An Energy-Aware Real-Time VM-Provisioning Framework For Heterogeneous Data Centres

Increasing power efficiency is one of the most important operational factors for any data centre providers. In this context, one of the most useful approaches is to reduce the number of utilized Physical Machines (PMs) through optimal distribution and re-allocation of Virtual Machines (VMs) without affecting the Quality of Service (QoS). Dynamic VMs provisioning makes use of monitoring tools, historical data, prediction techniques, as well as placement algorithms to improve VMs allocation and migration. Consequently, the efficiency of the data centre energy consumption increases.

In this thesis, we propose an efficient real-time dynamic provisioning framework to reduce energy in heterogeneous data centres. This framework consists of an efficient workload preprocessing, systematic VMs clustering, a multivariate prediction, and an optimal Virtual Machine Placement (VMP) algorithm. Additionally, it takes into consideration VM and user behaviours along with the existing state of PMs. The proposed framework consists of a pipeline successive subsystems. These subsystems could be used separately or combined to improve accuracy, efficiency, and speed of workload clustering, prediction and provisioning purposes.

The pre-processing and clustering subsystems uses current state and historical workload data to create efficient VMs clusters. Efficient VMs clustering include less consumption resources, faster computing and improved accuracy. A modified multivariate Extreme Learning Machine (ELM)-based predictor is used to forecast the number of VMs in each cluster for the subsequent period. The prediction subsystem takes users’ behaviour into consideration to exclude unpredictable VMs requests.

The placement subsystem is a multi-objective placement algorithm based on a novel Machine Condition Index (MCI). MCI represents a group of weighted components that is inclusive of data centre network, PMs, storage, power system and facilities used in any data centre. In this study it will be used to measure the extent to which PM is deemed suitable for handling the new and/or consolidated VM in large scale heterogeneous data centres. It is an efficient tool for comparing server energy consumption used to augment the efficiency and manageability of data centre resources.

The proposed framework components separately are tested and evaluated with both synthetic and realistic data traces. Simulation results show that proposed subsystems can achieve efficient results as compared to existing algorithms.