融合三支决策的预测式容器伸缩优化策略

Predictive Container Scaling Optimization Strategy Integrating Three-Way Decisions

摘要: 当前Kubernetes中基于固定阈值的响应式伸缩策略存在无法根据集群负载动态调整扩缩容力度以及存在的时间滞后性等突出性问题。针对该问题，利用负载预测模型在检测实时工作负载的同时对未来的工作负载进行预测，同时，根据历史负载变化情况并结合三支决策对负载波动进行三分，即低负载波动期、正常负载波动期、高负载波动期，针对每一种负载波动期，对伸缩力度进行细粒化，最后基于负载波动期和预测的CPU利用率执行相应的容器伸缩决策。通过对比Kubernetes原生算法和同类算法，所提出的伸缩优化策略能够有效地应对负载波动，降低SLA违约率，在保证QoS的同时降低了资源的浪费。

Abstract: The current responsive scaling strategy based on fixed thresholds in Kubernetes has outstanding problems such as the inability to dynamically adjust the expansion and contraction intensity according to the cluster load and the existence of time lag. To address this problem, the load prediction model is used to predict future workloads while detecting real-time workloads. At the same time, load fluctuations are divided into three categories based on historical load changes and three decisions, namely, low load fluctuation period and normal load. During the fluctuation period and high load fluctuation period, the scaling intensity is fine-grained for each load fluctuation period, and finally the corresponding container scaling decision is executed based on the load fluctuation period and the predicted CPU utilization. By comparing the Kubernetes native algorithm and similar algorithms, the proposed scaling optimization strategy can effectively cope with load fluctuations, reduce SLA default rates, and reduce resource waste while ensuring QoS.