Dynamic Index Pruning for Effective Caching

1
Dynamic Index Pruning for Effective Caching
Yohannes Tsegay Andrew Turpin Justin
Zobel School of Computer Science IT RMIT
University, Melbourne, Australia
Introduction
Gradual stop
Query t1, t2, t3

• Information Retrieval system make use of an
  inverted index to allow efficient query
  resolution.
• An inverted index is built of two main
  components, a vocabulary of all the unique terms
  in the collection, and for each unique term, an
  inverted list of documents in which that term
  occurs.
• The size of inverted lists grows proportionally
  to the size of the collection.
• On average, larger collections require more
  computation time and resources to evaluate a
  query.
• A significant amount of time spent evaluating
  queries using inverted list is performed fetching
  lists from disk.
• To reduce the amount of time spent accessing
  disk, search engines use inverted list caching.
• Caching large inverted lists can be
  counterproductive, as other cache items will be
  evicted to make room for the large lists,
  resulting in a large number of cache misses.

• Process blocks in decreasing order of impact
  weight.
• Add new documents to A, if its impact weight is
  min_score_doc(A).
• To ensure documents in A are in the correct
  rank, continue processing more impact blocks. Do
  not add, only update accumulators already in A.

5
4
2
1
t1
4
3
1
t2
t3
5
2
1
min_score_doc(A)
Accumulators (A)

• Stop processing further blocks if, after
  processing a large number of documents, the order
  of documents in A does not change. In this case A
  is said to have become stable.

Impacts
Query evaluator
Query evaluator
Each block contains equi-impact documents
Inverted list cache
dog
7
5
4
2
On disk index
On disk index
Query evaluation without caching
Query evaluation with caching
Documents are stored in blocks. Blocks are sorted
in decreasing order of impact
Contributions
Test data

• 1. Scheme to cache dynamically pruned inverted
  lists, instead of full inverted lists.
• This has the advantage of not only increasing the
  number of items maintained in cache, it also
  reduces the number of accumulators used to
  process a query.
• 2. Gradual stop, a method for dynamic pruning
  inverted lists based on impacts.
• 3. Cost aware cache eviction policies which take
  into consideration the cost of reading an
  inverted lists into cache.

• We used two large TREC collections, WT100g (100
  GB) and GOV2 (425 GB).
• Each collection has a corresponding query log, 2
  million queries from MSN search engine and 2
  million queries from Excite query log.
• The last one million queries from each log were
  run against their corresponding collection, using
  the Gradual Stop. The inverted lists used to
  process those queries were cached. Term hits and
  byte hits were counted.

Results
1. How did the proposed pruning method perform?
2. How did caching pruned lists perform?
Cost aware cache eviction policies

• Current cache eviction polices make use of
  recency and access frequency of items in cache.
• Two items of the same recency/access frequency
  are about to be evicted. One costs twice as much
  to read from disk as the other, which should we
  keep to reduce disk access?
• The cost of re-reading an inverted list should
  be considered before evicting it from cache.
• Cost per byte
• Greedy Dual Size with Frequency

Term hit rate percentage of inverted lists found
in cache to the total number of lists processed
for the GOV2-MSN data set
di disk access cost si size of inverted
list ti age of inverted list in cache fi
number of times list was accessed while in
cache L cost of the last item evicted,
initially 0
Acknowledgments Thanks to Microsoft for
providing the MSN query log. This research was
supported by the Australian Research Council.
CIKM, November 2007