Demand-Driven Dynamic Resource Scheduling Framework In Clouds

With the development and maturity of cloud computing technology,more and more users begin to rent and make use of resources from clouds.Besides,a lot of enterprises start to provide cloud resources to users.We could abstract three main roles from complex cloud environments.They are cloud resource providers(CRPs),cloud application providers(CAPs)and cloud application users(CAUs).This paper aims at designing a demand-driven dynamic resource scheduling framework in clouds from perspectives of the above roles.This framework could automatically execute the provision activity,placement activity and consolidation activity for virtual resources in the process of resource scheduling in cloud.CAUs will put forward the demands on performance,availability,security and some other metrics to cloud applications.On the other hand,CAPs rent virtual resources from CRPs in order to provide computing capabilities support to cloud applications,so they hope resource scheduling framework could help to provide virtual resource for cloud applications elastically in precondition of meeting the QoS constraints to cloud applications and it will be helpful for them to save cost.Furthermore,placement of virtual machines(VMs)will influence the availability of cloud applications and the communication overheads among VMs.So CAPs expects that the resource scheduling framework could take above two factors into consideration when generating placement scheme.From the perspective of CRPs,they wish that the scheduling framework could optimize the resource utilization rate of physical infrastructures in precondition of making them neither in the state of overload nor underload by taking advantages of resource consolidation activity.This paper extracts the main requirements from above three roles and separate them from each other and regards these requirements as the constraints to resource provision,placement and consolidation activities in the resource scheduling process.Based on above information,we design a demand-driven dynamic resource scheduling framework and provide a group of implements for it.In terms of workload prediction,this paper proposes a model which combines the advantages of both time series prediction model and regularity based prediction model.It could predict the trend of workload well and overcome the time-lag effects showed in time series prediction model by means of the underlying regularity in workloads.This model covers peak workloads well.In terms of virtual resources provision,this paper implements a strategy which combines the advantages of both proactive way and reactive way.With the help of the above prediction model,this strategy could make adjustments before peak or valley workloads come.When the prediction result is inaccurate,reactive calculation unit will react fast and it will help the amount of virtual resources become acceptable.In terms of virtual resources placement,this paper implements an availability-aware and communication overhead optimized placement strategy.In order to make quantitative analysis for the availability of cloud applications and the communication overheads among VMs,we build models for calculating above two metrics.Based on models built by us,the placement strategy proposed by us could take both above two factors into account when generating placement scheme.In terms of virtual resources consolidation,this paper implements a proactive and consolidation logic group oriented virtual resources consolidation mechanism.This mechanism partitions the physical machines(PMs)in target computing center into a group of consolidation logic groups.It helps to limit the actuating range of the consolidation and limit the time complexity of calculating the consolidation scheme indirectly.We also make use of the idea of separation and separate the consolidation activity into three procedures.They are identification,selection and mapping.We solve them in sequence.In addition,we make consolidation for virtual resources according to the resource utilization rate of PMs and VMs by taking advantage of proactive idea.It helps to break the limit caused by time-lag effects in static consolidation activities.This paper makes validation for the resource scheduling framework proposed by us in the real cloud environment.Experimental results show the following information.Compared to other models,the workload prediction model implemented in the framework achieves lower average relative error and higher prediction coverage rate;the provision strategy implemented in the framework could provide virtual resources for the cloud application elastically while enforcing the service level agreements(SLAs)in most of periods;the placement strategy implemented in the framework helps the cloud application always meets the availability constraint during theexperiment while trying to decrease the communication overheads among VMs;the consolidation mechanism implemented in the framework optimizes the resource utilization rate of physical infrastructures and reduce PMs overload situations with the prerequisite that the computing center provides enough computing resource for VMs.Thus we have made exhaustive validation for the framework proposed in this paper.