ISSN 1000-1239 CN 11-1777/TP

• 网络技术 •

### 面向大规模数据中心的常量度数互连网络研究

1. 1(解放军信息工程大学国家数字交换系统工程技术研究中心 郑州 450002);2(数学工程与先进计算国家重点实验室 江苏无锡 214125) (lu.feifei@meac-skl.cn)
• 出版日期: 2014-11-01
• 基金资助:
基金项目：国家“八六三”高技术研究发展计划重点基金项目(2009AA012201)；国家“八六三”高技术研究发展计划基金项目(2011AA01A203)

### Constant Degree Network for Massively Data Center

Lu Feifei1,2, Luo Xingguo1, Xie Xianghui2, Zhu Guiming2, Pu Xiaochuan2

1. 1(National Digital Switching System Engineering & Technological Research Center, the PLA Information Engineering University, Zhengzhou 450002); 2(State Key Laboratory of Mathematical Engineering and Advanced Computing, Wuxi, Jiangsu 214125)
• Online: 2014-11-01

Abstract: An important challenge on designing data center networking (DCN) is how to efficiently interconnect a large number of servers. Traditional tree-based structures are increasingly difficult to meet the design goals of data centers. Recently, a number of novel DCNs are proposed. However, these DCNs expand the scale of data center mainly by increasing the number of servers network interface card (NIC) ports, which brings expanding limitation and managing complexity. Consequently, it is meaningful and challenging to design a scalable structure for data centers, using only the commodity servers with fixed number of NIC ports and low-end, multi-port commodity switches. To address this problem, this paper proposes a novel DCN structure with constant degree called CH, which utilizes fixed number of NIC ports and commodity switches to interconnect large population of servers. The structure is server-centric, and leverages the expansibility and performance using multi-level interconnection. They own two potential benefits, i.e., the expansibility and equal degree. Theoretical analysis and experiment results show that CH has excellent topology properties and can provide large data center with multi-pattern data traffic with high bandwidth and high fault-tolerance, but without increasing the number of NIC ports. Moreover, the modularity makes the structure have good deployablility and maintainability.