MPI并行作业在云计算平台上的测试研究

Research on the Testing of MPI Parallel Computing Tasks Based on Cloud Computing Platform

随着并行空间计算任务的不断增多,传统的MPI服务器集群容易出现排队时间长、拒绝服务甚至系统瘫痪等情况。利用虚拟化、作业调度等技术构建云计算平台上的MPI虚拟集群可以提升MPI服务性能,从整体上缩短排队等待的时间,从而使得服务质量QoS(Quality of Service)得到保证。通过在OpenStack上部署MPI虚拟服务器集群、利用Torque实现MPI作业的调度管理的基础上,使用DEM(Digital Elevation Model)等高线生成算法组成MPI作业队列,对传统MPI物理集群与MPI虚拟集群进行性能对比分析,结果显示了云计算平台上MPI并行环境在面对大量任务作业时的优势。

With the number of the MPI tasks scheduled in the queue rising steadily, to certain tasks, such phenomena will be easi.ly to appear, e.g., waiting long time, being delayed or even denied to service. Utilizing the virtualization, job scheduling and relat.ed technologies in the field of cloud computing can enhance the QoS(Quality of Service) in the MPI server cluster, and the men.tioned problem would be solved to a certain degree. In the experiments, some Geo-spatial MPI parallel tasks, the parallel contour generation algorithm from DEM with MPI, are scheduled by Torque PBS to the different computing platforms respectively, i.e.,the combined platform with cloud computing(OpenStack) and parallel computing, and the traditional parallel computing plat.form. From the simulation tests results, it can be easily found that the combined computing environment has dominant advantag.es than the traditional finally.