Study on Case-Based Parallel Programming for Multi-threads

我们之前基于案例的并行编程研究的主要问题是，与线程和同步相比，任务划分的可重用性较低。在本研究中，基于并行算法类对任务划分进行了详细的分类，并对任务划分的组成部分进行了研究。通过分析针对每个并行算法类的情况开发的22个并行程序来分类任务划分。除了前面的三个类（block, cyclic, copy: level 1）之外，还提供了显示简单/重复，固定/变量，除法数等更详细信息的level 2，以及在level 2基础上显示必要数据的level 3。通过引入这些分类，可以提高病例检索的精度，从而可以检索到更多相似的病例。此外，一个自动检测程序，删除不必要的行，除了程序框架，并插入注释，显示什么应该被描述。因此，案例适应的工作量大大减少。此外，还通过以太网将16台PC连接在一起，形成了PC集群。在集群上安装了PVM，并进行了PVM并行编程实验。目前使用8台pc连接快速网卡、Myrinet，安装集群系统软件SCore开发分布式共享内存环境。简单的OpenMP并行程序在4台PC集群上使用Myrinet和SCore正确执行。

The major problem of our prevoius case-based parallel programming research was that the reusability of task division was low compared to threads and synchronization. In this research, task division was classified in detail based on parallel algorithm classes, and components of task division was investigated.Task division was classified by analysing 22 parallel programs developed for cases at every parallel algorithm class. In addition to previous three classes (block, cyclic and copy : level 1), level 2 which shows more minute information such as simple/duplicate, fixed/variable, and the number of division, as well as level 3 which shows necessary data based on level 2 were provided.The precision of case retreival increased so that more similar cases could be retrieved by introducing these classifications. Moreover, an automatic detection program that deletes unnecessary lines except program skelteons, and inserts comments which show what should be described. Consequently, the labor of case adaptation greatly reduced.Futhermore, a PC cluster was developed by connecting 16 PCs with ethernet. PVM was installed on the cluster and an experiment of PVM parallel programming was tested. At present, 8 PCs are connected with fast network cards, Myrinet, and distributed shared memory environment is developed by installing cluster system software SCore. Simple OpenMP parallel programs were executed correctly on the 4 PC cluster with Myrinet and SCore.