Buddy配置之發散式檔案系統設計

Abstract

本篇論文提出一種 " Buddy 配置之發散式檔案系統 " 之設計方法。 Buddy 系統以其在記憶體配置上之簡潔與快速而聞名。其能將檔案之實體 分佈以循序方式置放，因而當存取該檔案時其 seek 總時間將會降低。依 此特性，我們提出 Buddy 配置之發散式檔案系統，該系統係修改 Unix 之檔案系統並結合 restricted buddy 配置策略，使其成為發散式檔案系 統上之一磁碟處理機(Disk Processor)。發散式檔案系統之基本概念乃是 將作業系統中的檔案處理子系統，從一個邏輯上的處理行程轉變為一個實 體上的處理元件。藉由將工作切分給兩階層的處理器(檔案處理器(File Processor) 與磁碟處理器)，發散式檔案系統不僅能夠提供檔案的 pipeline 處理，也能成為多重結構化之檔案系統。目前我們已建立一個 Buddy 配置之發散式檔案系統的雛型，並在此雛型上進行系統效能的量測 。其結果顯示該雛型於處理循序存取時表現良好，但於隨機存取時系統效 能不佳。對於某些特別的環境，如 video server 等其工作負載( workload)多為循序存取者適用之。 This thesis presents the design of Buddy-allocated Diverging File System (DFS). Buddy system is well-known for its simplicity and speed in memory allocation. It can maintain the layout of a file in sequential and consequentially reduce the total seek time when accessing a file. With this property, we propose the Buddy-alloocated DFS which try to modify Unix file system combining with restricted buddy allocation policy and make it be a Disk Processor (DP) in DFS. The basic concept of DFS is to change the role of file system from a logical process in operating system to a physical processing component of operating system. By partitioning jobs into two-level processors (File Processor and Disk Processors), DFS can not only provide a file processing pipeline but also have the capability of being multi-structured. We have built a prototype of Buddy-allocated DFS and perform some evaluations on it. The results show that Buddy-allocated DFS prototype is performing well for sequential accesses but will suffer penalty in random requests. It can be used for some special environments such as video server which workloads are dominated with sequential accessing.