Robust approach for host-overload detection based on dynamic safety parameter

Abstract

Host-overloading detection is an important phase in the dynamic Virtual Machines (VMs) consolidation process. Using machine learning to predict the future workload on a host, is a very promising technique to avoid the overload host situation. In this work, we propose a novel approach for overloaded hosts detection, based on neural network and Markov model. The neural network is trained on a workload data set composed of VMs CPU-utilization history. The trained model is then used to predict the future usage for a given Physical Machine(PM), by summing up the predicted utilization of all its VMs. The confidence of this prediction is measured through a dynamic safety parameter, based on Markov model. The obtained results show that our approach outperforms the state of the art algorithms such as: MAD, IQR and LRR.