Design And Implementation Of Application-driven Optical Interconnection Network Node For Computing System

With the rapid development of big data services and distributed computing systems,compared with traditional data center(DC)networks,current cloud DC networks are featured with larger traffic volume,more complicated and diversified traffic pattern,in which unicast,multicast and broadcast traffic coexist,and meanwhile,different from end-to-end features of business in traditional networks,the new communication features are not only multipoint to multipoint,but also highly relevant among flows(Coflow),as a result of distributed applications.These obvious changes cause the growth of bandwidth demand,and the scheduling control becomes more complicated for flows,which brings new challenges to the current network communication system.In order to solve the network communication performance bottlenecks caused by these changes,and improve the communication efficiency,what can be done is to optimize the flow scheduling in the network.However,reasonable and efficient scheduling needs to classify these mixed flows in advance.Therefore,this thesis first analyzed the flow characteristics of applications in DC,and then based on OBRN(Optical Burst Ring Network)architecture oriented DC interconnection,proposed an application-driven optical interconnection network node,which can achieve flow classification under various scenarios and business adaptation after classification with small flow firstly scheduling policy.Finally,we verified classification accuracy and the improvement of network communication performance with designed optimized scheduling through employing the high-speed FPGA and fast/slow optical switches to build network switching platform.The test results show that the accuracy of flow classification can reach 92.5%in certain conditions.Meanwhile,compared with the scheduling ways without flow classification,the optimized scheduling of small-flow-first can greatly decrease the flow completion time,and accelerate the network communication,which shows that our design and implementation of the optical interconnection nodes are effective for flow optimized scheduling in complex scenarios.