COPA: A Combined Autoscaling Method for Kubernetes

Abstract

Autoscaling is one of the major features of Cloud Computing aiming to improve the Quality-of-Service(QoS) in response to fluctuating workloads. Existing state-of-the-art autoscaling methods for Kubernetes focus on single scaling mode, that is, only horizontal scaling and only vertical scaling. For horizontal scaling, a high resource usage rate cannot be guaranteed sometimes; and for vertical scaling, microservice instances appear a performance ceiling that does not grow indefinitely as the supply of resources increases. In this paper, we propose a novel combined scaling method called COPA. Based on the collected microservice performance data, real-time workload, expected response time, and microservice instances scheme at runtime, COPA uses the queuing network model to calculate a combined scaling scheme that aims to minimize the default cost and resource cost. We evaluated our approach in a Kubernetes cluster, and compare it with existing state-of-the-art autoscaling methods under four different workload types. Such experiments show a reduction of ×1.22 for resource cost while ensuring the QoS as compared to the baseline method.