Data Center Traffic Prediction And Scheduling Based On Deep Learning And SDN

With the continuous deployment and application of cloud computing,high-definition streaming media,and 5G services,the scale of the data center network carrying its services and the number of servers included are also increasing.While the rapid development of data centers has promoted global digital transformation,it has also brought new opportunities and challenges.At present,people know very little about the traffic characteristics of data centers,which hinders the efficient operation and maintenance of the network to a certain extent;at the same time,with the popularity of distributed computing and storage services in most existing electrical switching data,neither the transmission delay nor the system energy consumption meet the performance requirements of the current application on the data center.With the development of optical switching technology,the optical-electrical hybrid switching data center network has gradually attracted the attention of the industry due to its advantages of large bandwidth,low latency,and low power consumption.However,optical circuit switching takes a relatively long time to establish and release connections,and the scheduler needs to conduct unified scheduling by analyzing the reported traffic information.The process of scheduling and configuration increases the overall data transmission delay.Therefore,in order to better grasp the network status and optimize data transmission,traffic prediction technology can be used to predict the traffic characteristics of the data center at the next moment by learning past traffic information,so as to configure wavelengths and ports in advance to realize the average network delay and improve network performance.In view of the above situation,this paper first proposes a data center random traffic generation model for data flow size,destination address,and interarrival time,and then proposes a data center traffic prediction model based on multi-task learning LSTM neural network,and finally scheduling simulation verification is carried out in the SDN architecture,as follows:1.Through the study of data center traffic characteristics,the characteristics of data frame size,data stream size,destination address,and interarrival time are extracted,and the data center random traffic generation model is implemented according to these characteristics,and the data set generated by the model follows the Pareto distribution in the size of the data flow,and follows the lognormal distribution in the interval time,which conforms to the general traffic characteristics of the data center..2.According to the data center traffic information data set generated by the first part of the model,the LSTM neural network based on multitask learning(mt-LSTM)is used to predict the proportion of traffic going to each server in the next 15 ms in the 30 ms data set.Compared with the LSTM algorithm,the convergence speed of mt-LSTM is faster,and the mean square error is reduced by 20.4%,which has a higher accuracy rate.3.In the SDN architecture,the measurement of the two-layer packet sending and one-way transmission delay is realized,and based on the traffic direction predicted by the second part of the traffic prediction model,a new flow table is set on the switch in advance to complete the bandwidth scheduling,and reduce the end-to-end delay.When the bandwidth is 0.1M,the average delay is reduced by 91.2%;when the bandwidth is 0.2M,the average delay is reduced by 55.8%,and the network performance has been optimized to a certain extent.