Research On Kubernetes-oriented Automatic Container Elastic Scaling Technology

With the rapid development of cloud computing technology and container technology,container-based service applications are expanding and the number of containers is also increasing.Container cloud platform needs to monitor,schedule and manage all containers.At present,container cloud platforms represented by Kubernetes(K8s)mostly use reactive or planned automatic scaling strategies.These passive strategies suffer from scaling delays.Current research does not solve the problem of underprediction,which leads to insufficient application service capacity,resulting in higher request timeout rate and rejection rate,and ultimately lower service quality.In view of the above problems,the paper has carried out in-depth research from the following three aspects:(Ⅰ)In the study of horizontal scaling strategy of container cloud,the existing researches on the problem of network application access request loss caused by underprediction mainly focus on the prediction of underprediction,and there is no policy scheme to intervene after the occurrence of underprediction.A horizontal scaling strategy for container cloud called PFCS is proposed.The strategy on the basis of horizontal scaling strategies have been established on the policy framework to design,optimize the function of each module and data exchange relationship among various modules,mainly from to owe prediction after the intervention,reduce the quality of service due to owe to predict continued falling problem,integration of the two sets of independent load prediction model and a set of emergency expansion algorithm.(Ⅱ)In order to further reduce the probability of scaling delay caused by the load prediction time during the operation of container cloud,a microservice load prediction model based on the characteristics of Kubernetes container cloud platform was proposed,called TSBT.This model achieves the fast prediction of microservice load mainly through the specific optimization of deep learning model Transformer.The model processes and analyzes historical data to predict short-term load changes in the future.Compared with the existing mainstream load prediction models,the prediction time of the model is reduced by91% and the load prediction efficiency is improved under the condition that the prediction accuracy is not lower than the other models.(Ⅲ)An emergency scaling algorithm called ESA is proposed to solve the problem of load underprediction.The algorithm can effectively detect the occurrence of underprediction.In general,only a set of load prediction model is used to reduce system energy consumption.The emergency prediction model can be activated immediately after the underprediction is detected to provide more prediction results for telescopic decision making and shorten the response time for underprediction.Experimental results show that the proposed algorithm can reduce the request rejection rate by 18.6～19.8% and guarantee the service quality when the system suffers from load underprediction.