Research On Network Traffic Scheduling Mechanism Of Data Center Based On Network Behavior Prediction

Traffic scheduling has always been an important research direction in data center networks.Network behavior refers to indicators that can reflect the local or global performance of the network(i.e.,link delay,jitter,packet loss rate,and end-to-end delay,etc.).Among them,the delay,jitter,packet loss rate and throughput of each hop are called basic network behavior;the end-to-end delay of packets,the completion time of flow,the utilization rate of network resources and the congestion degree of the whole network are called global network behaviors.Precise network behavior modeling can help optimize traffic scheduling in the network to ensure quality of service.However,the existing modeling methods have shortcomings such as insufficient precision,low generalization ability and high overhead.At the same time,the existing traffic scheduling mechanism has some shortcomings,such as lack of timeliness,insufficient consideration of global network congestion,inadequate adaptability of scheduling mechanism.In order to solve these problems,the main work of this paper is as follows:(1)We propose a Link Delay Model(LDM).Firstly,we obtain the network behavior dataset through network measurement and feature engineering.And we select some common links by clustering.Secondly,we use an improved Graph Neural Network to learn the intrinsic relationship between basic network behavior of the selected link and the global Network behavior.As a use case,we propose an accurate end-to-end delay prediction method based on link delay.Experimental results show that LDM can predict end-to-end delay more accurately when all links are used,and the correlation coefficient R~2 reaches 0.969.Compared with the queuing model and Route Net,R~2 is increased by 73%and 11%,respectively.In addition,in the generalization ability evaluation of LDM,for unknown flow scheduling strategy,the Average Relative Error(MRE)of the LDM reaches 0.285,which is obviously better than the queuing model and Route Net.When using a few common links,the prediction performance of LDM is similar to Route Net,and the overhead is reduced by 78%.(2)Based on the accurate prediction of end-to-end delay by LDM,we extend LDM and design LDM-based Traffic Scheduling System(LDM-TSS)to optimize network traffic scheduling in data center network.System realizes the prediction of the global network behavior in the controller,and plans the path for the packets in advance,and finally realizes the decision-making and packets forwarding in the switches.Firstly,by analyzing the global network congestion situation,we determine the indicators that can effectively guide traffic scheduling are path delay or path congestion level.Then we design a packet scheduling algorithm based on prediction of Link Delay Model and implement decision-making in local programmable switches.The scheduling algorithm can effectively plan up to many paths in advance according to the proportion of packets.The experimental results show that LDM-TSS can effectively optimize the traffic scheduling in the data center network.Compared with ECMP and Let Flow,LDM-TSS reduces the average Flow Completion Time by 42.0%and 35.7%,respectively.(3)When the traffic pattern in the data center network changes,in order to ensure that LDM-TSS can effectively optimize the traffic scheduling.We design a model self-renewal scheme in LDM-TSS to enhance the adaptability of the system.The prediction model in LDM-TSS can be further learned on the basis of the original model with the change of network environment to maintain accuracy and stability.The experimental results show that for the complex and changeable data center network,compared with the LDM-TSS without the model self-renewal scheme,the average FCT is reduced by 7.2%by adding the model self-renewal LDM-TSS.