From: ahemavathy (ahemavathy_at_hotmail.com)
Date: Sun Feb 22 2004 - 17:15:56 PST
The paper is about the design, implementation and evaluation of Andrew File System - a distributed file system. The paper traces the system from its prototype and drawbacks to the refined system after changes were made to the prototype to improve performance. It was interesting to read how they had invented a way to cleverly overcome each of the drawbacks. Finally the paper compares the system to the IBM NFS and does some benchmarking. The results were clearly in favor of the AFS.
The AFS is for a client - server model. The clients should have their own disks. The Vice is the program on the server side and all file systems calls on the client have to go through the Venus. Venus has two caches : one for data and the other for status. Basically whenever the user requests a file on the remote server, Venus checks in its cache if it is already there. If not then it performs a network call and fetches the file and also caches it for future references. To maintain cache coherency, the system uses a callback mechanism where the server notifies the clients in case the cache data is out of date. By following this mechanism, the system reduces the network traffic and also the load on the server. Changes to the file are notified to Vice when the file is closed. The Venus participates only during the open and close calls of files.
Every file has a fid composed of the 1)volume number - which indicates the file's volume (Usually every user is given a Volume) 2) vnode - which is basically a index to obtain the inode and 3) a uniquifier - unique identifier. Volume to server mapping is maintained in Volume location database. Moving volumes between servers is made easy..only the database has to be updated.
It was a very good paper which exposes the limitations of old distributed file systems. AFS's prime strength is its scalability.
This archive was generated by hypermail 2.1.6 : Sun Feb 22 2004 - 17:14:47 PST