Optimization Research On Bursty Workload Oriented Resource Configuration In Cloud

The cloud computing technology provides users with large-scale computing capabilities of relatively lower price,bringing convenience for the Internet age.Nowadays,the rapidly developing Internet and advancing communication technologies are counter-productive to the vast number of customers,thus rendering the number of their submitted requests more and more frequent.Any social hotspot can cause a sudden increase in network traffic at any time,resulting in an instantaneous increase in workload.Once a task burst surges,the scalability and quality of service of the cloud system is likely to be affected.Scholars in related fields have carried out a lot of research on workload prediction methods and resource reservation and deployment schemes: it is necessary to avoid cloud providers deploying too many resources and causing unnecessary waste of costs;while configuration of cloud resources should be conducted to prevent such accidents.On the basis of these,this paper proposes a bursty-workload-oriented cloud resource configuration architecture,which includes three modules: task characterization model,workload trend prediction model and resource reservation and configuration model.The task characterization model extracts the attributes of historical tasks and accurately locates the unique information of spiky tasks.The workload trend prediction model sets the binary-state prediction for each task cluster,especially the task cluster that is more prone to sudden change,and selects the appropriate prediction method to estimate the workload amount according to the tasks arriving rate in clusters;in the resource reservation and configuration model a prediction-based heuristic algorithm is developed.There are four contributions in this paper:(1)an overall dynamic optimization architecture for bursty workload is proposed,which has a running monitor to synchronize the data of each module,optimizing parameters in real time;(2)the clustering method for task characterization is improved,where Mahalanobis distance is used in k-means clustering method to eliminate the correlation between task attributes to achieve the purpose of highlighting suspected sudden tasks;(3)normal and bursty states of arriving tasks are distinguished,where the time series method ARIMA and the trend extrapolation method are adopted accordingly to predict the task fluctuations in a targeted manner,with a control knob that monitors the increment of tasks and triggers prediction methods alternation for different workload scenarios;(4)a prediction-based reservation scheme is employed,equipped with a heuristic algorithm for resource migration and configuration,which saves energy without compromising the scalability of cloud computing and improving its robustness and service efficiency against workload bursts.This paper uses the Google cloud traces for model validation experiments.By comparing with the existing scheduling algorithms,this paper finally obtains a prediction accuracy of up to 90.05%,a capacity overflow rate of less than 5%,and a task completion guarantee rate improved by 25.8%,resource utilization improved by 18.2% and total energy consumption savings reduced by 17.3%,which are better than all or part of other scheduling algorithms,further confirming the availability and effectiveness of the solution to workload burst proposed in this paper.