Applying the Shared-Memory Parallel Programming Model to the Highly Parallel Cluster Computing Environment

为了使它们在高度并行的集群计算环境中保持一致，本文采用共享内存并行编程模式，一方面是为了降低并行编程成本，另一方面是为了获得足够的并行执行效率。首先，我们建立了一个由8个节点组成的集群系统作为平台，进行各种实验。并对OpenMP和MPI进行了研究，前者是共享内存并行编程模型的代表，后者是目前最流行的消息传递库。然后，我们计划开发一个转换器，将OpneMP程序转换为MPI分布式内存并行程序。OpenMP和MPI模型的差异主要集中在每个数据属性上。换句话说，在前者中，线程之间有共享数据要共享，这成为减轻程序员负担的重要因素。另一方面，在MPI中，只有每个进程本地的私有数据，如何管理这种差异是最大的问题。我们还研究了UPC作为共享内存并行编程模型的另一个例子，但是，包括默认数据属性是私有的，得出的结论是，从标准化程度和普及程度来看，UPC无法超越OpenMP。我们使用本研究开发的翻译器将几种OpenMP示例程序转换为MPI程序，并评估了它们的并行执行效率。因此，对于具有高局部性的程序，我们可以确认我们的转换器在并行执行效率上表现出足够的性能，甚至高于分布式共享内存系统的虚拟实现方法。

The aim of this study was to make them consistent in highly parallel cluster computing environments, one of which is to reduce parallel programming cost, and the other is to get enough efficiency in parallel execution, by employing shared-memory parallel programming paradigm.At first, we built a cluster system consisting of eight nodes as a platform to carry out various experiments. And we studied OpenMP and MPI, the former is the representative of the shared-memory parallel programming model, while the latter is the most popular message passing library. We then planned to develop a translator which converts an OpneMP program to an MPI distributed-memory parallel program. The difference of OpenMP and MPI models is focused to each data attribute. In other words, in the former, there are shared data to be shared between threads, and that becomes the big factor to reduce the programmer's burden. On the other hand, in MPI, there are only private data which are local to each process, and it was the greatest problem how to manage this difference. We also examined the UPC as another example of the shared-memory parallel programming models, but, including the default data attribute being private, reached the conclusion that the UPC cannot surpass OpenMP from a point of the degree of standardization and the popularity.We used the translator, which we developed in this study, to convert several kinds of OpenMP sample programs into MPI programs and evaluated their parallel execution efficiency. As a result, for the programs which have a high level locality, we could confirm that our translator showed enough performance on the parallel execution efficiency that was even higher than the virtual implementation methods of distributed-shared-memory systems.