1. Andrew File System (AFS)
CSS 534
Sourish Chatterjee, Xiaolin Ma

2. Agenda Overview AFS Design Goals and Choices Architecture Features
Venus and Vice
Features
Authentication
Access List
Logical Volumes
Filename Resolution
Cache Management and Concurrency
Viewpoints - Past, Now, Future
Comparison with other Distributed File System

3. Overview Developed jointly by Carnegie Mellon University and IBM
First wide distributed file system for large number of users
3 major revisions of design and implementation AFS-1 (1984), AFS-2 (1986), AFS-3 (1989).
Further development was transferred to Transarc Corporation (Now IBM)

4. Goals and Area of focus Goal was to build a system which is
Highly available
Scalable
High in performance
Fault tolerant
Areas of Focus was on
Hardware and Operating system
Network
File System
User Interface
Message system
When designing the Andrew File System, the goal was to support at least 7000 workstations, which means one workstation for each student or staff.
The file system is supposed to have high availability, which means users can get access to the files they want anywhere on campus at anytime.
And the file system should be scalable. It means it’s easy to expand and friendly to new technologies in the future.
A high performance and good fault tolerance are also required, to offer users a fast and reliable experience.
Designers mainly focused on these five areas. They are So I am going to talk about their designing choices on these five areas

5. Design Choices Workstations with Unix BSD 4.2 TCP/IP
Serving and Updating Whole File
Reduce communication overhead
Enough granularity
Many small servers instead of one large server
System control m/c for large AFS
50:1 ( Client : Server)
At the very beginning, designers decided to use a high-performance workstation rather than the cheaper machines like PCs to meet the need of some resources consuming tasks.
Designers chose Unix, or more specifically, the Unix BSD 4.2, because it is an operating system with good availability and some features like virtual memory and network support.
The network is built up gradually by connecting the pre-exist Ethernets of each department. So finally, a hierarchical network with a backbone and large number of local networks was formed. And TCP/IP was selected as the standard protocol.
File system( 1 make the file system compatible with UNIX at the system call level similar command 2 whole file strategy 3 )

6. Architecture N E T W O R K Workstations Servers User Program Venus
Vice
Unix Kernel
Local Cache
Unix Kernel
File Storage
File Storage
User Program
Venus
Unix Kernel
Local Cache
Vice
User Program
Unix Kernel
Venus
File Storage
File Storage
Unix Kernel
Local Cache

7. Architecture(Contd..) Venus (Workstations)
User Program
User Program
User Program
User Program
Andrew Cache Manager
To Vice
File System call
File System call
File System call
File System call
File System call diverted to Andrew Cache Mgr
Unix Kernel
Threads
Local File Cache
User Program remains suspended until the file fetch/store is completed through AFS

8. Architecture(Contd..) Vice (File Server)
File Manager
Auth Manager
Update Manager
Status Manager
From Venus
Internal Lightweight Process Threads
Unix Kernel
Single UNIX process
File Storage
Authenticates, Check permissions, Transfers Logical Volume, Updates File version, Keeps Workstation interests for callback, Collect other server(s) status

9. Authentication Manager
Features
Authentication
Workstation
File Server
Vice
Once a user logs into a workstation, the workstation goes through an authentication procedure based on the user's ID and password.
The login program sends encrypted message to the authentication manager running on the file server. If the message is valid, a token will be produced and sent back to the workstation. With this token, the cache manager on Venus can establish connections to the file servers freely.
The tokens can be passed on to other workstation to support remote login and the tokens expire after 25 hours.
Login Program
Venus
Authentication Manager
Credentials
Token
Unix Kernel
Unix Kernel
Local File Cache
File Storage
File Storage

10. Features Access Lists Inadequate traditional Unix access control mode
New access mechanism needed
Introduction of Access lists
Read the contents of files in the directory
Look up(or list the names of ) files in the directory
Insert files into the directory
Write(or modify) files in the directory
Lock files in the directory
Delete files from the directory
Administer the directory
The access control lists work on directory level
Andrew file system was designed to support the entire campus. The traditional access control mode of Unix cannot meet the need of a distributed file system, so the designers introduced a new mechanism to supplement the traditional one, which is the access control list mechanism.
Permissions to access files or directories are set in the form of access control lists. The control list only work on the directory level. There are seven types of permissions that can be given to users for a directory. They are:

11. Admin commands from Update manager
Features(Contd..)
Logical Volumes
Logical aggregation of files.
File is the lowest granular unit.
Used for fundamental operations like backup, load or space balancing between file servers by admin.
Admin commands from Update manager
Logical volumes are special aggregates of files. Doesn’t deal with blocks. The smallest unit inside Logical Volume is whole file.
Almost exclusive to the administrators and operators, end user seldom use it. They are used for backup, Load Balancing and moving a set of files from one file server to other
Actual Files on the Disk
Actual Files on the Disk
Actual Files on the Disk
Logical Volume A
Logical Volume B
Logical Volume C

12. Filename Resolution Conventional UNIX 4.2 BSD No inode notion in Vice.
Fixed File Name , inode and its path (variable)
namei routine - to locate file using iNode and file name. Performance overhead
No inode notion in Vice.
Hence 96 bit fid was introduced
32 bit Volume #
32 bit VNode #
32 bit Uniquifer
Logical Volume ID
Index of file info Array
unique id for a fid

13. Cache Management N E T W O R K 1 2 3 Workstation X Vice (W) File A X A
Venus
Workstation X
Vice
(W) File A
Workstation
File
Status X A
latest version of file Y B Z
Invalidated
File A
Cached File A
Venus
Workstation Y
(W) File B
Invalidate File B after 3
File B
Cached File B
Invalidate callback for Workstation Y
Venus
Workstation Z
(W) File B
Cached File B

14. Concurrency
Multiple workstations can perform same operation on a single file in server
No implicit locking on files
Application programs to cooperate in performing synchronization
Empirical study indicate very less percentage of multiple workstation writing a single file

15. Viewpoints - Past, Now, Future ?
Users’ viewpoint: closely similar to the UNIX FS
Noticeable wait to fetch a file
failure mode
access list
Operators’ viewpoint: views AFS as a collection of Logical Volumes which is very different from UNIX FS
Now AFS is available as a Open source product (OpenAFS)
Works on UNIX, Linux, MacOS X and Windows
Better user experience because of high network speed. Very small wait time
In AFS3 files which are larger than 64KB are send in 64KB chunks
Future ?
End users’ viewpoint: All the distributed implementation, files transfer or caching mechanism are transparent to end users. From their viewpoints, AFS is very similar to the UNIX, but allows them get a uniform set of files at any workstation on campus.
But the users may see two main differences between AFS and UNIX local file system.
First is
Operators’ viewpoint: unlike the end users, the operators see AFS completely different from the traditional UNIX file system. Usually they see the file system as a collection of logical volumes. Logical volumes are the unit of operations like create, adjust, clone, move, and backup.
Future - Looking at future, AFS might not be a great choice for big data. As the future of computing heavily relies on big data there are other file system(s) which are designed to suit those needs. We will end our presentation with some comparison between some other distributed file systems.

16. Comparison with other File Systems
AFS
NFS
HDFS
Target Users
Global/Large User set
Workgroup
Larger than AFS
Domain
Multi domain support
Single domain
Multi domain
Client Server Trust
Kerberos authorization
Client identity is trusted by default
Client identity is what O/S tells. No Kerberos Auth.
Compatibility with O/S
Same System Calls as of O/S
Different calls. Mainly used for non interactive programs

17. References
John H. Howard - Carnegie Mellon University
John H. Howard, Michael L. Kazar, Sherri G. Menees, David A. Nichols, M. Satyanarayanan, Robert N. Sidebotham, Michael J. West - Carnegie Mellon University
, Prof Dr. Stefan Leue University of Freiburg
, Monali Mavani, Member, IACSIT

18. Questions ?