| With the rapid development of information technology,the data volume in many fields have grown from TB level to PB level and it is still with the exponential growth rate,which causes that the data information and the complexity of data are under rapid expansion.In order to get more information of data,many new technologies of processing big data have been proposed,for example,cloud computing which as a large-scale distributed frame has been widely used,it leverages the virtualization technology to integrate the resources of computing,storage and network,and provides a novel paradigm for the provision of resources on-demand.Meanwhile,the parallel computing model provides an effective parallel distributed strategy for big data applications,and shields the complex details such as fault tolerance,data distribution,load balancing,which makes the application can be efficiently deployed on cloud nodes.Big data,cloud computing and parallel computing model have formed a complementary relationship.The performance indicators of cloud computing center can not only guide the users to balance the relationship between the time cost of processing the application and the number of cloud resources,and can guide the cloud provider to adjust the cloud configuration to guarantee Qo S with the minimum resources.Therefore,how to analyze the performance of cloud computing center for big data applications has been an important research issue.In this thesis,the main contributions of this thesis are as follows:(1)A single parallel computing model is mostly suitable for a type of big data application,so the diversity of application can be smoothly transformed into the diversity of parallel computing models.Meanwhile,the cloud computing center provides the infrastructure for parallel computing models,therefore,the complexity of cloud resources can be transformed into the complexity of resources which supports to deploy the parallel computing model.Based on the parallel computing model,this thesis presents a performance analysis strategy of cloud computing center for big data applications.(2)Our proposed analysis strategy is to simulate various parallel computing models into a queuing system,which leverages Markov chain to describe the execution behaviors of different parallel computing models in a unified specification for shielding the diversity of big data applications,and also utilizes probabilistic analysis to define the impact factors of performance.In addition,we proposed a compare model,which directly simulates the cloud computing center into a queuing system.We compares the usability and accuracy with the performance analysis strategy.(3)Existing performance models of parallel computing models are mostly used to predict the processing time of big data application,and the performance models of cloud centers are mainly used to predict the performance indicators.Our proposed queuing system of parallel computing models considered the above two factors,which can simultaneously predict the performance indicators of cloud centers and the time cost of processing big data application.(4)When dividing a big data application into multiple parallel subtasks,the number of subtasks is defined as a discrete random variable,and the value of which is determined by the parallel computation model.Meanwhile,we considered the following performance factors: the CPU and bandwidth competition among VMs which run on the same PM,the performance attenuation of PM when the number of VMs increasing,the interaction of intermediate data from different subtasks.(5)Based on the performance analysis strategy,we proposed a hybrid parallel computing model with performance predictable,which makes Map Reduce as the basic model and combines the BSP iterative idea.In addition,based on the service rate of cloud nodes,we optimize the program of data copy placement through the computing capacity ratio of cloud nodes.Meanwhile,we proposed a load-aware CPU scheduling with the CPU consumption value and allocation value. |
PDF Full Text Request