SKYPEER: Efficient Subspace Skyline Computation over Distributed Data

1

• SKYPEER Efficient Subspace Skyline Computation
  over Distributed Data
• Akrivi Vlachou, Christos Doulkeridis, Yannis
  Kotidis
• Department of Informatics
• Athens University of Economics and Business
• ICDE 2007
• 69721040 ???

2
Outline

• Introduction
• Skyline Computation in P2P Networks
• SKYPEER Algorithm
• Experimental Evaluation
• Conclusion

3
Introduction

• The skyline query is used to find a set of non
  dominated data points in a multi-dimensional
  dataset, and most previous work has assumed a
  centralized setting
• The past assumption are hardly applicable to
  large-scale P2P systems.
• Relying on a super-peer architecture we propose a
  threshold based algorithm, called SKYPEER and its
  variants
• For efficient subspace skyline processing, we
  extend the notion of domination by defining the
  extended skyline set
• Skyline queries help users make intelligent
  decisions over complex data

4
Skyline Computation in P2P Networks

• In general, super-peers maintain information
  about the peers they have been assigned, so that
  at query time, they can process a query without
  having to contact all peers.
• In this work we assume that the super-peer
  topology is predefined and we focus on the
  optimization of interactions among super-peers
  and peers.
• Preliminaries and Definitions
• Querying peer Pinit, henceforth referred to as
  initiator of a query
• All peers send their local datasets to Pinit
  where a centralized skyline algorithm is executed

5
Skyline Computation in P2P Networks
6
Skyline Computation in P2P Networks

• Locally evaluate as many parts of the query as
  possible.
• Each super-peer needs to collect from the
  associated peers only the skyline points of all
  subspaces.
• Given the locally stored extended skyline, each
  super peer individually processes a subspace
  skyline request and transmits the results to the
  query initiator.
• This approach is considered as baseline, and will
  be henceforth referred to as naive.

7
(No Transcript)
8
Skyline Computation in P2P Networks

• Extended Skyline
• we adjust the dominance definition to compute all
  necessary values co-instantaneously during the
  skyline calculation.
• Definition 1.
• For any dimension set U, where U ? D, p
  ext-dominates q if on each dimension di ? U, pi
  ltqi. The ext-skyline (ext-SKYU) is set of all
  points that are not ext-dominated by any other.

9
Skyline Computation in P2P Networks
10
SKYPEER Algorithm

• Mapping
• each d-dimensional point p is transformed to a
  one dimensional value f(p) based on the formula
• Observation 5.
• Let psky be a skyline point in a subspace U.A
  point p for which the following inequality holds
  cannot be a skyline point in subspace U.

11
SKYPEER Algorithm
12
SKYPEER Algorithm

• Merging
• Each peer executes a local subspace skyline
  computation
• The initiator peer computes the overall subspace
  skyline result by merging the local results

fig2
13
SKYPEER Algorithm

• Optimization
• 1. Threshold propagation
• (i) Fixed Threshold
• Pinit calculates its threshold t for q(U, t)
  and forwards the threshold value to all
  super-peers.
• (ii) Refined Threshold
• Pinit calculates and sends its threshold to
  its neighboring super-peers, which do not
  forward it immediately to other super-peers, but
  rather they first compute the subspace
  skyline, calculate the new threshold t, and then
  forward q(U, t).

fig1
fig2
14
SKYPEER Algorithm

• 2. Merging strategy
• (i) Fixed Merging
• All super-peers forward their computed
  subspace skyline back to Pinit, and Pinit is
  responsible for merging the results and computing
  the resulting subspace skyline for q(U).
• (ii) Progressive Merging
• Each super-peer merges the results it receives
  with its locally computed subspace skyline,
  before sending the results back to the
  super-peer from which it originally received
  the query.

fig1
fig2
15
Experimental Evaluation
16
Experimental Evaluation
17
Conclusion

• 4 variants are more efficient than naive
• Subspace?