有限目錄快取記憶體資料一致性協定效能改進之研究

Abstract

快取記憶體利用減少平均記憶體存取的延遲來增加多處理器的效能，不過 快取記憶體在多處理器的環境底下會導致資料一致性的問題。快取記憶體 資料一致性協定利用保存每一份區塊內的資料彼此間的一致性來解決資料 一致性的問題，但其缺點是此種協定會增加額外的網路傳輸。本篇論文提 供了另外的一些方法來減少網路的傳輸量。我們採用了有限目錄之快取記 憶體資料一致性協定，此協定有易於擴充的優點，容易在大型規模的系統 底下發展。我們使用了不同的置換策略及不同的編碼方式來改善原有的有 限目錄之一致性協定，好的置換策略可以減少網路上的傳輸。不同的編碼 方式可以等效的增加目錄的容量，也可以減少網路上的傳輸，再加上利用 事件產生模擬程式及追蹤導向的模擬程式來證明我們的效能改進。 Caches enhance the performance of multiprocessors by reducing average memoryaccess latency, but caches in a multiprocessing environment also introduce the cache coherence problem.Cache coherence protocol prevents this problem by maintaining a uniformstate for each cached block of data. The full-map protocol uses a directory, whose size is equal to the number of the processors in eachmain memory block.The overhead of the main memory increased severely when the number of theprocessors become large. In this thesis, the limited-directoryprotocol which uses a directory of fixed size, usually much smaller thenthe number of the processors, is considered. The limited- directory protocol does not have the scalability problemand can be applied to large scale multiprocessor system easily.Because of the fix size of the directory, theprotocols will generate invalidation messages when the number of sharingprocessors is large than the size of the directory.Thus, the limited-directory protocols willproduce extra network traffic for invalidation than the full-map protocolwill. This thesis proposes some enhancements for the limited-directory protocol to reduce the network traffic. We applied alternative replacement policies and special coding for the directory. A better replacement policy will decrease the network traffic forinvalidation, and a different coding scheme will equivalently enlarge thecapacity of the directory.We use an event generator to predict the performance gain, and atrace-driven simulation to show the improvement of the system performance over the original protocol.