函数式语言并行实现中并行颗粒度的动态控制

DYNAMIC GRANULARITY CONTROL FOR PARALLEL EXECUTION OF FUNCTIONAL PROGRAMS

本文针对并行处理中并行颗粒度的控制问题，将惰性计值的延迟思想引入并行进程的生成过程中，并与文献［１，２］的ＬＥ计值语义相结合，提出了函数式语言的并行ＬＥ计值语义．由于采用ＩＨＲＥＡＤ来实现对并行归约子进程的延迟处理，使得并行计算进程在运行时可以动态增加其并行颗粒度，从而减少了一些不必要的细粒度并行性的开发．与文献［５］相比，由于基于负载的动态内联思想在延迟处理操作中的引入亦进一步减少了很大一部分延迟操作，进而提高了并发程序中细颗粒度并行性的开发效率．基于ＴＲＡＮＳＰＵＴＥＲ多处理机的实验显示本文方法具有一定的应用前景．

This paper proposes a lazy process creation scheme for controlling the granularity in the parallel implementation of functional programming languages.Based on the LE(Lazyu-Eager) evaluation semantics,a parallel LE evaluation semantics is put forward.With the introduction of THREAD and load based inline,process creation can be delayed and the cost of process delay operations minimized.resulting in the efficient exploitation of the fine grain parallelism in programs.Experiment results on the Transputer system show that the approach is promising.