Improved dynamic modeling for controlled server queues

Resource provisioning for applications hosted in the cloud is a difficult task due to inherent performance variability in the infrastructure. Control theory has proven to be an efficient tool to increase the predictability of cloud applications. However, a prerequisite for a successful control design is an adequate model of the involved dynamics. In this paper we focus on modeling of controlled server queues that are subject to actuators, such as frequency scaling or admission control. We show that today’s models are only applicable to specific server types, characterized by queuing disciplines, and propose a model structure that can be applied for more general settings. Our structure is nonlinear, yet simple enough to allow for control design. We compare our approach to state-of-the-art models in an extensive simulation campaign, showing the superior versatility of our model. We also evaluate the model using measured data from a cloud-based face detection algorithm run in Kubernetes. Furthermore, we use our model in control design examples to show the insights that can be gained. We identify a critical frequency range where the characteristics of the involved service time distribution affect the control design, and where a more advanced controller structure might be needed. Finally, we present a feedback linearization control design based on our model that is evaluated using both simulations and a cloud-based application.