Predictive-Proactive Congestion Avoidance for Cross-Domain Data Center Interconnect

The inherent cross-domain nature of data center interconnect (DCI) presents a critical challenge in satisfying the transmission requirements of intelligent computing services such as ChatGPT and DeepSeek. However, existing transmission control solutions often suffer from inaccurate network state analysis and limited real-time adaptability to dynamic cross-domain traffic, resulting in low resource utilization and poor performance. To this end, this paper proposes a Predictive and Proactive Congestion Avoidance mechanism (PPCA) for DCI. Firstly, a digital twin network (DTN)-inspired congestion avoidance framework is proposed to support virtual mirroring of physical traffic states and efficient closed-loop traffic management, thereby enhancing the proactivity and stability of network operations. Then, a neighbor-assisted traffic prediction method is developed based on the temporal convolutional network (TCN) model. It integrates neighbor information to capture spatiotemporal dependencies effectively, enabling precise virtual mirroring of traffic states. Furthermore, an exploration-exploitation-based flow scheduling algorithm is designed to generate the optimal scheduling strategy for the predicted traffic, thereby avoiding congestion. Finally, extensive simulations are conducted with real-world traffic replaying, and results demonstrate that PPCA achieves superior abilities across 5 classical baselines in mitigating network congestion, balancing traffic load, and handling burst traffic.