数据中心网络多维感知的流调度模型及机理研究

With the rapid development of the cloud computing, data center, which acts as a platform carrying computing resource, network resource and storage resource, is deployed within enterprises, government and research institute, so social communities have shown an increasing reliance on the data center. Traditional data center consisted of servers and Ethernet switches, which severely limit the scalability of data center due to the shortcoming of Ethernet network. In Ethernet, ARP does not scale well, MAC learning leads to traffic floods and STP has poor convergence. To meet the programmability, performance and scalability and the implementing of the efficiently scheduling of network flow and maximized using of network resource, the project proposes multidimensional awareness-based dynamical flow scheduling, and focuses on the technologies of the transmitting of multiple-path flow. Firstly, the conception of link quota is introduced to illustrate the requirement of application, the configuration of network resource and the policy of the management. Secondly, by the theory of Nash equilibriums, multidimensional awareness-based dynamical scheduling for fitting the requirement of application with given network resource is proposed. Within the mechanism of openflow switching, a hybrid forwarding of link quota-based and destination MAC-based forwarding is studying to support a scalable data center network. Finally, by the support of multiple link quota and multiple path, a parallel transmitting and controlling of multiple subflow is proposed to improve the aggregation bandwidth of application. The dynamical flow scheduling based on multidimensional status is a fined grained scheduling, which could improve the usefulness of network resource and has important theoretical and practical significance in the designing of the data center.

数据中心网络细粒度流调度是实现网络资源精确控制和高效利用的重要方法，而随着云计算技术的发展和成熟，数据中心网络规模在不断扩展，数据中心应用需求更趋多样化和控制策略更加复杂化，现有网络资源动态调度技术难以适应上述发展趋势，亟待深入探索和研究流级调度基本机理，并取得突破。项目提出多维状态信息感知的动态流调度模型，对大规模数据中心网络多路径传输等机理展开研究：为了统一刻画用户需求和网络资源，设计了多维状态信息感知和聚合机制；借鉴Nash均衡思想，设计了应用需求与网络资源匹配的基于多维状态信息的动态流级调度机制；利用集中式网络管理器，实现管理策略与控制规则的映射以及对节点的控制；基于Openflow规则交换实现混合式转发，支撑大规模数据中心网络细粒度控制；提出多微流多路径并行传输，有效增加网络节点间聚合带宽。项目的研究成果为优化大规模数据中心网络资源配置提供新机制，具有重要理论及实践意义。

以大规模数据中心网络作为研究对象，从可扩展的数据平面抽象、数据平面多模式转发机制、无损网络传输模型、数据中心网络拥塞的监测与定位、基于拥塞控制的动态网络带宽分配等方面展开研究，主要突破了以下关键技术：首先提出了普适的数据平面抽象方法，从数据平面计算资源、存储资源和网络资源多维度进行抽象和统一表述：其次提出了数据平面混合转发机制，设计了面向存储的转发服务和面向规则的转发服务；提出了无损网络传输模型，从用户、网络和管理更好的支持网络可感知、应用可区分、策略可管理；提出了分布与集中相结合的方法对大规模数据中心网络进行拥塞的监测和定位，设计的贪心算法以尽可能少的探测分组数覆盖网络链路从而达到有效网络监控；提出了基于无损网络中限速表拥塞控制，设计实现了一种动态的带宽划分算法解决云计算数据中心的多实体竞争网络带宽资源的问题。