A Dynamic Resource Scheduling Method For Web Applications In Cloud Computing Environment

The Web based online interactive application is widely used in Cloud computing.And efficient dynamic resource scheduling algorithms for Web applications in Cloud computing have also been one of research hotspots.In this paper,a Spot instance and On-demand instance based heterogeneous cloud resource dynamic scheduling algorithm is designed for single-layer Web applications,which aims to minimize the resource rental cost while satisfying the Service Level Agreement.The existing fault-tolerant model provides a basic scheduling framework in heterogeneous Cloud computing.However,it takes no consideration of the stochastic Spot prices and can not guarantee Spot resource stability and the expected low cost in rental periods.And the required resources are overestimated by the margin based resource estimation method in the fault-tolerant model resulting in resource waste and high cost.In view of these problems,this paper develops some fault-tolerant based dynamic scheduling policies which take Spot price prediction,queuing model and feedback control theory into consideration.Firstly,the existing Spot price prediction models are introduced and an improved Hidden Markov Model with combined states is proposed;Secondly,the existing margin based and queuing model based resource estimation methods are summarized,and a queuing model based feedback control method is developed;Thirdly,a Spot price prediction and minimum rental interval limitation based fault-tolerant scheduling strategy is proposed and a queuing model and feedback control based fault-tolerant scheduling algorithm is also developed;Finally,based on the Elastic Sim platform,a Web application simulation platform is developed,and numerous experiments are conducted to evaluate the performance of the proposals.The details are as follows:(1)As for the poor performance of Hidden Markov Model with single state in longterm prediction,a Hidden Markov Model with combined states is developed.Single states are combined to build the initial prediction model,which improves the correlation between states and the after-effect of states.And then the initial model is optimized to predict the probability distributions of combined states and observations to help predict future Spot prices.Experimental results on multiple Spot types show that the proposed prediction model can improve the prediction accuracy to some extent.(2)Due to the lack of real-time correction of the queuing model based resource estimation method,a queuing model based feedback control method is proposed to make resource estimation.This method makes use of feedback control theory to correct the estimation error of queuing model.According to the deviation between actual delay and the target delay of Web tasks according to SLA,the control adjustment factor is calculated by the feedback controller,which helps correct the current task arrival rate.And then the queuing model is called to estimate required resources.Experimental results show that this method can effectively reduce the rental cost and control real delay within target range.(3)Aiming at the shortcomings of the fault-tolerant strategies,optimized dynamic scheduling strategies are proposed.A Spot price prediction and minimum rental interval limitation based fault-tolerant scheduling algorithm is proposed,which selects Spot types by the predicted costs in rental intervals to improve the Spot resource stability.And the minimum rental interval limitation avoids frequent updating of rental plans;What is more,a queuing model and feedback control based fault-tolerant scheduling algorithm is proposed to reduce the resource estimation error and guarantee the average response time.(4)Based on the Elastic Sim platform,a Web application simulation platform is developed.And numerous experiments are conducted based on the real Wikipedia access data and Spot price history from Amazon Elastic Cloud Computing to evaluate the performance of the proposals in this paper.