Research On The High Performance Optical Data Center Network

With the development of cloud computing and big data,it has brought dramatically changes to data center,which becomes the main platform to support those applications.Specifically,the changes include the following five aspects: first,there is an increasing trend to integrate computing,networking and storage as a unified system.Second,a single data center now holds tens even hundreds of thousands of servers.Third,the servers in data center work in collaborative ways to address the complicated tasks.Fourth,the virtualization technologies are widely adopted in data center.Fifth,the volume of traffic within data center has reached tens of zettabytes.All these changes present great challenges to the infrastructure of data center.Especially in cloud computing environment,the network,which connects all servers and storage devices,undertakes great roles including collaborating distributed tasks,delivering the heavy east-west traffic,and supporting various virtualization technologies.The performance of data center network has a significant impact on the efficiency of application and completion time of tasks.Traditional data center network cannot meet the increasing demand of cloud services because of the core-layer congestions,limited scalability,poor flexibility,and high cost.To solve these problems,new data center network architectures,including fat tree,VL2,BCube,DCell,Fi Conn,Cam Cube,etc,have been proposed.Although above designs can greatly improve the network performance in terms of End-to-End delays and throughput,these benefits are obtained at the cost of using more cables and switches.Thus the above new designs suffer from complicated topologies,significant wiring problem,excessive amount of devices,and high power consumption.Compared to the electrical designs,the optical interconnection technology has the characteristics of high bandwidth,low transmission delays,low power consumption,transparent delivery.It shows great potential to meet the communication requirements of cloud computing.Thus the optical interconnects have gained attentions in recent years.However,as most existing optical network designs directly transplant the transmission technique,interconnect structure,and switching nodes from the telecom backbone,they may suffer from the following drawbacks: first,in the design of a switching node,most existing switches either experience the overly long switching delay or have a low port density.Second,in the design of the network structure,many prior optical interconnects shows a poor performance in scalability and connectivity.Third,in the design of the control system,almost all existing designs uses the centralized control plane,which then suffer from some implementation issues such as the complex scheduling algorithm and long processing delay.These problems impose negative effects on the network performance,and makes optical interconnects play a supporting role in connecting the servers.A majority of traffic is still processed by electrical devices.The problem of power consumption,device cost,and wiring complexity has not been settled yet.To this end,this dissertation makes research on the design of high performance optical data center network.It tries to solve the existing issues,including the limited scalability,high blocking probability,and long control delay,by introducing a novel optical interconnect scheme.The main works and research achievements of this dissertation are as follows:1.A survey on the research process of the optical interconnection technique for data center is presented in the dissertation.This survey mainly summarizes the latest research achievements in the following five aspects: the network architectures,switch structure,switching schemes,SDN-based optical interconnects,and other research issues.Additionally,a strong emphasis is laid on the study of optical network architectures.Various optical interconnects,including the optical/electrical hybrid networks,the centralized all-optical networks,and the distributed all-optical networks,are described.A comparison of these optical interconnects,in terms of the switching schemes,scalability,technique feasibility,equipment cost,and so on,is made after introducing all existing optical interconnects.Finally,this survey proposes a designing approach to solve existing drawbacks in current optical designs.2.The existing optical circuit switching networks suffer from some problems including the poor scalability,high blocking ratio,and difficulty in acquiring the global traffic information.To solve above issues,a new optical/electrical network with homogeneous topology is proposed.By using the Clos based multistage topology,this design expands the optical connections from the core layer to the aggregation layer.More traffic can be delivered by the optical circuits.To establish an optical connection over multiple hops,a distributed control protocol is proposed.Moreover,the WDM technique is adopted to enhance the connectivity of the optical network.To make a better utilization of all available transmission paths and multiple wavelengths,a novel optical circuit switching strategy,which sets up the connection from the destination node to the source node,is proposed.Finally,the simulation result shows that compared to electrical fat tree,the proposed hybrid network can achieve 1.96 times higher throughput and saturation point when the traffic pattern is mixed with both elephant and mice flows.3.To make a better adaptation to the fast optical switching technique,further,considering the different properties of circuit switching and packet switching,a novel optical/electrical hybrid interconnects called THOR is proposed.It introduces different topologies for optical and electrical network portion to better exploit the unique properties of two different switching schemes.By employing the hypercube topology,THOR provides optical connections directly to hundreds of thousands of servers,and also achieves lower power consumption and buffer capacity compare to electrical and prior hybrid networks.A distributed control system is developed to support the fast optical circuit switching strategy.Moreover,this control system can be integrated into the electrical network to reduce the cost and control overhead.To realize the wavelength-based multi-hop optical switching,a new optical switch structure combing the heterogeneous switching modules is designed in THOR.This optical switch is able to forward message in both space and wavelength domain,using a more cost-efficient way.Further by exploiting the properties of this optical switch and the network topology,an optical path-shared routing algorithm is proposed to reduce the blocking ratio of the optical circuit network.The simulation result demonstrates that under the realistic traffic pattern,THOR is able to achieve 90% of the nonblocking bandwidth.As to the average flow completion time(FCT),THOR achieves significant improvement for both elephant and mice flow.Moreover,by extending the optical circuits into server-level,THOR has cut down at least 53% and 23% power consumption compared to fat tree and traditional hybrid network.4.There is a great challenge to implement the buffering strategy in optical packet switching(OPS)network because of high cost and requirement for reliable delivery.To solve this problem,a new switching scheme called multi-hop Negative Acknowledgement(NACK)is proposed.By exploiting the bidirectional transmission property of optical switching elements,further combining with the resource reservation strategy,a feedback channel can be established when an optical packet is forwarded to the destination node.Then via this feedback channel,the multi-hop NACK scheme can realize reliable delivery in a buffer-less OPS network,by returning the collided packet back to the source node.Based on this NACK scheme,a novel OPS network called Peta Scale is designed to achieve the scalability,reliability,and high-performance simultaneously.Exploiting the property of the topology,a WDM based communication strategy is developed to reduce the blocking ratio and optimize the link utilization.To support this communication strategy,a new optical switch structure is designed.Finally,the performance of Peta Scale is evaluated and compared to fat tree and H-LION.The simulation result reveals that under the uniform traffic pattern,Peta Scale achieves better delay and throughput than H-LION,and can deliver high bisection bandwidth that is 82.4% of the non-blocking network fat tree.Under the local traffic pattern,Peta Scale is able to outperform both fat tree and H-LION.