Inciting Cloud Virtual Machine Reallocation With Supervised Machine Learning and Time Series Forecasts

To meet the demand for resource oversubscription in cloud computing environments, future Infrastructure-as-a-Service (Iaas) cloud platforms need to provide extremely high consolidation levels of idle or underused virtual machines, while responding quickly to changes or spikes in virtual machine (VM) workloads. In this paper, we present a survey of virtual machine based placement automation solutions for cloud computing while discussing their applicability to the IaaS problem domain. We then proceed to discuss the design, implementation and evaluation of a prototype solution to the IaaS dynamic VM orchestration problem as designed with the survey results in mind. The novel VM orchestration solution we present is predicated on a first of a kind, automated, black-box, hypervisor-based, runtime over-consolidation detection framework that makes use of supervised machine learning techniques. The machine learning techniques are coupled to a hierarchical, light-weight, VM resource forecasting technique. This forecasting is based on time series analysis and a neural network which operates on blackbox virtual machine observations to allow anticipatory remediation for resource contention. Our approach, dubbed “Incite” requires no alterations to the VM, has virtually no runtime overhead as measured on KVM running on a modern hardware system, and is therefore applicable in Infrastructure-as-a-Service clouds where providers generally have no control over customer VMs. Keywords; v i r tual machines , load balancing, overconstraint detection, cloud computing, IaaS, machine learning, support vector machines, decision tree classifiers, neural network. I. INTRODUCTION AND MOTIVATION 
 The concepts of virtual machine migration, and grid-like virtual machine computing have existed for some time, yet the state of the art in VM dynamic placement is not sufficient to meet the needs of certain types of cloud computing data centers (where the author asserts that the perfected application of said dynamic placement technologies seems most promising). 
 A subset of dynamic virtual placement research that relates strongly to live virtual machine migration is optimal consolidation, otherwise referred to as VM-packing. This activity is an attempt to increase resource utilization and simultaneously reduce capital expenditure. For example, a data center will often attempt to dynamically maximize consolidation ratios, generally expressed as the average ratio of virtual machines to hosts, in order to reduce power consumption by powering off unused physical hosts during times of decreased computational demand. 
 Cloud computing environments span many physical hosts, and typically VMs are not tightly packed after lengthy periods of deployment. A cloud provider may ask, “Does it make conceptual sense to defragment a cloud?” From a power consumption and resource utilization point of view, the answer is certainly yes [POWER]. However, from a throughput perspective, the answer is not so clear. It all depends on whether one is overtaxing the hosts by over-consolidating one’s virtual machines or otherwise starving virtual machines for resources under