TRACK-ALIGNED EXTENTS: MATCHING ACCESS PATTERNS TO DISK DRIVE CHARACTERISTICS

1
TRACK-ALIGNED EXTENTSMATCHING ACCESS
PATTERNSTO DISK DRIVE CHARACTERISTICS
STILL INCOMPLETE

• J. SchindlerJ.-L.GriffinC. R. LumbG. R. Ganger
• Carnegie Mellon University

2
Key Ideas

• Track-based disk access could improve performance
  of modern disk drives by up to 50
• Track boundaries can be detected at a reasonable
  cost
• This knowledge can be used by existing file
  systems without creating hardware dependencies
• The technique results significant improvements of
  application performance

3
MOTIVATION

• Modern disk drives virtualize their storage space
  as a flat array of fixed-size blocks
• Prevents OS from taking into account actual
  characteristics of disk drive
• Sole feasible optimization technique is
  increasing block size
• Introduces several problems
• Has limited benefits

4
Increasing block size
5
Increasing block sizes (contd)

• Achieving good disk efficiency requiresvery
  large blocks
• Only 50 efficiency for random writes with 256K
  blocks
• Aligning blocks with track boundaries achieves
  much higher efficiencies
• Best results obtained when block size is multiple
  of track size (264KB in example)

6
TRACK-BASED DISK ACCESS

• Cannot keep increasing disk request sizes because
  large disk requests
• Have higher latency
• Require more I/O buffer space
• Work best when applications access sequentially
  very large files
• Require collocation of related small files

7
Disk characteristics

• Head switches occur anytime a single request
  accesses blocks that span two tracks
• Cause a delay of 0.6 to 1ms
• Are not likely to diminish in the near future
• Zero-latency access allows disk firmware to read
  data blocks in the order they can be read rather
  than in the requested order

8
Example

• Disk firmware wants to read sector 200-299
• Disk head located before sector 250
• Firmware will read
• Fist sectors 250 to 299
• Then sectors 200 to 249

9
TRACK-AWARE SYSTEM DESIGN

• Locating track boundaries is not easy because
• Outer tracks have more sectors than inner tracks
• Tracks have spare blocks
• Defect handling schemes vary widely among disk
  makes and models
• Need be performed once
• Defects only appear during first 48 hours of
  operation

10
Allocation and access (I)

• To use track boundary information, file system
  must support variable extents
• Extent-based file systems specify ranges of LBNs
  allocated to each file (extents)
• Should always specify extents that fit track
  boundaries

11
Allocation and access (II)

• Block-based file systems, such as FFS, group LBNs
  into fixed-size groups of 2n sectors called
  blocks
• Must ensure that blocks will never span track
  boundaries
• Wastes less than 5 of disk space
• This space could still be used for other purposes

12
Allocation and access (III)

• Should also
• Extend or clip prefetch and write back requests
  based on track boundaries
• Use disk command queuing to fully take advantage
  of zero-latency disks
• Current SCSI and IDE/ATA controllers do not allow
  out-of-order delivery to or from the host

13
Prototype implementation

• Prototype includes
• Two techniques for detecting track boundaries
• A modified version of FreeBSD FFS

14
Detecting track boundaries (I)

• First technique
• Identifies discontinuities in access efficiency
• Could be done by linearly increasing number of
  sector in each I/O request
• Uses instead binary search algorithm and tries to
  predict size of next extent
• Still very slow (4 hours for 9GB disk)

15
Detecting track boundaries (II)

• Second technique
• Specific to SCSI disks
• Extract disk information through SCSI commands
• Exploits the regularity of disk geometry
• Much faster ( lt one minute per disk)
• Does not always work when disk mapping scheme is
  not knwon

16
FreeBSD FFS overview

• Each block has
• An lblkno (logical block number) specifying its
  offset from the beginning of the file
• A blkno (physical block number) , which is an
  abstract representation of disk addresses used by
  the OS
• Each blkno corresponds to a range of contiguous
  disk sector numbers

17
Free BSD modifications

• Exclude from all allocation decisions all blocks
  that span track boundaries
• Mark them as used in the free block map
• FFS clustered read-ahead algorithm
• Accesses runs of blocks between excluded blocks
  with a single request
• Must still take care of track boundaries without
  excluded blocks

18
Evaluation

• Using
• Two disks supporting zero-latency access
• Two disks supporting it
• Also used Disksim simulator to simulate disks

19
Disk performance

• Measured request head time, that is, amount of
  time that the head is dedicated to a request
• Biggest improvements were for disks supporting
  zero-delay access
• Efficiency improvements of up to 50
• Also reduced standard-deviation of response times

20
Other experiments

• With a modified FFS file system
• Not that different timings
• With a video server
• Track extents resulted in much lower startup
  latency at high arrival rates
• A log-structured file system