Comb, Needle, and Haystacks: Balancing Push and Pull for Information Discovery Xin Liu Department of Computer Science University of California, Davis Joint.

Comb, Needle, and Haystacks:Balancing Push and Pull for Information Discovery

Xin LiuDepartment of Computer Science

University of California, Davis

Joint work with Q. Huang & Y. Zhang, PARC

Berkeley, 04/20/05

Comb-Needle Query Structure

Objective Simple, reliable, and efficient on-demand

information discovery mechanisms Constraints

Limited communication capacity and battery power

Berkeley, 04/20/05

Where are the tanks?

Berkeley, 04/20/05

Pull-based Strategy

Berkeley, 04/20/05

Pull-based Cont’d

Berkeley, 04/20/05

Push-based Strategy

Berkeley, 04/20/05

Comb-Needle Structure

Berkeley, 04/20/05

Application Scenarios

On-demand information query Any node can be the query entry point Queries may be generated at anytime Events can happen anywhere and anytime Examples:

Firefighters query information in the field Surveillance

Assume sensor nodes know their locations

Berkeley, 04/20/05

When an Event Happens

Event

Berkeley, 04/20/05

When a Query is Generated

Event

Query

Event

Berkeley, 04/20/05

Tuning Comb-Needle

Berkeley, 04/20/05

Reverse Comb

Query

Event

When query frequency > event frequency

Berkeley, 04/20/05

The Spectrum of Push and Pull

Pull Push

Global pull +Local push

Global push +Local pull

Push & Pull

Inter-spike spacing increases

Reverse comb

Relative query frequency increases

Berkeley, 04/20/05

Comparison

Berkeley, 04/20/05

Simulations

Radio model Path loss and random error

Topology model Regular grid with random shifts

Routing Constrained Geographical Flooding (CFG) for

random topology Based on simulator Prowler

Berkeley, 04/20/05

An instance of connectivity

Berkeley, 04/20/05

Simulation Cont’d

f_e=1f_q=0.1

Berkeley, 04/20/05

Simulation Cont’d

f_e=1f_q=1

Berkeley, 04/20/05

Random Topology

Berkeley, 04/20/05

Constrained geographical flooding

Needles and combs have certain widths

Berkeley, 04/20/05

Success Rate

Berkeley, 04/20/05

Power consumption

Berkeley, 04/20/05

A few issues

Adaptive scheme Reliability Single fixed query entry point Yes-or-No query

Berkeley, 04/20/05

Adaptive Scheme

Comb granularity depends on the query and event frequencies

Nodes estimate the query and event frequencies Important to match needle length and inter-spike

spacing Comb rotates

Load balancing Broadcast information of current inter-spike spacing

Berkeley, 04/20/05

An illustration

Regular grid Communication cost: hop counts No node failure Adaptive scheme

Berkeley, 04/20/05

Event & Query Frequencies

Berkeley, 04/20/05

Tracking the Ideal Inter-Spike Spacing

Berkeley, 04/20/05

Simulation Results

Gain depends on the query and event frequencies Even if needle length < inter-spike spacing, there is a

chance of success. Tradeoff between success ratio and cost

99.33% success ratio and 99.64% power consumption compared to the ideal case

Berkeley, 04/20/05

Strategies for Improving Reliability

Local enhancement Interleaved mesh Routing update

Spatial diversity Correlated failures Enhance and balance query success rate at

different geo-locations

Berkeley, 04/20/05

Spatial Diversity

Query

xEvent

Berkeley, 04/20/05

Fixed-Node Query

Only one fixed query entry point Depends on relative frequency Depends on the length of the query

E.g., 5 seconds vs. 30 minutes

Berkeley, 04/20/05

Numerical illustration

Berkeley, 04/20/05

Binary query

Is there a tank in the field? Ans: Yes or No. If not delay sensitive

Sequential query process Optimal comb width is shorter

Intuition: can stop earlier

Berkeley, 04/20/05

Numerical illustration

Berkeley, 04/20/05

Summary

Balance query cost vs. event report cost Adapt to system changes

Pull Push

Global pull +Local push

Global push +Local pull

Push & Pull

Relative query frequency increases

Berkeley, 04/20/05

Future work

Data compression A more realistic model for communication

cost Build a fixed comb structure for random

networks for better success rate What if no/limited location knowledge? Consider delay tradeoff Accommodate sleep-awake pattern

Berkeley, 04/20/05

Joint work with P. Mohapatra, C. Chuah, P. Cheng

On the Deployment of Wireless Sensor Networks

Berkeley, 04/20/05

Many-to-One Communication

Berkeley, 04/20/05

Network Deployment

Many-to-one communication Data from all nodes directed to a sink

node/fusion center Unbalanced traffic load Uneven power consumption

Limitations on network lifetime if uniformly distributed “Important” nodes in the route die quickly

Capacity bottleneck and Power bottleneck Desire for long-lived sensor networks

Linear and planar networks

Berkeley, 04/20/05

Precise placement With access Expensive nodes Higher layer of a hierarchical structure

Random placement No access Cheap nodes Lower layer of the hierarchy Coverage and connectivity properties

Precise vs. Random Placement

Berkeley, 04/20/05

Maximize coverage area Given the desired lifetime and # of node

available Maximize the lifetime of the network

Given the number of nodes and coverage area Minimize the number of nodes required

Given the coverage area and the desired lifetime

Consider large networks with long lifetime requirements

Objectives

Berkeley, 04/20/05

Why linear networks? Applications: Traffic monitoring, border line control,

train rail monitoring, etc. Abstract model for narrow-and-long applications

Duck island Tractability, insights for general cases

Highly asymmetric traffic load & location-dependent power consumption

Focus on communications What options do we have?

Linear Networks

Berkeley, 04/20/05

Possible Solutions

More energy for nodes with heavier load

More nodes in the area closer to the sink

Nodes closer to each other

Load balancingPlacement involves

topology control, routing, power allocation

Berkeley, 04/20/05

System Model

Berkeley, 04/20/05

Total energy constraint: (n-1)E Energy can be arbitrarily allocated among

nodes The network dies when no energy left

Thus,

i

Total Energy Constraint

Berkeley, 04/20/05

Problem Formulation

Numerical results as benchmark

Berkeley, 04/20/05

Homogenous initial energy allocation Observation: longer hops consume

more energy “jump” may not be a good idea

Observation: we do not want residual energy when the network dies. Power consumption per unit time should

be the same for all nodes Consider large T (desired lifetime)

A Greedy Algorithm

Berkeley, 04/20/05

A Greedy Algorithm

Berkeley, 04/20/05

Numerical Result

Berkeley, 04/20/05

Performance Analysis

Lifetime, power, and coverage

=4, 19% more node to double lifetime

=4, 138% more node to double coverage

Berkeley, 04/20/05

Extensions

Miscellaneous power consumption PT=c1+ R d

PR = c2

Transmit at max power at max rate to near nodes Similar results hold Intuition: shorter links, higher rate, less time for T/R.

Non-uniform traffic density Estimation errors on traffic density during the

deployment

Berkeley, 04/20/05

The effect of arbitrary energy allocation is negligible

Greedy algorithm Compensate for nodes with heavy load by

reducing communication distance Performs very well and adapts to various

conditions 2-D case Data aggregation

Summary

Berkeley, 04/20/05