Model-Based Deadtime Compensation of Virtual Machine Startup Times

Scaling the amount of resources allocated to an application according to the actual load is a challenging problem in cloud computing. The emergence of autoscaling techniques allows for autonomous decisions to be taken when to acquire or release resources. The actuation of these decisions is however affected by time delays. Therefore, it becomes critical for the autoscaler to account for this phenomenon, in order to avoid over- or under-provisioning.

This paper presents a delay-compensator inspired by the Smith predictor. The compensator allows one to close a simple feedback loop around a cloud application with a large, time-varying delay, preserving the stability of the controlled system. It also makes it possible for the closed-loop system to converge to a steady-state, even in presence of resource quantization. The presented approach is compared to a threshold-based controller with a cooldown period, that is typically adopted in industrial applications.