以可移植性溝通函式庫與原始碼轉換在嵌入式多核心平台上執行MPI及Brook程式

Abstract

近年來，由於多核心平台在嵌入式系統上的普及性，程式的平行化變得更重要。程式平行化的方式非常多樣化，許多程式撰寫者利用MPI製作平行化程式在叢集（cluster）系統上運行，有些程式撰寫者使用類似Brook的資料平行（data parallel）語言開發適合在圖形處理器（GPU）上執行的串流（streaming）程式，但這些平行程式無法在嵌入式系統上直接執行。 有鑑於此，在這篇論文中，我們探討在嵌入式系統上執行MPI以及Brook程式的方法。首先，我們在IBM Cell、PAC Duo、TI Davinci等嵌入式多核心系統以及x86多核心平台上建構可移植性的訊息傳遞函式庫（message-passing library），稱之為MSG函式庫。接下來，我們提出（一）藉由轉換MPI程式庫中常用的函式，以及（二）將Brook程式編譯為利用通訊程式庫來傳送資料，讓MPI和Brook程式得以利用MSG函式庫在上述多核心平台上執行。我們實作這兩項轉換技術，驗證這個方法的可行性。初步實驗結果發現，部分的測試用的MPI和Brook程式，可以自動地利用我們設計的機制在IBM Cell平台上順利執行，並且有不錯的效能。

Recently, multicore platforms have become popular for embedded applications, and the importance of parallel programming is increasing. Many parallel programs have been written in MPI for clusters. The Brook programming language is an example of data parallel programming models, which are popularly used to develop streaming programs for graphical processing units (GPU). However, these existing parallel programs cannot run on most embedded multicore systems because (1) the resources are limited in many embedded systems, and (2) the mechanisms for interprocessor communications are very different among embedded multicore systems. This thesis proposes a framework for running existing parallel programs on multiple embedded multicore systems. First, we developed and implemented a portable message-passing library, called the MSG library as an interprocessor communication mechanism across multicore platforms including IBM Cell, PAC Duo platform, TI Davinci and x86. In order to run MPI applications, we designed a converter to transform the essential and most frequently used functions in the MPI into MSG functions. For Brook programs, we modified the Brook compiler to transfer data with message-passing functions in MSG. We verified our approach with experiments on the IBM Cell platform. The results showed that some MPI and Brook programs can be automatically converted by our tools to run on the IBM Cell with good performance.