Aruna MS Comps...

28
Aruna Ravinagarajan System Energy Efficiency Lab Aruna Ravinagarajan Advisor : Prof. Tajana Simunic Rosing CSE Dept., University of California, San Diego

Transcript of Aruna MS Comps...

Page 1: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Aruna Ravinagarajan

System Energy Efficiency Lab

Aruna Ravinagarajan

Advisor : Prof. Tajana Simunic Rosing

CSE Dept., University of California, San Diego

Page 2: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Structural Health Monitoring (SHM)

� SHM: Process of monitoring a structure over time and identifying damage

System Energy Efficiency Lab

damage

� A wireless sensor network (WSN)…

� Monitors a physical space or object� Environment

� Humans and animals

� Structures

� Remote Location:

Needs long lasting energy source

Page 3: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

SHM – How is it done?SHM – How is it done?

System Energy Efficiency Lab

Page 4: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

SHM – How is it done?SHM – How is it done?

System Energy Efficiency Lab

Stuart G Taylor, Kevin M Farinholt, Eric B Flynn, Eloi Figueiredo, David L

Mascarenas, Erik A Moro, Gyuhae Park, Michael D Todd and Charles R

Farrar, 2009. A mobile-agent-based wireless sensing network for

structural monitoring applications. Meas. Sci. Technol. 20 045201 (14pp)

Page 5: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

PZT’s

Shimmer

SHiMmer for Structural Health Monitoring

System Energy Efficiency Lab

Page 6: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

� Localized embedded node for detection of damage in structures

� Local analysis on SHiMmer is mandatory due to data size, 4 MB per test

� High DSP computational workload

� Max power 370mW at 400 MIPS

� Peak power consumption 1.1W when sensing/actuating

� Powered only with energy harvesting

SHiMmer for Structural Health Monitoring

System Energy Efficiency Lab

Powered only with energy harvesting

Page 7: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Energy Harvesting

Challenges:

• Solar energy is not uniformly distributed over time

• Daily weather and seasons change the total input energy

System Energy Efficiency Lab 7

The task scheduler needs to The task scheduler needs to manage energy consumption manage energy consumption and accuracy of computationand accuracy of computation

Page 8: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Motivation

� Problem Definition:

� WSNs powered only via energy harvesting

� Operating with severe energy constraints

� Too much data to continually transmit

� Localized processing a must

System Energy Efficiency Lab

� Key challenge:

� Minimize energy costs while maximizing accuracy of computation

Page 9: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Related work

� Related work:

� Energy Harvesting in WSNs

� Guarantee energy neutrality and adapt duty cycle [Kansal et al., DAC ’06]

� Task Scheduling in WSNs

� Power Management with discrete service levels [Moser et al., ISLPED ’09]

� Adapting Task Utility in externally triggered WSN [J.Steck et al., INSS ’09]

System Energy Efficiency Lab

� Adapting Task Utility in externally triggered WSN [J.Steck et al., INSS ’09]

Page 10: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Steady State Mode

•Periodic Lifetime Monitoring

•Identifies damage that accumulates over long period of time

System Energy Efficiency Lab 10

• Effective bridge maintenance costs millions of $$$ every year

• By performing steady state monitoring, reduces burden on bridge’s annual maintenance

Page 11: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

External Request Mode

2007, Minneapolis Bridge

1967, Silver Bridge

Event

System Energy Efficiency Lab 11

2007, Minneapolis Bridge

http://www.engineeringcivil.com/civil-engineering-disasters-collapse-of-bridges.html

Page 12: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

External Request Mode

Event

Extra measurements are required Extra measurements are required

System Energy Efficiency Lab 12

Execution Time Constraint

� Given a time limit, what is the highest level of data accuracy?

Data Accuracy Constraint

� Given a minimum data accuracy, how long will it take to execute tasks?

Extra measurements are required Extra measurements are required to verify the structure’s integrityto verify the structure’s integrity

Page 13: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Challenge and Contribution

To run an intensive workload, Task Scheduler needs to manage:

•Energy Consumption

•Accuracy of computation

System Energy Efficiency Lab 13

Contribution:

• Regression based algorithm to optimize resource utilization

• Applying DVFS techniques scaled by available energy

Page 14: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Active Ultrasonic SHM

• 16 PZT sensors provide 120 different sensing paths

• 120 waveform data give about 4 Mbyte

System Energy Efficiency Lab 14

• Filtering signal using FFTs

• Convolving filtered set of data with baseline signals

• Damage detection consists of combining data using a correlation function

• A higher number of measurements increases detection and localization accuracy

Page 15: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Setting up Regression Model

BlackFin integrated DVFS defines three different DSP working modes:

• Active High: Vcore =1.20V, fDSP =300 MHz, PMax =370mW

• Active Low: Vcore =0.85V, fDSP =150 MHz, PMax =220mW

• Idle: Vcore =0.85V, fDSP =75 MHz, PMax =35mW

System Energy Efficiency Lab 15

Page 16: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Setting up Regression Model

The energy and time consumption, accuracy for every task in the task graph

can be determined based on the number of paths, frequency and power

System Energy Efficiency Lab 16

Page 17: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Setting up Regression Model

System Energy Efficiency Lab 17

Page 18: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Steady State Algorithm

• Steady State activity performed every 15 min (900 sec) : Tslot

• Eavailable represents the available energy in buffer

• Based on regression model, DVFS mode is automatically selected by the scheduler in order to maximize number of path

System Energy Efficiency Lab 18

selected by the scheduler in order to maximize number of path measurements with respect to energy availability

Page 19: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Steady State Algorithm

Available EnergyThe available energy is calculated as:

where Eth_min is the minimum amount of

tconsumedtharvestedtbuffertbuffer

tht

buffertavailable

EEEE

EEE

−+=

−=

−1

min_

System Energy Efficiency Lab 19

where Eth_min is the minimum amount of

energy to execute an external request

Page 20: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Steady State Algorithm

Available Energy

Estimate Execution TimeExecution time is estimated using

relation obtained from regression :

tavailableEtT ⋅=

System Energy Efficiency Lab 20

where tDSP and eDSP are model

coefficients that depend on the

DVFS mode (high/low frequency)

{ }lowHighDSPDSP

DSP

tavailable

DSPtexecution

ffte

etT

,, ∈

⋅=

Page 21: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Steady State Algorithm

Available Energy

Estimate Execution Time

System Energy Efficiency Lab 21

Select DVFS modeThe DVFS mode is selected to max

Npaths according to estimated limits:

{ }lowHighDSP

availableiexecution

iexecution

DSP

texecution

DSP

ffe

EfE

fEt

Te

,

)(

)(

<

=⋅

Page 22: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Steady State Algorithm

Available Energy

Estimate Execution Time

System Energy Efficiency Lab 22

Select DVFS mode

Execute tasks according to

task graph for Npaths

Transmit Result

Page 23: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

External Request Algorithm

• External Requests (ER) set two constraints:

• Maximum Execution Time

• Minimum SHM Accuracy

Available Energy

Max Time OR Min Accuracy

System Energy Efficiency Lab 23

• Minimum SHM Accuracy

• DVFS mode selection performed in manner similar to steady state

Select DVFS mode

Execute tasks according to

task graph for Nrequired paths

Transmit Result

The DVFS mode is selected

according to estimated limits:

requiredthavailable

data

DSPrequired

DSP

constrain

DSPrequired

EEE

a

AccuracyeE

t

TeE

>>

⋅=

⋅=

min_

min

min

max

min

OR

Page 24: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Experimental set-up

Inputs to the system:

• Variable solar energy conditions to the energy harvester

• Task sequence as shown in Task Graph

• DVFS active frequency and power modes

• Periodic activity time slots: T = 900s

System Energy Efficiency Lab 24

• Periodic activity time slots: Tslot = 900s

• Constant number of External Requests per day: 25

• Results compared to an iterative search algorithm for task scheduling in SHM2

J. B. Steck et al. "Adapting Performance in Energy Harvesting Wireless Sensor Network for Structural Health Monitoring Applications," 6th International Conference on Networked Sensing Systems, 2009

2

Page 25: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Number of measurements increase

• 50% measurement increase with regression algorithm

System Energy Efficiency Lab 25

algorithm

• 15-20% further increase with DVFS

Page 26: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

• 16% accuracy increase with regression algorithm

Measurement accuracy improvement

System Energy Efficiency Lab 26

• 27% accuracy increase with DVFS

Page 27: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

• 95% of external request served with regression

Increase in number of served ER

System Energy Efficiency Lab 27

with regression algorithm

• 20% of energy saving with DVFS

Page 28: Aruna MS Comps PDFversioncse.ucsd.edu/sites/cse.ucsd.edu/files/cse/assets/studentaffairs/docs/Aruna... · SHM – How is it done? System Energy Efficiency Lab Stuart G Taylor, Kevin

Conclusions

With our task scheduler, SHiMmer for SHM analysis achieved:

� Up to 85% increase in the number of daily measurements

� Up to 27% increase in result accuracy

� Up to 95% of external request service requests processed

System Energy Efficiency Lab 28

The algorithm improves performance through an efficient combination of:

� The adoption of the regression algorithm, optimizing the usage of available resources

� An efficient usage of DVFS