Study on parallelizing compiters with a granularity tuning mechanism

具有粒度调整机制的并行编译器研究

• 批准号: 10480057
• 负责人: YUBA Toshitsugu
• 金额: $ 8.9万
• 依托单位: The University of Electro-Communications
• 依托单位国家: 日本
• 项目类别: Grant-in-Aid for Scientific Research (B)
• 财政年份: 1998
• 资助国家: 日本
• 起止时间: 1998 至 2001
• 项目状态: 已结题
• 来源: https://kaken.nii.ac.jp/en/grant/KAKENHI-PROJECT-10480057/
• 关键词: parallel processing; parallel computers; task granularity; parallelizing compiler; granularity tuning; coarse-grain parallelism; SMP cluster; Parallel program; doacross

We aim to establish the fundamental technology of parallelizing compilers with a granularity tuning mechanism for efficient parallel processing. The parallelizing compiler generates a parallel object program, which will be executed in the shortest time on a parallel computer, by matching its hardware characteristics with parallel properties of a given application program. The main results are as follows :1. A new static parallelizing scheme is proposed, in which apart of a target sequential program is divided into parallel tasks with optimal granularity by using the LogP model as an abstract parallel machine. Do loops, do across loops and recursive function calls are applied for granularity tuning, and evaluation studies are carried out on some different parallel computers.An efficient execution mechanism for coarse-grain parallel processing is proposed for distributed memory parallel computers. The mechanism can be realized to transform a given sequential program into a coarse-grain task graph with execution start conditions as well as data reaching conditions as a function of parallelizing computers.3. A parallelizing compiler with a coarse-grain parallel processing function is experimentally constructed. The conventional Open MP compiler is utilized to translate the coarse-grain task graph with Open MP primitives into an executable parallel C program.4. A novel software distributed shared memory (SDSM) is proposed, which reduces memory consistency overhead by reflecting characteristics of application programs at a middleware level. The SDSM scheme is implemented in such parallel machine environment as a heterogeneous workstation cluster and an SMP-type personal computer cluster.

我们的目标是建立一个有效的并行处理的粒度调整机制的并行编译器的基本技术。并行化编译器通过将硬件特征与给定应用程序的并行特性相匹配，生成并行目标程序，该并行目标程序将在并行计算机上以最短的时间执行。主要研究结果如下：1.提出了一种新的静态并行化方案，该方案利用LogP模型作为抽象并行机，将目标顺序程序的一部分划分为具有最优粒度的并行任务。采用Do循环、do跨循环和递归函数调用进行粒度调整，并在不同的并行机上进行了评估研究，提出了一种适用于分布存储并行机的有效粗粒度并行处理执行机制。该机制可以实现将给定的顺序程序转换为具有执行开始条件以及数据到达条件的粗粒度任务图作为并行计算机的函数.实验构造了一个具有粗粒度并行处理功能的并行编译器。利用传统的Open MP编译器将带有Open MP原语的粗粒度任务图翻译成可执行的并行C程序.提出了一种新的软件分布式共享内存（SDSM），它通过在中间件层反映应用程序的特性来减少内存一致性开销。SDSM方案在异构工作站集群和SMP型个人计算机集群等并行机环境中实现。

项目成果

福岡岳穂, 本多弘樹, 弓場敏嗣: "OpenMPによる粗粒度タスク並列実行方式"情報処理学会研究報告 2000-HPC-82-12. Vol.2000,No.73. 65-70 (2000)

Takeho Fukuoka、Hiroki Honda、Toshitsugu Yumiba：“使用 OpenMP 的粗粒度任务并行执行方法”日本信息处理学会研究报告 2000-HPC-82-12，Vol.2000，No.73 (2000)。

Osawa Noritaka and Toshitsugu Yuba: "Lazy and Differential Replication in a Recoverable Distributed Shared Memory System, Lecture Notes in Computer Science"Proc. International Conf. on High-Performance Computing and Networking, Springer-Verlag. Vol.1401.

Osawa Noritaka 和 Toshitsugu Yuba：“可恢复分布式共享内存系统中的惰性和差异复制，计算机科学讲义”Proc。

高畠志泰, 大沢範高, 弓場敏嗣, 佐藤三久, 山口喜教: "細粒度並列アーキテクチャ用SISALコンパイラにおける並列粒度調整方式"情報処理学会論文誌. Vol.39,No.6. 1709-1717 (1998)

Shiyasu Takabatake、Noritaka Osawa、Toshitsugu Yumiba、Miku Sato、Yoshinori Yamaguchi：“用于细粒度并行架构的 SISAL 编译器中的并行粒度调整方法”，《日本信息处理学会会刊》第 39 期，第 6 卷，1709-。 1717 (1998)

Hideaki Hirayama, Hiroki Honda and Toshitsugu Yuba: "Distributed Shared Memory with Log Based Consistency for Operations with Commutative Law or Associative Law"Systems and Computers in Japan. Vol.32,No.8. 10-19 (2001)

Hideaki Hirayama、Hiroki Honda 和 Toshitsugu Yuba：“具有基于日志一致性的分布式共享内存，用于交换律或结合律运算”日本系统和计算机。

高畠.大沢.弓場 他: "細粒度並列アーキテクチャ用SISALコンパイラにおける並列粒度調整方式" 情報処理学会論文誌. Vol.39,No.6. 1709-1717 (1998)

Takabatake、Osawa、Yumiba 等人：“用于细粒度并行架构的 SISAL 编译器中的并行粒度调整方法”，日本信息处理学会汇刊，第 39 卷，第 6 期，1709-1717（1998 年）。