Research on hierarchical macro-data-flow parallel processing on SMP-cluster

本研究的目的是在具有层次化硬件并行性的对称多处理器集群上，开发一种利用粗粒度任务间的并行性来实现层次化宏数据流并行处理的方案。为了在SMP集群上实现宏数据流并行处理，本研究的研究者提出了一种在运行时确定数据传输的发送方-接收方对的方法。本研究的创新之处在于在实际系统上用该方法实现了宏数据流，并将该方法扩展到层次化宏数据流并行处理上，并在SMP集群上实现了层次化宏数据流处理，并对该机制进行了性能评估。研究结果表明，基于SMP集群的层次化宏流并行处理方法实现了高效率的并行处理，扩展了基于软件分布式共享存储的宏数据流并行处理方法，并在IPSJ会刊、IEEE集群计算国际会议、高级计算系统与基础设施研讨会(SACSIS)和部分IPSJ SIG技术报告上发表了研究成果。

The objective of this research is developing the scheme to realize the hierarchical macro-data-flow parallel processing, which exploits the parallelism among coarse-grain tasks, on the SMP (symmetrical multi processor)-cluster which has hierarchical H/W parallelisms.In order to realize macro-dataflow parallel processing on SMP-cluster, "data reaching conditions", a method to make the sender-receiver pair of a data transfer determined at run time, has previously been proposed by the investigator of this research. The originality of this research lies in both implementing macro-dataflow with the method on real systems and extending the method to HIERARCHICAL macro-data flow parallel processing.The implementation and evaluation of the hierarchical macro-dataflow processing on SMP-cluster have been done on the SMP-cluster parched by this grant. The results of this research show that the method for hierarchical macro-dataflow parallel processing on SMP-cluster achieves the high efficient parallel processing.In addition, we have expanded the method to realize the macro-dataflow parallel processing on software distributed shared memory and showed the effectiveness of the method.The results of this research have been published on the IPSJ transaction, IEEE International Conference on Cluster Computing, Symposium on Advanced Computing Systems and Infrastructures(SACSIS) and some IPSJ SIG Technical Reports.