Research On Dynamic Scaling Technology Of Container Cluster Based On Combined Prediction Model

Kubernetes is one of the most influential cloud platform container management tools today.Although the platform has been widely used and recognized,its original scaling strategy is relatively passive in the face of today's changing needs and complex situations.The strategy still has certain flaws.Aiming at the complex characteristics of server load data and the relatively passive nature of Kubernetes' responsive scaling strategy,an Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN),channel attention mechanism(SENet)and Temporal Convolutional Network(TCN)are proposed hybrid prediction model CEEMDANSE-TCN.By actively predicting the load situation of the container cluster at the next time,and performing corresponding pre-scaling operations according to the predicted value,the Kubernetes cloud platform can respond faster to changes in server load,and in this active way,the server can be prevented from being in a long-term situation.High-pressure state and resource waste,so as to achieve the purpose of improving the service quality of the platform and reducing the maintenance cost of the cloud platform operator.Aiming at the problem of unnecessary scaling of server container clusters caused by data jitter,this paper proposes an improved dynamic weight scaling strategy DWHPA based on the original Kubernetes algorithm.The strategy comprehensively considers the two data of the actual index and the predicted index,selects the historical data by the window method with the specified step interval to obtain the analysis matrix,and analyzes the matrix through the Critic weight method to dynamically determine the weight relationship between the two indicators.The obtained weight and the corresponding formula get the final expected Pod number.In this way,the impact of data jitter on the expected value and dynamic weight is reduced,and the overall trend of server load changes during data jitter is more objectively reflected to avoid container clusters.Invalid scaling,improve container cluster resource utilization.In this paper,Kubernetes is used to build a cloud platform container cluster for experiments,and the Google cluster dataset is used as the experimental data.The experimental results show that the prediction accuracy of the CEEMDAN-SE-TCN model is better than other methods under the same conditions.At the same time,through two sets of comparative experiments,it is verified that the dynamic scaling strategy proposed in this paper can respond in time and in advance when dealing with data transformation,and perform corresponding expansion and contraction operations on the server cluster to avoid misoperations caused by data jitter.the effectiveness and feasibility of this strategy.