Veronica C. Lepore

Distributed Systems

Professor Getschmann

April 1998

The distributed file systems advancement is probably one of most focused area of Distributed Operating Systems. The Andrew File System (AFS) is one Distributed File System that has become a standard in the business and research worlds. This paper will give the reader an overview of the AFS. AFS was designed as part of the Andrew project, a combined effort of IBM and a research team from Carnegie-Mellon University. Location transparency and location independence are strengths of AFS' design, which can work with a large number of workstations.

The Andrew File System (AFS), developed at Carnegie-Mellon University (CMU), Pittsburgh, under a sponsorship from IBM, is one of the distributed network file systems that enables files from any AFS machine across the country to be accessed as easily as locally stored files. AFS uses client/server technology. AFS servers are transparent to the users who use an AFS client to access the information that resides on the servers, but the information is accessible just as if it was stored on the users’ AFS client machine. The resources can be shared through networks, both local area and wide area. "Since AFS hides the underlying network, working in AFS is just like working on your" ("The Andrew File System.", 1998) AFS client's file system, but with access to many more files. The file system appears to the user as a giant virtual hard disk or mega-directory made up of subdirectories from across the entire continent. AFS is designed to support a large number of networked workstations, providing users, application programs, and system administrators with the benefits of a shared file system. Sun's NFS (Network File System) is similar to AFS. The advantages of using AFS over NFS will be discussed shortly.

With AFS, almost all of the users’ data can be cached on their disk(s). "AFS chooses entire files to be the basic unit of data movement requiring every used file to be copied entirely to the local disk and copied back whenever they are modified" (NHMFL, 1997). AFS is able to provide strong system consistency while maintaining good performance. The ability to cache files can save network bandwidth since they can be used again. AFS has been the starting point for the Andrew Project "at CMU that includes the ability for users to operate in a shared file system while totally disconnected for extended periods of time i.e. portable computers" (NHMFL, 1997). AFS is designed so that workstations have enough processing power, and do not use the server processing power whenever possible.

Today, AFS is marketed by Transarc Corporation and has been chosen by the Open Software Foundation as the basis of its Distributed File System (DFS). Open Software Foundation is "a foundation created by nine computer vendors, — Apollo, DEC, Hewlett-Packard, IBM, Bull, Nixdorf, Philips, Siemens, and Hitachi — to promote 'Open Computing'" (Többicke, 1996). The purpose of establishing a foundation is to develop a common operating system and interfaces, based on developments of Unix and the X Window System that will be compatible for a wide range of different hardware architectures.

The AFS file system uses a hierarchical tree structure. The lowest level of the structure, i.e., the worldwide root AFS directory, is called /afs. Subdirectories of the worldwide AFS root directory are called AFS cells. These cells contains each cell "representing an independently administered portion of file space" (Többicke, 1996). In UNIX file system under the root /afs directory, the cells connect to each other to form one enormous file system. It eliminates the need to transfer files between machines.

The worldwide AFS system is built up by a number of cells. The users are members of the cellothers.foo.com. In each cell there are a few servers and several clients. Each client has an extra directory, /afs, under which the entire AFS file system resides. Under /afs there is a directory for each defined cell, i.e. the cell resides in /afs/others.foo.com. This can be shortened to /afs/cell. A bit further down are the user directories. AFS has a mechanism for limiting the disk space a user’ files can occupy.

A user's designated portion of the AFS file system is a subdirectory called a volume. A client AFS program through a Kerberos authentication/password procedure mounts volumes. The files and programs in a volume can be stored on local AFS server(s) or on a remote AFS file server(s) on the Internet. From the user's point of view at the desktop, it seems as if the volume is physically mounted locally. The local volume and other AFS volumes are tracked by the AFS servers which make them appear as subdirectories of one virtual mega-directory or disk. With this kind of feature, it means that it is unnecessary for an AFS user to use the FTP program to transfer files between directories in the global AFS file space. The transfer of files between directories is possible once the proper directory permissions have been set up by a volume's owner; for instance, an AFS user can simply use basic UNIX file system commands such as cp to copy files from a permitted volume to their personal AFS volume. An AFS user does not have to know the physical location of the files she/he wants to access even though the file could be on a number of machines. The user just requests the file by its ordinary UNIX filename and AFS finds the file automatically.

To gain access to remote cells, the permission is set up with an AFS Access Control List (ACL), "an easy-to-use tool that makes it possible to share directories, files, and programs among AFS users" (NHMFL, 1997). There are also local cells that users can access in their local area networks.

The access lists, which are associated with a certain directory, consist of seven fields, rlidwak — read, lookup, insert, delete, write, administer, and lock. There are four levels: none, read (rl), write (rlidwk) and all. As ACLs differs from the standard UNIX privilege system, it is quite easy to set up directories that a specified group of people, or single individuals, can access. It is also easy to prohibit people from accessing files.

The biggest difference between ACLs and traditional UNIX file permissions is that an ACL is used to apply permissions to an entire directory in an AFS volume. ACL cannot be used to apply permissions to individual files within a directory by using the UNIX chmod command. The ACL of "a parent AFS directory is also automatically inherited by any new child subdirectories created in or moved into the parent directory" (Sandell). The only way to change the ACL of the child subdirectory is to separately apply a different ACL to it.

There are three default system-wide groups, which are: