發散式檔案系統之同作控制與復原管理之設計

Abstract

本篇論文提出一個具有共用性與可靠性的發散式檔案系統，它提供整合式 的同作控制與復原管理。本系統是由傳統的作業系統中分離出來，成為完 整而且獨立的檔案系統。它藉由階層式的處理機組，提供單元性交易服務 與穩定性儲體服務。我們以記錄導向(record-oriented) 的方法來設計此 系統，本方法與已往的作法有下列不同之處： 1. 彈性化的同作控制單 位(granularity of concurrecy control) 2. 快速的系統當機復原( system crash recovery) 3. 低成本的交付處理(commit-processing) 4. 動態式的檔案重組(file reorganization)我們對此方法建立一個交付 成本模式，並製作一個發散式檔案系統雛型，藉此來測量與分析本系統的 效能。 This thesis proposes a shareable and reliable Diverging File System(DFS) which supports integrated concurrency control and recovery management. DFS is a complete and independent file system which is diverged from traditional operating system. It consists of hierarchical processors to provide an atomic transaction service and a stable storage service. We present a record-oriented approach to design our system. Our approach is different from earlier proposals in the following ways: (1) flexible granularity of concurrency control (2) fast system crash recovery (3) low commit processing cost (4) dynamic file reorganization A cost model of transaction-commit processing is built. A prototype of DFS based on our approach is implemented. The performance of the prototype is evaluated and analyzed.