Performance evaluation of virtual machine live migration for energy efficient large-scale computing

Abstract

Large-scale computing systems must overcome a number of di culties before they can be considered a long-term solution to information technology (IT) demands, including issues with power use and its green impact. Increasing the energy e ciency of largescale computing systems has long posed a challenge to researchers. Innovations in e cient energy use are needed that can lower energy costs and reduce the CO2 emissions associated with information and communications technology (ICT) equipment. For the purpose of facilitating energy e ciency in large-scale computing systems, virtual machine (VM) consolidation is among the key strategic approaches that can be employed. Virtual machine (VM) live migration has become an established technology used to consolidate virtualised workload onto a smaller number of physical machines, as a mechanism to reduce overall energy consumption. Nevertheless, it is important to acknowledge that the costs associated with VM live migration are not taken into account in the context of certain VM consolidation techniques. Organisations often exploit idle time on existing local computing infrastructure through High Throughput Computing (HTC) to perform the computation. More recently the same approach has been employed to make use of cloud resources in large-scale computation. To date, the impact of HTC scheduling policies within such environments has received limited attention in the literature as well as the trade-o between energy consumption and performance. Also, the bene ts of using virtualisation and live migration are not commonly applied in High Throughput Computing (HTC) environments. In this thesis, we illustrate through trace-driven simulation the trade-o between energy consumption and system performance for a number of HTC scheduling policies. Furthermore, the thesis demonstrates the way in which various workloads can a ect the time of VM live migration. We use a real experiment to explore the relation between various workload characteristics and the time of VM live migration. In order to understand what factors in uence live migration, we investigate three machine learning models to predict successful live migration using di erent training and evaluation - vii - sets drawn from our experimental data. Through this thesis, we explore how virtualisation and live migration can be employed in HTC environment and used as a fault-tolerance mechanism to reduce energy consumption and increase the utilisation of a single computer in a large computing infrastructure. We propose various migration policies and evaluate them through the use of our extensions to HTC-Sim simulation framework. Moreover, we compare the results between the policies as well as the system where migration is not considered. We demonstrate that our responsive migration could save approximately 75% of the system wasted energy due to job evictions by user interruptions where migration is not employed as a fault-tolerance mechanism.