Research On High Performance Computing Cloud Platform Evaluation And Resource Management Optimization Technology

The IaaS cloud platform has characteristics of application for computing resources on-demand and customization of execution environment,which has attracted more and more organizations to build high-performance computing clusters based on cloud technology.Although the elastic resource usage model of the HPC cloud platform improves the resource use ratio as well as the flexibility of the system,the virtualization technology also introduces system overhead.Once the cloud platform is built unreasonably,HPC applications will suffer from large performance losses.To this end,it is necessary to delve into the construction of cloud platform that is suitable for HPC cloud platform and test and its performance comprehensively.In addition,resource optimization configuration is an important factor that restricts the performance of HPC cloud platforms,among which the virtual machine placement is a key factor.However,there is currently a lack of research on multicore-aware virtual machine placement multi-objective optimization strategies.This paper focuses on the design and performance testing of HPC cloud platform,and virtual machine optimization placement algorithm,mainly carried out the following work:Firstly,the domestic and international research of high-performance virtualization technology and HPC cloud platform are deeply analyzed,and the feasibility of HPC cloud model is demonstrated.By comparing the similarities and differences between the traditional cloud computing platform and the HPC cloud platform,the potential advantages and technical problems of HPC cloud are expounded.The features and components of the lightweight virtualization container technology Docker and the cloud managing systems OpenStack are deeply analyzed,and then an HPC cloud platform construction scheme based on Docker+OpenStack is demonstrated.Secondly,based on OpenStack and Docker,the NUDT_SCICloud platform is built and the performance is tested and verified.With the use of NAS parallel benchmark program and the Weather Research Forecast,the performance metrics such as integer and floating point calculation,I/O communication,and scientific calculation program execution efficiency are selected to compare the performance of the real physical cluster and the virtual machine cluster built on NUDT_SCICloud,which verifies the feasibility of a HPC cloud platform.The test results are analyzed in depth,and the additional performance overhead of virtualization for scientific applications is discussed.The problems and possible improvement directions of NUDT_SCICloud HPC cloud platform are pointed out.Finally,on account of the traditional virtual machine placement algorithm does not consider the problem that the physical servers have several cores,a multi-core aware virtual machine placement algorithm based on multi-objective optimization is proposed.Unlike the traditional placement algorithm,which only considers the CPU as a one-dimensional resource,the algorithm abstracts the multi-core CPU of a physical server into a multi-dimensional vector.Aiming at three optimization objectives:energy balance,SLA violation rate and resource utilization in HPC cloud environment,a multi-core aware virtual machine placement algorithm based on ant colony optimization is proposed.The results of CloudSim simulation experiments show that the proposed algorithm can balance the conflicting optimization objectives and obtain a better Pareto dominant solution than the traditional placement algorithm.