Operating Systems

1
Operating Systems

• Distributed Files
• 12/04/00

2
Introduction

• Why is it necessary?
• File server vs. file service

3
Design

• File service
• Directory Service

4
File Service interface

• What is a file?
• Attributes
• mutable vs. immutable files
• protection capabilities vs. access control lists
• upload download vs. remote access

5
(No Transcript)
6
Directory Service Interface

• Hierarchical file systems
• trees vs. dags vs. graphs (which allow cycles)
• Must all machines have the same view of the
  directory hierarchy?
• Is there a global root directory?

7
(No Transcript)
8
(No Transcript)
9
DSI contd

• Naming
• location transparency vs. location independence
• 3 approaches
• machine path e.g. /machine/path or machinepath
• mounting remote file systems onto local file
  hierarchy
• single namespace for all machines

10
File Sharing

• Unix semantics
• operations to files are immediately visible to
  other processes
• Session semantics
• changes arent visible until file is closed
• immutable files
• no updates allowed
• Transactions - atomicity

11
(No Transcript)
12
Implementation

• File usage patterns - Satyanarayanan 1981
• Most files are small
• Reading more common than writing
• Reads/writes generally sequential
• Most files have a short lifetime
• File sharing is unusual
• Average process only uses a few files
• Distinct file classes with different properties

13
Structure

• Differences between clients and servers
• Keep the file and directory service on different
  machines?
• Stateful vs. stateless servers

14
(No Transcript)
15
Stateful vs. stateless servers

• Shorter message requests
• Better performance
• Read ahead
• Idempotency easier
• file locking possible

• Fault tolerance
• No Open/Close needed
• no sever space wasted on tables
• no limits to of open files
• No problems if a client crashes

16
Caching

• 4 locations to store files
• server disk
• server main memory
• client disk
• client main memory

17
Caching - Server disk

• Plenty of space
• files accessible to all clients
• only 1 copy -- no consistency problems
• performance -(

18
Caching - Sever main memory

• Should the basic unit be whole files or disk
  blocks?
• What algorithm should be used if/when the cache
  fills up?

19
Caching - client

• Space vs. performance
• elimination of network transfers
• sharing/consistency problems
• 3 options for caching in main memory
• in user process address space
• in kernel
• separate user level cache
• RPC issues

20
(No Transcript)
21
Cache consistency(on client)

• Which semantic to use?
• What to do when 2 processes write the same file
  back?
• 4 approaches
• write through
• delayed write
• write on close
• centralized control algorithm

22
Replication

• What is it?
• Why?
• Reliability
• handle server crashes
• split workload
• Replication transparency

23
Approaches to replication
24
Creating replicated files

• Primary copy replication
• Voting
• Voting with ghosts

25
Voting