Distributed Quad-Tree for Spatial Querying in

Wireless Sensor Networks (WSNs)

Murat Demirbas, Xuming LuDept of Computer Science and Engineering,

University at Buffalo, SUNY, NY 14260

{demirbas, xuminglu}@cse.buffalo.edu

URL: http://www.cse.buffalo.edu/~demirbas/

iComp

2

In-network querying in WSNs

In contrast to traditional WSN applications that perform only data collection, new generation of WSN applications require in-network information querying

– Disaster relief applications– Battlefield applications

Where is the nearest enemy tank?

A soldier queries the WSN via a

palm device. Energy-efficiency and

latency suffers drastically if queries

are always routed to a centralized

basestation over many hops

In-network querying should satisfy

• Distance sensitivityCost of querying for nearby events should be lower

• Low-cost constructionCostly bottom-up constructions (via flooding) should be avoided

• Graceful resilienceNode failures should not impact performance disproportionately

3

Our contributionsWe present an in-network querying infrastructure that satisfies all these requirements for event querying

– Distance sensitivity: the cost is at most (stretch factor) times the distance d to the nearest event

– Low-cost maintenance: stateless, minimalist infrastructure, bottom-up construction is avoided

– Graceful resilience: single mote failures are masked and performance degrades proportionately wrt the severity of holes (failures of motes in a region)

22

4

Distributed Quad-Tree (DQT)

033 122 211 300

003 013 023 033 102 112 122 132 201 211 221 231 300 310 320 330

000 001 002 003 100 101 102 103 200 201 202 203 300 301 302 303

• We embed a DQT over the WSN : DQT is a multi-resolution structure, but for a lightweight representation of DQT we use an encoding trick

– (xs,ys) at NW and (xe,ye) at SE be the two endpoints of the area. Assume DQT have i levels. The area of each level 1 box of partition is w*l, where width w=(ye-ys)/2i and length l=(xe-xs)/2. Then, DQT address of a node (x,y) can be calculated as:

– Based on the DQT node id, a node determines which level it is at, which children, neighbor, parent it has

• To achieve low-cost construction, we exploit location info at the nodes

– Nodes know the boundary coordinates of deployment, and calculate which DQT node id they fall into using their coordinates

– We chose the clusterheads closer to the basestation to avoid backward links during querying and data collection

5

Event IndexingIndexing of event information

– A query from node 011 to node 100 may route to higher level clusterheads such as node 013 and node 033

– The solution is to use sibling links to nearby intermediate nodes.

– A sibling link is the link between a node and its neighbors in each direction (so each may at most have 8 sibling neighbors).

– The sibling links only exist between nodes on the same level in the structure

A node at level i maintains the event information of its cluster, as well as the event information of its neighbors

6

Event QueryingEvent querying algorithm

• NN query: Finding the data object which is closest to the querying object given a set of objects

• If query point is not the current location (the initiator of the query and the query point belong to two different nodes), query is routed to the query point via GPSR

• If matching answer is not found, query is sent to next parent progressively

(until root is reached)

• The result is returned to the initiator

Query Propagation PathP

Q

M

d1

d

Distance stretch factor: d1/d

7

Graceful resilience

• DQT achieves resilience via:– its stateless nature, and– using GPSR for routing

• More specifically :– Mote failures do not often lead to

failure of level 1 node & are masked– GPSR routes around coverage holes,

& delivers the message to a boundary node if destination is inside hole

– Since DQT is stateless, any node can act as a proxy node for another

• Simulations show that s’ increases slowly wrt failure rate

Stretch factor under different failure rate

s = ratio of DQT querying cost to dist (q, p)s’ =ratio of DQT querying cost to GPSR cost (q, p)

8

Simulation (ns2)Querying success rate

Stretch factor with node failure: Case 1 Stretch factor with node failure: Case 2

Settings: 16x16 nodes, unit distance 200m, transmission range 250m

Case 1: Failures happen before the event advertisement.

Case 2: The event has already been published in the structure before the failure

happens.

9

Related work

• Distance Sensitive Information Brokerage protocol (Funke et al[2006].)

– achieves distance-sensitivity via hierarchically partitions

– relies on communication protocol and needs costly construction

• DIMENSIONS (Estrin et. al. [2002]) – provides a unified view of data

processing and in-network querying via wavelet compression and waveRoute routing protocol

– sacrifices flexibility space to achieve multi-resolution capability

• Geographic Hash Tables (Ratnasamy

et al. [2002] ) – stores and retrieves information

using a geographic hash function– querying is not distance sensitive

• Distributed Index for Features in Sensor Networks (Greenstein et al. [2003])

– addresses complex querying, not restricted to event querying

– costly to construct and update– special purposed routing

10

Our current work: Model-based querying in DQT

– Only supporting event type of querying is very limited

– Complex querying is very costly without prior knowledge of data

– We use modeling to capture the correlations of sensor values and reduce the cost of querying

• We use multi-resolution modeling, and a query is answered with approximate values accompanied by confidence levels

The original temperature data

Our modeling of the data at high levels