The rapid growth of Data Center Network (DCN) traffic has brought new challenges, such as limited bandwidth, high latency, and packet loss to existing DCNs based on electrical switches. Because of its theoretically unlimited bandwidth and faster data transmission speeds, optical switching can overcome the problems of electrically switched DCNs. Additionally, numerous research works have been devoted to optical wired DCNs. However, static and fixed-topology DCNs based on optical interconnects significantly limit their flexibility, scalability, and reconfigurability to provide adaptive bandwidth for traffic with heterogeneous characteristics. In this study, we propose and conduct performance evaluations on a reconfigurable optical wireless DCN architecture based on distributed Software-Defined Networking (SDN), Deep Reinforcement Learning (DRL), Semiconductor Optical Amplifier (SOA), and Arrayed Waveguide Grating Router (AWGR). Our architecture is called ODRAD (which stands for Optical Wireless DCN Dynamic-bandwidth Reconfiguration with AWGR and Deep Reinforcement Learning). A Mininet simulation model is established to further verify the reconfigurability of the ODRAD network for various server scales. Based on experimental verification, ODRAD achieves an average end-to-end server latency of under a load of 99%. Compression results demonstrate a 17.36% improvement in packet rate latency performance compared to RotorNet and a 15.21% improvement compared to OPSquare at a load of 99% as the ODRAD network scales from 2,560 to 40,960 servers. Furthermore, ODRAD exhibits effective throughput across different routing protocols, DCN scales and loads.

中文翻译：

---

ODRAD：具有 AWGR 和深度强化学习的光学无线 DCN 动态带宽重新配置

数据中心网络（DCN）流量的快速增长给现有基于电气交换机的DCN带来了带宽有限、时延高、丢包等新的挑战。由于其理论上无限的带宽和更快的数据传输速度，光交换可以克服电交换DCN的问题。此外，还有大量研究工作致力于光纤有线 DCN。然而，基于光互连的静态和固定拓扑DCN极大地限制了其灵活性、可扩展性和可重构性，无法为具有异构特征的流量提供自适应带宽。在本研究中，我们提出并针对基于分布式软件定义网络 (SDN)、深度强化学习 (DRL)、半导体光放大器 (SOA) 和阵列波导光栅路由器 (AWGR) 的可重构光学无线 DCN 架构进行性能评估。我们的架构称为 ODRAD（代表带有AWGR 和深度强化学习的光纤无线 DCN 动态带宽重新配置）。建立Mininet仿真模型，进一步验证ODRAD网络对于各种服务器规模的可重构性。经过实验验证，ODRAD在负载下实现了平均端到端服务器延迟为99%。压缩结果表明，当 ODRAD 网络从 2,560 台服务器扩展到 40,960 台服务器时，在 99% 的负载下，与 RotorNet 相比，数据包速率延迟性能提高了 17.36%，与 OPSquare 相比，数据包速率延迟性能提高了 15.21%。此外，ODRAD 在不同的路由协议、DCN 规模和负载下表现出有效的吞吐量。