Roger ZimmermannCOMPSAC 2004, September 30

1
Spatial Data Query Support in Peer-to-Peer Systems

• Roger Zimmermann, Wei-Shinn Ku, and Haojun Wang
• Computer Science Department
• University of Southern California
• Los Angeles, CA 90089
• COMPSAC 2004

2
Outline

• Motivation
• Introduction to DHTs (CAN)
• Technical Approach
• Results
• Conclusions and Future Research

3
Motivation

• Spatial data sets are used for many applications,
  e.g., GIS, CAD,
• P2P systems provide a distributed platform that
  is very scalable.
• Pros
• Scalability, no central point of failure
• Cons
• Very dynamic (unreliable), topology maintenance
  required

4
Motivaton (cont.)

• Question how to use P2P systems for spatial data
  sharing.
• Query Challenges
• Unstructured P2P systems querying by flooding is
  not efficient
• Structured P2P systems based on DHTs (Chord,
  CAN) only efficient exact match queries are
  supported
• E.g., search files based on their
  names/titlesput(key, value) get(key) return
  value

5
Motivation (cont.)

• Spatial queries are usually range queries
• Intersect, overlap
• Nearest neighbor(s) (kNN)
• DHTs are not suitable without modification

6
Distributed Hash Tables (DHT)

• DHT systems Content Addressable Network (CAN),
  Chord, Pastry, etc.
• Using DHT to allocate large data sets to many
  nodes with no central control
• Data objects are near uniformly distributed
  through a hash function, resulting in superb
  scalability and load balance
• Each node only maintains a small routing table to
  know its neighbors
• Locating a particular data object requires
  O(logN) search steps on average

7
Content Addressable Network (CAN)

• A scalable indexing mechanism in a P2P network
• Creates a logical d-dimensional Cartesian
  coordinate space
• Divides the space into zones, where each zone is
  controlled by a node in the system
• Zones are dynamically partitioned or merged as
  nodes join and leave
• Each Zone is addressed with a Virtual Identifier
  (VID), which is deterministically calculated from
  the location of the zone

8
Content Addressable Network (CAN)
Example A 2-D space partitioned into 7 CAN zones
9
Content Addressable Network (cont)

• Node Operations
• (e.g., Insertion)
• Find a bootstrap node first

10
Content Addressable Network (cont)

• 2. Randomly choose a point in the CAN plane and
  route the new node from the bootstrap node to the
  chosen location

11
Content Addressable Network (cont)

• 3. The new node arrives at the destination zone
  covering that point. The destination zone is
  split into two zones, each controlled by one node
  (old and new)

12
Content Addressable Network (cont)

• 4. Update the neighborhood zone routing
  information

13
Content Addressable Network (CAN)

• Data Object Operation (e.g. Insertion)
• Generate a key based on the object identification
  and insert data object as a ltkey, valuegt pair
• Map the key into a point P in the CAN plane by
  using a uniform hash function
• Store the ltkey, valuegt pair at the node which
  owns the zone within which the point P is located
  
• To retrieve the value, the same hash function is
  applied to the key in order to regenerate the
  point P and find the zone owns that point, the
  zone will return the value to the client

14
Storing Spatial Data w/ DHTs

• Hash function distributes data objects evenly
  within the space to achieve a balanced load
• Spatial locality information needs to be
  preserved for range queries. Applying a hash
  function to spatial data will destroy locality
• Related work explored storing R-tree or Quad-tree
  based index on DHT
• Harwood et al. Hashing Spatial Content over
  Peer-to-Peer Networks
• Mondal et al. P2PR-tree An R-tree-based Spatial
  Index for Peer-to-Peer Environments

15
Spatial Range Query Design for P2P Systems

• Mapping a physical space to a CAN space
• Propose a new hash function to map spatial data
  objects onto nodes over a modified CAN system
• Purpose allow efficient spatial data query
  execution while at the same time considering load
  balance
• Calculating the location of zones in the logical
  space Virtual Identifier (VID) tree for mapping
  purpose

16
Spatial Range Query Design for P2P Systems

• Approach
• Object key is generated with three different
  components
• (a) Scatter region address based on the spatial
  locality of the object preserves spatial
  locality.
• (b) Zone address randomized achieves load
  balance
• (c) Object identifier (hashed)
• The scatter region size is fixed and predetermined

17
Spatial Range Query Design for P2P Systems
(cont.)

• The value of zone bit string is decided randomly
  and the object identifier is the data content
  hash result
• The VID tree is created with its height
  determined by the scatter region size
• The maximum number of zones is 2(ab)
• The relationship between data locality and load
  balance can be determined along a spectrum

18
Spatial Range Query Design for P2P Systems
(cont.)

• Scatter region(11000)e.g.a5 bits

000
001
010
011
100
101
110
111
11
10
01
00
19
Spatial Range Query Design for P2P Systems
(cont.)

• Zonese.g.b4 bits

000
001
010
011
100
101
110
111
11
10
01
00
20
Spatial Range Query Design for P2P Systems
(cont.)

• System Operation and Spatial Range Query
• Node Operation
• Bootstrap mechanism
• Node join mechanism
• Zone split and the search threshold
• Balance the number of data objects in each zone
• The zone being selected must be larger than the
  minimum zone size (1/2(ab))
• The threshold is the upper bound on the number of
  search hops to find a zone to split
• Data Object Insertion
• Data Object Deletion
• Spatial Range Query

21
Spatial Range Query Design for P2P Systems
(cont.)
Spatial Range Query Step 1 The querying node
launches a spatial range query.
22
Spatial Range Query Design for P2P Systems
(cont.)
Spatial Range Query Step 2The node determines
the overlapping scatter regions.
23
Spatial Range Query Design for P2P Systems
(cont.)
Spatial Range Query Step 3The node multicasts
the query to the overlapping scatter regions.
24
Spatial Range Query Design for P2P Systems
(cont.)

• Step 4
• The range query is multicast within all
  overlapping scatter regions (M-CAN).
• Recall data is randomized within each scatter
  region, so an exhaustive search is necessary
• Choice of scatter region size
• Large good load balance uniform within a
  scatter region
• Small exhaustive search covers less area

25
Conclusions and Future Research Directions

• We proposed a hash function to preserve both
  spatial locality information and constrained load
  balance
• The proposed mechanism works will with CAN P2P
  architecture
• We are currently running simulations to test our
  approach

26
Thank you! Questions?