Pacer: Automated Feedback-Based Vertical Elasticity for Heterogeneous Soft Real-Time Workloads

Abstract

Cloud computing can be used to provide a virtualized platform for running various services, including soft real-time applications such as video streaming. To satisfy an application's real-time requirements, CPU resources are often allocated for the worst case, resulting in system under-utilization or overpaying to the cloud provider under the pay-as-you-go model. To solve this problem, we present Pacer, a framework that provides application developers a platform to implement custom virtual machine-level resource allocation algorithms that utilize real-time application-specific performance feedback from applications running in virtual machines. We also present two example resource allocation algorithms for Pacer that are based on additive-increase-multiplicative-decrease and self-tuning PID control. We apply Pacer to video stream object detection applications to show that Pacer can save more than 50% CPU utilization and use CPU resources more efficiently, while still meeting deadlines for real-time applications.