在多工作線架構上之程式分割技巧

Abstract

多處理(Multiprocessing) 系統是高效能運算的一個關鍵方法。但一般多 處理系統有記憶體存取延遲(latency) 過長和處理器間通訊 (communication) 及同步 (synchronization) 延遲過長的問題。多工作 線 (Multithreading) 是解決此問題的一個方法。在此同時﹐我們需要一 個好的編譯器 (compiler) 來發揮多工作線架構的特性。本論文針對傳統 語言(像 C 與 Fortran) 在多工作線架構上的編譯進行研究﹐設計出一套 適合多工作線架構之編譯器使用的中間碼(intermediate form) 及提出一 些利用多工作線架構特性的程式分割(program partition) 的方法。此中 間碼能使編譯作業更具效率﹐而程式分割則主要以發揮程式中最大平行度 為考量。我們在理想模式及較為真實的模式下評估這些程式分割方法效率 ﹐及使用時機。 Multiprocessing is a key approach to high-performance computing. However, multiprocessing has problems of long memory latency and long inter-processor communication and synchronization latency. Multithreading has been proposed as a means for solving these problems. However, an effective compiler is needed to exhibit the full power of a multithreaded architecture. This study investigates several important aspects of a parallel compiler for multithreaded architecture on traditional languages (such as C and Fortran). We designed an intermediate form for compilers of multithreaded systems and proposed several program partition strategies. After evaluated them on both an ideal and a more realistic model we found it is effective.