細微平行度邏輯語言並行處理之研究

Abstract

本論文主要在探討如何利用超純量技術建構一套高性能的Prolog系統。論 文中說明系統建構時所遭遇的效能瓶頸，解決方式，及相關的設計和驗證 。最後，並以模擬的方式，找出最適合的系統參數。一個Prolog超純量系 統主要面臨的系統效能瓶頸有二（一）大量的系統狀態儲存所造成的大量 記憶體存取動作。吾人設計了一套暫存器窗管理方式，將原本的系統控制 資訊全部儲存在暫存器窗中，大大減少了系統記憶體存取量。（二）大量 的分歧指令妨礙了指令的擷取。吾人首先分析了分歧預測準確度和超純量 系統效能之關係，並特別針對Prolog語言之遞迴特性設計了一套分歧指令 預測方法--PAM 。它在模擬中有97% 之準確度，大大增加了系統平行度。 根據以上的研究，吾人依不同的成本需求設計了兩套參數不同的Prolog 超純量系統架構，其平行度分別可達3.85和2.48。 The problem associated with how to implement a Prolog superscalar system is concerned in this dissertation. This dissertation discusses the performance bottlenecks of the implementation and designs for solving them. There are two performance bottlenecks in designing a high performance Prolog superscalar system: (1) In Prolog execution, a great deal of bookkeeping is needed, SORWT is designed for solving this problem. It includes all the control information in a windowed register file, thus greatly reduces the memory access operations. (2) Branch instructions stall the prefetching of instruct- ions across basic block, which greatly reduce the perform- ance of a superscalar system. A new branch prediction method PAM is designed to solve this problem, it provides a pred- iction hit rate of 97% in Prolog programs. We combine above design methods and superscalar techniques to construct the Prolog superscalar system. We design two machines with the different system configurations, their speedups are 3.85 and 2.48.