sMAP: Integrating and Managing Physical Data
@scale: Insights from a Large, Long-Lived Appliance Energy
WSNStephen Dawson-Haggerty, Steven Lanzisera, Jay Taneja, Rich
Brown, and David Culler
Computer Science Division, University of California,
BerkeleyEnvironmental Energy Technologies Division, Lawrence
Berkeley National Lab
1Motivation
April 17, 2012IPSN 2012: Beijing, China2US Department of
EnergyRepresentative sample of plug-load power and energy
Capture traces, usage patterns, and models
@scaleApril 17, 2012IPSN 2012: Beijing, China3
90k ft2/4 floor building1 year deployed5000 plug-loads460
plug-load meters7 edge routers650m data points
Evaluate networking dynamicsScale to hundreds of meters over
multiple floorsMaximize accuracy of inexpensive meters
Manual device inventory resulted in about 5,000 plug-load
devices, more than 10 per occupant.
ACme meters previously presented at SenSys.3Study
methodologyApril 17, 2012IPSN 2012: Beijing, China4
Pick up where Jiang 09 leaves off4System architectureApril 17,
2012IPSN 2012: Beijing, China5
6LoWPAN at the edge access network data closet
One routing domain among the meter network
No app-specific code on the LBRs: implementation of the IP
architecture
Full adoption of the IP architecture
5NextApplication design
Network lessons
April 17, 2012IPSN 2012: Beijing, China6
Interaction
April 17, 2012IPSN 2012: Beijing,
China7http://www.screenr.com/ydh8
$ Ifconfig tun0$ telnet :: 6106Helplinks $ tcpdump I tun0$ ping6
fec0::936$ blip-shell fec0::936readreadcfg pwrealcalreadcfg
dest
7Application designApplication/IPv6 allows scripted interaction
with a large number of motesUpload calibration tablesModify
reporting destinationChange MAC parametersAvoid
reprogrammingInteraction uses standard tools
April 17, 2012IPSN 2012: Beijing, China8Time
StampDescriptionLocalTimeCounter since rebootGlobalTimeUNIX
timeSequenceMonotonic data sequence numberInsertTimeTime at
database
Limited development/engineering time available focus on whats
important8System architecture: HYDRO principlesMaintains multiple
next-hop optionsManage explore/exploit tradeoffHorizontally
scalable with multiple Load Balancing R0uters (LBRs)April 17,
2012IPSN 2012: Beijing, China9
Published 2010, similar to RPL non-storing mode.
Trickle-ized, density-sensitive state propagation
9Networking data set5-minute snapshots Top four links from each
network node, as reported to the edgeLink and device churn are
commonMean network degree is at least 16, diameter is about
4.5Analysis performed March-April 2011
April 17, 2012IPSN 2012: Beijing, China10
View the link topology, filtered through the lens of what the
routing protocol did.10Network data yieldApril 17, 2012IPSN 2012:
Beijing, China11
Most days exhibit 5th percentile data yield > 99%
Device dynamicsWhen do devices come on/off?Device reset detector
based on LocalTime rolloverApril 17, 2012IPSN 2012: Beijing,
China12
Also energy science implications for motes turning on and
off
12Exploration is ongoing
April 17, 2012IPSN 2012: Beijing, China13Path length and router
degree show diurnal and weekly variation
Methodology: construct directed graph based on reported topology
at each slice. Examine degree, mean shortest path.13
Is there a single, stable routing tree?April 17, 2012IPSN 2012:
Beijing, China14
Exploration of new potential candidate links is continuous
Diameter increases by factor of 2 using only stable links
stable linksall linksLt: reported link set at time t
Do we end up with a single set of stable links?
L = U(L_t)
14Lessons: networking and managementNo such thing as a static
networkScriptability/automated management is keyData reliability is
not all about the wireless part: Internet and practical
considerations Back up or replicate your databaseLocal buffering at
the end points, middle-boxes?Meters walk awaySometimes the whole
building is turned offHorizontal scalability is a must
April 17, 2012IPSN 2012: Beijing, China15
You need to recover when everything is turned on at once.15What
makes up building 90 plugs energy?16
Computers50% of energyDisplays10% of energyTask Lighting 7%
Networking 6%Other 7%Imaging10% of energyMisc. HVAC10% of
energyTimer controlled plug strips?75 MWh/year30% of non-computer
plug total 6% of building totalComputer power management?150
MWh/year60% of computer total 12% of building total
April 17, 2012IPSN 2012: Beijing, China
75 MWh = $11k/year, Cost @ $25/unit = $10kComputer power
management 150 MWh + 150 MWh from power management of other
devicesWhy is other so high mid-day: 18 microwaves, 10 toasters, 40
coffee makers, 14 on demand hot water, 10 electric tea kettles
16ConclusionsToolkit for domain scientists needed802.15.4e,
RPL-based networkingCommon hardware problemExploding the instrument
is possible!Further standardization: CoAP90% solutionsOverall
theme: careful simplicity
April 17, 2012IPSN 2012: Beijing, China17
What does a large, relatively successful deployment like this
really tell us?Productization is still a challenge for academics:
either need a company or we need to collaborate more on
infrastructure
17QuestionsSpecial thanks to Sara Alspaugh, Alice Chang, Iris
Cheung, Albert Goto, Xiaofan Jiang, Shelley Kim, Margarita Kloss,
Judy Lai, and Ken LutzApril 17, 2012IPSN 2012: Beijing, China18
BACKUPSApril 17, 2012IPSN 2012: Beijing, China19
Network co-development and deployment April 17, 2012IPSN 2012:
Beijing, China202005: Redwoods2007: RFC4944: 6LoWPAN2008: Full
sensornet IP architecture proposed2008: BLIP/Contiki 6LoWPAN
released2009: GreenOrbs (1000 Nodes)2010: Collection Tree
Protocol2012: RFC6553: RPL, IEEE 802.15.4e2012: @scale
Source for 76M is PG&E
20How Common is Computer-Display Power Management?
2140 Hour Work WeekPM w/ breaksRarely power down monitor83% of
monitors use power management15% use it with breaks for days at a
time2% do not use it
April 17, 2012IPSN 2012: Beijing, China
19 of 115 or 21What is the Distribution of LCD Computer Display
Energy Use?22
N=118April 17, 2012IPSN 2012: Beijing, China
How Common is Desktop Computer Power Management?
2340 Hour Work Week39% rarely powered down44% managedApril 17,
2012IPSN 2012: Beijing, China
Findings and Next Steps Bldg 90 network demonstrated
large-scale, end-to-end WSN and collected a lot of useful dataIT
equipment should be focus of office energy management programsUsing
data for LBNL-wide plug-load managementInventory and meter
installation are labor-intensiveExploring using public (homeowner
& building occupant) participation for data collectionIntegrate
metering & communications into productsRobust sensor network
needs more engineering? Evaluate commercial products now
availableElectricity only part of buildings energy
problemDeveloping low-cost WSN for gas and water metering24April
17, 2012IPSN 2012: Beijing, China
IntroductionEnergy science goalsComputer science goalsRelated
WorkDeploymentsStudy overviewSystem design &
architectureResultsApril 17, 2012IPSN 2012: Beijing, China25
MELS: Miscellaneous Electric LoadsLarge, rigorous study of
miscellaneous electric loads (mostly plugs)Roughly 1/3 of building
energy consumptionDifficult to study due to large number of small
consumers
April 17, 2012IPSN 2012: Beijing, China26
Methodology: Multipoint CalibrationApril 17, 2012IPSN 2012:
Beijing, China27
Automated 20-point calibration on every meter 3-part piecewise
calibration
90th percentile error is