Research On Key Issues Of HPC Application In Cloud

Cloud computing provide users with service of large-scale of computing and storing resources which is on-demand, pay-as-you-go and scalable in an instantly-available way. In contrast with Cloud computing, the limited resources in traditional computing model may not satisfy the demand of scientific applications to enormous resources in HPC (High Performance Computing). Moreover, there are some shortcomings in traditional way, such as high cost in investment and it is inconvenient in using resources, numerous computing resources are always unused in non-peak situations. Therefore, it is valuable to research complex scientific application in Cloud environment.Through migrate a FE-based3D Tooth Contact Analysis software "ZaKo3D" into Windows Azure, The thesis study several key issues to aim at HPC in Cloud, include application deployment, execution and job scheduling.(1) HPC application deployment model and parallel framework in CloudIt is difficult to move and deploy application in Cloud for scientists who is not familiar with complex computer configuration because of the differences and the complex configuration of the user interface in these Cloud platforms. In this thesis. we present a Cloud deployment model for HPC application and develop a case study which deploys the "ZaKo3D" application in the deployment model. It also establishes a framework to perform the HPC application in parallel in Cloud.(2) Performance and cost analysis of HPC application and applied scenarios in different computing modelIn contrast to traditional cluster, Cloud can save cost of investment. But the performance configurations of HPC platform also need to meet some requirements in application executions. On one hand costs will be increased in a high performance configuration, on the other hand low performance cannot afford HPC application. Hence, it is a valuable topic to study how to balance the relationship between the costs and performance configurations. This thesis investigate the total costs (including procurement, maintenance and manpower effort) of cluster and Cloud by developing cost-quantitative for Cloud platform and in-house cluster respectively, and evaluate performance and cost, it study the available rule for users to make the best decision to choose HPC platforms in rational combination of the price and performance within their capability.(3) Dynamic scheduling method in CloudScheduling a large amount of tasks in parallel on the Cloud nodes cannot always maintain the promised cost-efficiency due to the different workloads arising on these Cloud nodes, caused by some unforeseen situations. Generated overhead and load imbalances between nodes lead to numerous paid resources lay idle. This thesis proposes a dynamic parallel task scheduling method by employing a Master-Worker pattern in Cloud. The main idea of our work is that we schedule tasks on Cloud compute resources depending on the actual workload of each process instead of static-scheduled load. Dynamical scheduling method can efficiently decrease the overall runtime of MPI parallel computing. Moreover, in Cloud computing, it can improve Cloud resource utilization and deliver cost-efficiency.The researches of this thesis demonstrate that HPC in Cloud have better scalability. Cloud application deployment model provides users with the easy method to study HPC in Cloud. This study can give a reference to users to choice reasonable resources when develop their HPC applications according to requirements of application and themselves situation. The application execution in parallel in Cloud by dynamic scheduling method show better cost-efficiency than traditional method.