Research Of Cloud Workflow Scheduling Based On Time Series Prediction

Recent years have witnessed the prosperity of the cloud computing paradigm,and nowadays,scientific workflows could achieve higher cost-effective execution empowered by the elastic,on-demand resource provision scheme and pay-as-you-go pricing model of advanced cloud computing technologies.Scientific computing tasks are usually orchestrated as workflows to be executed by the cloud platform,however,the increasingly scientific problem size and high demand on cost-effectiveness make the scientific workflow scheduling on Infrastructure-as-a-Service(Iaa S)cloud a hot issue and attracts a lot of research interest.Most of the existing studies are based on the assumption that the performance of Virtual Machines(VMs)are static and time-invariant,and thus ignore the fluctuation and instability of cloud performance and transmission delay,especially when faced with peak hours that resource pool changes frequently.Therefore,targeting the deadline-constrained cost-effective workflow scheduling problem,instead of assuming the invariant cloud performance,this thesis considers the fluctuant performance of cloud VMs and its corresponding data transmission delay.It also leverages the time series prediction technologies and presents a predictive Krill-Herd-based scheduling approach.To overcome the drawbacks of cloud VM performance and transmission delay evaluation of current studies,this thesis builds a time-varying performance-based predictive cloud workflow scheduling model.In this model,real-time performance of cloud VMs and transmission are captured and predicted by an Autoregressive-Integrated-Moving-Average(ARIMA)approach,then it takes the predictive VM performance and transmission delay information as input to the Krill-Herd-based scheduling algorithm to yield cost-effective schedules under the user-defined deadline constraint.A series of case studies based on the performance and transmission delay of real-world public Iaa S cloud providers(e.g.,Tencent Cloud,Huawei Cloud,and Amazon EC2)is conducted to evaluate the effectiveness and efficiency of the proposed approach.Compared with the traditional studies,which consider static and time-invariant VM performance and transmission delay,the proposed ARIMA-based predictive Krill-Herd scheduling algorithm could achieve a better performance in terms of both completion time and total cost.Besides,the proposed approach could also achieve a more stable performance on completion time and lower violation rate of Service-Level-Agreement(SLA)than those studies which only consider predictive VM performance.