Providing High Reliability in a Minimum Redundancy Archival Storage System

1
Providing High Reliability in a Minimum
Redundancy Archival Storage System

• Deepavali Bhagwat
• Kristal Pollack
• Darrell D. E. Long
• Ethan L. Miller
• Storage Systems Research Center
• University of California, Santa Cruz

Thomas Schwarz Computer Engineering
Department Santa Clara University Jehan-François
Pâris Department of Computer Science University
of Houston, Texas
2
Introduction

• Archival data will increase ten-fold from 2007 to
  2010
• J. McKnight, T. Asaro, and B. Babineau, Digital
  ArchivingEnd-User Survey and Market Forecast
  2006 - 2010. The Enterprise Strategy Group, Jan.
  2006.
• Data compression techniques used to reduce
  storage costs
• Deep Store - An archival storage system
• Uses interfile and intrafile compression
  techniques
• Uses chunking
• Compression hurts reliability
• Loss of a shared chunk ? Disproportionate data
  loss
• Our solution Reinvest the saved storage space to
  improve reliability
• Selective replication of chunks
• Our results
• Better reliability compared to that of mirrored
  Lempel-Ziv compressed files using only about half
  of the storage space

3
Deep Store An overview

• Whole File Hashing
• Content Addressable Storage
• Delta Compression
• Chunk-based Compression
• File broken down into variable-length chunks
  using a sliding window technique
• A chunk identifier/digest used to look for
  identical chunks
• Only unique chunks stored

variable chunk size
sliding window
w fixed size window
chunk end/start
window fingerprint
chunk ID(content address)
4
Effects of Compression on Reliability

• Chunk-based compression ? Interfile dependencies
• Loss of a shared chunk ? Disproportionate amount
  of data loss

Files
Chunks
5
Effects of Compression on Reliability..

• Simple experiment to show the effects of
  interfile dependencies
• 9.8 GB of data from several websites, The
  Internet Archive
• Compressed using chunking to 1.83 GB. (5.62 GB
  using gzip)
• Chunks were mirrored and distributed evenly onto
  179 devices, 20 MB each.

6
Compression and Reliability

• Chunking
• Minimizes redundancies. Gives us excellent
  compression ratios
• Introduces interfile dependencies
• Interfile dependencies are detrimental to
  reliability
• Hence, reintroduce redundancies
• Selective replication of chunks
• Some chunks more important than others. How
  important?
• The amount of data depending on a chunk (byte
  count)
• The number of files depending on a chunk
  (reference count)
• Selective replication strategy
• Weight of a chunk (w) ? Number of replicas for a
  chunk (k)
• We use a heuristic function to calculate k

7
Heuristic Function

• k Number of replicas
• w Weight of a chunk
• a Base level of replication, independent of w
• b To boost the number of replicas for chunks
  with high weight
• Every chunk is mirrored
• kmax Maximum number of replicas
• As replicas increase the gain in reliability
  obtained as a result of every additional replica
  reduces
• k rounded off to the nearest integer.

8
Distribution of Chunks

• An archival system receives files in batches
• Files stored onto a disk until the disk is full
• For every file
• Chunks extracted and compressed
• Unique chunks stored
• A non unique chunk stored only if
• The present disk does not contain the chunk
• For this chunk, k lt kmax
• At the end of the batch
• All chunks revisited and replicas made for
  appropriate chunks
• A chunk is not proactively replicated
• Wait for a chunks replica to arrive as a chunk
  of a future file
• Reduce inter-device dependencies for a file.

9
Experimental Setup

• We measure Robustness
• The fraction of the data available given a
  certain percentage of unavailable storage devices
• We use Replication to introduce redundancies
• Future work will investigate erasure codes
• Data Set
• HTML, PDF, image files from The Internet Archive.
  (9.8 GB)
• HTML, image (JPG and TIFF), PDF, Microsoft Word
  files from The Santa Cruz Sentinel (40.22 GB)
• We compare the Robustness and Storage Space
  utilization of archives that use
• Chunking with selective redundancies and
• Lempel-Ziv compression with mirroring

10
Details of the Experimental Data
11
Weight of a Chunk

• When using dependent data (byte count) as a
  heuristic
• w D/d
• D sum of the sizes of all files depending on a
  chunk
• d average size of a chunk
• When using the number of files (reference count)
  as a heuristic
• w F
• F number of files depending on a chunk

12
Robustness, Effect of varying a, wF, b1,
kmax4, The Internet Archive
13
Robustness, Effect of varying a, wD/d,b0.4,
kmax4, The Internet Archive
14
Robustness, Effect of limiting k, wD/d, b0.55,
a0, The Internet Archive
15
Robustness, Effect of varying b, wD/d, a0,
kmax4, The Internet Archive
16
Robustness, Effect of varying b, wD/d, a0,
kmax4, The Sentinel
17
Choice of a Heuristic

• Choice of a heuristic depends on the corpus
• If file size is indicative of file importance,
  choose wD/d
• If files importance is independent of its size,
  choose wF
• Use the same metric to measure robustness

18
Future Work

• Study reliability of Deep Store
• With a recovery model in place
• When using delta compression
• Use different redundancy mechanisms such as
  erasure codes
• Data placement in conjunction with hardware
  statistics

19
Related Work

• Many archival systems use Content Addressable
  Storage
• EMCs Centera
• Variable-length chunks LBFS
• Fixed-size chunks Venti
• OceanStore aims to provide continuous access to
  persistent data
• uses automatic replication for high reliability
• erasure codes for high availability
• FARSITE a distributed file system
• Replication of metadata, replication chosen avoid
  the overhead of data reconstruction when using
  erasure codes
• PASIS, Glacier use aggressive replication as a
  protection against data loss
• LOCKSS provides long term access to digital data
• uses peer-to-peer audit and repair protocol to
  preserve the integrity and long-term access to
  document collections

20
Conclusion

• Chunking gives excellent compression ratios but
  introduces interfile dependencies that adversely
  affect system reliability.
• Selective replication of chunks using heuristics
  gives
• better robustness than mirrored LZ-compressed
  files
• significantly high storage space efficiency --
  only uses about half of the space used by
  mirrored LZ-compressed files
• We use simple replication. Our results will only
  improve with other forms of redundancies.

21
Providing High Reliability in a Minimum
Redundancy Archival Storage System

• Deepavali Bhagwat
• Kristal Pollack
• Darrell D. E. Long
• Ethan L. Miller
• Storage Systems Research Center
• University of California, Santa Cruz

Thomas Schwarz Computer Engineering
Department Santa Clara University Jehan-François
Pâris Department of Computer Science University
of Houston, Texas