Adaptive Sleep Scheduling for Energy-efficient Movement-predicted Wireless Communication

Adaptive Sleep Scheduling for Energy-efficient Movement-predicted

Wireless Communication

David K. Y. Yau

Purdue UniversityDepartment of Computer Science

Objective

• Reducing energy consumption of battery powered devices, e.g., Laptops and Handhelds, in wireless networks– Wireless communication is power

intensive– Node movement can be exploited to

reduce energy use in communication

Movement Prediction

• Observation: Reduced distance between communicating peers ⇒ Reduced transmission power requirement ⇒ Energy saving

– Assuming network interface has transmission power control capability

– Single hop communication – obvious– Multi hop communication – expected

Power Saving Strategy

• If likely to move closer to the target, postpone communication for a future time– Assuming application can tolerate some

delay k.

• Needs movement prediction– Based on movement history– Consumes energy itself

Network Structure

• Mobile nodes are moving within a rectangular plane

• We divide the network into virtual grids• Each grid has a unique grid ID

Assumptions

• Each node knows its own position – GPS• Each mobile host maintains a sequence of n previous grid IDs

• Simplifying assumption– Target is fixed– Every mobile node knows the target’s location

• Fixed target assumption can be relaxed– Both communicating peers are mobile

Notations

• History of node h:Sh = {x1, x2, …, xn}

• A window of size l (for i ≤ n-l+1):W(i,i+l-1) = {xi, xi+1, …, xi+l-1}

• Distance between two grids i and j: d(i,j)

Example Average Distance (AD) Algorithm

• [TMC 05] Calculate the average distance between a mobile node and the target over all windows of size k in the mobile node's movement history as:

∑ ∑+−

=

+−

=

=1

1

1

),(1kn

j

kj

jii yxd

kavg

If the current distance between the mobile node and the target is greater than avg, then the mobile node decides to postpone the communication, or else it communicates immediately.

• Less Computational overhead

• Takes into consideration the actual distance

AD Algorithm Example

Energy Use in Movement Tracking

• Continuous movement tracking requires device to be turned on– Idle device can consume significant energy

• Movement sampling– Device put to sleep between sampling

instants– More frequent sampling => higher

accuracy– How often and when to sample?

System/Communication Energy Tradeoff

Communication energy

Sleep period

System energy

Total energy

Sleep period

Sleep periodOptimalSleep period

Adaptive Wakeup Algorithms

• Speed based adaptation (SBA)– Faster movement more frequent

updates

• Delay budget based adaptation (DBA)– Less time until deadline more

frequent updates

Adaptive Wakeup Algorithms (Cont’d)

• Position based adaptation (PoBA)– Position estimated at next wakeup instant,

based on nodal speed and direction– Wakeup chosen at ``good’’ position estimate

• Performance based adaptation (PeBA)– Distance savings compared between current

and last updates– Decreased saving reduced sleep

Sensor Implementation

• Berkeley Mica mote – TR1000 916.5 MHz network interface– 256 level power control (radio output voltag

e proportional to square of input data pin current)

– GPS would add 12—24 mW in operation• TinyOS

– Sleep by snooze component– Power control by pot component

Implementation Architecture

Positioning System

Operating System Services

Application

Transmission Scheduler

SleepScheduler

Power Manager

Power Manager

• Transmission scheduler– Implements postponement algorithm

• Variable length of movement history • Variable delay budget (application specific)

– Buffers packets until decision to transmit• Adaptive wakeup scheduler

– Implements adaptive wakeup algorithm for position sampling

– Wakes up at sampling instant– Adjusts next sleep period

Measurement Setup

ConstantPower Supply Multimeter

Sensor

R+

+ +

-

- -

Energy Estimation by Component

• Multimeter setup measures total system energy use

• Energy breakdown by selectively turning off system components / activities– Radio network interface turned on / off– Packet actually sent / not sent (but delet

ed from transmission queue)– Energy difference between configuration

s gives estimates of component energy use

System Parameters

• Input parameters– Length of history maintained– Application delay budget– Mobility scenario– Fixed vs adaptive wakeup

• Output parameters– Energy use and percentage saving (total and comp

onent)– Percentage distance saving– Actual postponement delay– Number of wakeups for position sampling

Average Speed of Mobility Scenarios

Scenario m/s Km/s Miles/s

Walker 2 7.2 4.5

Runner 3 10.8 6.8

Bicycle 7 25 15.6

Vehicle (local)

20 72 45

Vehicle (highway)

30 108 67.5

Airplane 80 288 180

Percentage Distance Saving vs Sleep Period and Delay Budget

Percentage Energy Saving vs Sleep Period and Delay Budget

Percentage Distance Saving (Bicycle)

Percentage Energy Saving (Bicycle)

Number of Wakeups (Bicycle)

Actual Postponement Delay (Bicycle)

Percentage Distance Saving vs Mobility Scenario

Percentage Energy Saving vs Mobility Scenario

Number of System Wakeups vs Mobility Scenario

Performance Comparison between Sampling Strategies

Scheduler Power (mW)

Distance saving

Wakeups Delay

Baseline 18.04 85.2% 0 4.83

Fixed 1 9.64 85.2% 5.71 4.83

Fixed 2 9.63 53.3% 4.67 7.34

Fixed 4 8.87 40.7% 3.54 10.16

Fixed 8 8.90 5.1% 1.80 6.40

DBA 8.19 44.8% 3.17 9.66

Conclusions

• Node movement prediction can reduce the energy cost of wireless network communication

• However, need to balance against energy cost of movement prediction

• Adaptive sampling schedule works well based on operating conditions– Simulations + measurements on sensor prototype– Saves substantial energy by putting device to sleep– Outperforms fixed sleep period in general, since

optimal sleep period is hard to determine a priori