Research for a tuning supporting system for parallelization using informations over a network

在‘97财政年度，我们研究了QCD(量子颜色动力学)、量子化学(分子轨道)、CFD(数值流体动力学)、油藏、地震波分析等几个应用程序，发现用户自己的应用程序并行化的领域太少了。这是对自动并行化潜在巨大需求的经验证明。因此，从今年开始，我们开始设计一个子系统，可以根据数据引用关系对给定的源程序进行分析，并三维显示给定程序的行为。并对系统的功能和用户界面进行了改进。也就是说，改进后的系统可以展示程序结构、数据依赖关系和对应的源程序片段，将这三个查看功能结合在一起，建立了更紧密的联系。用户可以很容易地、任意地移动到程序的任何部分，并在调试时选择程序的任何部分进行并行化研究。在‘99财政年度，也就是本研究项目的最后一年，我们继续工作，并提出了一种用于自动并行化编译器的UIR(通用中间表示)。此外，我们还设计了UIR功能的可视化算法，并在原型系统中实现了该算法。现在，我们的可视化系统给出了一些对给定程序的数据进行划分的提示，这被认为是并行化中最困难和最重要的问题之一。

In the '97 financial year, we investigated several application programs, such as QCD (quantum color dynamics), quantum chemical (molecular orbit), CFD (Numerical Fluid dynamics), Oil reservoir, seismic wave analysis, etc. and found that there is too few fields for the users to parallelize their own application programs by themselves. This fact is an empirical proof of the potential huge needs for automatic parallelization.Therefore from this year, we set to design a subsystem which can analyze a given source program and display three-dimensionally the behaviors of the given program according to the data referencing relationship.In the'98 financial year, we designed and implemented each module of this system. And we improved the functionality and user interface of this system. Namely the improved system can show the program structure, data dependence relationships and the corresponding source program fragment in such a way that these three viewing functions are combined and built up a closer connection. The users can easily and arbitrarily move to and choose any part of their program to investigate for parallelization on tuning up.In the '99 financial year, the last year of this research project, we continued the work and could propose a UIR (universal intermediate representation) for automatic parallelizing compilers. Moreover, we designed the algorithm for visualing function for UIR and implemented it in our prototype system. Now our visualization system show some hints to partition the data of the given program which is recognized as one of the most difficult and important problems on parallelization.