1. 1 Environmental Energy Management in Pervasive Computers Aman Kansal, John Curnutte, Gautam Kachroo CS 218, Prof Mario Gerla

2. 2 What are sensor networks? Networks of numerous deeply embedded devices which sense (and control) the environment –Wireless, low energy, low datarate, autonomous Berkeley’s deployment at Great Duck Island At Huntington Botanical Gardens UCLA’s seismic array (under construction) NESL, UCLA’s localization array

3. 3 Energy Scarcity Batteries are too big Batteries don’t last forever Methods exist to extract energy from the environment –Solar –Vibrational –Wind –biochemical Smart Dust, Berkeley Prototypes from IASL, UWE, Bristol.

4. 4 Environmental Energy is Different from Battery Energy Supply varies with time –Need to adapt performance Supply varies in space –Different nodes get different energy Supply is repetitive (does not die out) –Opportunity to last forever

5. 5 The Problem Observation: Performance depends on how efficiently the available resources are used Question: How can the networking performance be adapted to the spatio- temporal characteristics of the energy availability? –Make the network last forever (until the hardware is outdated/damaged) –Scheduling: Who does more work

6. 6 OUTLINE Hardware Implementation Theoretical Framework What performance can be supported eternally? Algorithm Design How to achieve it within practical constraints of sensor networks?

7. 7 Source Characterization Leaky bucket like model: Apart from maximum flow, add a constraint for minimum flow too. Definition: E(t) is a (  -  1 -  2 ) source if for all T:

8. 8 Harvesting Theory THEOREM 1: If a system is powered by a (     ) source has energy storage capacity >= (      operates at constant power level  then it utilizes the energy source fully can survive forever. - Larger battery gives no performance gain - At least  rate of energy consumption can be supported

9. 9 Proof Outline Proof follows by contradiction –Maximum energy we can use = everything provided by environment Maximum  = average rate of energy –No energy from the environment is wasted: ensured by first condition Assume Initialization phase during which battery charged to  2 Consider the first instant when some small energy  E overflows (in time  t), then E(  t) =  t +  E Consider the preceding time duration t in which the battery had filled up  1 capacity: E(t) =  t +  1 Then E(t+  t) =  t +  1 +  t +  E >  t +  t) +  1 which is a contradiction –Similarly using other condition can prove that  can always be supported

10. 10 Not Operating at Constant Power Assume a leaky bucket consumer (  ’,  ) –No constraint on minimum energy usage Theorem 2: If a (  ’,  ) consumer powered by (  -     ) source, has battery size    , it can survive eternally   Workload Sensor Node Energy Consumption Overflow: Energy Wasted

11. 11 Scheduling Issues Need to determine      ’  from measured energy availability and load –Can then provide estimate of feasible performance region Each node knows only its own energy and a distributed scheduler is desired

12. 12 Algorithm Design Estimate the     parameters of the source –  estimated by a running average until difference between recent maxima and minima is within an acceptable margin Works for a cyclic source (such as the sun) –Can estimate     by locating largest underflow and overflow Done only at design time to estimate required battery size

13. 13 Performance Control Several options: sleep, Dynamic Voltage Scaling, radio range control etc. Sleep Duty Cycle –Chosen as supported by most hardware –radio range does not affect mote radio power consumption significantly Sensor Node Sensor SNOOZE: Processor and Radio sleeping Sleep Timer Event Detected Timer Expired

14. 14 Communication in the presence of Sleep Node can wake up if it has data to send How does a sleeping node receive data? –No time synchronization required between nodes Transmitter Beacon Wait for ACK Receiver Wake to listen for Beacon Sleep Beacon Heard ACK received Data generated DATA

15. 15 Optimal Routing for Maximum Lifetime is Impractical Optimal routing can be formulated as a linear programming problem Needs complete information of traffic flow Needs global knowledge of routes Needs global knowledge of energy characteristics at each node

16. 16 Networking Method: Practical Design Routing Algorithm for a event monitoring sensor network –Single Sink (base station) Multiple Sources (nodes with events) –Base station sends INIT Receiver sends ACK and forwards INIT Nodes measure energy and calculate latency

17. 17 Networking Method: Practical Design Path latencies important in sleepy network –But sending all latencies to base station reduces scalability In-network processing to compute path latency –Receive path latencies from children –Forward highest plus own latency

18. 18 Pseudo-code Initialize parameters T sleep = 0, childset = empty, childlatencylist = empty, parent = UNK, is_leaf = FALSE Spawn thread to estimate L. Generates L estimated event when finished. If received INIT message If parent == UNK parent = sender-id from INIT message Send ACK Rebroadcast INIT message with own ID Start ACK timer If received ACK message childset = childset U sender-id from ACK Delete ACK wait timer If ACK wait timer expires, set is_leaf = TRUE If L estimated event received If is_leaf == TRUE send LATENCY message to parent Adjust T sleep and duty cycle for this L Else If childlatencylist is complete Calculate L and send LATENCY message Adjust duty cycle and T sleep Else wait for childlatencylist to complete If received LATENCY message store latency value in childlatencylist

19. 19 Simulation Study Network spread in an outdoor environment with 50 nodes Solar data –Real data was collected with a single node for 9 days. –Generated random data distributed uniformly around this, nodes assume to receive this energy Latency performance of our protocol (fully distributed) compared to lowest latency route –Lowest Latency Route: found using Bellman-Ford algorithm (link latency used as cost) –high message overhead

20. 20 Simulation Study Frequency (%) LATENCY of a randomly deployed network Latency (s)

21. 21 Worst Case Path Latency (s) Distributed Routing Optimal Routing Results averaged over 20 random topologies Simulation Study Node Density

22. 22 Implementation Mote sized Solar Cells Battery Chemistry Specific Charging circuitry Environmental and Battery Energy Monitoring Voltage Level Convertor Customized node and sensor board Interfaces MK-2 Mica 2 Generic Heliomote (NESL, UCLA): Harvesting hardware added to Berkeley mote for energy harvesting POWER ENERGY DATA

23. 23 Implementation Our network uses uses available Heliomotes and some standard motes TinyOS Device Drivers for new harvestor hardware provide simple user interface: –async command result_t getData() –async event result_t dataReady

24. 24 Measured Data: Single Node Time (10 minute resolution) Using the solar data collected for 9 days using heliomote and the mote consumption characteristics:  = 23.6mW  1 = 1.463 MJ  2 = 1.856 MJ  = 153 J According to Theorem 2: battery required is only  1 +  2 = 3.32 MJ = 922.43mAh at 3V (about 1/2 of AA, could use smaller AAA battery) Note: Larger battery will not help sustainable performance Intensity  (t) E(t)

25. 25 Network Monitoring for Test- bed Apart from protocol messages, send each node characteristics to root node for monitoring Assumptions –Adapt to instantaneous current for demo –Mote range large, two hop created by address blocking 1 2 3

26. 26 Demo

27. 27 Future Work Setting up a network with more helio- motes Network tests in outdoor environment Meeting specified latency bounds Routing for other scenarios –Any node to any node –Energy aware routes

28. 28 Thank you