Research On Container Auto-scaling Based On Kubernetes

With the wide application of cloud computing and artificial intelligence,the role of container technology is increasingly important.Docker is one of the most widely used container technologies,which solves the problems of low resource utilization and inconsistent software stack environment in traditional virtual machine technologies.In the cloud computing environment,large-scale deployment of containers is required.The cluster management solution came into being,Kubernetes is a leader in container cluster management systems which is widely used in the industry.This paper analyzes Docker container technology,Kubernetes cluster architecture and autoscaling technology.Then the Kubernetes auto-scaling strategy is studied and the existing auto-scaling strategy is analyzed in the two stages of capacity expansion and capacity reduction.Finally,an optimization solution is proposed.Due to the problem of response delay in the expansion phase,a load-based expansion strategy is proposed.The method can predict the load of pods combining exponential smoothing and grey prediction and adjust the number of Pods based on the predictions.It can achieve the purpose of capacity expansion before the peak load and reduce the application request response.Aiming at the lack of considering the Pod resource factor in the shrinking stage,combined with the calculation method of state priority in existing strategies,this paper proposes an optimization shrinking strategy that comprehensively considers state factors and resource factors.Then,these two optimization strategies are implemented.Finally,the Kubernetes distributed cluster is built and experiments are conducted on two optimization solutions,The experimental results show that the optimization expansion strategy can effectively solve the response delay problem and reduce the application request response time compared with the existing capacity expansion strategy.Compared to the existing scaling policy the problem of unbalanced cluster load can be optimized and the resource fragmentation of nodes can be reduced,as well as the service quality of the cluster can be improved.