1. Matching Data Dissemination Algorithms to Application Requirements John Heidermann, Fabio Silva, Deborah Estrin Presented by Cuong Le (CPSC538A)

2. Outline Introduction Summary of Diffusion Algorithms Applications Performance with Different Diffusion Algorithms Systematic Evaluation Conclusions

3. Introduction Data dissemination approaches in sensor networks have adopted application-specific, data-centric communications protocols: –to reduce overhead by avoiding levels of abstraction. –to support application involvement in communication. Application-specific constraints and optimizations greatly reduce communications cost by replacing communication with computation in the network As number of protocols and sophistication of applications grows, choice of communication algorithms becomes a problem

4. Application-specific requirements Sensor network applications have different needs: –Different traffic patterns (many-to-one, many-to-many, one-to-many, one-to-one) –Different data rates (fixed and variable, frequent and infrequent) Applications must be robust to change: –Wireless links come and go –Nodes fail or move How can communication be robust but also efficient for many different applications ?

5. Multiple Diffusion Routing Algorithms To address application-specific requirements by matching routing algorithms to application requirements.

6. Diffusion Routing Algorithms Overview localized algorithms, named data, and support for in-network processing. adopt a declarative, publish/subscribe API. key API abstractions: –data is identified by a set of attributes. –data producers (or sources) generate data it by publishing. –data consumers (or sinks) subscribe to data, it is the business of the diffusion implementation to insure that data travels from publisher to subscriber efficiently.

7. Multiple Diffusion Routing Algorithms Two-Phase Pull Diffusion One Phase Pull Diffusion Push Diffusion GEAR

8. Two-phase pull diffusion A subscriber, or data sink, identifies data by a set of attributes propagating through the network in an interest message. Nodes establish gradients, states indicating the next-hop direction of other nodes. The first data message sent from the source is marked as exploratory and is sent to all neighbors that have matching gradients. When exploratory data reaches the sink, the sink reinforces its preferred neighbor, establishing a reinforced gradient towards the sink. Periodically floods data sink’s interests and exploratory data

9. GEAR Adds support for geographically scoped queries If nodes know their locations, then geographic queries can influence data dissemination Replaces network wide communication with geographically constrained communication

10. Push Diffusion Reverses the roles in the publish/subscribe API Floods only exploratory data messages

11. One-phase pull diffusion Subscriber based system that avoids one of the two phases of flooding in two-phase pull Only floods interests No exploratory messages

12. Applications Performance with Different Diffusion Algorithms

13. Push vs. two-phase pull Push reduces message count by ~60% compared to two phase pull

14. Push with vs. Push without GEAR GEAR reduces message count by ~40%

15. Systematic Evaluation

16. Evaluation Methodology Identify test application classes from experience –BAE tracking many-to-many -> benefit from push –PARC IDSQ One-to-many, one-to-one -> benefit from GEAR and push –James Reserve Data Collection Many-to-one->benefits from one-phase-pull Describe performance differences for applications designers –Use systematic emulation and simulation studies to explore design space: Use different diffusion algorithms. Varying number of sources and sinks. Varying topologies (clustered vs. un-clustered)

17. Algorithm performance with different numbers of source and sinks

18. Varying the number of sinks with one phase pull

19. Using geographic information

20. Results Summary Push works best with many sinks and few active sources One-phase pull works best with many sources and few sinks The break-even point between these algorithms depends upon control message frequency and data rates In networks with more than a few dozen nodes, the benefits of geographically scope queries outweigh other algorithmic choices Sending at very low data rates (lower than the control traffic interval) has a relatively high control overhead, unless Geographic scope or control rates are changed