Research On Load Balancing Strategy Of Data Center Network Based On SDN

Massive computing tasks and service migrations have resulted in a substantial rise in Data Center Network(DCN)traffic due to the fast development of cloud computing and mobile applications.The problems of link congestion and packet loss caused by uneven link loads have arisen.Software Defined Network(SDN)provided a new idea to solve the uneven load of DCN links with the advantages of separation of data forwarding and control,programming control,and the ability to dynamically obtain network status and achieve refined routing control.Moreover,the SDN multi-controller architecture made the scaling and management of DCN more flexible.However,the multi-controller architecture may suffer from controller load imbalance due to the evolving of DCN traffic.Therefore,this thesis investigates the load balancing strategy of DCN based on SDN.The main research contents are as follows.A Link Load Balancing Strategy based on Link Preference(LLBSLP)is proposed for the problem of unbalanced link load in the data plane of DCN based on SDN.The SDN controller dynamically detects the link status of the data plane to get the bandwidth utilization of each link,and then maps the link bandwidth utilization to the link preference.The traffic on the link selects the next hop link based on link preference to realize the even distribution of the traffic in the link.To achieve dynamic routing control,the mapping relationship between link bandwidth utilization and link preference changes dynamically with network traffic.The simulation results show that LLBSLP can effectively improve the link load balancing effect,increase the overall bandwidth utilization rate of the link,and reduce the packet loss rate.A Controller Load Balancing Strategy based on Deep Reinforcement Learning(CLBSDRL)is proposed for the problem of controller load imbalance in the control plane of DCN based on SDN.Modeling analysis for the network is performed by using the Markov Decision Process(MDP)to obtain system state,migration action set,and system reward.The Q-values of switch migration actions are obtained by fitting approximate function using Double Deep Q-Network(DDQN).The DDQN is then trained by optimizing the Q-network parameters using empirical replay techniques.After training,the DDQN analyzes the Q-value in the current system state,and chooses the migration action that corresponds to the maximum Q-value.According to the simulation results,CLBSDRL successfully balances the controller load and significantly reduces the balancing time.The load balancing impact of the LLBSLP and the CLBSDRL working together is verified using a simulation test platform.The findings demonstrate that the two proposed strategies can significantly improve the impact of link load balancing and controller load balancing,increase the average link bandwidth utilization and the average controller load ratio.