With the fast deployment of datacenters (DCs), bandwidth-intensive multicast services are becoming more and more popular in metro and wide-area networks, to support dynamic applications such as DC synchronization and backup. Hence, this work studies the problem of how to formulate and reconfigure multicast sessions in an elastic optical network (EON) dynamically. We propose a deep reinforcement learning (DRL) model based on graph neural networks to solve the sub-problem of multicast session selection in a more universal and adaptive manner. The DRL model abstracts topology information of the EON and the current provisioning scheme of a multicast session as graph-structured data, and analyzes the data to intelligently determine whether the session should be selected for reconfiguration. We evaluate our proposal with extensive simulations that consider different EON topologies, and the results confirm its effectiveness and universality. Specifically, the results show that it can balance the trade-off between the number of reconfiguration operations and blocking performance much better than existing algorithms, and the DRL model trained in one EON topology can easily adapt to solve the problem of dynamic multicast session reconfiguration in other topologies, without being redesigned or retrained.

中文翻译：

---

使用图形感知深度强化学习自适应地重新配置弹性光网络中的组播会话

随着数据中心（DC）的快速部署，带宽密集型组播服务在城域网和广域网中越来越受欢迎，以支持DC同步和备份等动态应用。因此，这项工作研究了如何在弹性光网络 (EON) 中动态地制定和重新配置组播会话的问题。我们提出了一种基于图神经网络的深度强化学习 (DRL) 模型，以更通用和自适应的方式解决多播会话选择的子问题。DRL 模型将 EON 的拓扑信息和组播会话的当前提供方案抽象为图结构数据，并分析数据以智能地确定是否应选择会话进行重新配置。我们通过考虑不同 EON 拓扑的广泛模拟来评估我们的提议，结果证实了其有效性和普遍性。具体而言，结果表明，它可以比现有算法更好地平衡重新配置操作的数量和阻塞性能之间的权衡，并且在一个 EON 拓扑中训练的 DRL 模型可以轻松适应解决动态组播会话重新配置的问题。其他拓扑，无需重新设计或重新训练。