Key points are not available for this paper at this time.
Kubernetes’ native horizontal scaling strategy, horizontal pod autoscaler(HPA), suffers from a single monitoring metric that does not accurately measure workload. In addition, there are response delays in the expansion stage, and the deletion of redundant Pods in the capacity reduction phase may adversely affect load balancing. In order to improve the efficiency of computing resource usage, this paper designs an automatic horizontal scaling system architecture based on the Monitor-Analyse-Planning-Execution (MAPE) loop to address the above problems. The system also equips the horizontal scaling system with the resource deletion strategy RRS, which enables the system to handle bursty workloads better. The main contribution of this paper is to propose a passive strategy based on weighted metric thresholds and an active strategy based on ARIMA prediction. The results of multiple experiments show that the ARIMA-prediction-based strategy has the problem of decreasing prediction accuracy in the late stage of the experiment, demonstrating the limitations of the ARIMA model in the real-time scaling system. The weighted-metric-threshold-based strategy outperforms HPA in that the timing, number, and replica variation of Pod scaling better matches workload variation and improves response latency. This paper provides a reference for researchers, developers, and others using Kubernetes to optimize the performance of HPA.
Yunyun et al. (Sat,) studied this question.