Scale Scheme For Energy Consumption Optimization In Cloud

Since cloud computing is advanced and convenient, it is more and more widelyusedinindustry. However,itbringsanotherimportantproblem,thefastgrowingenergyconsumption.Serversindatacentersareconsumingtremendouselectricenergynowadays. How-ever, most servers are in idle state most of the time. If we decrease energy consump-tion, the performance of server will also decrease, which is not the customer desirable.Usually, customer will sign Service Level Agreement (SLA) with data center provider,which sets the minimum performance of server. To lower energy consumption of datacenters and satisfy SLA requirement, turning of server when they are in idle state isintroduced. However, it takes time to turn servers back on. During the interval, theresponse time of servers could be prolonged that the SLA requirement is violated.Inthispaper,wefrststudypreviousstudyaboutenergymodel. Weconductexper-iments and get an energy model of a single server which is related with CPU frequencyand utilization. Then we use queue model to study the relationship of CPU frequency,workload arrive rate and power consumption in a server cluster. We fnd there existsan optimal frequency in cluster which makes the power consumption minimum in aspecifc workload, meanwhile SLA requirement satisfy.Withthe changeofworkload, dynamicturning on/ofserversandscalingfrequen-cy can reduce the whole cluster power consumption. Up to now, nobody has combinedthe two methods efectively. We will continue this research. Based on the study [1], wepresent the Enhanced AutoScale (EAS) technique, which brings optimal frequency toAutoScale. EAS combines turn on/of servers and dynamic voltage frequency scaling(DVFS). In EAS, DVFS is used to make up the performance loss when idle servers were turned of. Two kind of EAS, Centralized EAS (CEAS) and Distributed EAS(DEAS) are proposed. We also demonstrate the algorithm and implementation.Theexperimentresultsshowthattheproposedapproachescanreducetheresponsetime with little energy overhead. Our method has even50%more Performance-per-Watt (PPW) than AutoScale in some workload.