairness index of 0.99, while reducing the delay to 60% of the baseline, cutting jitter by half, and saving 43% of buffer usage. In terms of scalability, it supports 10,000 concurrent long flows with latency below 0.7 s. The Warbler has a low cost and strong adaptability with no need for precise tuning, demonstrating its potential in dealing with heterogeneous CCAs.