弹性光网络中面向网络功能虚拟化的资源管控机制研究

With the tremendous bandwidth in optical fibers, elastic optical network (EON) becomes one of the most attractive topics in optical communication area, due to its agile spectrum management. For the same reason, realizing efficient spectrum management and control turns into an extremely challenging as well as important issue in EONs. Meanwhile, network function virtualization (NFV) is another revolutionary innovation technology in telecommunication industry since it can shake up the traditional go-to-market strategies fundamentally. As for these two, it has been widely accepted that EONs can be as a promising physical networking infrastructure for NFV to better satisfy the quality-of-service (QoS) requirements of virtual network function (VNF) services. However, it should also be noted that NFV brings new challenges to the spectrum management and control mechanism in EONs. ..Therefore, in this project, we study how to realize efficient resource management and control for NFV services in EONs. Basically, we propose VNF service chain (VNF-SC) orchestration algorithms for both bandwidth-intensive and data-intensive VNF-SC requests to realize customized VNF-SC services in EONs, utilize machine learning methods for managing and controlling network resources and VNF-SC services more intelligently, and establish a software-defined EON (SD-EON) testbed including both control and data planes to demonstrate VNF-SC orchestrations in a real system and eventually validate the effectiveness of the research in this project.

得益于其巨大的带宽资源和灵活的资源分配，弹性光网络成为当前光网络通信领域的研究热点。针对弹性光网络的频谱灵活性，研究高效的资源管控机制变得极具挑战性和重要性。与此同时，网络功能虚拟化是通信领域里另一项具有变革性的创新技术，能够颠覆性地改变传统的市场策略。弹性光网络被认为能够为网络功能虚拟化提供一种最有效的网络基础设施，为保障高效的虚拟网络功能服务提供可靠的支撑。然而，网络功能虚拟化为弹性光网络的资源管控机制带来新的挑战。. 据此，本项目针对弹性光网络中面向网络功能虚拟化的资源管控机制开展研究，提出针对带宽密集型和数据密集型虚拟网络功能服务链请求的虚拟网络功能编排算法，以实现在弹性光网络中提供定制化的网络服务，并且研究如何利用机器学习对网络资源和虚拟网络功能服务进行智能化地管理与控制，最后搭建基于软件定义弹性光网络架构的虚拟网络功能服务链编排管控系统，验证研究成果的有效性。

本项目研究弹性光网络（Elastic Optical Networks, EONs）互连的数据中心网络（Data Center Networks, DCNs）中面向网络功能虚拟化的资源编排机制，主要分为理论算法和系统平台两部分。首先，在IT和带宽资源编排方面，我们提出基于博弈论的资源定价与编排框架，设计了基于深度强化学习的资源定价与编排算法。然后，在VNF-SC服务编排方面，我们提出了基于深度强化学习的自适应VNF-SC服务框架，它利用深度学习对VNF-SC请求的动态分布进行预测，并且通过在网络中提前部署VNF-SCs，来降低VNF-SC请求的阻塞率，最重要的是，它可以根据请求的动态分布灵活地调整网络服务周期，使得请求服务质量和网络管控的复杂度之间能够取得一个好的平衡。最后，在系统开发方面，我们设计并实现了一个知识定义网络编排（Knowledge-Defined Network Orchestrator, KD-NO）系统。具体地说，它首先预测DC中虚拟机未来对IT和带宽资源的需求，然后根据预测得出最优的网络配置。为了完成预测分析，我们的KD-NO系统收集关于DCN资源利用的遥测数据，分析数据的空间和时间相关性，并预测未来虚拟机的工作负载和流量。然后，我们提出了一种基于DRL的网络自动编排在线解决方案，从低级知识中提取高级知识，动态获取最优网络配置。我们在实际的DCN测试台上实现了KD-NO系统，并验证了其有效性。