Efficient Cloud Auto-Scaling with SLA Objective Using Q-Learning

Abstract

Threshold based cloud auto-scaling is one of the most common methods used to scale cloud applications. A major drawback of this method is that the thresholds are set manually by the user in an ad hoc fashion, not optimally, and specially crafted for a specific application behavior, leading to SLA failures. We present Q-Threshold - A novel algorithm for adaptively and dynamically adjusting the thresholds with no need for user configuration while meeting SLA objectives. In this context we present new methods for improving reinforcement Q-Learning auto-scaling with faster convergence, reduced state space and reduced action space in a distributed cloud environment. We demonstrate the effectiveness of our methods both on simulations and on real applications.