Low Power Wireless Sensor NetworksLow Power Wireless Sensor Networkshttp://www-mtl.mit.edu/research/icsystems/uamps

Rex Min, Manish Bhardwaj, Seong-Hwan Cho, Eugene Shih, Amit Sinha, Alice Wang, Anantha Chandrakasan

Massachusetts Institute of TechnologyMassachusetts Institute of Technology

Emerging Networked ApplicationsEmerging Networked Applications

Integrated PDAs Home/Office Networking(e.g., Bluetooth)

Medical Monitoring

Equipment MonitoringSensor Networks

Integrated Integrated systemsystem--onon--aa--chipchip to sense, process and to sense, process and collaborate collaborate

The MIT The MIT µµAMPSAMPS ProjectProject

n A universal substrate for power aware data gathering from a massively distributed wireless network

Sensor& A/D StrongARM RF

Tx/Rx

Battery/DC-DC Conversion

µ-OS (Power Aware Control)

Remote Basestation

System RequirementsSystem Requirements

n Sensor Types: Low Rate (e.g., acoustic and seismic)

n Bandwidth: bits/sec to kbits/sec

n Transmission Distance: 5-10m (< 100m)

n Spatial Densityo 0.1 nodes/m2 to 20 nodes/m2

n Node Requirementsn Small Form Factor

n Required Lifetime: > year

n Operational Diversity:

...from the environmento Event arrival

rate/type

o Ambient noise

o Signal statistics

...from network roleso Sensor

o Relay

o Data aggregator

...from user demandso Tolerable latency

o Result SNR

o Pr(Detection)

Integrated SensorIntegrated Sensor--NodeNode--onon--aa--ChipChip

n Integration is the key enabler for massively distributed wireless sensing

µ-PROC&

DSP

MULTIPLEOUTPUTDC-DC

MEMS

MICROBATTERY

A/D

MEMORY

What is the best computation/communication fabric?What is the best computation/communication fabric?How coupled should protocol design be to the fabric?How coupled should protocol design be to the fabric?

RF

n Diversity in operating scenarios: number and type of events, signal statistics, desired quality, latency, etc.

n Cannot achieve Esystem = Eperfect at all pointsoOptimize at important scenarios (Esystemi di is high)

1−

=

∑∑

Scenariosiperfect

Scenariosisystem

PA dE

dE

i

i

η

Scenario

Eperfect

Ene

rgy

Esystemdi

Power AwarenessPower Awareness

Power Aware Node ArchitecturePower Aware Node Architecture

Leakage current Workload variation

Bias currentStart-up time

Standby currentLow duty cycle

Capacity variations

Efficiency variations

Desired result quality variations

Available energyVoltage scheduling

RadioSA-1100A/DSeismic Sensor

Acoustic Sensor

ROMRAM

DC-DC Conversion

Battery

Power

Protocols

Algorithms

µOS

n Graceful energy scalability across a diversity of operating conditions and energy-quality trade-offs

OS Directed Power ManagementOS Directed Power Management

offonsleepsleeps3

offoffsleepsleeps4

rxonsleepsleeps2

rxonsleepidles1

tx, rxonactiveactives0

RadioSensorMemoryARM

Battery and DC/DC converter

Sens

or

A/D

Rad

io

Sensor Node

Memory

StrongARM

µ-OS

• OS must decide suitable transitionpolicy based on observed history

-200

0

200

400

600

800

1000

1200

-10 0 10 20 30 40 50 60

s0

s1

s2s3

s4

Pow

er (m

W)

Transition Latency (ms)

Deeper sleepLower powerMore overhead

Idle Mode Leakage PowerIdle Mode Leakage Power

n Leakage dominates switching energy for low duty cycles

n A major concern for event-driven operation (PDAs, sensors, etc.)

)/(10 SVleakage

TI −∝

Leakage and Switching PowerLeakage and Switching Power

0.13 µ, 15mm die, 1V

1% 2% 3% 5% 8% 11% 15% 20%26%

10

20

30

40

50

60

70

30 40 50 60 70 80 90 100

110

Temp (C)

Po

wer

(Wat

ts)

Leakage

Active80

900.1µ, 15mm die, 0.7V

6% 9% 14% 19%26%

33%41%

49%

56%

10

20

30

40

50

60

70

30 40 50 60 70 80 90 100

110

Temp (C)

Po

wer

(Wat

ts)

80

90LeakageActive

Need to Develop Techniques for Leakage ControlNeed to Develop Techniques for Leakage Control

Courtesy of Vivek De (Intel)

Low Duty Cycle RadioLow Duty Cycle Radio

n Start-up time dominates the energy for small packet sizes

n Innovative radio design required…

10

100

1000

10000

10 100 1000 10000 100000

Packet size (bits)

Ene

rgy

Per

Bit

(nJ)

20mW Electronics Power1mW Transmit power @ 1Mbps

Startup Costs are Fundamental Startup Costs are Fundamental ––Latency not just a function of user requirement Latency not just a function of user requirement

DVS on SADVS on SA--11001100

SA-1100

Control

µOS

VoutController

Power

5 V

Voltage request, 0.9 - 1.6 V

5

1.6V limiter

5

Digitally adjustable DC-DC converter powers SA-1100 core

µOS selects appropriate clock frequency based on workload and latency constraints

SA-1100 requests a voltage appropriate for its clock frequency

MIT DVS PCBMIT DVS PCB

StrongARMEvalualtion

Board

Software Voltage SchedulingSoftware Voltage Scheduling

n Operating system predicts and schedules the voltage

n Adapt power supply to deliver “just enough performance”

Data from StrongARM-1100

StrongArmSA-1100

DC-DC Regulator Controller

5

CP

U C

ore

Po

wer

, 0.9

-1.6

Vo

lts

Buck Regulator

MOSFET control

DVS DemonstrationDVS Demonstration

oUser adjusts number of filter taps

o Frequency/Voltage adjusted appropriately (via eCOS based µOS)

Frequency /Voltage

Workload (filter taps)

Computation vs. CommunicationComputation vs. Communication

Compute, Don’t CommunicateCompute, Don’t Communicate

1E-11

1E-10

1E-09

1E-08

1E-07

1E-06

1E-05

1E-04

1E-03

1 10 100 1000 10000

Energy for Electronics + Transmit

R2 Propagation LossLimit (no electronics)Assuming 10pJ/bit/m2

En

erg

y (J

)

Distance (m)

n Computation: 1nJ/op (µ-Processor) and Communication (@10m): 150nJ/bit

n @10 m: ~150 instructions/transmitted bit on a low-power processor

n @10m: > 1Million instructions/transmitted bit using dedicated hardware

Protocol ArchitecturesProtocol Architectures

Source Destination

Router

Multi-hop Routing Example(ignoring electronics)

• 1 hop over 100 m: 100nJ/bit• 10 hops of 10 m:

10 × 1 nJ/bit = 10nJ/bit

n Particular attention must be placed on multiple access schemes

n Scheduled vs. Reactive routing (synchronous vs. asynchronous)

Similar TradeSimilar Trade--off to Onoff to On--chip Interconnectchip Interconnect

Distributed DSP using DVSDistributed DSP using DVS

Ecomp(Vdd=1.5V) = 7 * Efft +Ebf + ELOB

= 27.27 mJ

Parallelizing the FFT means we can reduce the supply voltage and frequency

Ecomp(variable Vdd) = 15.16 mJ

FFT is operated at .9 V

BF & LOB is operated at 1.3 V

n Approach 2 :FFT is done at node and transmitted to C-H

A/D

Sensor 6

FFT

A/D

Sensor 2

FFTA/D

Sensor 1

FFT

Cluster Head

BF LOB

Sensor 7

n Approach 1 : All computation is done at C-H

A/D

Sensor 1A/D

Sensor 2

A/D

Sensor 6

FFT

Cluster Head

BF LOB

Sensor 7

Energy Efficient Link LayerEnergy Efficient Link Layer

0

500

1000

1500

2000

2500

(15,

7,2)

(31,

6,7)

(31,

11,5

)

(31,

16,3

)

(63,

7,15

)

(63,

16,1

1)

(63,

24,7

)

(63,

39,4

)

(63,

45,3

)

EncodeDecode

§ Energy scalability through variation of error-correction scheme

§ Computation-communication tradeoff between coding and Tx power for BER reduction

Energy

Ene

rgy

per b

it (n

J)

n Self-powered operation is a real option if the power dissipation can be scaled to 10’s - 100’s of µW oMechanical vibration (e.g., machine-mounted sensors)o Electromagnetic fields (RF)

n A major opportunity exists in developing energy scavengers(generator and associated electronics) for extracting useful energy from ambient sources

Energy ScavengingEnergy Scavenging

Generator RegulatorVDD

Load Electronics

Energy ScavengingEnergy Scavenging

MEMS Generator

PicoJouleDSP

Power Controller

[Amirthrajah00]

n Scavenge energy from mechanical vibrations to power micropower sensor systems

n Power delivered ~ 10µW

Hardwired Fabrics enable No Power Signal ProcessingHardwired Fabrics enable No Power Signal Processing

Node PrototypeNode Prototype

sensor/processor board radio baseband

n Version 1 prototype with COTS components

n Future nodes will feature custom chipsets

Node and Network APINode and Network API

nn Enable and encourage endEnable and encourage end--user to operate network in a user to operate network in a powerpower--aware manneraware mannero Sufficient abstraction to hide complexity of distributed wireless

network

oGet-optimize-set paradigm to maintain network state

nn Functional interface, object abstractions, and behavioral Functional interface, object abstractions, and behavioral semanticssemanticsoGather and set state of nodes, links, network

o Facilitate data exchange between node and basestation

oRealize a user’s desired operating point for the network

o Visualize network state

oBuilt-in and customizable energy models for energy, delay, etc.

n Just-in-Time computing through supply optimization minimizes energy dissipation

n Leakage is a first order issue – active leakage management at the architecture, circuit, and device levels are critical

n Focus must shift from computation to communication-centric design

n Protocols must be fabric and domain aware o Energy per operation (mW/MIPS) will scale with technologyoCommunication costs (nJ/bit) will not scale at the same rate

SummarySummary

Low Energy Sensor Design Requires a SystemLow Energy Sensor Design Requires a System--level level Approach Approach –– Tight Coupling Between Fabrics, Tight Coupling Between Fabrics,

Algorithms and ProtocolsAlgorithms and Protocols