Open MP Compiler with Task Scheduling Function

针对OpenMP程序在分布存储并行机上的并行执行和任务调度问题，提出了数据到达条件的新概念。粗粒度任务并行处理方案采用“执行开始条件”来表示粗粒度任务之间的并行性，通过将任务动态分配到处理器上来实现宏数据流处理。在诸如PC集群的分布式存储系统上实现宏数据流处理需要用于在分配给不同处理器的任务之间进行显式数据传输的新功能。作为函数之一，根据运行时确定的使用-定义链生成发送方-接收方对是至关重要的；然而，执行开始条件不包含该函数所需的信息。提出了数据到达条件的概念，通过扩展传统的数据到达定义，建立了数据传输的发送方和接收方对，并提出了通过“数据预加载”来减少数据传输开销的方法，以及在“软件分布式共享内存”上实现宏数据流的方法。在实际集群系统上的实现和评估表明，利用“数据到达条件”实现的宏数据流在分布式存储系统上是有效的。这些研究结果已经在IPSJ Transaction、JSPP、IPSJ SIG Note上发布。

The objective of this research is to develop the methods of the parallel execution and the task scheduling for OpenMP programs on the distributed memory parallel computers.In this research, the novel concept "data reaching condition" was developed. The coarse grain task parallel processing scheme using "execution start condition", which represents the parallelisms among coarse grain tasks, realizes a macro-dataflow processing by dynamic-assignment of tasks onto processors. An implementation of the macro-dataflow processing on a distributed memory system, such as PC-cluster, requires new functions for explicit data transfer between tasks assigned to distinct processors. As one of the functions, it is vital to make a sender-receiver pair based on the use-definition chain determined at run-time ; however, the execution start condition has no information needed for the function. The concept "data reaching conditions" is used to make a sender-receiver pair of a data transfer, by extending the concept of the conventional reaching definition.Moreover, the method to reduce the data-transfer overhead by "data preload" and the method to implement the macro-dataflow on "software distributed shared memory" were developed.The implementation and evaluation on the real cluster system shows the effectiveness of the macro-dataflow on the distributed memory system by using "data reaching condition".These results of this research have been published on the IPSJ Transaction, JSPP, IPSJ SIG Note.