迴圈引線平行化方法之研究

Abstract

在本篇論文中提出在多處理機系統編譯器中引線平行化的方法．我們針對 迴圈中的循環 (iteration) 間的平行度加以分析，進而改變迴圈的架構 ，使其成為可以平行執行的引線 (thread)，而可在數個處理機上同時的 執行．在產生平行引線的同時，我們還分析了各種自我排程 (self- scheduling) 方法並提出了一種我們對自我排程方法的改進．本文中將對 這些方法、原理做詳細的說明。目前這些方法已建立在我們所發展的 PCE (Parallel Compilation Environment) 中，同時也在 IBM RS/6000 工作站上完成初步之測試工作。經由若干標竿程式 (Benchm- ark Programs) 之初步測試與評估，均有良好的效果。這對超純量多處機 系統編譯器的設計提供相當重要的參考。 The multi-threaded system design has become more and more and more important trends in various kinds os high performance computer system design. In this thesis, we have proposed an efficient method to generate parallel threads. In our method, we analyze a loop structure first, and then construct it into parallel threads running on several processors simutaneously. At the same time, several self-scheduling schemes have been evaluated. In addition, a new self-scheuling approach has been proposed in order to take the most advantage of loop-carried dependent loops. We have implemented these method in our PCE (Parallel Compil- ation Environment) and running on IBM RS/6000 workstation. We ha- ve also evaluated our method by using several benchmark programs. The results show that some reasonable execution speedup can be obtained.