莊裕澤臺灣大學:資訊管理學研究所鄭雅玲Cheng, Ya-LingYa-LingCheng2010-05-052018-06-292010-05-052018-06-292008U0001-1707200817053300http://ntur.lib.ntu.edu.tw//handle/246246/179854隨著電腦網路的快速成長以及電腦儲存設備成本的降低，越來越多個人電腦使用者能夠負擔大容量的硬碟與高速網路服務。一個使用者擁有多個儲存設備，並且將重要或常用的檔案備份在不同主機上，已經是非常普遍的現象；進一步透過網路有效率而且可靠地存取、共享、備份檔案的使用者需求也逐步升高。然而，現有的分散式檔案系統大多是為了區域網路中的檔案分享而設計，它們的設計目的是提供集中式的檔案儲存空間，以降低當時昂貴的儲存設備成本及簡化管理工作。這些早期的分散式檔案系統無法順暢的支援廣域網路中的檔案分享與存取。此外，這些分散式檔案系統只能提供固定的檔案介面，缺乏使檔案介面可以再被繼承、延伸以達到客製化的彈性。本篇論文中，我們設計一個支援廣域網路中檔案分享、存取、備份的物件導向分散式檔案系統。本系統是一個整合現有實體檔案系統的邏輯性檔案系統，對本地端檔案及遠端檔案提供一致的存取方式。我們選擇將系統建置在MOSP上，此乃鑒於MOSP開發網路物件導向系統的方便性，以及它整合異質性平台與網路通訊協定的能力，能幫助我們達成支援無縫隙、無界線檔案分享的目標。此外，我們提出了一個新的副本管理設計用以解決廣域網路中檔案分享的效能表現及可靠度等問題。With the rapid growth of computer networks and the price-reduction of computer storage, more and more personal computer users are able to afford large hard drives and high speed internet services. Now, it is common for personal users to possess more than one storage device on different computers and backup their critical files. This is an increased demand to access, share and backup files remotely through the Internet or other networks with reasonable performance and reliability. However, most of the current distributed file systems are designed for LANs, and intended to provide centralized file servers for use in Intranet environments to eliminate storage costs and management requirements. They do not gracefully handle inter-platform file access, sharing and backup over the WAN. Besides, most current distributed file systems provide fixed interfaces, not offering the flexibility for programmers or service providers to extend interfaces to meet customized requirements.n this thesis, we are developing an object-oriented distributed file system that supports file access, sharing and replication over the WAN. The system is a logical file system that can integrate all kinds of underlying physical file systems, providing an identical way to manipulate local and remote files. We build our system on MOSP because it provides a simple way to develop an object-oriented networked system and it accommodates heterogeneous platforms and network protocols, which helps achieve our goal of supporting boundless and seamless file sharing over the WAN. In addition, we propose a new replication management mechanism, which addresses the performance and reliability concerns.Contentshapter 1 Introduction 1hapter 2 Related Work 6.1 Protocol 6.1.1 Brief Summary 9.2 Distributed File System 9.2.1 System Architecture 10.2.2 Replicated Distributed File System 12.2.3 RAID Distributed File System 17.2.4 Brief Summary 19.3 File Synchronization 21.3.1 Brief Summary 25hapter 3 System Design 26.1 Overview 26.2 Object-Oriented Design 30.2.1 System Model 31.3 Replication Management 33.3.1 Components of Replication Architecture 35.3.2 Setting up a Replication Environment 41.3.3 Deployment of the Replication Architecture 46.3.4 Basic Operations 48.3.5 Consistency Control and File Synchronization 55.3.6 Fault Tolerance 59hapter 4 Performance Evaluation 62.1 Experiment Environment 62.2 Experiment Result 63hapter 5 Conclusion and Future Work 74.1 Conclusion 74.2 Future Work 76eference 77ppendix 80ist of Figuresigure 2-1 (a) An MOSP Java binding example 8igure 2-2 (b) An MOSP Java binding example 8igure 2-3 (c) An MOSP Java binding example 8igure 2-4 System Architecture of NFS [1] 11igure 2-5 Bayou System Model [5] 14igure 2-6 An example of dependency check criteria in Bayou [9] 15igure 2-7 Per-file striping [27] 18igure 2-8 Per-client striping [27] 18igure 2-9 Groove Communication Model [10] 23igure 3-1 File binding 27igure 3-2 A sample MIDL document for a file object 29igure 3-3 File inheritance over networks 30igure 3-4 Class Diagram of AbstractFile 32igure 3-5 Class Diagram of FileHandle 33igure 3-6 Composition relationship between File and FileHandle 33igure 3-7 Composition relationship between FileHandle and VolumeFileHandle 33igure 3-8 Composition relationship between VolumeFileHandle and ProxyFileHandle 33igure 3-9 Replication Architecture 35igure 3-10 The volume file handle request process 40igure 3-11 The proxy file handle request process 41igure 3-12 Volumes configured on a user computer 43igure 3-13 A replication environment 43igure 3-14 Flowchart of volume link to root process 44igure 3-15 Message sequence chart for file reading with single volume server 48igure 3-16 File reading from multiple volume servers 49igure 3-17 Message sequence chart for file writing 51igure 3-18 Traditional file copy over networks 53igure 3-19 Direct file copy 54igure 3-20 The validation check process to determine whether a replica is current or obsolete 57igure 4-1 Bandwidth utilizations for 2 GB file reads with a 100 Mbps Ethernet connection 66igure 4-2 Eclipsed time for 2 GB file reads from a volume server with 100Mbps upload bandwidth 66igure 4-3 Eclipsed time for 2 GB file reads from a Linux volume server to a Windows client with 100Mbps bandwidth over the WAN 67igure 4-4 Network consumption of the client computer for the 2 GB file copy on NFS 70igure 4-5 Network consumption of the client computer for the 2 GB file copy on our system 70igure 4-6 Average download rate for 10 MB file reads with different upload ability of the volume servers 72igure 4-7 Average download rate for 60 MB file reads with different upload ability of the volume servers 72igure 4-8 A drop in network utilization for a transparent failover 73igure 7-1 The AbstractFile interface 80igure 7-2 The File interface 80igure 7-3 The Directory interface 80igure 7-4 The RaidFile interface 81igure 7-5 The FileHandle interface 81igure 7-6 The RaidFileHandle interface 82igure 7-7 The Root Interface 82ist of Tablesable 2-1 Feature comparison of MOSP and other programming models[19] 7able 4-1 A MOSP outgoing message for read the second 16 KB chunk data of a file 67able 4-2 A MOSP return message with 16 KB data 68able 4-3 Average percentage of MOSP overhead with different chunk sizes 68able 4-4 Eclipsed time for file copy in pure Linux configuration (Unit: second) 69able 4-5 Eclipsed time for file copy in Windows/Linux configuration (Unit: second) 70application/pdf773947 bytesapplication/pdfen-US分散式檔案系統副本管理延伸式檔案系統物件導向式系統廣域網路檔案共享Distributed File SystemReplication ManagementExtensible File SystemObject-Oriented SystemWide-Area Network File Sharinghttp://ntur.lib.ntu.edu.tw/bitstream/246246/179854/1/ntu-97-R95725044-1.pdf