1. Energy-Efficient Monitoring of Extreme Values in Sensor Networks Loo, Kin Kong 10 May, 2007

2. Contents Motivation Settings and assumptions Algorithms Experiment results

3. About this work A SIGMOD 2006 paper by Silberstein, Munagala and Yang

4. Motivation Development in sensor technologies imposes new challenges to old problems The MAX (or MIN) query is important in monitoring extreme conditions in a sensor network

5. Sensor network Consider a network of sensors, where each node is battery-operated contains a sensor has some computing power communicates with each other wirelessly reports its value to the “ root ” directly or indirectly may have a limited range of communication R u1u1 u2u2 u3u3 u4u4 u5u5 u6u6

6. Sensor network (cont ’ d) The root may keep the state reported by nodes can issue commands (e.g., init, query) to nodes performs computation The goal is to implement the MAX query, that returns the exact maximum value (v max ) and the node (u max ) that reports the value, and minimize the energy consumed by the network R u1u1 u2u2 u3u3 u4u4 u5u5 u6u6

7. Considerations Energy consumption of a node is dominated by communication between the node and other nodes and the root Hence, one should reduce the amount of communication to save energy R u1u1 u2u2 u3u3 u4u4 u5u5 u6u6

8. Message exchange Communication within the network is made in the form of “ packets ” TypeDescription Boot Initial application and query Trigger Node sending its own value Query Root initiating fetching of values Reply Response to Query ThresholdUpdate Update to node constraints MaxDesignate Designate node as current max MaxOff Notify node that it is no longer max

9. Assumptions The MAX query is processed repeatedly over a series of “ rounds ” Each node generates a value in each round Each round is long enough for all necessary messaging to occur to complete the query In case any node (say, u i ) communicates with the root indirectly, u i must be registered with some “ parent ” node u j at any time i.e., the routing tree is more or less static unless a parent node becomes unavailable

10. Topology-oblivious algorithms Assumes that all nodes directly communicate with the root; or, intermediate nodes only work as a router Three algorithms Temporal Suppression (TS) Range Caching (RC) [1] SLAT [1] Olston et al. Adaptive precision setting for cached approximate values. In SIGMOD 2001, CA, USA, 2001.

11. Temporal Suppression (TS) It is not necessary that a node sends its value in every round; do so only when there is a change Rootu1u1 boot reply time trigger change a round

12. Range Caching (RC) TS may not save a lot of message traffic Consider a case when v max = 100 If, for a node u i, v i changes from 10 to 20. It is still unlikely that u i becomes u max but TS reports the change. If the value of a node is far from being the maximum, the node sends its value only when a major change is observed

13. Range Caching (RC) (Cont ’ d) Each node u i maintains a range [lb i, ub i ] so that v i =(lb i +ub i )/2 A node does not report its value if the value remains within the range The node u max has the range [v max, v max ] The root keeps track of the range of every node Remark: the paper does not explicitly mention how to set the initial range Let ’ s assume that to be [v i -r, v i +r] for u i for some fixed r The range of each node will expand or contract as the process rolls out

14. RC – initial round The root sends a boot to every node Each node replies its value and its range The root computes v max and informs the corresponding node v 1 =v max range of v 2 range of v 3

15. RC – subsequent rounds If, for some u i, v i is out of its range [lb i, ub i ], u i sends its value v i to root If v max <ub x for some non-reporting node u x, the root fetches the value of u x Range setting: If u i spontaneously reported its value, expand its range by a factor  > 1 If u i reported by request of the root, contract the range by the factor  with a 50% probability v 1 =v max range of v 2 range of v 3

16. SLAT Single Level Adaptive Threshold Each node u i keeps a threshold  i, which is also known to the root Within a particular round, u i,  i  v max Remark: in rounds except the initial one, a node u i reports its value v i only if v i >  i Initially: each node u i sends its value v i to the root each node u i sets  i = v i

17. SLAT – subsequent rounds Stage I – node initiated reporting A node u i reports it value v i if the node is u max and its value has changed the node is not u max and v i >  i Stage II – root initiated querying Find v* as the maximum of all reported values (and v max if u max did not report) If, for some unreported node u x, v* v* Remark: this is needed only if the value of u max has decreased

18. SLAT – subsequent rounds (cont ’ d) Stage III – threshold setting if u i reports in Stage I, set  i = the new v max, which will be sent by the root when it is find if u i is queried in Stage II and u i reports its value, set  i =v i if u i is queried in Stage II and u i does not report, set  i =v i and inform the root of the new  i

19. SLAT If the maximum value drops, the root may query individual nodes R u1u1 u2u2 u3u3 u4u4 u5u5 u6u6 v=40 v=20 v=45 v=30 v=25 v=32 v=38 Q Q R R

20. Topology-aware algorithms Assumes that some nodes communicate with the root indirectly Two algorithms SLAT-A HAT

21. SLAT-A Single Level Adaptive Threshold with Aggregation Observation: if a child node reports a value which is smaller than that of the parent node, the value of the child node cannot be MAX and so can be dropped However, if packets are dropped, the root cannot tell whether a reported value is that of a parent or a child R u1u1 u2u2 u3u3 u4u4 u5u5 u6u6 v=20 v=40 v=45 v=30 v=25 v=32 v max =45

22. SLAT-A The root does not keep thresholds of individual nodes If the maximum value drops, the root may need to query a whole sub-tree R u1u1 u2u2 u3u3 u4u4 u5u5 u6u6 v=20 v=40 v=45 v=30 v=25 v=32 v=38

23. HAT Hierarchical Adaptive Thresholds In SLAT-A, it is possible that a whole sub-tree is queried when the maximum drops Each parent node keeps the thresholds of its child nodes If node u i is a child of u j,  i   j

24. HAT (cont ’ d) If the maximum value drops, the root queries u 4 u 4 knows u 6 does not exceed 20, so the query is not propagated to u 6 R u1u1 u2u2 u3u3 u4u4 u5u5 u6u6 v=20 v=40 v=45 v=30 v=25 v=32 v=38  6 =20

25. Policies for setting HAT thresholds The threshold of a node u i can be set anywhere between max( child, v i ) and  parent between The ultimate goal of threshold setting is to maximize the time between successive threshold settings

26. Experiments Network simulator settings: 200 nodes in a 400m x 400m area radio range is 50m Each test is run for 21 rounds, and the result of the first round is discarded

27. Experiment results Random behaviour All nodes change with some probability by random amount

28. Experiment results Randomly rising

29. Experiment results Uniformly falling All nodes fall in every round by a percentage

30. Experiment results Highest values dropping

31. Conclusion The paper explores existing techniques on finding max and adapted them to the environment of sensor networks The paper introduces novel methods SLAT, SLAT-A and HAT Experiment results show that HAT is the most energy-efficient among the algorithms compared

32. Q & A

33. Thank you