A Comparison of File System Workloads

1
A Comparison ofFile System Workloads

• D. Roselli
• J. Lorch
• T. Anderson
• University of California, Berkeley

2
Motivation

• Wanted to find common file usage patterns
• Focused on how workload and file system
  parameters affect disk behavior
• Cache sizes
• Memory mapping
• Delayed write policy
• File layout policies

3
Trace collection

• Three Unix environments
• 20 workstations used for undergraduate classes
• 13 workstations usedby faculty, graduate students
  and staff working on research projects
• One web server
• One set of eight desktops running Windows NT

4
Trace collection issues

• Modified Unix kernel in order to record full
  pathnames of all files being accessed
• Auditing subsystem of HP-UX only recorded file
  names as specified by the user
• Often relative to current directory

5
Estmating block lifetimes (I)

• A previous study measured block lifetimes as
  delete time minus create time
• This delete-based method
• Did not take into account the lifetimes of
  blocks that were not deleted during measurement
  interval
• Underestimated block lifetimes

6
Estimating block lifetimes (II)

• Create-based method partitions trace into two
  parts
• First part is used to collect information on
  block creation times
• Proceed as in delete-based method for blocks
  that were deleted during total observation period
• Assume that other blocks have lived for at lest
  the duration of the second part of the trace

7
Example
D2
X
Time
Trace starts
Trace ends
First part of trace ends
Create Block B
If block B was never deleted while trace
was collected, its lifetime must be greater than
D2
8
Conclusions (I)

• Different systems show
• Different I/O loads
• Different block lifetimes and lifetime
  distributions

9
Conclusions (II)

• Block overwrites
• Cause the most significant fraction of deleted
  blocks
• Show substantial locality
• A small write buffer is sufficient for nearly all
  workloads
• Ideal write delay vary with workload
• Standard 30 second delay
• Slightly longer delays

10
Conclusions (III)

• Even small caches can sharply reduce disk read
  traffic
• Very large caches will not cause disk traffic to
  be dominated by writes
• A small number of memory mapped files are shared
  among many active processes
• Should keep file in memory as long as it is
  memory-mapped by any process