Research On The Resource Deployment And Task Scheduling In Cloud Computing

Cloud computing is regarded as the carrier of the next generation information technology.Data storage, processing, analysis, decision, tasks scheduling and resource management areoperated in the cloud; in other hand, there are many tasks to be dealt with, a huge amount of datato be transmitted, users with different needs and different resource types, so, the new and higherrequirements about the resources deployment and task scheduling are put forward: they not onlymeet the needs of users to obtain the high quality service, improve the resource utilization of thesystem as much as possible, ensure service providers to obtain the biggest benefit, but also are inhigh efficiency and energy saving from the perspective of sustainable development and greenenvironmental protection. Therefore, resource deployment and task scheduling is an importantissue in cloud computing research and implementation to be solved.This thesis targets data deployment, task scheduling, performance optimization, performanceenergy-aware resource management and other important research fields in cloud computing.Correspondingly, according to the types and characteristics of the resources, the needs of the userto obtain high performance experience and task scheduling, etc, the resource deployment and taskscheduling are researched about the applications of data-intensive, performance and cost-sensitive,performance and energy balance, respectly. The main contributions of this dissertation as follows:（1）Research on the data deployment and task scheduling based on associated amountin cloud computingIn data-intensive application, one hand, there is a large amount of data to be transmitted manytimes; on the other hand, network bandwidth is limited, therefore, the data transmission efficiencyis low. In this paper, firstly, the author proposed the dependency model based on the maximumamount of data dependency and used the model to form the affinity matrix, then that affinitymatrix via finite elementary transformation matrix can be converted to clustering matrix was beproved too. Secondly, designed the bond energy algorithm based on the dependency model whichclusters the largest relevant data together. Finally, designed the K partitioning algorithm based onthe maximum measured value model to partition the clustered matrix to K parts, then, accordingthe results of partitioning, each part was deployed to correlated datacenter by the task scheduler.The simulation results show that the proposed model and algorithms can reduce the times andamount of the transferring data between different datacenters and improve the performance of thesystems.(2) Task scheduling optimization based on particle swarm optimization in cloudcomputingOn the one hand, the different data centers have different charge standard to users, on theother hand, the bandwidth between different data centers is limited and the capacity of processingis various in different data center, so the different data deployment and task scheduling canobviously affect user fees and performance. In this paper, tasks scheduling are studied in multi-data centers environment to optimize the performance of the user experience and cost. In thestudying, tasks scheduling are mapped to task processing interacted graph. Based on the graph,mathematical model based on task scheduling was proposed; then, using the small position valuerule convers the continuous vectors to discrete vectors and designed the particle swarm algorithmto optimize the task scheduling and the user experience. In order to prevent the particle swarmoptimization algorithm traps in local optimum, then, designed the particle swarm optimizationalgorithm based on the variable neighborhood search algorithm. Finally, through a large number ofexperiments, the optimal benchmark value, degree and step was found. Simulation results showthat the proposed model and algorithm can obviously not only optimize the processing time,transferring time and processing cost and transferring cost, but also improve the processingperformance and the user experience.(3) Research on the optimizing method for performance and energy-aware in cloudcomputingFor the prevalence problem that the resource utilization lows and the waste of energyconsumption is very much, this paper studies the dynamic consolidation of virtual machine in datacenter environment to optimize the performance and energy consumption, improve thecomprehensive performance of the virtual data center and meet the customer service qualityrequirements. In this paper, firstly, studied the energy model of the computer, then improved theoriginal energy model and proposed an new energy model (Exponent Model); secondly, in light oflocal regression analysis to predict the utilization rate of the CPU and then determining the virtualmachines is overloaded or not and need migrate or not; thirdly, proposed the variable mean value,minimum utilization rate and the first quartile method to estimate that the computer is underloaded, then, determine the virtual machine to be shifted or not; fourthly, proposed the minimummigration time, maximum CPU utilization and minimum CPU utilization method to determinewhich virtual machine can be migrated. In the end, designed the energy aware best fit algorithm todeploy the virtual machines and then to optimize the energy consumption and performance.Simulation results show that the under load detection method of variable mean value, theoverloaded detection method of local regression analysis, the virtual machines selection method ofthe maximum CPU utilization of the computer and the energy aware best fit algorithm not onlyimprove the performance but also optimize the energy consumption of the system.(4) Research on the dynamic performance optimization based on queueing system incloud computingAs the service quality requirements and the arrival rate of customer are different, the numbersof servers is provided are also different in the cloud computing data center. In order to study theoptimal number of servers change rule with the customer service quality requirements and thearrival rate, firstly, this paper studied the queueing theory problem of multi-requesting customersand two service windows that are not equally in ability; by analyzed to prove the condition ofsteady-state solution existing and theoretically deduced, experimentally verified the specificexpression of the parametersLq,Ls,Wq andWs. Based on the above results, the mult-requestingcustomers and multi-windows services queueing theory is studied, then, designed the synthesizeoptimizing strategy, model and algorithm to optimize the performance of the mult-requestingcustomers and multi-windows services queueing theory, then simulated them; the simulation resultverifies that the proposed methods are better than classic method of first come first serve and ashort service priority queuing method; this method can significantly reduce customer service time mean waiting in line, waiting in the service of the average queue length, and be able to servicemore customers at the same time.