多處理單元電腦系統之指令排程設計與其在AD100電腦系統之應用

Abstract

改良計算機運算速度始終都是設計者首要考慮的重點，而使用具有平行化架構或是超純量(Superscalar)架構之電腦系統常是改良運算速度的方法之一。因而，如何將指令碼經過排程而能平行化執行，將成為提升這種系統工作效益必須有的功能。本論文說明了對於實際的電腦系統指令進行排程可能遭遇到的問題，這些問題包括:(1)指令必須相鄰(Ad」aacent)執行的限制(2)管線化(Pipeline)多週期指令的排程問題。本論文提出一基於軌跡排程法的延伸方法，使其適合於處理以上之排程問題，並且實際製作了此一排程法。 我們應用所提出來的排程法於航太領域普遍使用的Aninn電腦系統，同時發展了一套C語言平行編譯器雛型，配合我們製作的排程法，可以將部份的C語言敘述編譯並壓縮成平行執行的AD100指令碼。初步實驗證實以我們所設計的系統進行自動化排程，不僅可以解決上述兩項排程問題，並且在我們測試的16種運算中除了除法與Liverm)oreloop7之外全比現有的ADSIM軟體快。

Improving the computing speed is often the primary object to the computer system designer. One of the sulution method is to use the parallel and superscalar processing architechtures. Futhermore, it is also required to schedule the sequential instruction codes into parallel ones to maximize the performance of those systems. There are two important functions have been proposed in this thesis: those are (1) scheduling the instructions which must be executed adjacently (2) scheduling the instructions operated in multi-cycle and pipelined processing. We have applied our scheduling algorithm to the AD 100 computing system, which is widely used for the simulation of aero-craft systems. For our developing requirement, we developed a prototype of parallel C compiler, which can compile the C codes and produce parallel object codes for AD 100. Our experiments have shown that our method outperforms the existing compiler in AD 100, named ADSIM, in 16 tests except two, the division and livermore-loop-7.