Research On Intelligent Traffic Perception Mechanism For Cloud Data Center Traffic Management

The data center undertakes the tasks of data computing,storage,and ensuring the efficient transmission of data between servers,and is a key node for providing resource services and information sharing.As the underlying communication facility of the data center,the transmission performance of the data center network greatly affects the service quality of various services,so it has become a research hotspot in academia and industry in recent years.In order to reduce the transmission delay of the network,optimize the bandwidth utilization and improve the performance of corresponding services,many traffic scheduling algorithms have been proposed to achieve differentiated scheduling of service traffic with different delay requirements.Most of these flow scheduling mechanisms are based on the idea of Shortest-Job-First(SJF),which prioritizes the transmission of delay-sensitive short streams,and reduces flow completion time through reasonable bandwidth allocation.Usually,in order to determine the priority of each flow,these mechanisms need to effectively distinguish between long and short flows by means of traffic perception.Existing traffic sensing methods mainly include direct acquisition through applications,perception based on priority queues,and estimation of flow length based on machine learning algorithms,etc.However,these methods have shortcomings such as difficulty in parameter configuration,difficulty in actual deployment,and unreasonable design in a multi-service mixed environment.This thesis conducts in-depth research on the problem of traffic perception in traffic management.Combined with deep reinforcement learning,clustering and other artificial intelligence algorithms,this thesis has mainly achieved the following research results:1)An intelligent threshold sensing mechanism DeepAalo is designed for information-unknown Coflow scheduling.For scheduling scenarios with unknown information,most of the existing Coflow scheduling algorithms are based on a priority queue scheduling model,and use the amount of sent data as the basis for identifying long and short flows and dividing scheduling priorities.The thesis analyzes the problem that the static queue threshold adopted by the above mechanism makes it difficult to adapt to actual traffic changes,and designs a threshold-aware mechanism DeepAalo based on deep reinforcement learning.DeepAalo transforms the design of the priority queue threshold into a continuous learning process,and minimize Coflow completion time by automatically adjusting the threshold-to-flow mapping.Simulation experiments show that DeepAalo can effectively design queue thresholds according to traffic characteristics,optimize the completion time of Coflow,and has a good adaptability to the environment with dynamic traffic changes.2)An intelligent flow length estimation mechanism LFE for mixed flow scheduling is designed.The study found that effective flow size estimation can distinguish long and short flows and significantly improve the performance of traffic scheduling.Aiming at the shortcomings of the existing flow estimation mechanism in the mixed application flow environment,such as insufficient prediction accuracy,difficult deployment,and difficulty in quickly distinguishing long and short flows,this thesis designs a fast flow length estimation mechanism LFE based on a clustering algorithm.LFE combines PCA,FastDTW,DBSCAN,and GBDT algorithms realize rapid flow division and flow length prediction by identifying historical flow sequences with similar characteristics.Simulation experiments show that LFE has better flow prediction accuracy,can quickly identify long and short flows,and optimize the completion time of mixed flow scheduling.