Power-aware meta scheduler with non-linear workload prediction for adaptive virtual machine provisioning

Abstract

Infrastructure cloud typically involves provisioning of dynamically scalable and virtualized resources to cloud users. It is a fact that the resource demand in the cloud is highly dynamic in nature. To meet the dynamic demand from the cloud consumers, over-provisioning of resources is the common solution. This ultimately increases power consumption when the demand is normal or drops below average. On the contrary, under-provisioning of resources may lead to Service Level Agreement (SLA) violations. To balance between power conservation and performance issues, it has been realized that forecasting the demand for computing resources is essential in cloud environment. Hence we proposed a prediction based adaptive resource provisioning methodology incorporating statistical predictor in our earlier work. In order to improve prediction accuracy, we have proposed a recurrent neural network called Non-linear Auto Regressive network with eXogenous input (NARX) based prediction in this paper. The proposed NARX predictor makes near-accurate run time estimate of resource demand as it is able to learn hidden patterns and trends in the historical data representing resource demand and hence it helps to realize better power conservation. This paper shows that the proposed predictor integrated with adaptive provisioning resulted in 27.25 % more power saving compared to its statistical counterpart. © 2014 Springer International Publishing Switzerland.