From Smart Dust to Reliable Networks

1. From Smart Dust to Reliable Networks Kris Pister Prof. EECS, UC Berkeley Founder & CTO, Dust Networks

2. Outline • Conclusion • Background • The Science Project • Market • The Hype • Technology • Challenges • Status • Applications • Open Research Problems

3. Conclusion • The market is real • Industrial Automation, Building Automation • $100M? in 2006, $500M by 2010 • Net impact of academic research on this market to date: zero to slightly negative • Open problems in “ADA*” • Metrics/optimal reliability and power consumption • Time synchronization • Deeply duty cycled 802.11 end-points • Certification, binding, commissioning *ADA = Academically Dull Applications

4. Grand Challenge A B C Reliably, at low power

5. Smart Dust Goalc. 1997

6. Smart Dust, 2002 RECEIVER OPTICAL IN SENSORS ADC FSM 375 kbps 16 mm3 total circumscribed volume ~4.8 mm3 total displaced volume 8-bits PHOTO TRANSMITTER OPTICAL OUT 175 bps 1V 1-2V 3-8V 1V 1V 2V SOLAR POWER

7. UCB “COTS Dust” Macro Motes Services David Culler, UCB Networking TinyOS Rene 00 Mica 02 Dot 01 Demonstrate scale • Designed for experimentation • sensor boards • power boards NEST open exp. platform 128 KB code, 4 KB data 50 KB radio 512 KB Flash comm accelerators WeC 99 James McLurkin MS Small microcontroller - 8 kb code, 512 B data Simple, low-power radio - 10 kbps EEPROM storage (32 KB) Simple sensors

8. University Demos – Results of 100 man-years of research Motes dropped from UAV, detect vehicles, log and report direction and velocity Intel Developers Forum, live demo 800 motes, 8 level dynamic network, 50 temperature sensors for HVAC deployed in 3 hours. $100 vs. $800 per node. Seismic testing demo: real-time data acquisition, $200 vs. $5,000 per node vs.

9. Sensor Networks Take Off! Industry Analysts Take Off! $8.1B market for Wireless Sensor Networks in 2007 Source: InStat/MDR 11/2003 (Wireless); Wireless Data Research Group 2003; InStat/MDR 7/2004 (Handsets)

10. Wireless Sensor Networking Decision Systems • Significant reduction in the cost of installing sensor networks • Enables new class of services • Increases sensor deployment Monitoring Systems Control Systems Enterprise Applications Digital Sensors and Actuators Serial Devices Analog Sensors and Actuators Physical World

11. $748,000,000 in ‘03 WDRG, 2003

12. Cost of Sensor Networks Mesh Networking Computing Power Installation, Connection and Commissioning Sensors $ Time

13. Low Data Rate WPAN Applications (Zigbee) PERSONAL HEALTH CARE BUILDING AUTOMATION CONSUMER ELECTRONICS security HVAC AMR lighting control accesscontrol TV VCR DVD/CD remote PC & PERIPHERALS INDUSTRIAL CONTROL asset mgt process control environmental energy mgt mouse keyboard joystick RESIDENTIAL/ LIGHT COMMERCIAL CONTROL patient monitoring fitness monitoring security HVAC lighting control access control lawn & garden irrigation

15. Mesh Systems

16. Dust Networks • Founded July 2002 • Angels, In-Q-Tel, ~$1.5M • 28 employees in Jan 04 • Series A Feb 2004 • Foundation • IVP • Series B Feb 2005 • Crescendo • Cargill

17. Network Architecture • Goals • High reliability • Low power consumption • No customer development of embedded software • Customer visibility into all aspects of network operation/status/health • Minimal/zero customer RF/networking expertise necessary • Challenges • 1W emitters in regulated but unlicensed RF bands • Extreme computation and communication resource constraints • MIPS, RAM, bps

18. What do OEMs and SIs want? ^ and scientists and and engineersand startups and grad students and…. • Reliability • Reliability • Reliability • Low installation and ownership costs • No wires; >5 year battery life • No network configuration • No network management • Typically “trivial” data flow • Regular data collection • 1 sample/minute…1 sample/day? • Event detection • Threshold and alarm

19. Reliability • Hardware • Temperature, humidity, shock • Aging • MTBF = 5 centuries • Software • Linux yes (manager/gateway) • TinyOS no (motes) • Networking • RF interference • RF variability

20. IEEE 802.15.4 & WiFi Operating Frequency Bands Channel 0 Channels 1-10 2 MHz 868MHz / 915MHz PHY 868.3 MHz 902 MHz 928 MHz 2.4 GHz PHY Channels 11-26 5 MHz 2.4 GHz 2.4835 GHz Gutierrez

21. 900 MHz cordless phone -50dBm Solid mote signal -20dBm

22. Spatial effect of multipath

23. Frequency dependent fading and interference From: Werb et al., “Improved Quality of Service in IEEE 802.15.4 Networks”, Intl. Wkshp. On Wireless and Industrial Automation, San Francisco, March 7, 2005.

24. Beware of static measurements and RF pathloss simulations • Site surveys need to be done over at least 24 hours • Simulation tool results need much more speckle Pictures from www.wirelessvalley.com

25. Radio Reliability in a Crowded Spectrum • UWB? • Unclear potential for duty cycling • DSSS doesn’t cut it • Helpful, but only about 10dB • +20 dBm doesn’t cut it • Helpful, but expensive in batteries • 802.11 & cordless phones • Must frequency hop • Time synchronization required… …but you probably needed that anyway. • Lots of channels, lots of bandwidth, better scaling, …

26. Network Types Why not use 802.11? Full Mesh Star Star-Mesh Powered mesh infrastructure Star-connected sensors No infrastructure Mesh-connected sensors X X

27. Zigbee 1.0 • Single channel networks are built into standard. This will be fatal for reliability. • Tree-based routing recommended by standard will likely not be adopted, especially given the single-channel radio. • No definition of duty cycling routers • Assumes powered routers, battery powered leaf nodes • No explicit prevention of router duty cycling – Zigbee 2.0?

28. Power-optimal communication A B A wakes up and listens B transmits B receives ACK A transmits ACK Worst case A/B clock skew • Assume all motes share a network-wide synchronized sense of time, accurate to ~1ms • For an optimally efficient network, mote A will only be awake when mote B needs to talk Expected packet start time

29. Packet transmission and acknowledgement Radio TX startup ACK RX Packet TX Radio TX/RX turnaround Mote Current Energy cost: 295 uC

30. Fundamental platform-specific energy requirements • Packet energy & packet rate determine power • (QTX + QRX )/ Tlisten • E.g. (300 uC + 200 uC) /10s = 50 uA

31. Idle listen (no packet exchanged) ACK RX Radio RX startup Mote Current Energy cost: 70 uC

32. Scheduled Communication Slots A B B TX, A ACK Ch 3 Ch 4 Ch 5 Ch 7 Ch 8 Ch 6 • Mote A can listen more often than mote B transmits • Since both are time synchronized, a different radio frequency can be used at each wakeup • Time sync information transmitted in both directions with every packet

33. Latency reduction A B B TX, A ACK • Energy cost of latency reduction is easy to calculate: • Qlisten / Tlisten • E.g. 70uC/10s = 7uA • Low-cost “virtual on” capability • Latency vs. power tradeoff can vary by mote, time of day, recent traffic, etc. Tlisten

34. Latency reduction A B G • Global time synchronization allows sequential ordering of links in a “superframe” • Measured average latency over many hops is Tframe/2 T2, ch y T1, ch x Superframe

35. Time and Frequency A C B Time One Slot Freq 902.5 MHz 903 MHz … 927.5 MHz One Cycle of the Black Frame • Graphs & Links are abstract, with no explicit time or frequency information. • Frames and slots are more concrete • Time synchronization is required • Latency, power, characteristic data rate are all related to frame length • Relative bandwidth is determined by multiplicity of links

36. Time and Frequency A C B Time Channel Cycle N+1 Cycle N Cycle N+2 • Every link rotates through all RF channels over a sequence of NCH cycles • 32 slots/sec * 16 ch = 512 cells/sec • Sequence is pseudo-random

37. 50 channels, 900 MHz 900MHz 930MHz

38. 16 channels, 2.4 GHz 2.4GHz 2.485 GHz

39. Configure, don’t compile SmartMeshTM Console IP Network XML SmartMesh Manager Mote ~100 ft Reliability: 99.99%+ Power consumption: < 100uA average

40. 50 motes, 7 hops 3 floors, 150,000sf >100,000 packets/day

42. Communication Abstraction Network Services Configurable Data Filter/Control IP Network XML B H C Analog I/O Digital I/O Serial Port A SmartMesh Manager E G F • Packets flow along independent digraphs • Digraphs/frames have independent periods • Energy of atomic operations is known, (and can be predicted for future hardware) • Packet TX, packet RX, idle listen, sample, … • Capacity, latency, noise sensitivity, power consumption models match measured data • Build connectivity & applications via xml interface

43. Available data IP Network XML SmartMesh Manager • Connectivity • Min/mean/max RSSI • Path-by-path info: • TX: attempts, successes • RX: idle, success, bad CRC • Latency (generation to final arrival) • Data maintained • Every 15 min for last 24 hours • Every day for last week • Lifetime • Available in linux log files or via XML

44. Micro Network Interface Card Network Services Configurable Data Filter/Control Analog I/O Digital I/O Serial Port mNIC • No mote software development • Variety of configurable data processing modules • Integrators develop applications, not mesh networking protocols • For compute-intensive applications, use an external processor/OS of your choice.

45. Energy Monitoring Pilot • Honeywell Service: monitor, analyze and reduce power consumption • Problem: >> $100/sensor wiring cost • Solution: • Entire network installed in 3 hours (vs. 3-4 days) • 9 min/sensor • Software developed in 2 weeks (XML interface) • 12 months, 99.99%

46. Chicago Public Health – Dust, Tridium, Teng Temperature and power monitoring

47. Tridium NiagraAX

48. Micro Network Interface Card Network Services Configurable Data Filter/Control Analog I/O Digital I/O Serial Port mNIC • No mote software development • Variety of configurable data processing modules • Integrators develop applications, not mesh networking protocols • For compute-intensive applications, use an external processor/OS of your choice. Sensor uP

49. Perimeter Security Passive IR Passive IR and Camera 1.5 in MEMS and GPS 2.5 in 2.5 in

50. SAIC Field Demonstration with the USBP • Two hundred 2nd generation sensor nodes • Geophone, passive IR, magnetometer, camera, radiation sensors • Packaged for long lifetime, easy deployment, and effective concealment • Deployed on the U.S. – Mexico border for onsite assessment and persistent unattended surveillance Demonstration in Naco, Arizona Improved electronic surveillance needed to supplement the limited physical security offered by fences