异构计算环境中循环级并行性提取及调度

Extracting Loop-level Parallelism and Scheduling in Heterogeneous Computing Environment

首先给出了循环最大加速比的定义，它是循环体中隐含并行性大小的度量。并且讨论了Ｄｏ循环的各种类型和它们的相关性以及隐含的并行性。在异构环境中，如何开发这些并行性，文中给出了几种方法，这就是异构计算消除串行瓶颈，异构Ｄｏ＿ａｌｌｌｏｏｐ调度，异构Ｄｏ＿ａ＿ｃｒｏｓｓ调度。无论在理论分析还是在实验结果上，这些方法所取得的计算性能都优于同构计算。

in this paper, we put forward the definition of loop maximum speedup, which is the metric of parallelism hidden in loop body. We also study the classes of Do iteration and their dependence as well as the parallelism they contain. How to exploit such parallelism under heterogeneous computing environment? The paper presents several approaches, which are eliminating serial bottleneck by means of heterogeneous computing, heterogeneous Do-all-loop scheduling and heterogeneous Do-a-cross scheduling. We find that, not only on theoretical analysis but also from experimental results, these schemes acquire better performance than in homogeneous computing.