Model Predictive Control for Energy-Efficient, Quality-Aware, and Secure Virtual Machine Placement

Abstract

Listen

Translate article

Modern datacenters rely on virtualization to deliver complex and scalable cloud services. To avoid inflating costs or reducing the perceived service level, suitable resource optimization techniques are needed. Placement can be used to prevent inefficient maps between virtual and physical machines. In this perspective, we propose a holistic placement framework considering conflicting performance metrics, such as the service level delivered by the cloud, the energetic footprint, hardware or software outages, and security policies. Unfortunately, computing the best placement strategies is nontrivial, as it requires the ability to trade among several goals, possibly in a real-time manner. Therefore, we approach the problem via model predictive control to devise optimal maps between virtual and physical machines. Results show the effectiveness of our technique in comparison with classical heuristics. Note to Practitioners —This paper is motivated by the success of cloud-based services. Specifically, we consider the case where the resources of datacenters are accessed through the infrastructure-as-a-service paradigm. To efficiently handle the load of requests and to tame operational costs, a proper optimization is needed. To this end, we focus on the selection of the best maps between virtual and physical machines. We propose a holistic placement framework for the deployment of user-requested virtual machines by pursuing different performance goals, such as counteract to hardware outages or reboots due to software aging issues, ensure proper security policies, maintain a suitable service level perceived by users, and reduce power requirements. The optimal strategies are computed with model predictive control, which allows to consider complex constraints and take advantage of future information. The proposed framework is tested in different scenarios characterized by a variety of workloads and traces gathered in a real cloud datacenter. Results indicate that our approach outperforms bin-packing techniques.