DEWDROP: AN ENERGY- AWARE RUNTIME FOR COMPUTATIONAL RFID Michael Buettner (UW), Benjamin Greenstein (Intel Labs, Seattle), David Wetherall (UW)
Key Question How can we run programs on embedded computers using only scavenged RF energy? Battery free, “invisible” sensing and computation is key to truly ubiquitous computing applications
Scenario: Activity Recognition for Elder Care Elders can stay at home longer if caregivers know they are safe If we know what (and how) objects are used we can determine activities • Taking medicine, making a meal What we want: A non-intrusive way to gather data on object use
Existing Solutions Cameras: Remote monitoring • Cons: Obvious privacy concerns “Mote” based sensor networks: Detect object use from accelerometer data • Cons: Batteries limit deployment - Size - Lifetime - Cost Infeasible to deploy motes on 10s of everyday objects
Proposed Solution: Computational RFID Sensing ¡and ¡ Power ¡and ¡ RFID ¡ ¡ Computa4on ¡ commands ¡ Reader ¡ Rich ¡Data ¡ • Builds on passive RFID technology • Readers transmit power and commands • Battery-free tags harvest RF to compute, sense, communicate • Prototype hardware now becoming available • Goal: RFID tag “sticker” form factor costing $1
Dewdrop: A Runtime for CRFIDs Enables CRFID tags to use the scarce available energy to run programs: With varied and non-deterministic energy needs When input power varies by two orders of magnitude Dewdrop runs programs at close to their maximum rate, and where they could not otherwise run
Outline • Intel WISP – A CRFID Tag • Challenges to Running Programs Efficiently • Dewdrop Design • System Evaluation
Intel W ireless I dentification and S ensing P latform • Features • 16-bit TI MSP430, 8K flash • 3D accelerometer, light, temp • 10 uF capacitor for energy storage • 4 m range with standard readers • Community • In use at 30+ universities, ~50 publications
WISP Applications • Exercise, sleep monitoring • [Borrielo 2008, Stankovic 2010] • Neural monitoring, medical implantables • [Yeager 2010, Halperin 2008] • Cold-chain, undersea neutrino detector • [Yeager 2007, Trasatti 2011] • RFID security • [Fu 2009, Kohno 2008] • CRFID programmability • [Ransford 2011, Gummeson 2010] • Most use WISPs < 1 m from reader where energy is plentiful
Challenges to Running Programs Efficiently 1. CRFIDs have miniscule energy stores 2. Programs have different energy needs 3. Platform inefficiencies 4. Energy is harvested even while executing
CRFIDs have miniscule energy stores Program starts Black-out X Threshold Program completes Time Reader turns on • Low power mode (~1uA) to store energy, maintain state Active mode (~100s of uA) to compute and sense • 100s of ms to charge, 10s of ms to discharge • Tags must store enough energy to complete program before beginning execution
Programs have different energy needs • Wide range of energy needs Heavy Light • E.g., Sense, sense and communicate E E • May be non-deterministic T T • E.g., RFID MAC protocol Non-deterministic • Run-to-completion E • E.g., communication, sampling sensors T • Tags run only one program at a time • Tags must store different amounts of energy when running different programs
CRFIDs have inefficiencies Wasted Time E Black-out Threshold T • The more stored energy, the longer it takes to store additional energy • CRFIDs use capacitors as they are small and can recharge indefinitely • Voltage regulation è inefficient to operate with more stored energy • Storing excess energy is inefficient
Energy is harvested even while executing Closer to reader E E T T • Received power supplements stored energy • Reader frequency hopping è power changes every 400 ms • The amount of stored energy required depends on the distance from the reader and RF environment
Challenges to Running Programs Efficiently: Implications
Storing the right amount of energy increases performance 1 Normalized Task Rate 0.8 0.6 Program runs most efficiently when capacitor is charged to 1.8V 0.4 0.2 0 1.5 2 2.5 3 3.5 Wake − up Voltage • Wake-up voltage: Determines the amount of energy stored before starting program • Light WISP program: sample accelerometer, 1.5 m from reader
No fixed threshold works for all programs 1 Normalized Task Rate 0.8 0.6 Program runs most efficiently when cap is charged to 2.5V 0.4 Program won’t run at all at 1.8V 0.2 0 1.5 2 2.5 3 3.5 Wake − up Voltage • Heavy, non-deterministic program: sample accelerometer and transmit value to reader, 1.5 m
No fixed threshold works for all distances 1 Normalized Task Rate 0.8 0.6 Less supplemental power at 3 m means tag should 0.4 charge cap to 3V 0.2 0 1.5 2 2.5 3 3.5 Wake − up Voltage • Heavy, non-deterministic program, 3 m from reader • CRFID must adapt to program needs and environment
Outline • Intel WISP – A CRFID Tag • Challenges to Running Programs Efficiently • Dewdrop Design • System Evaluation
Dewdrop: An Energy-aware Runtime Adaptively find the wake-up voltage that maximizes execution rate for the program and RF environment Two factors that reduce execution rate: • Not storing enough energy: Program fails and it takes time to recharge and execute again • Storing too much energy: Overcharging wastes time Constraint: Runtime operation must be simple • Every active cycle costs energy • No floating point, no hardware multiply/divide
Adapt to the Program and Environment • Goal: Maximize execution rate à Minimize time wasted from program failure and overcharging • Heuristic: Total waste is minimized when the wasted time from failures and overcharging is equal • On program complete: Update running average of time wasted overcharging • On program failure: Update running average of time wasted failing • If Avg overcharge > Avg fail : decrease wake-up threshold by β • Else: increase wake-up threshold by β
Heuristic results in a good operating point 1 Normalized Value 0.8 Dewdrop finds 0.6 this operating point 0.4 Response Rate 0.2 Charge Waste Fail Waste 0 2 2.2 2.4 2.6 2.8 3 3.2 Wake � up Voltage • Equalizing the sources of wasted time results in efficient program execution
Dewdrop Implementation 1. Low power wake-up • No hardware mechanism to wake up at specified voltage • Dewdrop polls capacitor voltage periodically until target is reached • Exponentially adapted polling interval is lightweight and accurate 2. Low power voltage sampling • Waking up to sample voltage consumes precious energy • We reduced the energy cost of voltage sampling by a factor of 4 More details in the paper
System Evaluation
Dewdrop makes good use of scarce energy 80 Sense (Dewdrop) Task Rate (per second) Matches performance Sense (HwFixed) for light program Light, Dewdrop 60 SenseTx (Dewdrop) SenseTx (HwFixed) Light, Hardware Doubles range for 40 Heavy, Dewdrop heavy program Heavy, Hardware 20 0 1 1.5 2 2.5 3 3.5 4 Distance (m) • Compare to efficient, but inflexible, hardware mechanism • State-of-the-art before Dewdrop • Execution rate should scale with received power: 1/d 2
Dewdrop finds an efficient operating point 1 X Wake-up voltages X X X Normalized Task Rate and rates found 0.8 by Dewdrop 0.6 Light, 1.5 m 0.4 Heavy, 1.5 m 0.2 Heavy, 3 m 0 1.5 2 2.5 3 3.5 Wake − up Voltage • Dewdrop finds wake-up voltage within 0.1V of best • Generally achieves > 90% of max rate for all distances
Dewdrop increases application coverage 100 Increased Dewdrop 80 Percent of Tags Coverage Hardware 60 40 20 0 30 29 28 27 26 25 24 Transmit Power (dBm) • Elder care scenario: 1 reader, tagged objects in apartment • 11 WISPs streaming accelerometer data (3 trials) • Dewdrop can run the program with much less power
Conclusion • Running programs using harvested RF energy is feasible • Batteryfree è small, perpetual, embeddable • Dewdrop makes CRFIDs more usable and useful • Technology trends will increase range and performance • Passive device range expected to continue doubling every 4 years • WISP 5.0 in development • WISPs and tools are available to the community • WISP hw/sw open source, USRP-based RFID reader
Questions • WISP Wiki: wisp.wikispaces.com • UW Sensor Systems Group: sensor.cs.washington.edu • www.cs.washington.edu/homes/buettner • buettner@cs.washington.edu
Recommend
More recommend
