A Fast File System for Unix

1
A Fast File System for Unix

• Marshall K. Mckusick, William N. Joy,
• Samual J. Leffler and Robert S. Fabry
• Computer Systems Research Group, UCB

Presented By Parang Saraf
CS 5204 Operating Systems, Virginia Tech
2
About the Paper

• Considered as one of the most fundamental papers
  in operating systems
• Have been cited around 930 times
• Describes a new file system

3
Traditional File System

• File System developed at Bell Laboratories
• A file system is described by its Super-Block
• Number of Data Blocks
• Count of maximum number of files
• Pointer to free list (linked list to all free
  blocks)
• Disk drive is divided into partitions
• Each disk partition may contain one file system
• A file system never spans multiple partitions

4
Traditional File System
5
Traditional File System Inode

• Each file has a descriptor associated with it
  Inode.
• Information includes
• Ownership of the file
• Time stamps marking last modification and access
  time
• Array of indices pointing to the data blocks
• Direct Blocks 8
• Indirect Blocks Singly, Doubly and Triply

6
Traditional File System Inode
7
Traditional File System Inode
8
Traditional File System Problem

• Inode information segregated from Data
• Long seek time from inode to its data
• Files in single directory are not typically
  allocated consecutive slots for inode information
• Many non-consecutive blocks of inodes are
  accessed when executing operations on inodes of
  several files in a directory
• Sub-optimum allocation of data blocks
• Small Block size 512 bytes
• Many Seeks Next sequential block is not on the
  same cylinder
• Limited read-ahead

9
Old File System

• Developed at Berkeley
• Increased Throughput
• Changing the basic block size from 512 bytes to
  1024 bytes
• Each disk transfer accessed twice as much data
• Less number to indirect blocks used
• Increased Reliability
• Staging modifications to critical file system
  information so that they could either be
  completed or repaired cleanly after a crash

10
Old File System Problem

• Old file system was still using just 4 of disk
  bandwidth
• Main problem Scrambled Free List

11
Old File System Problem

• Old file system was still using just 4 of disk
  bandwidth
• Main problem Scrambled Free List
• Initially ordered for optimal access
• Scrambled because files were created and removed
• Eventually becomes entirely random blocks
  allocated randomly
• On creation provides transfer rates up to 175
  kbps
• Rate deteriorates to 30 kbps after a few weeks of
  moderate use
• Possible Solution Dump, rebuild and restore /
  Fragmentation

12
New File System

• Each disk drive contains one or more file systems
• A File System is described by its super-block,
  located at the beginning of the disk partition
• Super-block is replicated to protect against
  catastrophic loss
• Block size is any power of two gt 4096 bytes
• Decided at the time of file system creation and
  cant be changed
• File Systems can have different block sizes

13
New File System Cylinder Groups

• Comprises of one or more consecutive cylinders

14
New File System Cylinder Groups

• Comprises of one or more consecutive cylinders
• Disk partition is divided into one or more
  cylinder groups
• Has associated book-keeping information
• A redundant copy of super-block
• Space for inodes
• A bit map describing available blocks replaces
  free list
• Summary information describing usage of data
  blocks

15
New File System Cylinder Groups

• Contains static number of inodes
• Allocated at file system creation time
• Default policy one inode for each 2048 bytes
• Book-keeping information begins at varying offset
  from the beginning of the cylinder group
• Redundant information spirals down into the
  cylinder
• Any single track, cylinder or platter can be lost
  without losing copies of the super-block

16
New File System Structure
17
New File System Key Contributions

• Optimizing storage utilization
• File System Parameterization
• Layout Policies

18
Optimizing Storage Utilization

• New 4096 size blocks transfers 4 times more
• Problem with large blocks
• Wasted space due to small files

19
Optimizing Storage Utilization

• Solution
• Divide the 4096 block into 2, 4 or 8 fragments to
  accommodate small files
• Fragment size is specified at the time file
  system is created
• Block map records the space available at fragment
  level

20
Optimizing Storage Utilization

• Free List vs Bitmap

21
Optimizing Storage Utilization

• Space allocation
• Space is allocated when a program does a write
  system call
• Three possible conditions
• Enough space left in an already allocated block
  or fragment
• File contains no fragmented blocks allocate new
  blocks and fragments
• File contains one or more fragmented blocks but
  has insufficient space to hold new data new
  block is allocated, old fragments are copied and
  new fragments are appended

22
Optimizing Storage Utilization

• Free space reserve
• Minimum acceptable percentage of file system
  blocks that should be free 90
• Only system administrator can allocate blocks
  after that
• Important for the layout policies to be effective
• After this the file system throughput is cut in
  half because of the inability to localize blocks
  in a file

23
Optimizing Storage Utilization

• Wasted space comparison
• Space wasted by 4096/1024 byte new file system is
  same as 1024 byte Old File System
• New file system uses less space for indexing
  large files
• Uses same amount of space for small files
• Free space reserve should also be counted as
  wasted space

24
File System Parameterization

• Optimum block allocation based on hardware
  parameters
• Speed of Processor
• Hardware support for mass storage transfers
• Characteristics of the mass storage devices
• Blocks are allocated on the same cylinder
• Block allocation depends on whether the processor
  has an input/output channel or not

25
File System Parameterization
Accessing which data is faster?
26
File System Parameterization
Accessing which data is faster?
Depends whether processor has I/O channel or not
27
File System Parameterization

• Rotationally Optimal Blocks
• Processors without I/O channels must field an
  interrupt and then prepare for a new disk
  transfer
• Disk rotates during this time
• Place blocks such that disk rotation is taken
  into account before the start of a new disk
  transfer operation
• Cylinder group summary information includes count
  of blocks based on different rotational positions
  8 positions
• Super-block contains a vector of lists called as
  
• Rotational Layout Tables Used by system when
  allocating new blocks

28
File System Parameterization
29
Layout Policies

• Layout policies divided into two distinct parts
• Global Policies
• Local Allocation Routines
• Two allocable resources
• Inodes
• Data Blocks

30
Layout Policies

• Global Policies
• Uses file system wide summary information to make
  decisions regarding the placement of new inodes
  and data blocks
• Tries to localize data that is concurrently
  accessed while spreads out unrelated data
• Inodes
• Places all inodes of files in a directory in the
  same cylinder group
• A new directory is placed in a cylinder group
  that has a greater than average number of free
  inodes and the smallest number of directories
  already in it ensures that files are
  distributed throughout the disk

31
Layout Policies

• Global Policies
• Data Blocks
• Tries to place all data blocks for a file in the
  same cylinder group
• None of the cylinder groups should ever become
  completely full
• Heuristic Solution redirect block allocation to
  a different cylinder group when a file exceeds 48
  kb and at every MB thereafter
• Ensures that cost of one long seek per MB is
  small
• New cylinder groups are chosen from those
  cylinder groups that have a greater than average
  number of free blocks left
• Finally it calls Local Allocation Routines for
  block allocation

32
Layout Policies

• Local Allocation Routines
• Allocates a free block as requested by the Global
  layout policies
• Uses a four level allocation
• First Level use the next free block that is
  rotationally closest to the requested block on
  the same cylinder

33
Layout Policies

• Local Allocation Routines
• Second Level if there are no free blocks on the
  same cylinder, a free block in the same cylinder
  group is selected

Cylinder Group
34
Layout Policies

• Local Allocation Routines
• Third Level if the cylinder group is full, use
  the quadratic hash function to hash the cylinder
  group number to find another cylinder group to
  look for a free block
• Fourth Level if the hash fails, use an
  exhaustive search on all cylinder groups
• Quadratic Hash
• is used because of its speed in finding unused
  slots in nearly full hash tables
• File systems parameterized to maintain 10 free
  space rarely use this

35
Performance

• Measured Throughput

36
Performance

• List Directory command performance
• For large directories containing many
  directories, disk access for inodes is cut by a
  factor of two
• For large directories containing only files, disk
  access for inodes is cut by a factor of eight
• Both reads and writes are faster in new file
  system
• Because larger block sizes are used
• The overhead of allocating is more but cost per
  byte allocation is same
• Reading rate is always at least as fast as
  writing rate
• Writes are slower for 4096 byte block as compared
  to 8096 byte block
• In old file system writing was 50 faster than
  reading

37
New File System - Limitations

• Limited by memory to memory copy operations
  required to move data from disk buffers in the
  systems address space to data buffers in the
  users address space
• Buffer alignment of both address space
• One block is allocated to a file at a time
• Pre-allocate several blocks at once and releasing
  unused ones on file closing

38
Functional Enhancements

• Long File Name
• File Locking
• Symbolic Links
• Rename
• Quotas

39
Long File Name

• Maximum length of file name is 255 characters
• Directories are allocated 512 byte units called
  chunks
• Chunks are broken into Directory Entries
• Contains information necessary to map the name of
  file with inode
• First three fields are fixed length inode
  number, size of entry and length of file name

40
File Locking

• Hard Lock always enforced when a program tries
  to access a file
• Advisory shared or exclusive locks requested by
  the programs
• System administrator privilege can override locks
• No deadlock detection is attempted

41
Symbolic Links

• A symbolic link is implemented as a file that
  contains a pathname
• Pathname can be relative or absolute
• On encountering a symbolic link while
  interpreting a component of a pathname, the
  contents of the symbolic link is prepended to the
  rest of the pathname

42
Rename

• Old file system required three system calls for
  renaming
• Target file could be left with temporary name due
  to crash
• New rename system call added that guarantees the
  existence of the target name
• Renaming works both on directory and files

43
Quotas

• Old file system any single user can allocate
  all the available space in the file system
• Quota restricts the amount of file system
  resources that a user can obtain
• Sets limits to both inodes and number of disk
  blocks
• Hard and soft limits

44
Key Take-Away points

• Substantially higher throughput rates large
  block size
• Flexible allocation policies
• Better locality of reference
• Less wastage
• Adapted to wide range of peripheral and processor
  characteristics

45
References

• Presentation on A Fast File System by
• Zhifei Wang www.cs.pdx.edu/walpole/class/cs533/
  spring2006/slides/191.ppt
• pdc-amd01.poly.edu/wein/cs6243/ppts/fastfile.ppt
• Sean Mondesire and Subramanian Kasi
  www.cs.ucf.edu/courses/cop5611/spring05/item/FFS.p
  pt
• www.scs.ryerson.ca/aabhari/File_System.ppt
• http//flylib.com/books/en/3.224.1.79/1/
• http//osr507doc.sco.com/en/HANDBOOK/graphics/hard
  disk.gif