Scheduling Architecture And Techniques For Hybrid Optical/Electrical Data Center Networks

With the extensive deployment of cloud applications,the communication patterns and business requirements of data centers have undergone fundamental changes.These changes have raised higher requirements for the data center network architecture.The traditional data center networks are unable to meet the needs of cloud applications in terms of bisection bandwidth,transmission delay,network scalability,fault tolerance,and resource utilization.Hybrid optical/electrical network architecture,however,has better performance than the traditional electrical network architecture in the above aspects.Those existing hybrid optical/electrical network scheduling techniques mainly focus on improving the optical switch utilization.And the existing scheduling algorithms take the network traffic as a whole,and try to transmit the sum of traffic in the network as soon as possible.In fact,however,there are different flows with different characteristics in the data center networks including both delay-sensitive short flows and highly-persistent long flows.And the average flow completion time is the key metric of the network performance.The existing scheduling techniques do not consider the different characteristics of different flows and only perform coarse-grained link scheduling.Thus,how to design a hybrid optical/electrical network scheduling scheme based on the flow characteristics in order to minimize the average flow completion time has become a key issue in improving the performance of hybrid optical/electronic data center networks.Based on the above problems,we perform the research in the following four aspects:(1)The development of scheduling techniques in hybrid optical/electrical data center networks are deeply researched,and the existing scheduling works are analyzed and summarized.We research the traffic characteristics and the optimization goal in the data center network as well as the challenges they bring to the scheduling method design.We also analyze the insufficiency of the existing scheduling methods.(2)A novel scheduling architecture for hybrid optical/electrical data center networks is proposed based on the traffic characteristics in data centers.We introduce the concept of flow scheduling into the hybrid optical/electrical networks.We combine flow scheduling and link scheduling methods to achieve the goal of minimizing the average flow completion time.In this hybrid network scheduling architecture,flow scheduling and link scheduling are two separate modules.The designs of the two parts are decoupled but they work together for the same optimization goal.The scheduling system has high design flexibility and good network performance.Finally,we use a packet-level network simulator to implement and compare different hybrid optical/electrical network scheduling methods.The simulation results show that our scheduling method is better than the existing ones,and the average flow completion time is the shortest.(3)A flow scheduling approach based on flow priority is designed and implemented.The flow scheduling uses the Shortest Flow First policy and sets the flow remaining size as the flow priority.Flows with smaller remaining size have higher priorities.Each local flow scheduler on ToR maintains information of the active flows produced by its connected hosts.Every time a flow arrives or finishes,the host notifies the flow scheduler to update the active flows set and perform scheduling.The scheduling algorithm calculates which flows to send according to each flow's priority and its network input and output ports.Then the flow scheduler informs the relevant hosts to restart or suspend the related flows.(4)A link scheduling approach based on traffic bundle priority is designed and implemented.The existing link scheduling methods do not consider the relative priorities of different flows,in which all flows are treated as a whole.And the traffic accumulation stage is usually long,which further increases the completion time of short flows.Given the above deficiencies,we introduce the logical concept of traffic bundle,which is used to evaluate the priority of traffic in each virtual output queue,and we design two definition methods of traffic bundle.The link scheduling algorithm preferentially assign the optical link for the virtual output queue whose traffic bundle has higher priority.Shortest Flow First policy can be approximated via the definition of the bundle priority.The scheduling cycle is relatively shortened,and the waiting time of short flows is further reduced.In addition,priority queues are used for the virtual output queues,and the packets are transmitted in the order of priority,which also follows the principle of Shortest Flow First.In summary,this thesis aims to optimize the scheduling techniques in hybrid optical/electrical data center networks.Based on the analysis of flow characteristics in data center networks,we propose a novel scheduling architecture and its detailed design in hybrid optical/electrical data center networks,and improve the network performance.Therefore,this thesis solves some practical problems in the scheduling of hybrid optical/electrical data center networks,and has both engineering value and theoretical significance.