Special Topics in Computer Science Advanced Topics in Information Retrieval Lecture 7 (book chapter 9): Parallel and Distributed IR

1
Special Topics in Computer Science Advanced
Topics in Information Retrieval Lecture 7 (book
chapter 9) Parallel and Distributed IR

• Alexander Gelbukh
• www.Gelbukh.com

2
Previous Chapter Conclusions

• How to accelerate search? Same results as
  sequential
• Ideas
• Quick-and-dirty rejection of bad objects, 100
  recall
• Fast data structure for search (based on
  clustering)
• Careful check of all found candidates
• Solution mapping into fewer-D feature space
• Condition lower-bounding of the distance
• Assumption skewed spectrum distribution
• Few coefficients concentrate energy, rest are
  less important

3
Previous Chapter Research topics

• Object detection (pattern and image recognition)
• Automatic feature selection
• Spatial indexing data structures (more than 1D)
• New types of data.
• What features to select? How to determine them?
• Mixed-type data (e.g., webpages, or images
  withsound and description)
• What clustering/IR methods are better suited
  forwhat features? (What features for what
  methods?)
• Similar methods in data mining, ...

4
The problem

• Very large document collections
• Google 4,000,000,000 pages
• Slow response?
• Solution parallel computing
• Google 10,000 computers

5
Parallel architectures
Data stream Data stream
Single Multiple
Instruction stream Single SISD classical SIMD simple
Instruction stream Multiple MISD (rare) MIMD many SISD
6
MIMD architecture

• The most common
• Can be
• tightly coupled
• loosely coupled
• Distributed
• Many computers interacting via network
• PC Clusters
• Similar to MIMD computers, but greater cost of
  communication
• very loosely coupled
• More coarse-grained programs

7
Performance improvement

• Time speedup S
• Ideally, N times (number of processors)
• In practice impossible
• The problem does not decompose into N equal parts
• Communication and control overhead
• lt 1 / f, where f is the largest separable
  fraction of theproblem
• Cost
• Per processor S / N

8
Two approaches to parallelism

• Build new algorithms
• E.g., neural nets
• Naturally parallel
• Problem to define the retrieval task
• Adapt the existing techniques to parallelism
• Allows relying on well-studied approaches
• We will consider this option

9
Ways to use parallelism

• Multitasking
• N search engines
• Good for processing many queries
• Problems
• A single query is not speeded up
• Bottleneck disk access (index)
• Possible solution replicating (part of) data.
  RAIDs
• Parallel algorithms
• IR data. Main question how to partition the
  data
• Document / index term matrix(terms can be LSI
  dimensions, signature bits, etc)

10
Possible partitionings

• Horizontal document partitioning. Union of
  results
• Vertical term partitioning. Basically, intersect
  results

11
Inverted files Logical partitioning

• Logical vs. physical document partitioning
• Logical for each term, use pointers into
  inverted file data for each processor, to
  indicate its portion

12
Inverted files Logical partitioning
Construction and updating

• Also parallel
• Construction
• Assign docs to processors
• Order docs such that each processor has an
  interval
• Process in parallel
• Merge. Each piece is ordered already

13
Inverted filesPhysical document partitioning

• Several separate collections, one per processor
• Separate indices
• Then the lists are merged (they are already
  ordered)
• Priority queue is used
• The result is not sorted Insertion is quick
• The maximal element can be found quickly
• First k elements can be found rather quickly
• Details in the book
• Consistent scores are needed
• Global statistics is needed. Can be computed at
  index time

14
Logical or physical partitioning?

• Logical requires less communication
• Faster
• Physical is more flexible. Simpler implementation
• Simpler conversion of existing systems

15
Inverted files Term partitioning

• Each processor processes a part of the inverted
  file
• The results are intersected (for AND)
• (or as appropriate for Boolean operations, OR and
  NOT)
• When term distribution in user queries is
  skewed,then document partitioning is better
• When uniform, term partitioning is better.
• Twice for long queries, 5 10 times for short
  (Web-like)

16
Suffix arrays

• Array construction can be parallelized
• merges are parallel
• Document partitioning is applied
  straightforwardly
• Each processor maintains its own suffix array
• Term partitioning can be applied
• Each processor owns a branch of the tree
  (lexicographicinterval)
• Bottleneck all processors need access to the
  entire text

17
(No Transcript)
18
Signature files

• Document partitioning straightforward
• Create query signature, distribute to each
  processor
• Merge results (using Boolean operations if
  needed)
• Term partitioning shorter signatures
• Merging and eliminating false drops is slow
• This method is not recommended

19
SIMD computers

• Single Instruction, Multiple data
• Uncommon
• Good for simple operations
• Bit operations in signature files
• Details in the book
• Ranking is supported in hardware in some
  computers
• If signature file does not fit into memory, can
  beprocessed in batches
• I/O overhead
• Use multiple queries with the same batch
• This improves throughput, but not response time

20
SIMD computers

• Inverted files are difficult to adapt to SIMD
• The inverted file is restructured
• Details in the book

21
Distributed IR

• MIMD with
• Slow communication
• Not all nodes are used for a given query
• Encryption issues
• Document partitioning is usually used
• Term partitioning imposes greater
  communicationoverhead
• Document clustering can be useful (to distribute
  docs by processors)
• Index clusters and then search only the best ones
• Another approach use training queries, then
  similarity of the user query to these

22
Research topics

• How to evaluate the speedup
• New algorithms
• Adaptation of existing algorithms
• Merging the results is a bottleneck
• Meta search engines
• Creating large collections with judgements
• Is recall important?

23
Conclusions

• Parallel computing can improve
• response time for each query and/or
• throughput number of queries processed with same
  speed
• Document partitioning is simple
• good for distributed computing
• Term partitioning is good for some data
  structures
• Distributed computing is MIMD computing with slow
  communication
• SIMD machines are good for Signature files
• Both are out of favor now

24
Thank you! Till May 17? 18?, 6 pm