1. Techniques for Building Long-Lived Wireless Sensor Networks
Jeremy Elson and Deborah Estrin
UCLA Computer Science Department
And
USC/Information Sciences Institute
Collaborative work with R. Govindan, J. Heidemann, and SCADDS of other grad students

2. What might make systems long-lived?
Consider energy the scarce system resource
Minimize communication (esp. over long distances)
Computation costs much less, so:
In-network processing: aggregation, summarization
Adaptivity at fine and coarse granularity
Maximize lifetime of system, not individual nodes
Exploit redundancy; design for low duty-cycle operation
Exploit non-uniformities when you have them
Tiered architecture
New metrics

3. What might make systems long-lived?
Robustness to dynamic conditions: Make system self-configuring and self-reconfiguring
Avoid manual configuration
Empirical adaptation (measure and act)
Localized algorithms prevent single points of failure and help to isolate scope of faults
Also crucial for scaling purposes!

4. The Rest of the Talk
Some of our initial building blocks for creating long-lived systems:
Directed diffusion - a new data dissemination paradigm
Adaptive fidelity
Use of small, randomized identifiers
Tiered architecture
Time synchronization

5. Directed Diffusion A Paradigm for Data Dissemination
Key features
name data, not nodes
interactions are localized
data can be aggregated or processed within the network
network empirically adapts to best distribution path, the correct duty cycle, etc.
1. Low data rate
2. Reinforcement
3. High data rate

6. Diffusion: Key Results
Directed diffusion
Can provide significantly longer network lifetimes than existing schemes
Keys to achieving this:
In-network aggregation
Empirical adaptation to path
Localized algorithms and adaptive fidelity
There exist simple, localized algorithms that can adapt their duty cycle
… they can increase overall network lifetime
Average Dissipated Energy
(Joules/Node/Received Event)
Network Size (nodes)
0.005
0.01
0.015
0.02
0.025
0.03 50
100
150
200
250
300
Diffusion without suppression
flooding
Diffusion with suppression
Omniscient multicast

7. Adaptivity I: Robustness in Data Diffusion
A primary goal of data diffusion is robustness through
empirical adaptation: measuring and reacting to the
environment.
20% node failure
Because of this adaptation,
mean latency (shown here)
for data diffusion
degrades only mildly
even with
10%-20% node failure.
10% node failure
no failures

8. Adaptivity II: Adaptive Fidelity
extend system lifetime while maintaining accuracy
approach:
estimate node density needed for desired quality
automatically adapt to variations in current density due to uneven deployment or node failure
assumes dense initial deployment or additional node deployment
zzz
zzz
zzz
zzz

9. Adaptive Fidelity Status
applications:
maintain consistent latency or bandwidth in multihop communication
maintain consistent sensor vigilance
status:
probablistic neighborhood estimation for ad hoc routing
30-55% longer lifetime with 2-6sec higher initial delay
currently underway: location-aware neighborhood estimation

10. Small, Random Identifiers
Sensor nets have many uses for unique identifiers (packet fragmentation, reinforcement, compression codebooks...)
It’s critical to maximize usefulness of every bit transmitted; each reduces net lifetime (Pottie)
Low data rates + high dynamics = no space to amortize large (guaranteed unique) ids or claim/collide protocol
So: use small, random, ephemeral transaction ids?
Locality is key: random ids much smaller than guaranteed unique ids if total net size large and transaction density small
ID collisions lead to occasional losses; persistent losses avoided because the identifiers are constantly changing
Marginal cost of occasional losses is small compared to losses from dynamics, wireless conditions, collisions…

11. Address-Free Fragmentation
AFF Allows us to optimize # bits used for identifiers
Fewer bits = fewer wasted bits per data bit, but high collision rate; vs.
More bits = less waste due to ID collisions
but many bits wasted on headers
Data Size=16 bits

12. Exploit Non-Uniformities I: Tiered Architecture
Consider a memory hierarchy: registers, cache, main memory, swap space on disk
Due to locality, provides the illusion of a flat memory that has speed of registers but size & price of disk space
Similar goal in sensor nets: we want a spectrum of hardware within a network with the illusion of
CPU/memory, range, scaling properties of large nodes
Price, numbers, power consumption, proximity to physical phenomena of the smallest

13. Exploit Non-Uniformities I: Tiered Architecture
We are implementing a sensor net hierarchy: PC-104s, tags, motes, ephemeral one-shot sensors
Save energy by
Running the lower power and more numerous nodes at higher duty cycles than larger ones
Having low-power “pre-processors” activate higher power nodes or components (Sensoria approach)
Components within a node can be tiered too
Our “tags” are a stack of loosely coupled boards
Interrupts active high-energy assets only on demand

14. Exploit Non-Uniformities II: Time Synchronization
Time sync is critical at many layers; some affect energy use/system lifetime
TDMA guard bands
Data aggregation & caching
Localization
But time sync needs are non-uniform
Precision
Lifetime
Scope & Availability
Cost and form factor
No single method optimal on all axes

15. Exploit Non-Uniformities II: Time Synchronization
Use multiple modes
“Post-facto” synchronization pulse
NTP
GPS, WWVB
Relative time “chaining”
Combinations can (?) be necessary and sufficient, to minimize resource waste
Don’t spend energy to get better sync than app needs
Work in progress…

16. Conclusions Many promising building blocks exist, but
Long-lived often means highly vertically integrated and application-specific
Traditional layering often not possible
Challenge is creating reusable components common across systems
Create general-purpose tools for building networks, not general purpose networks