Mechanical Transport of Bits - Part I

1. Mechanical Transport of Bits - Part I Duo (Steve) Liu David Lee May 3, 2004 EE206a

2. Current Problems Two Principal problems with data collection in sparse sensor networks: • High energy consumption in multihop routing between widely separated nodes • Routing hotspots near the destination of the data that shortens the effective lifetime of the network

3. What is a Data MULE? • A data MULE is a mobile entity known as a Mobile Ubiquitous LAN Extension. • A MULE can • Pick up data from the sensors within radio range • Buffer it • Drop off the data to wireless access point

4. How can a Data MULE help in Sensor Networks • In the MULE architecture sensors save energy by • short range transmission • no routing protocol overhead • low duty cycle of radio listening • Any mobile entity with an attached transceiver can move about collecting data.

5. MULE Architecture Overview

6. MULE Architecture: Top Tier • A top tier of WAN connected devices, such as access points. • Abundant power and unrestricted network connectivity. • Synchronizes data that is collected.

7. MULE Architecture: Middle Tier • Mobile transport agents (MULEs) • Provides the system with the scalability and flexibility for a relatively low cost, which is the intended goal. • MULEs have large storage capacities, renewable power, and able to communicate with both sensors and access points.

8. MULE Architecture: Bottom Tier • Wireless sensors that are randomly distributed. • The goal is to reduce the work performed by the sensors because of the constraints such as energy and radio range for communication.

9. Performance Metrics • Data Success Ratio (DSR) • This measures the effectiveness of data delivery. It is defined as the ratio of the total amount of data transferred to the MULE to the total amount of data generated. • Latency • This is the average time taken by data to reach MULE from the time of its generation. • Communication Energy • Energy required for communication by the motes. We assume that by saving energy required for communication, lifetime can be improved substantially.

10. Parameter Space • Sensor related • Data generation rate (λ) • Sensor buffer size • Duty cycle • MULEs related • MULEs arrival time • Buffer size • Access point related • Distribution and the number of access-points • Radio related • Radio model • Data transmission rate

11. Parameters (cont.) • Parameters that are sufficient to characterize the performance metrics • Sensor data generation rate (λ) • Sensor buffer size (SB) • Amount of data transferred between a MULE and a sensor, denoted by K • MULEs arrival at a sensor • A MULE’s visit to access-points

12. Analytical Model • Queuing Theory • The buffer size of sensor defines the capacity of the queue • If the buffer is full then any new data is dropped • The arrival of a MULE in a sensor’s range is considered as a discrete event • This event causes transfer of data from the sensor’s queue to the MULE • The sensor then waits for the next MULE arrival event to transfer the data • The amount of data that can be transferred is taken as a fixed quantity, denoted as K

13. Assumptions • The MULEs inter-arrival times (times between arrival of two MULEs) are independent and identically distributed. • At a given time only one MULE interacts with a given sensor and vice versa • Sparse sensor networks where no two sensors are in the same radio range • MULEs have sufficiently large buffers • No data transmission loss

14. Analytical Model Results • Stability Condition • The system is stable (the queue reaches a unique stationary regimes) iff λ/Kμ<=1, 1/μ is the arrival rate • Data Success Ratio • DSR = • Average queuing delay • Wq =

15. Analytical Model Results (Cont.) • If the MULE arrival distribution and the data generation process is Poisson, then: • It allows us to obtain closed form results and is reasonable under certain environments • Distribution of Q depend only on the ratio of μ and λ • The absolute value of μ and λ is not important • Average amount of data transferred • K = CT x B • B is the radio transmission rate, CT is the amount of time the MULE is in the radio range of sensor

16. Impact of Sensor Duty Cycle • Sensor periodically listens for DT (the time for discovery) seconds every BT (beacon interval) seconds • Duty-cycle = DT/BT • Effective MULE arrival rate • μ*= μ(CT-DT)/BT • If CT – DT >= BT, μ* = μ • Effective amount of data that be transferred • K* = K(1-BT/2CT) if CT – DT >= BT • K* = K/2(1+DT/CT) if CT – DT < BT

17. Effects of scaling μ

18. Effect of scaling K

19. DSR Under Different Mobility Models

20. Latency Under Different Mobility Models

21. Performance Results • The performance results determined using analysis were close to (within 5%) results of detailed simulation • DSR • DSR is less than 60% if the parameters are chosen such that stability condition just met • DSR can be made close to one by increasing μ or SB • When K is large, choosing SB and μ such that μ*SB > 5λ resulted in a DSR greater than 95% • When K is small, the sensor buffer occupancy and latency is quite large. However, the performance improves sharply by increasing K initially and eventually saturates when K* > 3 x λ/μ* • Mobility models • Mobility models which have high variance perform worse than more deterministic models • The performance of Poisson arrivals and random waypoint were almost same and similar to deterministic arrivals • This indicates that MULEs with fixed mobility pattern are not much beneficial than random waypoint kind of motion

22. Data MULE vs Ad-hoc Network • Average Energy Ratio • The ratio of the average energy consumed at a sensor in the ad-hoc network to the energy consumed in the MULE architecture • Hotspot Ratio • The ratio of the hotspot usage in the ad-hoc network to the hotspot usage in the MULE architecture

23. Data MULE vs Ad-hoc Network (Cont.)

24. Data MULE vs Ad-hoc Network (Cont.) • Overall MULE’s performance is much better • Couple of issues about the results • The results are somewhat biased since it does not consider the energy consumption due to MULE • Latency in the MULE network is much more than latency in ad-hoc network

25. Data MULE vs Ad-hoc Network (Cont.) • Advantages • Far less infrastructure • No routing overhead • The system is more robust with a sufficient density of MULEs • More system flexibility • Disadvantages • High latency • The system presupposes a sufficient amount of physical movement in the environment • Possibilities of unexpected errors such as MULE failure or inability to reach sensors due to change in terrain

26. Data Mule Path in the Network • Data Mule (observer) moves on a single path through static sensor nodes • Model 1: Nodes are independent and uniformly and randomly scattered in A • Model 2: Nodes are placed such that no 2 sensors are less than d meters apart

27. The Proposed Model • Assumptions • Observer is not power-constrained • All sensors are identical • v = speed of observer • T = amount of time for sensor to send all its data to observer • R = distance of sensor location to the path • Rmax = communication range • Rmax≥ (R2+(vT/2)2)1/2 • Outage: an unsuccessful communication

28. Model 1: Independent Uniformly Distributed Nodes • May not be possible to collect data from all sensors • Several nodes may be located close together • All try to send data to the observer simultaneously • Observer may not be able receive data from all • Solution: • Increase transmission range • Increase data rate • Unfortunately, this causes increase in transmission power

29. Waiting Time vs Outage • Futile to start communicating with a sensor node that will not stay within range long enough • Max wait time = (1/v)[(2Rmax2-d2)1/2-vT]

30. Performance Results

31. Performance Results

32. Model 2: Random Distributed Nodes with Min. Separation d • Outage may not be acceptable in certain applications • Zero outage can be guaranteed by appropriate choice of parameters • d ≥ [(2R)2+(vT)2]1/2

33. Power Comparison

34. Power Comparison

35. Case 1: Mobile Observer • R = 500 m • Communication bandwidth = 1 MHz • Energy consumed per cycle by a single sensor = 0.3 µJ

36. Case 2: Static Observer – Single Hop Communication Observer • Different sensors have different communication ranges • Assuming all sensors transmit at the same power • Sensors far away will transfer as a slower data rate • Avg. Energy Consumed per cycle – 111.8 µJ

37. Case 3: Static Observer – Multihop Communication • Power consumption depends on path taken • To simplify, assume sensors are located in a uniform triangular lattice • Avg. Energy Consumed per cycle – 0.88 µJ

38. Communication Protocol • Protocol needs to be designed such individual sensors have little responsibilities other than collect data and sending it when requested • Life of a Sensor network • Startup • Steady state • Failure

39. Startup • Sensor nodes and data mule do not know each other’s position • Cycle 1: Sensor periodically listens to the channel for the data mule • Cycle 2: Data mule traveling on a regular path needs to broadcast a beacon • Sensor sends RTS signal • After receiving the RTS, data mule stops beacon and sends CTS to sensor • Sensor transmits packet to data mule

40. Steady State • Data mule • Accurate position knowledge of sensors • Send wake signal to sensors known to be in range • Sensor • Predict when the data mule arrives • Monitor channel only when data mule is expected to be nearby

41. Failure • Data mule detects failures through inability of node to respond to wake signal • Reschedule remaining nodes • If too many nodes fail, the network will not be able to gather sufficient data

42. Summary • Data mule requires far less infrastructure and little routing overhead • Sensor nodes consume less power in communication with the data mule • Data mule useful when low data latency is not a requirement

43. References • Arnab Chakrabarti, Ashutosh Sabharwal, Behnam Aazhang, “Using Predictable Observer Mobility for Power Efficient Design of Sensor Networks,” IPSN 2003. • Sushant Jain, Rahul C Shah, Gaetano Borriello, Waylon Brunette, Sumit Roy, “Exploiting mobility for energy efficient data collection in sensor networks,” Proceedings of Modeling and Optimization in Mobile, Ad Hoc and Wireless Networks, (WiOpt) Mar, 2004. • R. C. Shah, S. Roy, S. Jain, and W. Brunette, “Datamules: Modeling a three-tier architecture for sparse sensor networks,” Proceedings of IEEE SNPA, Anchorage, Alaska, 2003.