公开(公告)号：US07526637B2

公开(公告)日：2009-04-28

申请号：US11453288

申请日：2006-06-15

申请人： Chang Hee Jung ,  Dae Seob Lim ,  Jae Jin Lee ,  Sang Yong Han

发明人： Chang Hee Jung ,  Dae Seob Lim ,  Jae Jin Lee ,  Sang Yong Han

IPC分类号： G06F9/00

CPC分类号： G06F8/456 ,  G06F9/5066 ,  G06F11/3404 ,  G06F2201/88

摘要： Provided is a parallel program execution method in which in order to reflect structural characteristics of a multithreaded processor-based parallel system, performance of the parallel loop is predicted while compiling or executing using a performance prediction model and then the parallel program is executed using an adaptive execution method.The method includes the steps of: generating as many threads as the number of physical processors of the parallel system in order to execute at least one parallel loop contained in the parallel program; by the generated threads, executing at least one single loop of each parallel loop; measuring an execution time, the number of executed instructions, and the number of cache misses for each parallel loop; determining an execution mode of each parallel loop by determining the number of threads used to execute each parallel loop based on the measured values; and allocating the threads to each physical processor according to the result of the determination to execute each parallel loop.The method significantly improves the performance of the parallel program driven in the multithreaded processor-based parallel system.

摘要翻译： 提供了一种并行程序执行方法，其中为了反映基于多线程处理器的并行系统的结构特征，在使用性能预测模型编译或执行时预测并行循环的性能，然后使用自适应 执行方式。 该方法包括以下步骤：生成与并行系统的物理处理器的数量一样多的线程，以便执行并行程序中包含的至少一个并行循环; 通过生成的线程执行每个并行循环的至少一个单个循环; 测量执行时间，执行指令的数量以及每个并行循环的高速缓存未命中数; 通过基于测量值确定用于执行每个并行循环的线程数，确定每个并行循环的执行模式; 以及根据确定的结果将线程分配给每个物理处理器以执行每个并行循环。 该方法显着提高了在基于多线程处理器的并行系统中驱动的并行程序的性能。