Load Balancing Mechanism Based On Traffic Perception For Data Center Networks

With the development of cloud computing and big data technology,the amount of network traffic presents explosive growth,which imposes higher requirements for data center networks(DCN)of the underlying infrastructure.DCN traffic has the following characteristics: a small number of elephant flows consume most of the network bandwidth,and the traffic shows strong burst.Especially in fat-tree structure DCNs,there exist multiple parallel transmission paths for servers attached with different edge switches to communicate,and each transmission path is composed of multiple links.However,due to unreasonable scheduling of bursty traffic some links features low utilization rate while other links bear unbalanced traffic distribution causing frequent congestion.It has become a critical issue of DCN optimization to effectively schedule the traffic among multiple transmission paths and achieve network transmission with high throughput and low delay.This thesis aims to address the network performance optimization problem based on traffic perception in DCN.The adaptive path selection optimization mechanism based on reinforcement learning,Top K elephant flow detection and traffic perception algorithm,and load balancing implementation scheme under SDN scenarios are studied consequently.The main work is as follows:(1)In order to solve the problem of unbalanced traffic distribution and insufficient utilization of parallel transmission path in DCN,an adaptive path selection optimization mechanism based on DQN(Deep QNetwork)model of deep reinforcement learning is proposed.Through the perception of network traffic and real-time link state,the powerful capability of reinforcement learning in delayed feedback and adaptive adjustment is exploited to choose optimal scheduling for network data flows.Furthermore,the experience playback mechanism and target neural network are employed to reduce the correlation between data and improve the model's convergence property,finally,realizing the adaptive optimization of network traffic.Experimental results show that the proposed mechanism can adapt to the changes of network environment quickly and allocate network resources reasonably and effectively.(2)In order to improve the scheduling efficiency of the DQN-based adaptive path selection optimization mechanism,an improved two-threshold Top K elephant flow detection algorithm is proposed,where the data flows with high transmission rate and large amount of traffic are defined as elephant flows,This algorithm uses D-Left Hash to maintain the elephant flow information,and an effective Count field is added into the hash units to enhance the fairness of the flow displacement,so that the flows in large traffic size can preempt the hash unit with higher probability.Experimental results show that,by reducing hash conflict,the algorithm can more accurately fulfill the detection of elephant flows,so as to find the data flows to be scheduled for the path selection optimization mechanism.(3)Based on the above studies,an adaptive load balancing scheme is designed and implemented on the platform of open source SDN controller Floodlight.Mininet network simulator and other tools are utilized together to conduct experimental verification.The results show that the DQN-based path selection optimization mechanism combined with the double-threshold Top K elephant flow detection algorithm can effectively improve the average link utilization of DCN.The optimized network performs significantly better than the equivalent multi-path algorithm ECMP and mixed routing mechanism SHR in terms of end-to-end throughput and time delay.