一种基于负载特征预测的容器云弹性伸缩策略

A Container Cloud Elastic Scaling Strategy Based on Load Characteristics Prediction

当前主流的容器云平台Kubernetes,利用基于阈值的弹性伸缩策略提供弹性伸缩服务时存在2个问题:一是不区分应用类型,采用单一指标不能准确衡量复合型应用负载情况;二是在应用遭遇暂时的负载增减时容易造成伸缩抖动,导致额外的系统耗费和资源浪费.针对上述问题,提出一种基于负载特征预测的容器云弹性伸缩策略,该策略利用不同的负载特征区分应用类型,对复合型应用类型采取多个指标衡量负载情况;为减少无谓的伸缩抖动,该策略同时利用应用的负载预测值和当前负载值共同进行伸缩决策.实验结果表明,与Kubernetes基于阈值的弹性伸缩策略相比较,该策略能更准确衡量CPU?MEM复合型应用负载情况,在应用遭遇暂时性负载增减时能减少无谓的伸缩抖动.

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.