WOLF A Novel Reordering Write Buffer to Boost the Performance of LogStructured File Systems

1
WOLF A Novel Reordering Write Buffer to Boost
the Performance of Log-Structured File Systems

• Jun Wang, Yiming Hu
• Proceedings of the FAST 2002 Conference on File
  and Storage Technologies
• Jan, 28-30, 2002

2
Log Structured File System

• Writes all modifications to disk sequentially in
  a log-like structure
• Speeding up both file writing and crash recovery
• Log is the only structure on disk
• Segment cleaner (garbage collector) needed

3
Problems of LFS (1/2)

• Garbage collection overhead(a.k.a. segment
  cleaning overhead)
• Especially under random updates and accesses
• Several schemes have been proposed to speed up
  garbage collection process

4
Problems of LFS (2/2)

• Proposed schemes
• Greedy/benefit-to-cost cleaning
• Hole-plugging policy
• Adaptive algorithm
• And more
• All concentrated to make cleaner efficiently
  after files are written to disk

5
New Scheme (1/2)

• WOLF
• reordering Write buffer Of Log-structured File
  system
• Uses two or more segment buffers
• When write data arrives, system sorts them into
  different buffers according to their expected
  longevity
• One for active data, another for less-active
  data,

6
New Scheme (2/2)

• WOLF (contd)
• Leads to bimodal distribution
• Segments are either mostly full ormostly
  empty
• Cleaner can select many nearly empty segments to
  clean and compact into a small number of segments
• By re-organizing data in RAM before they reach
  the disk, this scheme also make the system do
  less garbage collection work

7
LFS writing process 1
Three kinds of blocks
Valid data block
Empty block
Invalidated block
Data blocks first enter a segment buffer
Data
Buffer
Note This slide and following 3 slides are
copied from author(Jun Wang)s own FAST 2002
presentation slides.
8
LFS writing process 2
Buffer will be written to the disk when full

Disk
After a while, many blocks in segments are
invalidated, Leaving holes and require garbage
collection

Disk
9
WOLF writing process 1
Four kinds of blocks
Valid active block
Empty block
Valid inactive block
Invalidated block
Data blocks first enter one of two segment
buffers based on expected activities
Data
Buffer1
Buffer2
10
WOLF writing process 2
Buffer will be written to the disk when full

Disk
After a while, most blocks in active segments are
invalidated,While most in the inactive segments
are intact.

Disk
11
WOLF Design Write (1/2)

• Key problem
• How to separate active data from inactive data
  and put them into different buffers accordingly

12
WOLF Design Write (2/2)

• Heuristic adaptive grouping algorithm
• Each block in reorder buffer has a reference
  count (initially zero)
• Reference count is incremented when the block is
  accessed
• Count is reset to zero after certain period of
  time (timebar(initially 10min))
• Timebar could be adaptively tuned for various
  incoming accesses (doubled, or halved)

13
WOLF Design Read

• WOLF only changes the write cache structure of
  LFS
• Read operation are not affected
• WOLF read performance is similar to that of LFS
• But, when system is heavily loaded, WOLF should
  have better read performance because of its more
  efficient garbage collection process

14
WOLF Design GC

• use adaptive approach (Matthews SOSP97)
• Select eitherbenefit-to-cost cleanning
  orhole-pluggingdepending oncost-benefit
  estimates

15
Experiment Environment(1/2)

• WOLF simulator
• LFS simulator (port from Sprite LFS)
• Disk simulator

16
Experiment Environment(2/2)

• Workload model
• Four real world traces
• INS, RES, SITAR, HARP
• Four synthetic traces
• Uniform
• Hot-and-cold
• Ephemeral small file
• Transaction processing suite (TPC-D)

17
Performance Result (1/2)
Note This slide and following slide are copied
from author(Jun Wang)s own FAST 2002
presentation slides.
18
Performance Result (2/2)
19
T h e E n d
Any question?