Chombo中的AMRPossion应用的DAG并行化

发布时间：2018-03-15 03:01

本文选题：结构化网格　切入点：任务图并行　出处：《山东科技大学》2017年硕士论文　论文类型：学位论文

【摘要】：在过去的十年中,计算机系统不再依靠提高CPU频率来提高系统性能,集成了多个简单内核的多核和众核CPU已成为近十年来的主流技术。如何充分利用多核、众核系统资源获得好的性能是高性能计算应用的关键。结构化网格应用是高性能计算领域的一个典型应用,随着多核和众核平台的广泛应用,如何将已有的代码转换到这些新的并行平台上是一个巨大的挑战。在基于多核、众核平台的结构化网格应用中,由于数据划分而带来的相邻节点间的通信会给整个并行计算系统的性能带来显著影响,分析通信过程并实现通信与计算的重叠可以提高系统的性能。本文在多核平台上,以Chombo中的AMRPossion算例为目标,研究任务图并行方法。首先,分析了结构化网格应用中相邻数据块通信的特点。然后,对迭代计算中相邻数据块之间的依赖关系进行分析,区分本地数据和远程数据,本地数据通过共享存储器实现通讯,而远程数据必须通过节点间的网络实现通讯,是系统性能瓶颈。因此,利用计算与通讯的重叠来掩盖远程数据通信造成的开销,从而提高系统性能。最后,完成了AMRPoisso算例的多进程的任务图并行化,并在自建的多核平台上进行了测试,测试数据表明:与MPI + OpenMP并行模型相比,基于并行任务图DAG+MPI的混合并行程序可以有效地提高性能。
[Abstract]:In the past decade, computer systems have not relied on increasing the frequency of CPU to improve the performance of the system. The multi-core and multi-core CPU integrated with many simple cores has become the mainstream technology in the past decade. The key to the application of high performance computing is to obtain good performance of multi-core system resources. Structured grid application is a typical application in the field of high performance computing, with the wide application of multi-core and multi-core platform. How to convert existing code to these new parallel platforms is a huge challenge. The communication between adjacent nodes due to data partitioning will have a significant impact on the performance of the whole parallel computing system. The performance of the system can be improved by analyzing the communication process and realizing the overlap of communication and computation. Taking the AMRPossion example in Chombo as the target, the parallel method of task diagram is studied. Firstly, the characteristics of adjacent data block communication in structured grid application are analyzed. Then, the dependency relationship between adjacent data blocks in iterative computation is analyzed. Distinguishing between local data and remote data, local data communicates through shared memory, and remote data must communicate through networks between nodes, which is the bottleneck of system performance. The overlap of computation and communication is used to cover up the overhead caused by remote data communication, so as to improve the performance of the system. Finally, the multi-process task diagram of AMRPoisso example is parallelized and tested on the self-built multi-core platform. The test data show that the hybrid parallel program based on parallel task graph DAG MPI can effectively improve the performance compared with the MPI OpenMP parallel model.
【学位授予单位】：山东科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP38

【相似文献】