1. Introduction to Wireless Sensor Networks
Directed Diffusion
28 March 2005

2. C. Intanogonwiwat R. Govindan D. Estrin
Directed Diffusion : A Scalable and Robust Paradigm for Sensor Networks
C. Intanogonwiwat
R. Govindan
D. Estrin

3. Introduction and terminology …..
Availing cheap nodes for sensing, communication and computation
Deploying them in a region of interest to form a network and sensing environment phenomena ( events )
Events are transmitted ( directed ) from the sensing nodes( source ) to a destination ( sink ) for processing.

4. Simplified view….

5. Example of events …… Detecting variations in temperature
Seismic vibrations
Detecting any object like a four-legged animal in an area under inspection

6. Objective…… Making the routing algorithm
1)energy efficient : Optimizing radio communications, efficient routing and performing local computations
2)Scalable : Scale with an increase in the number of source and sinks
3)Robust : Handling node failures

7. Two ways of packet forwarding during routing….
Address Centric: The nodes route data independently without looking at the data content.
Data Centric: The nodes while routing data use aggregation functions to eliminate redundancy.
Our focus is data centric.

8. Assumptions …… Data centric routing
Achieving a desired global behavior through local interactions
Application aware – the task types are known at the time the sensor network is deployed

9. Directed Diffusion basics…..
A sink node expresses interest in a particular data and inserts it as a query in the network
Sensor nodes reply to this interest
An interest may look like “At every I ms for the next T seconds send me a location estimate of any four legged animal in sub region R of the sensor field “

10. Possible naming (structure) of an Interest…
type = four-legged animal
Interval = 10ms
Rect = [-100,200,300,400]
Timestamp = 01:22:35
expiresAt = 01:30:40
Consists of attribute value pairs – its like querying the network for a particular data

11. Interest propagation……..
Flooding
Geographic routing ( filtering out the interests on basis of the coordinate specification )
Using cached data to find out which neighbor had previously responded to similar interest
Any other intelligent way, depending on the application

12. Establishing Gradients…..
Done between every pair of nodes
Consists of a <rate, direction > pair
E.g. the gradient from A to neighbor B
rate : the inverse of the value of the Interval in the interest sent by B
direction : The link to B ( A might have many neighbors – a local naming is required )
They are used for sending back data to the sink – the path with the highest gradient is generally preferred

13. Simplified view….

14. The Algorithm…….
Initially the sink sends an exploratory interest ( with a low data rate i.e. high interval )
The sensors store it in an Interest cache and forwards it. Subsequent interests having same type,interval,rect values are suppressed – thus selective forwarding
Gradients set up between neighbors

15. Algorithm(cont……)
A sensor whose sensed value matches with the type in an interest samples the readings based on the stored interval and sends it to all the neighbors with which it has a gradient
The intermediate sensors route the data based on the gradient in that direction
Eventually the sink receives the sampled information through some neighboring node

16. Directed Diffusion…. Directional Flooding Interest Gradient Sink
Source

17. Directed Diffusion….
Interest
Gradient
Sink
Source

18. Directed Diffusion….
Interest
Gradient
Sink
Source

19. Directed Diffusion….
Gradient
Sink
Source

20. Data Caching…….
Helps in suppressing similar interests from different sinks
Helps in suppressing similar event information from different sources and helps in data aggregation

21. Data propagation…..
The sources send back the data along the paths which were set up
Interest
Reply

22. Reinforcement …..
The sink chooses a high quality( optimal path ) by choosing the appropriate neighbor
(using greedy strategy) and reinforces it by
1) sending an interest packet with a lower interval to that link
2) negatively reinforce non-optimal links

23. Reinforcement…..
The reinforced interest is forwarded by each sensor node till it reaches the source
The exploratory gradients exist which helps the network to be robust in case of node failures.

24. Example of reinforcement….
original interest
reinforced interest

25. Directed Diffusion…. Directional Flooding Interest Gradient Sink
Source

26. Directed Diffusion….
Interest
Gradient
Sink
Source

27. Directed Diffusion….
Interest
Gradient
Sink
Source

28. Directed Diffusion….
Gradient
Sink
Source

29. Directed Diffusion….
Reinforcement
Gradient
Sink
Source

30. Directed Diffusion….
Reinforcement
Gradient
Sink
Source

31. Directed Diffusion….
Reinforcement
Gradient
Sink
Source

32. Directed Diffusion….
Data
Gradient
Sink
Source

33. Negative reinforcement…..
Send interest packets with higher interval to faulty links or links with higher delay.
A measure to reduce redundant communication after finding out the optimal path

34. Directed Diffusion….
Data
Gradient
Sink
Source

35. Directed Diffusion robustness….
Data
Gradient
Sink
Source

36. Directed Diffusion….
Data
Gradient
Reinforcement
Sink
Source

37. Directed Diffusion….
Data
Gradient
Reinforcement
Sink
Source

38. Directed Diffusion….
Data
Gradient
Reinforcement
Sink
Source

39. Design considerations……

40. Multiple sources……
source
sink
Data aggregation…

41. Multiple sinks……
source
sink
sink

42. Evaluation metrics……….
Average delay : average one way latency between transmitting an event and receiving it at the sink
Average dissipated energy : ratio of the total dissipated energy per node to the number of distinct events seen by the sink
Event delivery ratio : number of distinct events received to the number originally sent
ns-2 simulator used for evaluation

43. Compared with ……
Flooding : unrestricted broadcast of events to the sink nodes
Omniscient multicast : Each source transmitting along the shortest path multicast tree to the sink nodes

44. DD performance graphs….

45. Impact of node failures….

46. Observations…….. Energy efficient – outperforms omniscient multicast
Robust and fault tolerant
Works only for query driven networks

47. Rumor Routing….. Two types of data delivery models ….
Push(Event driven): Sources push data to the sink
Pull (Query driven) : The sink pulling data from the sources
A hybrid approach – rumor routing
Rumor routing results in lesser number of transmissions than either of the above in certain situations.

49. Algorithm…. Sources on observing events create agents
Agents carry routing information and go on a random walk across the network
The have a fixed Time-To-Live
Routing information is carried to nodes in the form of rumors and recorded.
Agents also synchronize themselves with information from nodes.

50. Power of agents…

51. Rumor routing ….

52. Features…
Once a query finds a recorded rumor it gets a definite direction/route to the event source.
Saves on transmissions by avoiding groping around for the sources
Also can get quick information about the event instead of having to go all the way to the source.
Needs more storage, agent complexity involved

53. Thank you …