1 In-network data aggregation What is directed diffusion? Old way - - PDF document

1

11/29/2004 1

Directed Diffusion for Wireless Sensor Networking

With material from the subsequent paper, Building Efficient Wireless Sensor Networks with Low-Level Naming

By Eric Siegfried CSE 521

11/29/2004 2

Motivation

Content and data-centric

Where are nodes with X data over the next

5 minutes, at 1 second intervals?

Where is an object within some region?

Challenges:

Scalability Energy efficient Fault tolerant

11/29/2004 3

ID-based communication

Requires unique host ID addressing Application is end to end Integrating nodes into the network Tendency for less robust routes

11/29/2004 4

Building Efficient Wireless Sensor Networks with Low Level Naming

Cost of computation vs. communication

3000 instructions vs. 1 bit 100m

Attribute based communications

Linda, LIME (tuple space)

Typically expensive, not good for

resource constrained networks

11/29/2004 5

Routing issues

DSR and AODV recreate IP style

network

Preferable to be based on attribute-

value-operation tuples?

Can deal with rectangular region Can deal with interests, non specific node

11/29/2004 6

Filters

Access information about diffusion

(gradients, list of neighbors) for in- network aggregation

Can help to avoid flooding Peer to peer behavior, all can filter Power efficient

2

11/29/2004 7

In-network data aggregation

Old way

Binding a service to a geographical region

which lists the node identifiers

Elect one or more nodes to aggregate

New way

Name nodes via geographic attributes Run application specific filters, and inject

application specific code into the network

11/29/2004 8

What is directed diffusion?

Interests Data Messages Gradients Reinforcements

11/29/2004 9

Naming tasks

Interest type = four-legged animal interval = 20 ms duration = 10 sec rect = [-100, 100, 200, 400] Reply type = four-legged animal instance = [125, 220] intensity = 0.6 confidence = 0.85 timestamp = 01:20:40

11/29/2004 10

Interest propagation

Interests diffuse through the network,

periodically refreshed by the sink

Interests only contain information about

the previous hop

Checks cache, may create an entry Checks for gradient, may add one Reinforced path, delivery of data

11/29/2004 11

Gradient Establishment

value, and direction

Rate per hour Timestamp expiration

11/29/2004 12

Data propagation

Seeing data which meets a certain

intensity and confidence

Referenced in interest cache Gradient data rate vs. event data rate

3

11/29/2004 13

Reinforcement, making paths

Exploratory events Positive reinforcement Dealing with multiple sources and sinks

Include figure 3c, 3d

11/29/2004 14

Low

Exploratory Gradient

Event Event

Low Low Exploratory Request Gradient

Bidirectional gradients established on all links through flooding

11/29/2004 15

Local Behavior Choices

• 1. For propagating interests

In our example, flood More sophisticated behaviors possible: e.g. based on cached information, GPS

• 2. For setting up gradients

Highest gradient towards neighbor from whom we first heard interest Others possible: towards neighbor with highest energy

• 3. For data transmission

Different local rules can result in single path delivery, striped multi-path delivery, single source to multiple sinks and so on.

• 4. For reinforcement

reinforce one path, or part thereof, based on observed losses, delay variances etc.

• ther variants: inhibit certain

paths because resource levels are low

11/29/2004 16

Illustrating Directed Diffusion

Sink 11/29/2004 17

Illustrating Directed Diffusion

Sink 11/29/2004 18

Illustrating Directed Diffusion

Sink

4

11/29/2004 19

Illustrating Directed Diffusion

Sink 11/29/2004 20

Illustrating Directed Diffusion

Sink 11/29/2004 21

Illustrating Directed Diffusion

Sink 11/29/2004 22

Illustrating Directed Diffusion

Sink 11/29/2004 23

Illustrating Directed Diffusion

Sink 11/29/2004 24

Illustrating Directed Diffusion

Sink Source

5

11/29/2004 25

Illustrating Directed Diffusion

Sink Source 11/29/2004 26

Illustrating Directed Diffusion

Sink Source 11/29/2004 27

Illustrating Directed Diffusion

Sink Source 11/29/2004 28

Illustrating Directed Diffusion

Sink Source 11/29/2004 29

Illustrating Directed Diffusion

Sink Source 11/29/2004 30

Illustrating Directed Diffusion

Sink Source

6

11/29/2004 31

Illustrating Directed Diffusion

Sink Source 11/29/2004 32

Illustrating Directed Diffusion

Sink Source 11/29/2004 33

Illustrating Directed Diffusion

Sink Source 11/29/2004 34

Illustrating Directed Diffusion

Sink Source 11/29/2004 35

Illustrating Directed Diffusion

Sink Source 11/29/2004 36

Illustrating Directed Diffusion

Sink Source

7

11/29/2004 37

Illustrating Directed Diffusion

Sink Source 11/29/2004 38

Illustrating Directed Diffusion

Sink Source 11/29/2004 39

Local repair for failed paths

Reaction to corruption, degradation

Include figure 3d

Wasting resources to find the lossy link

11/29/2004 40

Negative reinforcement (path truncation)

Include figure 4a

Soft state to time out data gradients

Reset gradients to be exploratory

11/29/2004 41

Negative Reinforcement (loop removal)

Not always removed

Figure 4b vs 4c

Some useful paths even though looped

11/29/2004 42

Simulation

8

11/29/2004 43

Simulation(2)

11/29/2004 44

Simulation(3)

11/29/2004 45

Simulation(4)

11/29/2004 46

Conclusions

Nature of directed diffusion paper Weaknesses

Did non develop software architecture for

realizing attributes and in-network processing in an OS

Simulation did not include radio

propagation

Costly to have exploratory packets

11/29/2004 47

Micro-Diffusion

5.5 KB of memory Supports 5 gradients and 10 packets

(2B per packet) per node

Reduced in size, but header format is

compatible with full diffusion implementation

Filters currently not implemented for

micro-diffusion

11/29/2004 48

Nested Queries

Preference to second method, more power efficient.

9

11/29/2004 49

Energy Savings

Suppression uses less data per event with multiple sources

11/29/2004 50

Power Savings(2)

If no aggregation, each source pays

cost of full path, whereas aggregation pays only one hop prior to aggregation

Understanding aggregation is an

important area of future work

Saves around 40% traffic to aggregate Nested queries reduce loss by 15-30%

11/29/2004 51

Future work

How diffusion’s parameters map to

different needs

Trade offs between overhead and

reliability in exploratory messages, interests, and reinforcements

Adding feedback and congestion control

to diffusion

11/29/2004 52

Future work(2)

Energy aware MAC protocol needed New applications, and how signal

processing interacts (for example) with in network processing and filters

Impact on wired networks

11/29/2004 53

Credits

Dushanth Sandeep Gupta