Research On Container-based Workload Forecasting Model And Energy-efficient Scheduling

With the constant development of container technology in recent years,following IaaS,PaaS,and SaaS,a new mode called Container as a Service(CaaS)has issued in the field of cloud computing.CaaS refines the granularity of resource partitioning and further expands the scale of cloud computing,but on the other hand,it leads to the increasing energy consumption of cloud data centers.There is a close relationship between the load of the container and the energy it generates.Therefore,we can optimize the overall energy consumption of the system by the dynamic configuration and scheduling of resources according to the load status of the container.However,due to the frequent changes in the load state of the container in the cloud,the load value obtained from the monitoring software may have a large hysteresis.So,we need a model which can predict the load state of the container.In addition,due to the large scale of the container and the necessary consideration of both container and virtual machine during scheduling,so the solution for energy conservation in CaaS is more complicated.In summary,how to accurately predict the load of container and effectively schedule the container according to the predicted result is an urgent problem to be solved.Aiming to solve this problem,the main contributions of this paper are made as follows:(1)This paper investigates three kinds of single-valued prediction models widely used in the field of load forecasting,but these models have shortcomings such as error sensitivity and poor robustness,which is not conducive to scheduling decisions.Therefore,based on the traditional interval prediction model,we propose a trend-aware interval prediction model(SAC-GPSO-SVM)to predict the load of the container.The model classifies the load of different trends by spectral feature and autocorrelation coefficient analysis(SAC),and uses SVM method to make targeted prediction.In order to improve the prediction effect,the model also introduces particle swarm optimization(GPSO)with gradient information to optimize the hyperparameters in SVM.Results of extensive experiments on the public dataset show that SAC-GPSO-SVM can effectively narrow the prediction interval while improving the prediction coverage.(2)In order to ensure the quality of service and reduce the energy consumption of cloud data center,this paper proposes a container scheduling strategy based on dynamic scaling(DSCS).This strategy firstly increases or decreases the number of containers according to the scaling rule and the real-time prediction results of the container load from SAC-GPSO-SVM,which will ensure that all services are running with the minimum number of containers as much as possible.Then,through the container selection algorithm(MCor-ML)and host selection algorithm(CorHS)based on load correlation analysis in the strategy,the newly added container is effectively placed,and migrate/merging process of the container on the overload/underload host is realized at a minimum cost.Finally,by dynamically detecting system state,the strategy will destroy virtual machines without containers and put idle hosts into hibernation to reduce resource waste.Results of experiments on the CloudSim platform,which extends the function for container emulation,compared with other strategies,while ensuring the quality of cloud services(SLA violation rate is less than 5%),DSCS can achieve better energy saving effect,which is increased by 7.41%.