ARAScaler: Adaptive Resource Autoscaling Scheme Using ETimeMixer for Efficient Cloud-Native Computing

Abstract

The container resource autoscaling techniques offer scalability and continuity for microservices operating in cloud-native computing environments. However, they manage resources inefficiently, causing resource waste and overload under complex workload patterns. In addition, these techniques fail to prevent oscillations caused by dynamic workloads, increasing the operational complexity. Therefore, we propose an adaptive resource autoscaling scheme (ARAScaler) to ensure the stability and resource efficiency of microservices with minimal scaling events. ARAScaler predicts future workloads using enhanced TimeMixer (ETimeMixer) applied with the convolutional method. Additionally, ARAScaler segments the predicted workload to identify burst, nonburst, dynamic, and static states and scales by calculating the optimal number of container instances for each identified state. The offline simulation results using seven cloud-workload trace datasets demonstrate the high prediction accuracy of ETimeMixer and the superior scaling performance of ARAScaler. The ARAScaler achieved a resource utilization of approximately 70% or higher with few updates and recorded the fewest resource overload instances compared to existing container resource autoscaling techniques.