Research On Reconfigurable Optical Switching Network And Control System

The rapid development in data boosting applications,such as cloud computing,Internet of Things(IoT),5G-related services,and network applications,is posing unprecedented challenges to the conventional electrical switch based high performance data center interconnection network,especially in switching technology,network architecture,cost,and power consumption.To outbreak these restrictions,several novel network architectures have been proposed to improve the performance of high-performance data center networks,such as Fat-Tree,VL2,BCube,DCell,FiConn,CamCube,etc.However,due to the technical challenge to increase the pintensity of the Ball Grid Array(BGA)packaging technique,the link bandwidth based on electrical switching architecture has been limited by the ASIC I/O bandwidth.Although the throughput of the network can be improved through the multi-layer topology,the multi-layer transmission method increases the end-to-end delay of the network,with network complexity and equipment cost also dramatically increasing as the network scales.Besides,the variety in data center network(DCN)applications always result in the diverse,real-time and bursting traffic patterns.Confined by static interconnections,traditional electrical topologies cannot adapt to such traffic variety,leading to deterioration in network performance,especially when application requirements change.Optical switching networks have been extensively studied as a potential solution to overcome the limitations of bandwidth,latency,and cost in electrical networks via providing ultra-high capacity.Benefitting from the intrinsic high data rate and transparent data format,optical switches can provide theoretically unlimited bandwidth to overcome the bandwidth bottleneck.Moreover,optical switching networks eliminate the deployment of optical-electrical-optical(O/E/O)conversion,leading to the significant reduce in overall building cost and power consumption.On the other hand,optical switching technology combined with wavelength division multiplexing(WDM)technology also provides flexibly configurable and low-latency network services.The flat design of the architecture can meet requirements for the next-generation highperformance data center interconnection network.Despite the tremendous potential for optical switches,some issues still exist in practically deploying optical high performance data center:1)To take advantage of the fastest switching that optical switches can achieve in nanoseconds,the fast control mechanism is required to configure the switch to forward packets at the nanosecond level.2)In slot allocation based optical switching networks,the stationary scheduling sequence forwards rack-generated packets to the optical switch orderly.Meanwhile,the destination information is required to arrive in the switch controller in time to module the packet accordingly.Therefore,precise time synchronization between the rack and switch controller is key to ensuring the orderly operation of switch nodes.Moreover,to coordinate the ultrahigh rate and high bandwidth characteristics,the synchronization accuracy in optical switching networks needs to reach nanosecond-or even picosecond-level.3)Due to the lack of effective buffer in the optical domain,packet contention on the optical switch will cause a large number of packet losses.4)The traffic patterns of applications in the highperformance data center interconnection network have the characteristics of diversity.The different Quality of Service(QoS)requirements of different services result in heterogeneous requirements for network infrastructure.Therefore,high performance data center interconnect network need a reconfigurable and highly flexible connection to provide a suitable network framework for various applications.Targeting at aforementioned challenges,network architecture design,reconfiguration and routing algorithms,scheduling strategies and contention resolution are emphasized in this paper to realize the reconfigurable topology and the corresponding collaborative mechanism.The main work is summarized as follows:(1)A hierarchical and reconfigurable optical/electrical interconnection network(HRIN)adapted to the traffic pattern varietyIn this work,a reconfigurable optical/electrical interconnection network is designed to seize traffic patterns determined by various high performance computing(HPC)applications.The traffic matrix of the target task can be decomposed into multiple matrix groups for parallel execution,and the network topology of each cluster can be reconfigured according to the characteristics of the sub-flow matrix associated with the cluster.For inter-cluster cross-connections,we use a shuffle switching network to provide a single-hop optical bypass for nodes with high communication intensity to satisfy inter-cluster communication requirements and improve the path diversity.To improve the efficiency of architecture reconstruction,this paper also proposes a topology matching algorithm and a reconstruction optimization algorithm.Meanwhile,based on the Floyd algorithm,an improved routing strategy via harnessing the new link weight is proposed to point at link congestion problem.Simulation experiments verify the feasibility of the reconfigurable optical/electrical interconnection network architecture.(2)A fast switching and picoseconds-synchronized optical datacenter network(FSSAW)The fast flexible optical switches(nanosecond)and control systems have been designed,implemented,and evaluated to realize the real-world deployment of fast optical switches in data center networks.Based on arrayed waveguide grating router(AWGR)and White Rabbit Protocol(WR),the fast optical data center network with the picosecondsynchronization jitter is proposed.Cascaded AWGR,novel fast optical ToRs with multiple transceivers(TRXs)and SOA array,and distributed field-programmable gate array(FPGA)schedulers are deployed in the cluster.The number of SOA equals the number of TRX buffers allocated by destination.Based on the collected traffic volume of ToR,the FPGA scheduler can automatically reconfigure the optical bandwidth in slotbased round-robin scheduling to achieve fast optical switching.Moreover,electronic switches are deployed to forward the inter-cluster traffic and are enhanced with WR protocol(ES-WR)for the first time in DCNs to synchronize the time of ToRs.Experimental demonstrations and simulations verify the network performance of the ReSAW architecture in terms of packet loss rate,end-to-end latency,and throughput.(3)Reconfigurable optical switching network(ReSAW)based on traffic transmission priority schedulingIn view of the inflexibility of periodic round-robin scheduling in the FSSAW network,a non-contention reconfigurable scheduling scheme based on traffic forwarding priority is designed to achieve reconfigurable optical switched network(ReSAW).ReSAW features flexible optical bandwidth via allocating time slots and wavelengths and coordinating various traffic patterns.The FPGA-based distributed scheduling system calculates the delivering priority for the traffic in the next time slot based on the current traffic information.Besides,the scheduler reconfigures wavelengths according to traffic priority,that is,each transmitter(Tx)is only allowed to deliver the wavelength with the highest priority to destination ToR within the same time slot.Since the number of wavelengths that can be received cannot exceed the number of receivers(Rx)deployed on the ToR.Therefore,the scheduler only allows each transmitter(Tx)to send the wavelength with the highest priority to the destination ToR in the same time slot,and the wavelength with low priority will be reconfigured forwarding priority by the scheduler for the next time slot.The experiments verify the reconfigurable performance of the architecture in terms of end-to-end latency,throughput,and packet loss.Simulations are set up with different oversubscription(OV)and time slot lengths to evaluate the effectiveness of the proposed scheduling strategy.