From Data Center Resource Allocation to Control Theory and Back

Abstract

Continuously adjusting the horizontal scaling of applications hosted by data centers appears as a good candidate to automatic control approaches allocating resources in closed-loop given their current workload. Despite several attempts, real applications of these techniques in cloud computing infrastructures face some difficulties. Some of them essentially turn back to the core concepts of automatic control: controllability, inertia of the controlled system, gain and stability. In this paper, considering our recent work to build a management framework dedicated to automatic resource allocation in virtualized applications, we attempt to identify from experiments the sources of instabilities in the controlled systems. As examples, we analyze two types of policies: threshold-based and reinforcement learning techniques to dynamically scale resources. The experiments show that both approaches are tricky and that trying to implement a controller without looking at the way the controlled system reacts to actions, both in time and in amplitude, is doomed to fail. We discuss both lessons learned from the experiments in terms of simple yet key points to build good resource management policies, and longer term issues on which we are currently working to manage contracts and reinforcement learning efficiently in cloud controllers.