MASSIVE ARRAYS OF IDLE DISKS FOR STORAGE ARCHIVES

1
MASSIVE ARRAYS OF IDLE DISKS FOR STORAGE ARCHIVES

• D. Colarelli
• D. Grunwald
• U. Colorado, Boulder

2
Highlights

• Paper proposes
• To replace tape libraries by large non-redundant
  arrays of disks
• To cache on active drives
• Files that have been recently accessed
• Update logs for other files
• To keep other drives mostly inactive by spinning
  them down between accesses

3
Introduction (I)

• Robotic tape libraries are now the standard
  solution for archiving very large amounts of data
• Disadvantages include
• Slow access timesaverage search time of 41s for
  T9940 drives
• Not much cheaper than disk drives
• Could we replace tem by massive arrays of hard
  drives?

4
Introduction (II)

• Major limitation of hard drive solution is power
  consumption
• Almost ten times that of equivalent tape library
• Could power down disks that are not currently
  accessed
• 50 of data are likely to be never accessed
• 25 of data are likely to be accessed once

5
Introduction (III)

• Must be at least as reliable as tape libraries
• No need to use a redundant scheme
• Solution is Massive Array of Inactive Drives
• Paper investigates design issues through
  trace-driven simulations

6
Design Issues

• Two major design decisions
• Data migration or duplication (caching)
• File system or block-level interface

7
Migration or caching

• Migration would move hot data to active drives
• Migration uses disk space more efficiently
• Requires a map or directory mechanism that maps
  the storage across all drives

• Caching would cache read data and act as a write
  log for write data
• Keeps two copies of all cached files
• Maps or directories are proportional to size of
  cache

8
File system or block interface

• Could use file system information to cache entire
  files
• Would probably perform better
• Would require system modifications

• Would work with existing systems

9
MAID with caching
Passive drives(spin up/down)
Active drives (always on)
Passive Drive Manager
Cache Manager
Virtualization Manager
10
Design choices (I)

• Compared MAID-cache and MAID-no cache
• MAID-cache
• Caches read and writes on active drives
• Caching unit is chunk of 64 sectors
• Cache policy is LRU
• All writes are placed in the cache write-log
  where they wait to be committed to the
  non-active (passive) drives

11
Design choices (II)

• Must always check write log before reading data
  from the cache or the passive drives
• Passive drives remain on standby until
• A cache miss occurs
• The write log becomes too long
• Return to standby when spin-down inactivity time
  limit is reached
• Varying time limit is primary way to affect
  system performance and energy consumption

12
Simulation parameters

• Power management policy
• Always on
• Fixed-delay spin-down
• Adaptive spin-down
• Data layout
• Linear keep successive blocks on same drive
• Striped the opposite
• Caching/No caching

13
Simulation results

• Based on a supercomputer center workload
• All MAID configurations achieve similar power
  consumptions
• 15 to 16 of that of always on configuration
• MAID configurations w/o cache have average
  response times comparable to that of always on
  configuration
• Workload had little locality

14
Simulation results (II)

• Average response times of MAID configurations
  with cache much worse than that of always on
  configuration
• 0.680 to 0.720 s compared to 0.303 s
• Striped configuration with fixed spin-down delay
  has lowest average response time of all MAID
  configurations
• 0.309 s

15
Conclusion

• MAID can achieve average response times
  comparable to that of an always on configuration
  with a much lower power consumption

IMPORTANT In a more
recent paper, the authors found out that cached
configurations worked much better for workloads
exhibiting more locality of accessesthan their
supercomputer center workload