Research On Traffic Pattern Analysis Method For Data Center Network

With the development of 5G and the Io T,the number and types of applications are increasing,and the traffic in the network is growing explosively.In the era of micro service architecture development,high-speed and large amount of traffic threaten the network communication of data center,and the network security of data center is particularly important.In the data center network,with the application adopting micro-service development and distributed deployment,the internal network traffic transits from the north-south direction to the east-west direction.Subtle changes in workload and background traffic will lead to large differences in internal traffic.The traditional traffic analysis technology is more about classifying applications and protocols and monitoring network status,while the data center network is mostly internal communication traffic,which needs to monitor network status from a new perspective.This thesis analyzes the structural characteristics of data center network,and puts forward a traffic pattern analysis method for data center network.The main research contents and work of this thesis are as follows::In the traditional traffic classification,the accuracy of traffic classification based on port is low,and the user privacy leakage caused by traffic analysis based on DPI.Combined with the characteristics of data center network traffic transmission and unsupervised clustering technology,an adaptive network state clustering module is constructed.The adaptive network state clustering module combines the data cleaning technology in the data preprocessing stage to fill,remove and normalize the traffic samples;Aiming at the high dimension of traffic data in the network architecture of large data centers,PCA is used in the model to reduce the dimension and retain the main characteristics of traffic;The advantages and disadvantages of K-means and Gaussian mixture model clustering methods are compared and analyzed.Combined with BIC criterion,an adaptive state selection algorithm is proposed,and an adaptive Gaussian mixture model is introduced;Finally,different types of network states are output.The network state clustering module maps traffic characteristic samples to different states,and describes the network state as a whole.Aiming at the problem that the existing sequential pattern mining methods are more used to mine frequent patterns in relational databases than time-series traffic data,a wide inclusion traffic pattern extraction algorithm based on network state sequence is designed.The traffic pattern extraction algorithm improves the existing sequential pattern mining methods,and obtains the traffic patterns that represent the network status of the data center.The algorithm adds constraints to subsequence extraction,prunes redundant subsequences,integrates similar subsequences,evaluates the performance of subsequence extraction algorithm by using the overall inclusion of network state,and obtains subsequence sets that meet the conditions.These subsequence sets are traffic patterns,which can be used for pattern matching.The network state clustering module and the wide inclusion traffic pattern extraction algorithm are verified.The leaf ridge network structure is built in Mininet for simulation experiments.The packet statistics are obtained and the network states generated in different environments are compared.The results show that the network state clustering module can distinguish different network states.The algorithm uses the network state sequences of different environments to verify the wide inclusion traffic pattern extraction algorithm.The algorithm mines the frequent traffic patterns with priority.The mining results of the proposed traffic pattern mining algorithm have high inclusion of the network state and have application value.