實作一個具有迴圈分割功能的可移植平行編譯器

Abstract

為了要使電腦的速度能夠超越最快的循序處理器，平行機器已經成為近十 年來電腦的設計趨勢，而到目前為止已有許多平行電腦問世。除了硬體架 構的進步，新的作業系統亦支援多重引線，使得機器的平行能力能為使用 者所用。但由於缺少好的平行編譯器，使得系統效能的增進遇到了瓶頸。 在本篇論文中，我們在執行 OSF/1 作業系統的 Acer Altos 10000 實驗 機器上，實作了一個具有迴圈分割功能的可移植平行編譯器。我們亦定義 了 B Threads，為一個支援此平行編譯器所需之基本的引線相關函數及資 料型態的集合。因為這個平行編譯器被設計成高度模組化，所以要將之移 植到其他的平臺便非常容易。此平行編譯器會根據不同的迴圈分割演算法 ，將平行迴圈轉成多重引線程式。實驗結果顯示，當處理器個數為四時， 對矩陣乘法及 adjoint convolution 分別可以得到 3.75 及 3.46 的最 佳加速。 In order to achieve performance improvements over the fastest sequential processors, parallel computers have become the tendency of computer design in the past ten years. Many parallel computing systems have been developed and implemented. In addition to the advance in machine architecture, newly operating systems also support multithreading to achieve the parallelism supported by the host machine. But the lacking of good parallelizing compilers becomes the bottleneck of performance improvements of these system. In this thesis, we implemented a portable parallel Fortran compiler with loop partition on our experimental target system: Acer Altos 10000 system, running OSF/1 operating system. We also defined a minimal set of thread-related functions and data types, called B Threads, that is required to support the execution of this parallel compiler. Our parallel compiler is highly modulized so that the porting to other platform will be very easy. It partitions parallel loops into multithreaded codes based on several loop-partition algorithms. The experimental results show that the best speedups are 3.75 and 3.46 for matrix multiplication and adjoint convolution, respectively, when the number of processors is four.