关键词: 弹性伸缩, 容器云, 负载特征, ARIMA预测, 伸缩抖动

Abstract: Kubernetes, the current mainstream container cloud platform, has two problems when using the elastic scaling strategy based on threshold to provide elastic scaling services. First, it does not distinguish between application types, and it cannot accurately measure the load of composite applications with a single index. Second, when the application encounters temporary load increase or decrease, it is easy to cause scaling jitter, which leads to additional system consumption and resource waste. In view of the above problems, this paper proposes a container cloud elastic scaling strategy based on load feature prediction, which uses different load features to distinguish application types and adopts multiple indexes to measure the load of composite application types. In order to reduce unnecessary scaling jitter, the strategy uses both the predicted load value and the current load value to make scaling decisions. The experimental results show that compared with Kubernetes elastic scaling strategy based on threshold, this strategy can more accurately measure the CPUMEM composite application load and reduce unnecessary scaling jitter when the application is subject to temporary load increase or decrease.

Key words: elastic scaling, container cloud, load characteristics, ARIMA prediction, telescopic jitter