Development of Integrated System Software for Parallel Environment with Flexibility

研究目标是研究和开发具有柔性的先进信息处理集成系统软件，推进先进信息处理的发展。以下是本年度的研究成果。(1)多处理器操作系统K1原型。我们基于消息池开发了多处理器操作系统K1。消息池提供了操作系统内核和系统服务器任务之间的通信机制。通过这种机制，可以自然地实现操作系统负载的分配。消息池还支持多播通信。我们评估了在真实的多处理器工作站上实现的K1的基本性能。评估结果表明，K1提供的组播通信机制是有用的。(2)并行编译系统的研究。在大型多处理器系统中，数据分区和分配对系统性能的影响很大，因此我们对数据分区和分配进行了研究。提出了数据分区和分配的中间表达式DPG (Data Partitioning Graph)，并在DPG的基础上设计了数据分区算法CDP^2。我们还开发了包含线性依赖向量的循环的并行化方法。传统方法很难实现循环的并行化。我们通过标准化并行编译器内部的中间表达式，并使用面向对象的方法重新编写并行编译器，使现有并行编译器的结构清晰易懂。针对下一代并行化编译器，采用面向对象的方法研究了编译器系统的构造方法，并在多处理器系统上实现了编译器的原型。

The objectives of the research are study and development of the integrated system software for advanced information processing with flexibility, and progression of advanced information processing. The followings are results of the research in this year.(1) Prototyping the Multiprocessor Operating System K1.We developed the multiprocessor operating system K1 based on a message pool. The message pool provides a communication mechanism among tasks in OS kernel and system servers. By this mechanism, distribution of OS's load can be realized naturally. The message pool also supports multicast communication. We evaluated basic performance of K1 implemented on the real multiprocessor workstation. Results of evaluations give us the multicast communication mechanism provided by K1 is useful.(2) Studies of Parallelizing Compiler Systems.We studied on data partitioning and allocation because they affect system performance greatly on large scale multiprocessor systems. We proposed intermediate expression form DPG (Data Partitioning Graph) which is base of data partitioning and allocation, and devised data partitioning algorithm CDP^2 on DPG.We, also, developed the parallelizing method in loops including linear dependency vector. Parallelizing of the loops is difficult for traditional methods. We made the structure of the existing parallelizing compiler clearly understandable by standardizing internal intermediate expression in it and re-writing it using object oriented methods. For the next generation of parallelizing compilers, we studied the constructing method of compiler systems using object oriented methods, and implemented the prototype of the compiler on the multiprocessor system.