Lecture Outline

Aims and Objectives
Introduction
Distributed File System Design
Distributed File System Implementation
Example: Sun's Network File System

• Introduction

• Distributed File System Design

• Distributed File System Implementation

• Sun's Network File System

• File Service vs. File Server

  File Service: specification of what the file system offers its clients

  File Server: process which run on one or more machines and helps implement the file service

  System can have multiple file servers, offering different file services (e.g., UNIX and DOS)

  Users shouldn't need to know the details of how a request for a file is handled and neither should they need to know that there are several file servers

• Two parts: File service (operations on files - read, write, append) and directory service (mkdir, rmdir, file creation and deletion)

The File Service Interface

• File Properties

  Byte sequence vs. data structure

  Attributes (owner, creation/modified date, size, permissions)

  Immutable vs. mutable

  Protection via Capabilities vs. Protection via Access Control Lists

• File Service Models

  Upload/download

  

  Remote Access

  

The Directory Server Interface

• Independent of File Service Model

• Defines how user-attributed file names can be composed

• How can processes and machines have the same "view" of a directory hierarchy, e.g., remote mounting, in a distributed file system

• Naming Transparency

  Location Transparency

  • No clue as to where server is located

  Location Independence

  • Files can be moved without changing their names

• Two-Level Naming

  Symbolic Names vs. Binary Names

  Looking up a symbolic name and getting a list of binary names - refers to original and all backups, copies, etc. - Provides fault tolerence

• Semantics of File Sharing

  What happens when two users read/modify the same file simultaneously?

  UNIX semantics - mimics single processor system

  • READ following WRITE returns data written

  • 2 WRITES followed by READ gives last WRITE

  • Assumes absolute time ordering on events

  UNIX semantics can be achieved on distributed systems as long as files are not cached - but this results in a poor service

  Alterative - write cached file modifications to server immediately. Inefficient (e.g., temporary files)

  Session Sematics

  Relax UNIX semantics. "Changes to an open file are visible only to the process that modified the file"

  But what about a file that is modified by 2 or more processes? Which gets written last?

  And what about another process that attempts to read a file while it is being modified? - it won't get the latest version

  Possible solutions

  • All files are immutable - only create and read

  • If a file is overwritten, reads by others fail

  • Use atomic transactions

• Issues in determining what to implement

  How big are files on average?

  What's the average lifespan of a file?

  Do files tend to be shared?

Issues Regarding the Internal organisation of file servers and directory servers

• Should there be differences between clients and servers?

  No difference

  Client a server can run on same machine

  Clients and servers are fundamentally different

• Should the file and directory servers be separate or combined?

• Should servers maintain state information about their clients?

Caching

• 4 potential places to store files

  

• Server's disk

  + all files accessible to all clients

  + plenty of space

  + no consistency problems

  - network transfers / performance

• Server's cache

  + faster than disk access

  - network transfer

  - record might not be in cache

• Client's disk

  + no network transfer

  - slow performance

  - file consistency

• Client's cache

  + no network transfer if file/record is in cache

  + read ahead

  - file consistency

  Most popular option

Cache Consistency

• Write-through algorithms

• Write-on-close algorithms

• NFS supports heterogeneous systems, e.g., MS-DOS clients and UNIX servers

NFS Architecture

• NFS allows same machine to be both a client and a server

• No restriction on geographical locations of servers and clients

• NFS servers export directory trees

• Client accesses exported directories by mounting them

• Two clients that mount the same directory can exchange data through it

NFS Protocols

• Mounting Directories

  Client sends request for permission to mount a path name

  If request granted, server returns a file handle

• Directory and file access

  NFS is stateless

Next Lecture...

Distributed Operating Systems: Case Study: Amoeba