Research On Container Resource Scheduling For Kubernetes

With the increasing demand for cloud computing services,container technology appears and develops rapidly.As the most popular container management tool today,one of Kubernetes’ advantages is its automatic container scheduling and elastic scaling capabilities.However,Kubernetes scheduling strategy has some problems,such as insufficient evaluation factors,inability to reflect the real use of nodes and imperfect scoring rules,which easily lead to uneven allocation of resources in nodes.The elastic scaling strategy has the problem that the elastic scaling trigger is not timely.When the service volume rises and falls rapidly,the resource supply is not timely and the resource waste is easy to occur.To solve the above problems,this thesis studies container scheduling and elastic scaling,and proposes a Kubernetes-oriented container scheduling strategy and a load prediction model.The main work is as follows:(1)This thesis proposes a multi-objective container scheduling strategy.The scheduling strategy based on the Analytic Hierarchy Process(AHP)container is improved.Firstly,a multi-objective node evaluation model is established based on the four factors of CPU,memory,disk and network.Based on the real-time status of four types of resources after container deployment,AHP is used to set the relative importance of each factor,calculate the weight of each resource,and improve the scoring strategy.The closeness between the candidate node and the optimal node is used as the scoring index.The experimental results show that,compared with other scheduling strategies,the proposed scheduling strategy has a lower peak value of the average imbalance degree of cluster resources during container deployment,a shorter fluctuation time and a lower average imbalance degree of cluster resources after container deployment,which can effectively improve the load balancing degree.(2)This thesis proposes a load prediction model for container cloud resources.The model integrates Variational mode decomposition(VMD),Sparrow Search Algorithm(SSA)and Long Short-Term Memory(LSTM).First,VMD method is used to decompose the CPU utilization time series to reduce the complexity of time series.Then,the parameters of LSTM are optimized using SSA algorithm.Finally,the components of VMD method decomposition are input into the LSTM neural network optimized by SSA.The container resource load is predicted,which is used as the basis for elastic expansion of the container.The experimental results show that the method has high prediction accuracy and low Mean Absolute percentage Error(MAPE).The forecast trend of the elastic scaling policy based on the forecast method is consistent with the actual load,and capacity expansion and reduction can be performed in advance based on the forecast results.This thesis has 18 figures,11 tables and 90 references.