Research On Reconfigurable Data Center Optical Interconnection Network And Scheduling Strategy

With the rapid development of 5G,cloud computing,Big data,artificial intelligence and other applications,the network traffic of the data center deploying these applications is increasing explosively,which brings unprecedented challenges to the current data center network architecture based on electric switching technology.Researchers have proposed a series of new architectures such as Fat-tree,leaf-spine,Bcube,Dcell,Dragonfly,etc.to address the issues with the electronic switching architecture.Although these architectures can improve network performance,due to the limitations of CMOS devices themselves,electronic switching networks still face bottlenecks in bandwidth,energy consumption,and latency.Benefiting from the transparent characteristics of data modulation format and rate,and avoiding the optical-electrical-optical conversion,the data center optical interconnection architecture can provide applications with high bandwidth,low energy consumption and low delay services.However,the original architecture mostly used MEMS as optical switch.MEMS will introduce a high millisecond delay in the process of link configuration,and the loss in the process of using MEMS is serious.The emergence of optical switching devices such as AWGR has brought a technical basis for the realization of fast optical switching networks.The proposed architectures such as LIONS and Sirius have verified the possibility of AWGR as an optical switch.However,the internal traffic of the data center is strong,and the network topology designed by evenly dividing the optical bandwidth may have some link congestion and some links are idle.Therefore,it is necessary to implement a reconfigurable mechanism to reconfigure the network bandwidth and improve the bandwidth utilization.On the other hand,the optical interconnection architecture uses optical receivers to receive wavelength signals,but the number of receivers at a switch port is limited.If the number of wavelengths reaching a port at the same time exceeds its reception capacity,wavelength contention will occur.In order to avoid conflicts,the traffic to be sent needs to be uniformly scheduled at the sending end.In view of the above problems,this paper focuses on the research of the optical interconnection architecture and scheduling strategy of the reconfigurable data center.The main work of this paper is summarized as follows:Firstly,an optical interconnection architecture based on AWGR is proposed.The architecture cascades the small port AWGR to realize the optical switching within the cluster and uses the electric switch to realize the communication between the clusters.The traditional optoelectronic hybrid interconnection reconfigurable network generally adds or subtracts optical links dynamically in a specific network topology,and the process is complex.The advantage of this architecture is that it can divide time slots in real time according to traffic conditions to reconstruct link bandwidth,and can adapt to different traffic modes.The simulation results show that the packet loss rate after reconstruction has increased by 58.3%,the throughput has increased by 9.4%,and the average latency of the entire network has increased by 63.8%.Secondly,a scheduling strategy named ANTSS is proposed to reduce the network conflict rate,reduce packet loss and improve network performance.The controller can grasp the information of the current network traffic in real time,analyze these data flows,and then decide which data can be forwarded to minimize network conflicts and send more data flows,and ensure the real-time of the scheduling strategy through the time-slot pipeline mechanism.The simulation results show that using the ANTSS strategy can reduce network packet loss rate by 73.6%,increase throughput by 11.6%,and reduce latency between servers and ToR by 11.9%and 14.7%,respectively.