1. Energy-Conserving Data Placement and Asynchronous Multicast in Wireless Sensor Networks Sagnik Bhattacharya, Hyung Kim, Shashi Prabh, Tarek Abdelzaher Department of Computer Science University of Virginia ACM Mobisys’03 speaker ： Jenchi

2. Outline Introduction Related work Service Model Data placement Evaluation Conclusion

3. Introduction The primary function of sensor networks is the collection and delivery of sensory data Power is one of the most expensive resources In this paper develop a distributed framework that improves power conservation by application-layer sensor data caching and asynchronous update multicast The goal of the framework is to reduce the total power expended on the primary network function Communication is a prime candidate for optimization

4. Related Work The approach differs from traditional multicast routing Updates are propagated asynchronously in a lazy manner in accordance with consistency constraints The depth of the tree is determined by the update and the request rates, and it adapts itself to minimize the communication The work in an overlay multicast algorithm that works on top of the network layer, rather than traditional multicast routing that takes place at the network layer

5. Related Work (cont.) Data placement is also similar to some of the ideas used in the placement of web server replicas Data placement furthers this idea by using the property of location-awareness of the sensor nodes

6. Service Model A dense ad hoc wireless sensor network with multiple observers, spread over a large monitored area The observers’ attention is directed to a relatively limited number of key locales in the network Focus locales ： important events or activities are taking place Sensor nodes at each focus locale elect a local representative for communication with the rest of the world

7. Service Model (cont.) Our service adopts a publish-scribe model Each representative publishes sensory data about its focus locale to observers who subscribe to a corresponding multicast group to receive such data Update traffic is multicast from focus locales to receivers in an asynchronous manner Data caches are created at the nodes of the multicast tree Different observers may specify different period requirements for the same measurement

8. Service Model (cont.) Our middleware achieves four main functions It determines the number of data caches for each focus locale It chooses the best location for each cache such that communication energy is minimized It maintains each cache consistent with its data source at the corresponding focus locale It feeds data to observers from the most suitable cache instead of the original sources

9. Service Model (cont.) Key differences between this problem and the problem of caching in an internet context Internet The topology restricts the choice of cache locations Sensor network Is dense enough such that a data cache can be placed at any arbitrary physical location Internet The number of Internet proxy caches is typically much smaller than the number of different web sites Sensor network The middleware caching service runs on every sensor node The number of sensor nodes is larger than the number of focus locales

10. Problem formulation Focus locale (X, Y) sensor updates at (X, Y) occur at an average rate R update BS={BS 1, BS 2,…, BS M } is a set of M observers that request data from that locale with rates R req ={R 1,R 2,…,R M } Sensor ： (X,Y) BS 1 BS 2 BS 3

11. Problem formulation (cont.) The cost of message transfer between two nodes in the tree the power expended on a packet’s transfer on the shortest route multiplied by the packet rate 1 2 3 The center of gravity of the N input points

12. Problem formulation (cont.) The problem is that of constructing a minimum-cost weighted Steiner tree, which connects the sensor node to the observers

13. Data placement Is a distributed physical systems Each step of the algorithm reduces a measure of total energy until a minimum energy tree is found Use a distributed greedy heuristic that iteratively places each node at the center of gravity of its neighbors

14. Data placement (cont.) The algorithm each node on the multicast tree rooted at the sensor maintains a location pointer to its parent as well as a location pointer to each of its children Each child node maintains the maximum propagation rate, which is the maximum of all requested update rates of all observers served by that child Flurries of environmental updates that exceed some receivers’ requested rates are not propagated unnecessarily to those receivers

15. Data placement — Joining the Multicast Tree k New Node (observer) Join Request 1.The location of the observer 2.Its desired update rate R k

16. Data placement — Joining the Multicast Tree k New Node (observer) Join Request

17. Data placement — Joining the Multicast Tree k New Node (observer) Join Request

18. Data placement — Joining the Multicast Tree k New Node (observer) Join Request

19. Data placement — Joining the Multicast Tree k New Node (observer) Join Request

20. Data placement — Joining the Multicast Tree k New Node (observer) Join Request No children that are closer to the observer Nearest neighbor New link

21. Data placement — Copy Creation and Migration N k New Node (observer) Nearest neighbor Computes the center of gravity of itself and all its neighbors Node N computes the savings, if any, resulting from creating a new copy at that center of gravity If the savings from creating the copy exceed a threshold, the option of creating this copy is deemed viable

22. Data placement — Copy Creation and Migration Nearest neighbor creates downstream copy If N is the origin sensor k N Nearest Neighbor (Origin sensor) Computes the center of gravity of itself and all its neighbors Prospective Copy New Node

23. Data placement — Copy Creation and Migration Nearest neighbor creates downstream copy If N is the origin sensor k N Nearest Neighbor (Origin sensor) Computes the center of gravity of itself and all its neighbors Prospective Copy New Node

24. Data placement — Copy Creation and Migration Nearest neighbor creates upstream copy If N is not the origin sensor k N Nearest Neighbor New Node Computes the center of gravity of itself and all its neighbors Prospective Copy

25. Data placement — Copy Creation and Migration Nearest neighbor moves If N is not a fixed copy k N Nearest Neighbor New Node Computes the center of gravity of all its neighbors Prospective Move

26. Data placement — Copy Creation and Migration Nearest neighbor moves If N is not a fixed copy k N Nearest Neighbor New Node Prospective Move

27. Data placement — Copy Creation and Migration Nearest neighbor moves If N is not a fixed copy k Nearest Neighbor New Node Prospective Move N

28. Data placement — Copy Creation and Migration At most one copy is created for every newly joined member The algorithm creates at most m-2 copies where m is the total number of observers

29. Data placement — Leaving the Multicast Tree Observer K sends a leave() message to its parent N Node N stops forwarding messages to the departed observer N resets the maximum forwarding rate If N is a migratory mode, it computes the center of gravity of all remaining neighbors, and moves there if the savings exceed a threshold If there is only one child left for the migratory node, the node is deleted and its parent takes over its child

30. Data placement (cont.) Sampling R update To take the inverse of the average of the last five inter-arrival times

31. Evaluation Use Berkeley motes as the underlying distributed platform Each node has up to three sensors Runs on an 8-bit 4MHz micro-controller and has 128kb of program memory and 4kb of data memory NS-2 simulator network ： 200m×200m nodes ： 200 ≦ N ≦ 500 each node have a radio range of 20m Packet sizes ： 30 bytes Base station ： roughly 5% the number of nodes in the network Each node knows it own location Focus locale ： is generated at random The request rates are generated at random with a specific average throughout the experiment Energy consumption is measured in terms of Joules per node per flow Transmitting a single bit consumes 1 μ J and receiving consumes 0.5 μJ Use Geographic forwarding as a routing algorithm

32. Evaluation — Simulation Results To compare the performance of the data placement middleware against four baselines A simple unicast-based query-response model Update multicast (synchronous push model) Directed diffusion Update flooding

33. Evaluation — Simulation Results Comparing the energy consumption of the four baselines for different node densities Regular multicast is better than data placement! Because the overhead of data placement is offset by considerable savings when the average update rate increases beyond the request rate

34. Evaluation — Simulation Results The average energy consumption in the steady state after all observers have joined the tree Data placement is better! Because it does not send unnecessary updates

35. Evaluation — Simulation Results To measure energy consumed when a new observer joins the tree

36. Evaluation — Simulation Results Lifetime of nodes in a sensor network using data placement

37. Conclusion Data placement reduces energy consumption and increase the lifetime of a sensor network The algorithm places copies of the requested data and updates them so as to minimize the communication overhead and power consumption of data transfer The algorithm is completely distributed and requires very little local processing Data placement is a new approach for energy conservation in wireless sensor networks