Summary: | 碩士 === 國立臺灣大學 === 資訊工程研究所 === 81 === DE系統(Distributed Execution System)是一個由臺灣大學資訊工程研
究所發展出來的一套多電腦系統。它本身包含了三個主要部份:DE-DT
server ( DE - Distributed Task),DE-DSM server (DE -
Distributed Shared Memory)和 DE-Compiler。主要的目的是為了達成
在科學應用程式上平行處理的威力。我們的步驟是將傳統的FORTRAN程式
轉換成在DE-System下可以執行的平行碼,而使用者可以不必考慮加入任
何敘述。在這篇論文中,我們主要針對的是DE-Compiler中的資料區域性
分析。這部份是用來處理經過資料分割後,資料在DE-DSM中的問題。它將
由原來已被分割的資料與程式中,找出區域性的資料。同時也對純量變數
在平行程式中的處理方式做一些探討。這的資料區域性分析系統包含了三
個主要部份:純量變數的處理,屬性的分析(分析所有的存取可以得到正
確的屬性),屬性的流程分析(決定在什麼地方置放屬性敘述,什麼地方
改變屬性)。
DE-System (Distributed Execution System) is a multi-computer
system developed by the department of Computer Science and
Information Engieering of National Taiwan University. It
consists three major components, DE-DT server, DE-DSM server
and DE-Compiler. Its goal is to achieve the power of parallel
processing on scientific applications. The approach is to let
DE-Compiler translate sequential FORTRAN programs without any
user-defined statement into parallel executable code under DE-
System. In this thesis, we focus on the data locality analysis
pass of DE-Compiler that deals with attributes analysis after
data partition pass of DE-Compiler. It extracts the locality
of data from original program that is already partitioned and
distributed. It also takes care of the attribute of scalar,
which is a trouble on parallel program. The data locality
analysis system has three major parts: Scalar variables,
Attribute analysis, Attribute flow analysis.
|