数据中心Fat-Tree批量调度光包交换新架构

Along with the fast development of cloud services, data centers have attracted wide attentions as the underlying service enabling technology. The internal switch network of current data centers faces multiple critical issues, such as high power consumption, complex internal interconnects, high cost and unguaranteed switching performance, etc. There are also quite large research rooms on design flexibility, scalability and failure recovery issues as well. In this project, we adopt optical switches and interconnects to connect the server racks in data centers and consider the inherent reconfiguration overhead of the switch network. We propose to adopt batch scheduling based packet switching and traffic matrix decomposition under the Fat-Tree multi-stage switch architecture, and average the switch configuration matrices across multiple optical core switches, to achieve performance guaranteed switching among the servers and racks without packet loss and with bounded packet delay. Based on studying highly efficient new algorithms of batch scheduling and traffic matrix decomposition, we propose the idea of trading off between packet delay bound and the number of core switches, and formulate the corresponding mathematical models, such that we can carry out in-depth studies on architecture and cost optimizations for the internal switch network of data centers, as well as its survivability against core switch failures. Our work brings about a new architecture for the internal switch network of data centers, which is highly efficient, compact, power saving, low-cost, survivable, scalable, and flexible, with guaranteed switching performance.

随着云服务的快速发展，数据中心作为业务支撑平台受到了广泛的关注。当前数据中心内部交换网络面临能耗大、内部连接复杂、成本高、交换性能难以保证等问题，并且在设计灵活性、可扩展性和故障恢复能力等方面尚待进一步深入研究。本项目在数据中心服务器机架之间采用光交换和光连接，考虑其交换网络固有的重配置开销问题，提出在Fat-Tree多级交换架构下采用批量调度包交换和流量矩阵分解、并在多个核心光交换器上分摊交换配置矩阵的方法，来实现服务器和机架之间无丢包、延时绑定的交换性能。基于对高效的批量调度与流量矩阵分解新算法的研究，提出数据包延时与核心交换器数量折中设计的思想，并建立相应的数学模型，进而对数据中心内部交换网络结构与成本优化、核心交换器故障恢复进行深入研究，实现高效、紧凑、节能、低成本、可靠、可扩展、设计灵活、交换性能保证的数据中心内部交换新架构。

通过四年的研究，本项目主要研究了数据中心服务器机架之间采用光交换和光连接，考虑其交换网络固有的重配置开销问题，提出在Fat-Tree多级交换架构下采用批量调度包交换和流量矩阵分解、并在多个核心光交换器上分摊交换配置矩阵的方法，来实现服务器和机架之间无丢包、延时绑定的交换性能。基于对高效的批量调度与流量矩阵分解新算法的研究，提出数据包延时与核心交换器数量折中设计的思想，并建立相应的数学模型，进而对数据中心内部交换网络结构与成本优化、核心交换器故障恢复进行深入研究。