Research on Multigrain Parallel Processing on Multiprocessor Systems

点击翻译按钮获取中文摘要

In 1994, which is the final year of this research project on "Multigrain Parallel Processing, " an automatic parallelizing compiler was developed and evaluated on an actual multiprocessor system. The compiler hierarchically combines a coarse grain parallel processing scheme which was developed by this research project in 1993, a near-fine grain parallel processing scheme, and traditional loop parallelization scheme. It was confirmed that the compiler allows us to efficiently exploit parallelism in a source program on the actual multiprocessor system.The multigrain parallel processing scheme consists of the following steps :1.Process macrotasks, or coarse grain tasks, in parallel among multiple processor clusters by using a macro-dataflow scheme developed by this project.2.Process a loop in parallel, which is assigned to a processor cluster, among processors inside a processor cluster by using Do-all or D-across scheme if the loop is parallelizable.(3)Process a sequential loop or a basic block in parallel, which is assigned to a processor cluster, among processors inside processor cluster by using a near-fine grain parallel processing scheme.Furthermore, this project also developed a new optimization technique called "near-fine grain parallel processing scheme without explicit synchronization" which minimizes synchronization overhead in near-fine grain parallel processing.The above results were or will be published as 7 journal papers, 6 international conference papers, 1 chapter of an international book, 3 symposium papers with review process, 10 technical reports of societies in Japan, and 15 short papers in annual conference of societies in Japan.