Title: MOBICOM 2002
1 Directed Diffusion for
Wireless Sensor Networking C. Intanagonwiwat, R.
Govindan, D. Estrin, John Heidemann, and
Fabio Silva
2Contents
- Introduction
- Directed Diffusion
- Interest and Data Naming
- Interest Propagation and Gradients Set-up
- Data Propagation
- Reinforcement
- Simulations
- Conclusion
3Introduction
- Problem How can we get data from the sensors?
- Sensor network
- Frequent Node Failure
- Energy-Constraint
- Request Driven
- Task sink-gtsensors (query dissemination)
- Event sensor source-gtsink
- Data Centric
- Communication is for named data
- Diffusion closely resembles some ad-hoc routing
4Interest and Data Naming
- Interest/Query
- Type tank
- Interval 10ms (event data rate, 100 events per
second) - Rect -100, 100, 200, 400
- Timestamp 01 20 40
- ExpiresAt 01 30 40
- Data/Reply
- Type tank
- Instance 150, 220
- Location 125, 220
- Intensity 0.6
- Confidence 0.85
- Timestamp 012040
- Named using Attribute-Value Pairs
Duration10 min (time to cache)
5Interest Propagation and Gradients Set-up
- Sink periodically broadcasts interest
- Exploratory interest with a large interval
- Low data rate (few data packets are need in unit
time) - Neighbors update interest-cache and forwards the
interest - Flooding
- Directional flooding based on location.
- Directional Propagation based on previously
cached data - Gradients set-up
- Gradients are set up to the upstream neighbors
- Weight data rate
Interest(type) Timestamp Gradient1(data rate) Gradient2 .. Duration
6Exploratory Gradient
Exploratory Request Gradient
Event
Bidirectional gradients established on all links
through flooding
7Data Propagation
- If Event occurs,
- Search interest cache for matching interest
entry - Compute the highest event rate among all its
gradients, - and Sample events at this rate
- And Send data to the relevant neighbors
- Receiving node
- Find matching entry in interest cache, no match
silent drop - Check and add data cache (loop prevention)
- Re-send message with appropriate rate
(down-conversion)
8Exploratory events
Exploratory event initial interest? ?? event
Instance 150, 220
Source
Instance 150, 220
Sink
9Positive Reinforcement
- After sink starts receiving exploratory events,
Reinforces one particular neighbor for real data - Is achieved by data driven local rules
- Example of such a rule
- Receives previously unseen event from a neighbor
- Sink re-send original interest with a smaller
interval (higher data rate) - Receiving node also reinforce at least one
neighbor - Using data cache
- Example neighbor from which it first received
the latest event matching the interest
10Positive Reinforcement (Contd)
Source
11Positive Reinforcement (Contd)
Source
Instance 150,300
We reinforce that neighbor if it is sending new
events
12Positive Reinforcement (contd)
- Its possible more than one path being reinforced
- Selects empirically low-delay path
- When one path delivers event faster,
- Sink uses this path for high-quality data
13Negative Reinforcement
- Negatively reinforce a path
- To time-out data gradient unless it is explicitly
reinforced - To explicitly send negative reinforcement message
- Local repair for failed paths
- When C detects its failure, negatively reinforce
failed link and reinforce another path
14Simulations
- Vehicle tracking system in ns-2
- 3 Metrics
- Average dissipated energy
- Average delay
- One way latency between transmitting events and
receiving it - Distinct-event delivery ratio
- These metrics are studied as a function of
network size.
15Parameter setting
- Sensor field 50 nodes in 160m x 160m square
- Radio range is 40m
- Keep the average density of sensor nodes constant
- 5 sources and 5 sinks (? low load)
- Each source generates two events per second
- Rate for exploratory events is one event per 50
seconds - Window for negative reinforcement is 2 seconds
- 1.6Mb/s 802.11 MAC
- Energy model
- Idle time 35mW
- Receiving power 395mW
- Transmission power 660mW
16Average dissipated energy
- Multiple path
- Reinforcement is very aggressive
- Negative reinforcement is very conservative
- Listening energy
Omniscient multicast is idealized scheme, but has
no data aggregation.
17Average delay
Reinforcement rules seem to be finding the low
delay paths
18Event Delivery Ratio with node failures
Turn off 1020 nodes for 30 seconds,
repeatedly Each source sees different vehicles
19Average delay with node failures
20Average dissipated energy with node failures
21Negative reinforcement
22Duplicate suppression
23High idle radio power
ATT Wavelan 1.6W (for transmission), 1.2W (for
reception), 1.15W (for idle time)
24Conclusions
- Directed Diffusion is significant energy
efficient. - Directed Diffusion is stable under node failures.
- Performance depends on sensor radio MAC layers.
25Acknowledged problems
- Experiments did not evaluate operation under high
load - Reinforcing multiple routes leads to wasteful
excess transmissions - Experiments used the wrong MAC layer