Article Preview
TopIntroduction
Cloud computing, as an emerging technology, is featured by the ability of elastic provisioning of on-demand computing resources ranging from applications to storage over the Internet on a pay-per-use manner (Graubner et al. 2013). Cloud computing brings in numerous benefits for companies and end customers, e.g., end customers can invoke computational resources for almost all possible types of workload when resources are reachable and a large number of today’s Web services are deployed on the Cloud (Deng et al.2014), and published on the Internet (Gan et al. 2017).
Cloud data-centers are key enablers for the scalability of the cloud platform (Xia et al., 2015). However, inefficient resource utilization is a common problem in today’s cloud datacenters (Vogels, 2008) mainly because datacenters are usually designed and deployed to meet the peak work load but actually running at low load at most of the operational time. Datacenter owners thus have to afford huge cost for the investment into physical resources, which are often under-utilized. Both industry and academy are in a high need of methods to optimize energy efficiency of cloud datacenters. Recently, great efforts have been taken in this direction. The underlying requirement for energy-efficiency-optimization is to minimize energy consumption while meeting the performance promised to cloud users specified by the Service-Level-Agreement (SLA). Several techniques and methods, at both hardware and software level, are developed for the above-mentioned purpose. At hardware level, e.g., Dynamic Component Deactivation (DCD) and Dynamic Performance Scaling (DPS) are two such techniques, by which hardware components can enter the dormant state under the control of hardware timers. While at software level, the live migration techniques are developed to offload tasks and processes among different machines for load balancing, VM-consolidation, or risk-decentralization purposes.
Virtual machine consolidation moves virtual machines (VM) form under-utilized physical machines (PM) to better-utilized ones. Thus, PMs with non-load can be switched into dormant state and consumes less energy. Moreover, the remaining PMs alive can consequently take a more condensed workload and thus the cloud system achieves a high utilization rate. Unfortunately, VM consolidation itself consumes time, energy, and resources, especially when the destination PM is highly loaded. Performance and QoS degradation, which may further cause SLA violation, can often be observed when frequent migration activities are carried out. Hence, strategies must be designed in such a way that energy reduction and SLA commitment are both taken into account.
Note that various early contributions consider consolidation and migration plans to be independently carried out on PMs, i.e., every PM decides whether to migrate or receive VMs based on its own utilization rather than system utilization. Recently, some other works consider dependent consolidation plans but most of them assume invariant utilization rate of PMs in computing energy reduction of candidate migration plans, i.e., the gains of energy efficiency is calculated by comparing VM distributions before and after the consolidation plan is carried out but the utilization rate of every PM is assumed to be unchanged after the consolidation is accomplished. Such assumption may lead to considerable accuracy loss in calculating energy reduction of candidate consolidation plans and could further mislead the determination of optimal plan. In contrast, however, this work considers time-varying utilization rate of PMs and circumvent such accuracy loss by employing the number of PMs to be turned off within a consolidation period as the optimization objective instead of energy reduction.
In this study, a novel model is proposed to describe the performance-aware and SLA-constraint VM consolidation problem and develop a selection algorithm taking the predicted migration overhead (derived by the Pareto distribution) as inputs and another algorithm to generate the optimal matching plans based on preference scores of candidate VMs. For the model validation purpose, this study conducts a case study on the CloudSim simulation platform and show that our proposed algorithms achieve better energy saving efficiency.