LP-HPA:Load Predict-Horizontal Pod Autoscaler for Container Elastic Scaling

Abstract

In the cloud environment, application elastic scaling is very important. The number of copies can be dynamically adjusted according to load. A good elastic scaling scheme can not only ensure the stability of application, but also improve resource utilization of platform. The existing responsive scaling strategy of Kubernetes platform has many problems, which can not meet requirements of web system for service quality. This paper optimizes the default elastic scaling scheme in Kubernetes cluster, and proposes a container dynamic scaling scheme LP-HPA (load predict horizon pod autoscaling) based on load prediction. This scheme uses LSTM-GRU model to predict the application load, comprehensively considers predicted data and current data, realizes dynamic scaling of container, and ensures the service quality of application. Finally, by building Kubernetes cluster, this paper uses open source data set to verify the LP-HPA scheme. Experimental results show that our proposed scheme is better than Kubernetes' default scaling scheme in three scenarios: load rise, load drop and load jitter.