檔案系統效能改善之研究

Abstract

由於中央處理機和磁碟輸出入系統間的效能差異日益擴大，使得檔案系統 逐漸成為計算機系統效能的瓶頸。在一個檔案系統中有兩種類型的物件， 一種是資料(data)，另一種稱之為元資料(metadata)。資料是指真正檔案 中所儲存的內容，元資料則包括檔案存取控制及其他描述檔案內容特徵的 資料。最近的研究結果顯示，元資料的寫入佔了全部磁碟輸出入動作的 38-40%。因此我們提出並實作了一個元資料排序 (metadata ordering)及 非同步寫入的機制，大量減少了元資料的同步寫入。等到適當的時機，再 將這些寫入動作改以非同步方式來完成。效能測試的結果顯示，系統的效 能有相當幅度的改善。延遲寫入(delayed write)是一種用來改善 UNIX 檔案系統效能的方法。寫入策略(update policy)則負責決定在何時、以 何種方法，來將這些延遲寫入的資料寫回其所分配的磁碟位置。最近的研 究結果顯示傳統的寫入策略忽略了一個事實—不同的 I/O 需求有著不同 的特性。此一現象造成了許多時間上較急迫的交談式工作卻忍受著變異相 當大的反應時間。寫入策略之目的在控制一些在背景完成的寫入動作。這 些 I/O 在時間上通常較不急迫，不應該危害到那些較急迫之工作的完成 。在本論文中我們也設計並實作了一個有效的寫入策略。我們提出了寫入 負荷均化(burst declustering)以及機會性非同步寫入 (opportunistic asynchronous write) 們並發展了一個系統活動感應器 SAPRO(system activity probe) 用來監督系統活動以及磁碟排隊狀態，以動態調整我們 寫入策略的行為。依據效能測試的結果顯示，讀取反應時間在變異性以及 最壞情況上都有顯著的改進 (分別是40%及90%)。讀取的平均反應時間和 整體系統效能也有不錯的進步。 Due to the increasing gap between CPU and disk I/O speeds, file system is becoming the performance bottle-neck in computer system design. There are two types of entities in a file system: data and metadata. Data mean the actual contents of files. Metadata include access control and other descriptive about files. Previous research indicates that metadata writes account for 38-40% of disk I/O operations. We propose the design and implementation of a metadata ordering mechanism and its corresponding asynchronous write facility. With such a facility, we can eliminate large numbers of synchronous metadata writes, and have the flexibility of choosing a better way to these metadata modifications to disk asynchronously. Extensive performance evaluation shows that substantial performance improvement can be achieved under various benchmarks. Delayed write is a technique to improve the file system performance of UNIX operating systems. When write operations are delayed, an update policy decides how and when to write modified blocks to their assigned disk locations. Recent results show that conventional update mechanisms neglect the fact that different I/O requests naturally have different It makes time critical interactive jobs endure large variations response time. Update policy controls background write that are less time critical and should be performed under that do not violate the urgency of other time criti-cal jobs. In this thesis we also propose an effective update scheme. suggest new techniques : burst declustering and opportunistic asynchronous write. Besides, we develop a system activity sensor called SAPRO (system activity probe) that monitors the system activities and disk queuing status to adjust the behavior of our algorithm. Performance evaluation shows that our al-gorithm can alleviate the lengthy queuing delay, reduce the variance and case read response time significantly (40% and 90% The mean read response time and total system performance are also improved