面向云计算数据中心的光互连网络架构与路由设计

With the rapid development of data-intensive services and applications, cloud computing is employed to provide virtualized computing, storage and network services, through the centralized management and scheduling of networked computing resources. Data center networks (DCNs) interconnect thousands of servers inside the data center for transporting and switching high volumes of data among servers. Traditional DCNs are based on electronic switching fabrics and have the following problems: low bisectional bandwidth, poor scalability, high cost and power consumption, long packet delay, inefficient routing schemes and so on. Optical technology has been adopted in DCNs due to its high bandwidth capacity. However, current optical DCN solutions still have the problems such as: (1) Optical circuit switching (OCS) based proposals suffer from the relatively long path setup and switching delay in tens of milliseconds, which is not suitable for bursty traffic; in addition, traditional OCS networks follow the fixed coarse wavelength grids, which results in low spectrum utilization efficiency and poor network performance; (2) Optical packet switching (OPS) based proposals require complex buffer congestion managements or expensive optical components due to the lack of optical memories; (3) Hybrid optical switching based proposals need optical components that are not commercialized and require a large number of optical transceivers as the number of servers increases. To solve the problems mentioned above, this project studies the architecture and routing of optical interconnected data center networks for cloud computing. This project proposes a novel two-level optical interconnected DCN architecture, which consists of intra-cluster optical interconnection networks (OINs) and the inter-cluster backbone OIN. The intra-cluster OINs leverage the wavelength division multiplexing (WDM) routing and Ethernet technologies providing high-speed packet switching among nearby servers inside the cluster.It would increase network throughput and reduce packet delay. The inter-cluster OIN utilizing multi-carrier light sources and elastic optical networking technologies provide variable bandwidth pipes among clusters, which would improve the spectrum utilization efficiency and network performance. Considering the characteristics of data center traffic, we propose a novel routing scheme by combining the oblivious routing and demand specific routing methods in conventional DCNs through predicting the probability density functions of the traffic matrices. The research results of the project will provide new ideas and key technologies for designing future high-performance, highly-scalable and high-throughput green cloud data center networks.

为满足日益增长的数据密集型服务与应用的发展，云计算通过统一管理和调度由网络连接的计算资源，为用户提供虚拟的计算、存储和网络等服务。数据中心网络是云计算数据中心内互连大规模服务器，实现海量数据传输和交换的基础平台。针对传统电交换数据中心网络存在：等分带宽小、可扩展性差、成本和能耗高、流量路由效率低等问题，本项目研究面向云计算数据中心的光互连网络架构与路由设计，提出新型的二层光互连数据中心网络架构。连接相近服务器的集群内光互连网络混合使用光波分复用路由和以太网技术能够提供高速分组交换。服务器集群间主干光互连网络利用多波长光源和灵活栅格光网络技术提供可变带宽的光通道。本项目还通过对数据中心流量特性的分析，采用对流量矩阵进行概率密度函数估计的方法，提出融合流量无视路由和流量具体路由的新型路由机制和算法。本项目的研究成果为设计高性能、高吞吐量、高可扩展的绿色"云化"网络平台提供基础理论和关键技术。

云计算通过统一管理和调度由网络连接的计算存储资源，为用户提供虚拟的计算、存储和网络等服务。数据中心网络是云计算数据中心内互连大规模服务器，实现海量数据传输和交换的基础平台。针对传统电数据中心网络存在的：等分带宽小、可扩展性差、成本和能耗高、流量路由效率低等问题，本项目研究面向云计算数据中心的光互连网络架构与路由设计。..本项目从以下三个层次对光互连数据中心网络架构设计进行了研究。首先是数据中心服务器集群间主干光互连网络架构设计。我们提出了基于多波长光源的频谱高效灵活栅格光网络架构，并提出了频谱碎片感知的频谱分配算法，最后通过理论分析与仿真实验对提案架构与算法进行了性能分析与评价。其次是数据中心服务器集群内光互连网络架构设计。我们提出了基于阵列波导光栅路由器AWGR和电分组交换机的光电混合数据中心网络架构HWRE，并通过理论与数值分析了其成本、能耗以及网络性能如平均分组时延、波长请求阻塞率等。最后是数据中心机架内光互连网络架构设计，我们提出了基于光耦合器的软件定义数据中心机架内无源光互连网络SD-POIRN。同时设计了软件定义媒介访问控制SD-MAC机制避免机架内服务器发送数据产生碰撞，提高网络吞吐量，降低分组延迟。..本项目从以下两个方面对数据中心网络的流量路由与调度机制进行了研究。首先是流量路由。我们提出了面向传统树型以及CLOS型数据中心网络结构的软件定义混合路由机制SHR及分类路由机制SCR。SHR和SCR通过统计计算将数据流分为大流和小流，对大流采用自适应路由算法，对小流采用流量无视路由算法。仿真结果表明，与传统ECMP算法相比，SHR和SCR能够提高网络吞吐量、降低数据流丢弃率和分组端到端时延。其次是流量调度和拥塞控制。我们提出了一种基于流量预测矩阵的负载均衡流量调度机制应用至光电混合数据中心网络，以及差分流传输控制机制用于解决数据中心网络的Incast 拥塞问题，提高网络吞吐量，降低小流完成时间。..本项目同时对相关未来网络技术：软件定义网络和内容中心网络进行了研究。我们提出了基于软件定义的内容中心网络结构，基于K-均值算法和复杂网络度量的内容中心网络缓存机制，以及网络状态感知的内容路由机制。以上研究成果将为未来高性能、高吞吐量、高可扩展的绿色数据中心网络架构和路由设计提供技术方案。