Optical circuit switching networks dynamically adjust their topologies to meet varying communication patterns. Conventional methods rely heavily on comprehensive traffic data collection, centralized processing, and complex resource allocation strategies. However, as networks scale to hundreds of thousands of nodes, these centralized solutions become impractical, imposing overwhelming demands on the computational and storage capacities of controllers. This results in a significant degradation of network reconfiguration efficiency and responsiveness. To address these challenges, we propose DiReNet, a distributed architecture for joint topology and routing optimization in optical circuit switching networks. Unlike traditional traffic matrix-based centralized approaches, DiReNet leverages link modification requests as a more effective reconfiguration metric for large-scale networks. Each node independently monitors link utilization and triggers localized reconfiguration, using in-band control and paired transmission to reduce control overhead. DiReNet also features a flooding-negotiation routing mechanism with one-hop relay limits, coordinating traffic draining, flooding, and feedback to ensure efficient resource utilization and seamless operation during reconfiguration. Simulations show that DiReNet’s reconfiguration delay is nearly unaffected by network scale, remaining only 24.8 µs even with 70,000 servers. Under real-world workloads from Google data centers, Meta’s Hadoop clusters, and web search traffic, DiReNet outperforms the classical distributed scheme RotorNet with higher throughput and lower latency while reducing queue length requirements by 3.4× and reconfiguration speed requirements by 100×. Compared to classical centralized schemes (e.g., KM, iSLIP, and Solstice), DiReNet achieves up to 100× lower latency and queue length and reduces the flow completion time by 21.49%–73.7%.
Li et al. (Tue,) studied this question.