Anomaly-Aware Adaptive Congestion Control: Learning from Real-Time Deviations

Collie (Kong et al., 2023) systematically uncovers performance anomalies in RDMA systems, while Zhao et al. (2021) demonstrate the power of LSTM models to predict anomalies in software systems. However, current congestion control algorithms rarely adapt on the fly to detected or predicted anomalies—they largely react to packet loss or delay metrics, not higher-order behavioral deviations. This idea proposes a new class of congestion control algorithms that continuously monitor for deviations from expected network performance (using real-time anomaly detection, e.g., LSTM or similar) and adapt their congestion window, pacing, or other parameters immediately upon anomaly detection. This could mitigate cascading failures or long-tail delays in high-speed or critical networks. The novelty lies in closing the loop: instead of treating anomaly detection and congestion control as separate disciplines, we merge them for proactive, context-aware adaptation. This could lead to more resilient networks, especially in environments where traditional signals (loss, ECN, delay) lag behind root-cause anomalies.

References:

1. Collie: Finding Performance Anomalies in RDMA Subsystems. Xinhao Kong, Yibo Zhu, Huaping Zhou, Zhuo Jiang, Jianxi Ye, Chuanxiong Guo, Danyang Zhuo (2023). Symposium on Networked Systems Design and Implementation.
2. Predicting Performance Anomalies in Software Systems at Run-time. Guoliang Zhao, Safwat Hassan, Ying Zou, Derek Truong, Toby Corbin (2021). ACM Transactions on Software Engineering and Methodology.