Research On Power Consumption Model And Energy-efficient Scheduling In Cloud Computing

As the demands for compute resources keep growing rapidly,the scale and number of cloud data centers are ever-increasing,bringing about a prominent issue of excessive energy consumption.Deploying dedicated metering devices is costly and also causes many difficulties when the monitoring system needs to expand.In contrast,software-based systems built on power consumption models enable fine-grained,low-cost and highly scalable monitoring even in large-scale,heterogeneous data centers,which is urgently demanded by cloud service providers(CSPs).With the support of real-time power monitoring system,applying energy-aware dynamic scheduling to cloud data centers to reduce energy consumption has been prevailing.Aiming at improving the energy efficiency of cloud computing,this dissertation mainly makes three contributions as follows:(1)The methodology for system-wise power estimation and a number of mainstream component-level power consumption models are presented.The dissertation introduces the theoretical foundations of power consumption models for CPU,memory and hard disk driver,and figures out their drawbacks.Furthermore,an I/O-mode aware disk power consumption model is proposed based on experimental observation on disk power behaviors,followed by analysis and evaluation of all the surveyed models in accuracy.The evaluation was conducted using public server power data set as well as experiments on physical machines.(2)A new method for building “black box” power consumption model for the baremetal machine is proposed.The method leverages temporal relationship between time-series power consumption data while adapting Elman Neural Network(ENN)to improve accuracy and enable incremental training with time-series data sets.A mixture of CPU-intensive and I/O-intensive workload data is collected to form the training data set in order to enhance the model's generalization ability.(3)The dissertation proposes a Peak Efficiency Aware virtual machine Scheduling(PEAS)strategy that leverages real-time power information from the monitoring module.The mathematical definition of a new metric-peak efficiency is given,based on which the optimal utilization and optimal resource offers(by physical machines)are derived.In this dissertation,compute resource is abstracted and quantified in Compute Resource Units(CRUs).For virtual machine(VM)placement,a peak efficiency aware VM placement(PEAP)algorithm is adapted.PEAP prioritizes physical machines by power efficiency and strives to keep them running under optimal utilization.For VM reallocation,a peak efficiency aware cost-efficient VM reallocation(PEACR)algorithm is adapted.PEACR detects overloaded and under-utilized hosts with dynamic thresholds and reallocates VMs at a low cost.Results of extensive experiments on extended CloudSim 3.0.3 show that PEAS effectively outperformed three typical scheduling strategies(i.e.,PABFD+MM,PABFD+MMT and TVRSM)in energy consumption reduction and in the meantime,guaranteed(or even improved)system performance measured in average job execution time,and promoted cloud computing's Quality of Service(Qo S)through preventing a high risk of SLA(Service Level Agreement)violations.