Research On Structure Of Scalable Data Center Networks

Data centers gather software and hardware resources of data processing,which undertakes the massive data processing and storage,and provides the dynamic,efficient,reliable and inex-pensive information service for users on request.Architecture of data center networks is com-prised of a large number of servers with high-speed links and switching equipments,which de-termines the organizational form of data processing and storage,and directly impacts level of service for users.In this thesis,we focus on the research of structure in scalable data center networks,and design novel data center networks according to different application requirements and equipment characteristics.Based on the research status at home and abroad,to meet per-formance requirement of data center networks with different scales,this thesis focuses on three related issues of data center networks design,large-scale data center network structure,free-scale data center network structure and container data center network structure.The main contents are as follows:1,To solve the problem of structure solidification in large-scale data center networks,this thesis proposes FleCube which is directly built on multi-port servers.The proposed architecture is recursively constructed on division of multiple ports in a server,where a high-level FleCube is built from low-level FleCubes by means of complete connection.According to the struc-tural characteristics of FleCube,we propose divide-and-conquer routing and multi-path routing algorithm.FleCube enjoys a good design flexibility and scalability,which can meet the high-performance requirements of large-scale data center networks.2,To design large-scale data center networks on single-port and dual-port servers,this thesis proposes ComCell and DPCell,respectively.ComCell takes symmetry and flexibility of network structure into consideration,where divide-and-conquer routing and fault-tolerant routing algo-rithms were proposed.DPCell can achieve a large scalability and bisection width with a small network diameter.Both architectures can build large-scale and high-performance data center networks based on the characteristics of equipments.3,To improve network performance in flexible data centers,this thesis proposes Length Shuffle connection based on greedy strategy to built high-performance and flexible network ar-chitectures.The proposed connection follows a simple principle that greedily connects the fur-thest two top-of-rack switches with their available ports.This connection can reduce the number of long paths significantly,which results in a small average length of routing path.The Length Shuffle connection can work with random connection and regular connection,where it can ef-fectively reduce data transmission delay and improve network throughput.4,To satisfy requirements of high network performance and high fault tolerance in con-tainer data center networks,this thesis proposes a novel container data center network,DPClos.DPClos takes advantage of the folded 3-stage Clos structure and dual-port servers,which can provide double capacity of network bandwidth,a small network diameter and high fault toler-ance.DPClos provides a short data transmission delay and large network throughput,and it can meet performance requirements of container data center networks.