Research On Optical Network Architectures And Resource Optimization Mechanisms For Data Centers

Data centers are the information infrastructures of the new cloud service,including smart city,intelligent transportation,intelligent medical care and pre-cision agriculture.Data centers are the performance engines for cloud service and have important research and application value.The optical networks for data centers adopt optical switching technology,which greatly improves their capacity,and also effectively avoids the conversion of optical-electric-optical,and receives extensive attentions in recent years.At present,the research on data center optical networks mainly focuses on the network architectures.This is because the network diameter,data center traffic characteristics(high fan-out and cluster traffic)and network architecture features(reconfiguration and non-reconfiguration),drive the redesign of the traditional network architectures.At the same time,the resource optimization of the data center optical networks is also worth studying.The reasons are the data center optical networks need to support the new communication patterns,and the establishment of the light-paths will be restricted by their reconfiguration,which poses new challenges for resource optimization.This thesis starts with data center traffic characteristics,new communication patterns and network architecture features,and studies data center optical network architectures and resource optimization.Specifically,the main innovations and contributions of this thesis include:(1)Aiming at the problem of uneven port capacity caused by wave-length selective switch and the problem of excessive network diameter in multi-level interconnection,a low network-diameter data center optical network architecture based on wavelength selective switch is proposed(SWSS).SWSS is a non-reconfigurable two-level network architecture.The first is to form a basic building block(BCU)by interconnecting the wavelength ports of multiple wavelength selective switches(WSS)through a structure of complete graph,which solves the problem of uneven port capacity caused by WSS,and BCU supports the traffic localization;the second is to reduce the network di?ameter by interconnecting BCUs through the complete graph.Under random traffic,when the network load is Erlang=900 and the network size is N=4,the simulation results show that the blocking rate of SWSS is 25%lower than that of Fat Tree;under non-random traffic,with Erlang=500 and N=4,the simulation results show that the blocking rate of SWSS at the traffic localiza-tion factor of 4/1 is reduced by 42%in comparsion to that under random traffic,while that of Fat Tree is reduced by 7.6%.(2)Aiming at the problem of OSA architecture without considering cluster,low fan-out and low scalability,a high fan-out scalable and cluster-based optical interconnect(FSCOI)network architecture for data centers is proposed.FSCOI is a reconfigurable network architecture consisting of an optical switch and many clusters.Its high fan-out is obtained by two steps:the first is that the couplers responsible for sending,multiplex the optical sig-nals of all the top-of-rack(ToR)switches in the cluster;the second is WSS responsible for receiving,can communicate with all ToR switches within the cluster.In the case where the network size of the 48 ToR switches and the most data volume first(MDVF)algorithm are used,when the hotspot traffic load is HS(8,120,40),the simulation results show the throughput of FSCOI outper-forms that of OSA by up to 11%;when the hotspot traffic load is HS(8,120,60)and the percentage of intra-cluster traffic is 30%,the performance gain is of FS-COI is 25%,while that of OSA is 5.4%.(3)Aiming at the problem of parallel flow routing and spectrum al-location(CofRSA)in orthogonal frequency multiplexing data center opti-cal networks,an integer linear programming(ILP)is modeled,and most coflow size first(MCSF)sorting algorithm and greedy inserting(GI)al-gorithm are proposed.Within a data center,communication between data-intensive service often requires a set of parallel data flows,called coflow,and can express a variety of communication patterns.The ILP model not only con-siders the flow constraints between different coflows,but also considers the flow constraints within a coflow,and its goal is to minimize the number of spectrum used.At a smaller network of 6 nodes,ILP provides an optimal solu-tioin;at a larger network of 20 nodes,when the coflow width is W=4 and the traffic granularity is D=25,GI outperforms MCSF and first-come-first-served(FCFS)by up to 6.6%and 9.5%,respectively.(4)Aiming at the problem of user bandwidth reservation problem(KSTA)in the reconfigurable orthogonal frequency division multiplexing data center optical networks,an integer linear programming is modeled,and the most spectrum first(MSF)algorithm,most time first(MTF)as well as most data volume first(MDVF)algorithm are proposed,and sim-ulated annealing(SA)algorithm is applied.The goal of the ILP model is to maximize the network throughput,with constraints of spectrum continuity,time continuity,and request selection.At a smaller network of 8 ×8 optical switch,ILP provides an optimal solution;at a larger network of 16 × 16 optical switch,when Erlang=35 and the spectrum granularity is SG=17,SA outper-forms FCFS by up to 14%,whereas MSF,MTF,and MDVF outperform that by up to 4.6%,9.6%,and 11%,respectively.In summary,this thesis studies two key problems of data center optical networks.In terms of network architecture,the existing optical switching de-vices are adopted,and the network diameter,data center traffic characteristics and network architecture characteristics are considered,and two novel optical network architectures are constructed.In terms of resource optimization,a new communication pattern and network architecture characteristics are considered,which introduces new problems and are solved.The simulation results verify the feasibility and effectiveness of the research works and provide reference value for the extensive application of data center optical networks.