1. The Tenet Architecture for Tiered Sensor Networks O. Gnawali, B. Greenstein, K-Y. Jang, A. Joki, J. Paek, M. Viera, D. Estrin, R. Govindan, E. Kohler USC, UCLA SenSys 2006

2. 2 The Tenet Two-Tier Architecture  Motes and Masters  Multi-node data fusion done on masters  Masters program motes using tasks

3. 3 Example Task  Notify application when temperature > 50F  A task contains an arbitrary number of tasklets linked together.

4. 4 Efficiency Costs  Opportunity cost of multi-mote data fusion  Motes can still fuse locally-generated data Sensor data have high temporal but low spatial redundancy  More data routed to the masters  A well-designed WSN will have a small diameter  Higher congestion  Application parameters can be tuned, e.g., only high- confidence pursuers report to masters

5. 5 Five Design Principles  Asymmetric Task Communication  Master send mote tasks, mote send master reply, mote cannot initiate tasks (no inter-mote communication)  Addressability  Masters can talk to each other, any master can talk to any mote, a mote can reply to its tasking master  Task Library  Each task is a subset of a mote’s generic functionality  Robustness  Resilience to extensive network failures  Manageability  Tools must offer useful insight into network failures

6. 6 Tenet Task and Task Library  Focus on simplicity rather than expressiveness  A task is composed of tasklets, which are parameterized services  Linear composition  Tasklets maximize flexibility while remaining simple  Each task has a unique ID, a list of tasklets, and their parameters  Task library composed at compile-time due to TinyOS

7. 7 Tasklets  Can be composed into a wide range of tasks

8. 8 Task Data Structure  Tasks are dynamically allocated  Active Containers hold task data  Cloned when a tasklet repeats Attributes are 3-tuples:

9. 9 The Mote Runtime  Task-aware queues used by services (e.g., wait)  Tenet scheduler operates at tasklet-level granularity  Allows multiple tasks to execute concurrently

10. 10 Three Task Operations  Installation  Receive a task with a new ID  Modification  Receive a task with an existing ID and a body  Deletion  Receive a task with an existing ID and no body  All active containers associated with a task are destroyed

11. 11 Example Tasks  Blink  CntToLedsAndRfm  Ping and MeasureHeap  SenseToRfm

12. 12 Data Fusion Example  Take 10 samples, timestamp it  classify as interesting if 3 or more samples > 45  calculates the deviation from the running mean  displays the sample on the LEDs  sends the statistic, timestamp, and sample if interesting

13. 13 Network Subsystem Requirements  Must support different applications on tiered networks  Routing must be robust and scalable  Master-to-mote  Mote-to-master  Small memory footprint  Tasks must be reliably disseminated from any master to all motes  Results must be delivered with end-to-end reliability

14. 14 Addressing and Routing  Every mote and master has a globally unique 16-bit address  Motes use TinyOS address  Masters use last 16-bits of IP address  Master-to-master: IP routing  Mote-to-master: tiered routing  First route to nearest master, then to destination master  Use standard WSN tree-routing protocol like MintRoute

15. 15 Tiered Task Dissemination  Reliably floods tasks to all motes  Partial network re-tasking achieved using a predicate tasklet  Implemented in a generic packet flooding protocol called TRD  Reliably floods packets to all nodes (both motes and masters)  Based on beaconing

16. 16 Reliable Transport  Transmits responses from motes to masters  Three types  Best effort  Reliable transactional  Stream transport for high data rate applications  All use hop-by-hop retransmissions  The reliable protocols use a simplified version of TCP

17. 17 Summary of Novel Networking Mech.

18. 18 Evalution: Concurrency  How many tasks can a tmote support at once?

19. 19 Execution Time  Most CPU-intensive tasklet, GatherStatistics, can process 1200 samples in 14.8ms  CPU-bound max sampling rate is 81,000 samples per second

20. 20 Application: PEG  PEG = Pursuit-Evasion Game  One or more pursuers collaborate to corral one or more evaders  Use WSN to help pursuers detect non-line-of-sight evaders  Native implementation uses a leader  Multiple nodes sense the evader, leader fuses the data  Stress tests Tenet (no mote-level fusion)  Tenet implementation adjusts the detection threshold

21. 21 PEG Experimental Setup  56 tmotes, 6 stargates  Simplifications  Evader detected using RSSI  Radio transmit power limited to achieve multihop 9-hop diameter  One evader, one stationary pursuer on central master

22. 22 PEG Evaluation  Tenet has higher accuracy but higher latency  Tenet has lower message overhead

23. 23 Vibration Monitoring Case Study  Tenet used to implement Wisden DetectOnSet reduces network traffic Tenet simplifies programming

24. 24 Manageability  The following task can be used to capture the routing trees:  This can be used to evaluate the task dissemination latency:

25. 25 Robustness  Failure of a master forces routing algorithm to adjust

26. 26 Future Work  Near term  Actuation  Mote-tier storage  Bounded-latency communication  Long term  Impact of disconnection due to mobility  Authenticity  Data Integrity  Multi-user control and resource management

27. 27 Conclusion  Tenet simplifies programming while not significantly increasing overhead

28. 28 Application  Pursuit-Evasion  Pursuer mobile robots chase after evader robots with the help of a sensor network  Traditional implementation employs mote-tier data aggregation to reduce redundant evader reports

29. 29 Tenet Components  Task library  Composable tasklets  Reliable task dissemination protocol  Data transport mechanism (3)  Inter-tier routing subsystem

30. 30 Motivation  Most sensor network architectures support multi-node in-network data fusion  Increased complexity  Reduced manageability

31. 31 Tenet Design  Tenet limits multi-node data fusion to the masters  Can perform more sophisticated fusion algorithms  Larger context  better decisions