Research On The Traffic Management Technologies Of Data Center Networks Towards Application Performance Requirements

In recent years,with the extensive application of cloud computing and virtualization technology and the continuous development of emerging application mode,the structure,function,application model and scale of data center network has been changing profoundly.As the information infrastructure,data center has gradually been in the core status of data transmission,the computing and storage.Currently,large data centers have been established by domestic and foreign well-known IT companies,such as Google,Microsoft,Baidu,Tencent,to support online search,online shopping,social application,Map Reduce and other large-scale cluster computing key service.Data centers typically contain thousands of interconnected servers,and a large number of running services cause the exponential growth in network traffic,often resulting in packet loss,increased latency,reduced throughput.At the same time,the performance decline of data center network can seriously affect the performance of applications and quality of service(Qo S).In addition,due to the uneven size and unbalanced distribution characteristics of data center network traffic and soft real-time constraints of online applications,traditional Internet traffic engineering rigid routing and scheduling optimization model has been difficult to adapt to the development needs of emerging applications,which brings new challenges and problems in the data center network traffic management.This paper first summarizes the research progress of the data center network and the related traffic management technology,and analyses the main contradiction between the traffic flow characteristics of the data center network and the traffic optimization technology.Then,based on the key technical problems of how to provide service-oriented performance requirements of data center network traffic management,this paper mainly studies the SDN-based routing mechanism,improved TCP protocol of the transmission layer and virtual machine placement to optimize the traffic layout in the topology from low to high network level.The main research results of this paper are as follows:1.According to the low scalability of current large-scale software defined data center network traffic routing mechanism which cause network performance bottleneck,this paper proposes a data center network segment routing mechanism based on Open Flow.The mechanism makes use of edge switches to data stream size threshold test to distinguish the size of the flow.In order to meet the Qo S guarantee and network scalability requirements of different services,this paper proposes a segment routing algorithm for elephant flow.In view of the small flow,we propose an ant colony optimization based ECMP algorithm.Finally,compared to the traditional ECMP algorithm and Hedera algorithm,simulation results show that the mechanism reduces the overhead of controller,and has better network throughout and faster convergence of the algorithm.2.Aiming at this problem that current data center network transport mechanism lacks of comprehensive performance guarantee of soft real-time and high throughout for Online Data-Intensive(OLDI)applications,this paper proposes LSTCP,a dynamic priority-based slack-aware transport control protocol for data center network that adopts the Least Slack First(LSF)scheduling strategy to prioritize flows.Based on the feedback of Explicit Congestion Notification mechanism,LSTCP dynamically adjusts congestion window according to the precedence of the flow and the extent of network congestion,thus implementing the priority scheduling for emergency flows and earlier deadline flows.Experiment results show that LSTCP reduces the AFCT of short flows and guarantees the throughput of long flows compared with the traditional deadline-aware TCP.3.In view of the problem that data center network static virtual machine placement strategy is difficult to meet the demand of virtual machine dynamic resources,a dynamic virtual machine placement model based on Markov decision is proposed in this paper,which transforms the traffic optimization problem into dynamic virtual machine placement problem DVMPP.A Q-learning-based dynamic virtual machine placement algorithm Q-VMDPA is proposed,and Q-learning algorithm is used to solve the large search space.The experimental results show that compared with the current placement algorithm,Q-VMDPA reduces the congestion of the data center network and optimizes the network traffic distribution under reasonable running time overhead.