PARALLEL SKYLINE COMPUTATION ON MULTICORE ARCHITECTURE

1
PARALLEL SKYLINE COMPUTATION ON MULTI-CORE
ARCHITECTURE

• POSTECH DAG ???
• ????? ?? ??? ???

2
Changes
http//www.tomshardware.com/reviews/mother-cpu-cha
rts-2005,1175.html
3
Changes
CPU Core ? ??? ?? ??
4
(No Transcript)
5
Database operations?

• I/O intensive operations
• Disk reads/writes
• Random/Sequential access
• Little performance gain
• No CPU intensive DB operation?
• Skyline Query

6
Skyline Query

• Which objects are the most important or
  interesting?
• For one attribute
• For many attributes

7
The Skyline of Chicago
8
The Skyline of Hotels
9
Skyline Query

• A dominates B
• All attributes of A are better than or equal to
  their corresponding attributes of B.
• At least one attribute of A is strictly better
  than that of B
• Skyline Query
• Compute a subset of objects that are not
  dominated by any other objects from a given set.

10
Skyline Query

• The advantage of skyline query
• It is hard to determine the preference among
  attributes.
• Ex) ?? ??????? ??? 1km ??? ??? ??? 100 ???
  ???? 1.7? ????.
• Applications
• Hotel recommendation system
• Product recommendation system

11
Skyline Algorithms

• Sequential algorithms / no index structure
• BNL
• SFS
• LESS
• Sequential algorithms / index structure
• NN
• BBS
• ZSearch
• Parallel algorithms
• Several algorithms for distributed environment
• No algorithm for multi-core environment

12
Skyline Algorithms for Multi-core

• Distributivity of Skyline
• S(D union D) S(S(D) union S(D))
• Can compute a skyline from a set of local
  skylines
• Divide-and-Conquer approach

13
Skyline Algorithm for Multi-core

• Skeleton Programming Model
• Structured Parallel Programming
• Provide several patterns for parallel programming
• Ex) map, reduce,
• Simple, Safe Parallel Programming
• Simple vs. Expressive

14
Skyline Algorithm for Multi-core

• (parallel) map
• (parallel) reduce

A
B
C
D
E
F
G
H
A
B
C
D
E
F
G
H
A
B
C
D
E
F
G
H
A
B
C
D

R
15
Skyline Algorithm for Multi-core

• Two auxiliary functions
• sskyline
• Abbv. sequential skyline
• In-place skyline algorithm for local skyline
  computation
• smerge
• Abbv. sequential merge
• Merge two skylines into one skyline

16
Skyline Algorithm for Multi-core
A
B
C
D
E
A
E
C
D
A
E
C
D
C
E
D
C
E
D
C
E
D
C
E
D
C
E
D
17
Skyline Algorithm for Multi-core
A
B
C
D
E
F
G
H
18
Skyline Algorithm for Multi-core

• Now, we have
• (parallel) map
• (parallel) reduce
• (sequential) skyline
• (sequential) merge
• Then, (parallel) skyline is
• reduce merge (map skyline D,,D)?

19
Skyline Algorithm for Multi-core

• No!
• Why?

A
B
C
D
E
F
G
H

M
N
A
20
Skyline Algorithm for Multi-core

• Additional auxiliary function
• pmerge
• Abbv. parallel merge
• Merge two skyline into one skyline in parallel

Thread 2
A
B
C
D
E
F
G
H
Thread 1
21
Skyline Algorithm for Multi-core

• Now, we have
• (parallel) map
• (sequential) reduce
• (sequential) skyline
• (parallel) merge
• Finally, (parallel) skyline is
• reduce merge (map skyline D,,D)!!

22
Experimental Setup

• Dell PowerEdge server
• Two quad-core Intel Xeon 2.83 GHz CPUs
• 8GB main memory
• 4KB disk page
• Synthetic datasets
• Anti-correlated
• Independent

23
Experimental Result
24
Experimental Result
25
Criticism

• How about parallelizing existing algorithms?

26
Lets parallelize BBS
27
(No Transcript)
28

• Assume all points are indexed in an R-tree.
• mindist(MBR) the L1 distance between its
  lower-left corner and the origin.

29

• Each heap entry keeps the mindist of the MBR.

30

• Process entries in ascending order of their
  mindists.

31
(No Transcript)
32
(No Transcript)
33
(No Transcript)
34
(No Transcript)
35
(No Transcript)
36
Branched and Bound Skyline (BBS)

• Profiling
• It spends more than 90 of running time to
  compare a new candidate with existing skylines
• Lets parallelize this comparison
• Simple approach fails!!
• Optimization accumulate candidates

37
Experimental Result
38
Experimental Result
39
Conclusion

• A skyline algorithm for multi-core architecture
• pskyline
• Multi-core architecture can benefit DB operations
• It is not simple to exploit parallelism
• pbbs

40
Question?
41
Thank you