The Impact of Global Clustering on Spatial Database Systems

1
The Impact of Global Clustering on Spatial
Database Systems

• VLDB Conference 94
• Thomas Brinkhoff Hans-Peter Kriegel
• 2000/5/16
• ??? ???

2
Contents

• Introduction
• Queries in Spatial Database Systems
• The Storage of Spatial Objects
• The Cluster Organization
• Evaluation

3
Introduction(1/3)

• Characteristics of Spatial Databases
• Manage high numbers of objects
• A high variation in their complexity
• Need selective spatial access
• Spatial Access Methods
• Spatially adjacent objects are arbitrarily
  distributed

4
Introduction(2/3)

• Global Clustering
• A set of data pages representing spatially
  adjacent objects is stored on consecutive pages
  of the disk.
• Problem
• A global reorganization of all objects in the
  database is not reasonable in a dynamic
  environment

5
Introduction(3/3)

• Goals of this paper
• Obtain an evaluation of the importance of several
  techniques for global clustering
• The impacts of global clustering on spatial joins
  have not been investigated

6
Queries in Spatial DB Systems

• Point Query
• Window Query
• Spatial Join
• A ?? B
• Intersection join

i? j
7

• Introduction
• Queries in Spatial Database Systems
• The Storage of Spatial Objects
• Spatial Access Method
• Clustering
• Organization Model
• The Cluster Organization
• Evaluation

8
Spatial Access Method(1/2)

• Access Methods
• Organize a dynamic set of objects in secondary
  storage
• B-tree or linear hashing are not suitable
• SAM
• group spatial objects which are close to each
  other in data space close to each other in the
  data pages

9
Spatial Access Method(2/2)
10
Clustering(1/2)

• Access time
• Seek time latency time transfer time
• Goal
• Minimize the number of seek operations and the
  rotational delay in order to reduce access cost.

11
Clustering(2/2)
Page

• Internal Clustering
• Local clustering
• Global clustering

Object
Page
O1
O2
O3
O6
O1
O2
O5
O4
O3
12
Organization Models(1/3)-secondary organization

• Primary index for approximation
• Second index for spatial objects

13
Organization Models(2/3)-Primary organization

• The exact representation of the objects are on
  the datapages.

14
Organization Models(3/3)-for Global Clustering

• Combine sets of pages with larger storage units,
  Custer Units

15

• Introduction
• Queries in Spatial Database Systems
• The Storage of Spatial Objects
• The Cluster Organization
• Requirements
• SAM R-tree
• Cluster Organization
• Modification of R-tree
• Evaluation

16
The Cluster Organization(1/5)

• Requirements
• SAM with high quality space partitioning scheme
• Support insertion and deletion
• 3 queries should be efficiently supported
• Maximum cluster size exists
• A reasonable storage utilization
• SAM
• R-tree
• One of the most efficient variants of the R-tree

17
The Cluster Organization(2/5)-The Cluster
Organization

• The Cluster Organization
• Static definition of the size of a cluster unit
• Cluster all objects in a cluster unit whose MBRs
  are stored in one data page.

18
The Cluster Organization(3/5)-The Cluster
Organization

• Page size 4kb
• Entry size 46Bytes
• Smax1.5M Sobj

• Storage utilization 66
• Per Cluster unit 58 objects

19
The Cluster Organization(4/5)-Modification of
the R-tree

• Size of all objects in one cluster unit gt the
  maximun clsuter size Smax
• split data page and the cluster size
• Splitting
• No re-insertion

20
The Cluster Organization(5/5)-Processing
21
Evaluation (1/9)

• TEST DATA
• Seek time average 9msec
• Latency time average 6msec
• Transfer time average 1msec

22
Evaluation (2/9)

• Storage Utilization
• Bad Utilization of Cluster organization

23
Evaluation (3/9)

• Buddy System
• Each physical unit has the size Smax2-i (igt0)
• Each cluster unit has the buddy of the smallest
  possible size
• Cluster gt Buddy
• ? into a bigger buddy
• Split
• ? into a smaller buddy
• Restricted Sizes
• (Smax,, 0.5Smax, 0.25Smax)

24
Evaluation (4/9)

• Windows Queries
• Transfer the complete cluster unit
• This is the handicap until now

25
Evaluation (5/9)

• Geometric Threshold
• use the degree of overlap between the region of a
  cluster unit and the query window
• Use threshold T
• If (T lt degree) send page-by-page
• If (T gt degree) send the cluster unit

• The SLM-Technique
• Reading requested and non-requested pages

26
Evaluation (6/9)

• Windows Queries

• The SLM-technique is the best choice

27
Evaluation (7/9)

• Point Queries

28
Evaluation (8/9)

• Spatial Join
• a
• 86,094 pairs of intersect
• b
• 1.2 million pairs of intersect

29
Evaluation (9/9)

• Impact of Global Clustering on the performance of
  a Complete Spatial Join

30
Conclusion

• Global Clustering speeds up the access to spatial
  objects for large window queries as well as for
  spatial joins.
• Using a buddy system, it shows a good storage
  utilization.
• SLM-technique is the best choice for window
  queries.

31
(No Transcript)
32
(No Transcript)