Energy-Aware Resource Management Framework for Overbooked Cloud Data Centers with SLA Assurance

Abstract

Resource overbooking-based approaches have been adopted as one of the key technique for reducing power consumption of data centers by minimizing the number of physical machines (PMs) that need to be kept active. They do so by assigning virtual machine (VM) requests to a PM in excess of the PM's supported capacity, with the anticipation that these assigned VMs will not fully utilize their requested resources and hence the aggregate amount of needed resources will not surpass the PM's capacity. However, although resource overbooking improves PM utilization, it may lead to PM overloads where the actual VMs' demanded amount of resources exceeds the PM's capacity, thus resulting in violating service- level agreements (SLAs) of some of the hosted VMs. In this paper, we propose an integrated resource allocation framework for data centers that minimizes the number of active PMs through dynamic VM placement while ensuring that SLAs of admitted VMs are not violated by reducing the number of PM overload occurrences through VM resource prediction, resource scaling, and VM migration. Experimental evaluations based on real Google traces show that our proposed framework outperforms existing techniques by incurring lesser overload events with no to very few SLA violations of admitted VMs while yielding energy consumptions similar to those achieved under existing approaches.