The right tool for the job: Switching data centre management strategies at runtime

TLDR: This work considers the goals of minimizing Service Level Agreement violations and minimizing power consumption, and proposes three methods of dynamically switching strategies, and evaluates these methods through simulation.

Abstract: Applications are shifting into large scale, virtualized data centres that provide resources on a pay-per-usage basis. Data centres must minimize resource consumption while providing enough resources to meet application requirements. To meet highly variable application demands, a dynamic approach to virtual machine (VM) management is required. This involves three basic management operations: (i) placing newly arrived VMs, (ii) migrating (moving) VMs off of highly utilized machines to avoid performance degradation, and (iii) migrating VMs off of underutilized machines so that they may be shut down to save power. We define a management strategy to consist of a set of policies that guide these three operations. We consider the goals of minimizing Service Level Agreement violations and minimizing power consumption. Developing a management strategy to achieve both of these goals is challenging, as the goals are often in conflict. We propose achieving both goals through dynamically switching between two management strategies, each with a single goal, depending on current data centre state. We propose three methods of dynamically switching strategies, and evaluate these methods through simulation. Dynamic strategy switching offers improved results over a single management strategy.