权益分类	功能权益	普通用户	{{item.name}}会员
{{category.name}}	{{benefitItem.name}}

スーパーコピュータ用自動並列化コンパイラに関する研究

超级计算机自动并行编译器研究

基本信息

批准号：
05452354
负责人：
NARITA Seinosuke
金额：
$ 4.54万
依托单位：
Waseda University
依托单位国家：
日本
项目类别：
Grant-in-Aid for General Scientific Research (B)
财政年份：
1993
资助国家：
日本
起止时间：
1993 至 1995
项目状态：
已结题

项目摘要

Multitasking and microtasking have long been used for the parallel processing of Fortran programs on shared-momory type multiprocessor systems. Howerver, multitasking has shortcomings such as the difficulty for the specification of parallelism by the user and the scheduling overhead due to OS calls. Micortasking has been widely used for loop parallelization, but there still remain many loops such as complicated data dependencies among iterations and conditional branches outsides loops, which cannot be paralyzed automatically. To resolve those problems, we have proposed a method for the processing of macro dataflows. In this scheme, the compiler automatically decompose a given program into coarse-grain tasks and detects parallelisms existing among the coarse grain task by analyzing the earliest executable conditions.By using the scheduling routines generated by the compiler, the scheduling overhead can be kept minimum. In macro datafow processing, due considerations are given to minimize data transfer overheads, and more efficient parallel processing can be achieved by localizing onto local memories. The efficiency of the proposed parallel processing of coarse grain tasks through macrodata flow processing was confirmed on the prototype multiprocessor system OSCAR.It was also tested on commercial multiprocessor systems such as a Fujitsu FPP0500, an Alliant FX-4, an KSR1 and an NEC Cenju-3. It has been found that macro dataflow processing can provide more parallelism than conventional multitasking and microtasking. It was also confirmed that lower overheads can be achieved compared with conventional methods, hence faster speed for program processing.

在共享内存型多处理器系统上，多任务处理和微任务处理一直被用于Fortran程序的并行处理。然而，多任务也有缺点，比如用户难以规范并行性，以及操作系统调用带来的调度开销。微任务处理在循环并行化中得到了广泛的应用，但仍然存在许多循环，如迭代之间复杂的数据依赖关系和循环外的条件分支等，无法自动瘫痪。为了解决这些问题，我们提出了一种处理宏数据流的方法。在该方案中，编译器自动将给定的程序分解为粗粒度任务，并通过分析最早的可执行条件来检测粗粒度任务之间是否存在并行性。通过使用编译器生成的调度例程，可以将调度开销保持在最低限度。在宏数据流处理中，适当考虑最小化数据传输开销，并且可以通过本地化到本地内存来实现更有效的并行处理。在原型多处理机系统OSCAR上验证了采用宏数据流处理方法并行处理粗粒度任务的效率。它还在富士通FPP0500、Alliant FX-4、KSR1和NEC Cenju-3等商用多处理器系统上进行了测试。研究发现，宏数据流处理比传统的多任务处理和微任务处理能提供更高的并行性。与传统方法相比，可以实现更低的开销，因此程序处理速度更快。