| As the core carrier of cloud computing,the data center supports the oper-ation of actual business through the allocation of resources such as computing and bandwidth.The load balancing mechanism in the data center network can evenly distribute the traffic to each path,reduce bandwidth resource waste or form a system bottleneck,and improve the overall transmission performance.Commonly used Equal-Cost Multi-Path routing(ECMP)mechanism uses a static hash function to evenly distribute traffic between paths without being aware of network conditions.In the case of a mixture of elephant/mouse flows or asymmetric network topology,it may lead to unbalanced load among net-work links.The CONGA,Hermes,CLOVE and other mechanisms proposed in recent years have optimized network congestion detection and routing strate-gies to varying degrees.However,no matter whether it is full path probing or a fixed number of partial probing,there is a performance degradation caused by insufficient flexibility.Therefore,this thesis investigates the mechanisms in the field of load balancing in recent years,uses the Deep Reinforcement Learning(DRL)algorithm for dynamic path probing,and combines with prudent traffic scheduling strategy to propose a new load balancing mechanism named Dyno,which can not only improves the performance of the data center network,but also has strong generalization capabilities.The innovative design and contributions of this paper include the following aspects:ˇA dynamic path detection method based on the DRL model is proposed,and the State,Action and Reward of the model are designed to avoid the problems of ambiguity in congestion description,random actions,or use of non-real-time measurable indicators.ˇThe Model-Replication method is innovatively proposed to solve the train-ing problem in data center network scenarios,which greatly reduces the difficulty of DRL model convergence and inter-process communication overheadˇIn the process of model deployment,lightweight technology is used in the field of load balancing,and multiple decision trees are used to replace DRL model,which can ensure the timeliness of Dyno's decision-making at the cost of minimal loss of accuracy.ˇIn view of the possible performance decline of reinforcement learning when the environment changes drastically,a Model-Rotation method is proposed to greatly improve the generalization and robustness of Dyno.This thesis uses Dyno to conduct a comprehensive comparison experiment with the current typical load balancing mechanisms.The results show that under a symmetric topology to the model training,Dyno can reduce the average flow completion time by up to 16.9%and an average reduction of 8.5%under differ-ent network loads compared with CONGA(the optimal load balancing mech-anism under the Leaf-Spine topology).Even under different traffic patterns,network loads and asymmetric topologies,Dyno still maintains a balanced per-formance in the processing of long and short flows,again leads CONGA with an average advantage of 6.9%for short flows,5.9%for long flows and 5.3%for overall flows,and surpasses all other comparison mechanisms. |
PDF Full Text Request