1. 1 Balancing Push and Pull for Efficient Information Discovery in Large-Scale Sensor Networks Xin Liu, Qingfeng Huang, Ying Zhang CS 6204 Adv Top. in Systems-Mob. Comp Presentation By Morgan Yeh

2. 2 Overview Introduction Related Work Balancing ‘Push and Pull’ Simulation Environment Research Results Conclusion

3. Introduction Many emerging sensor network applications involve the dissemination of observed information to interested clients Propose a “comb-needle” query support mechanism that integrates both push and pull data dissemination and analyze its performance in large-scale wireless sensor networks

4. Introduction The combing structure is dynamic Granularity adjusts dynamically based on query and event frequencies to minimize communication cost Combs are finer and needles shorter when the query frequency is relatively low compared to the event frequency, and vice versa

5. Related Work Reducing the number of potentially redundant forwardings in the flooding process using neighborhood topological information or via probabilistic retransmissions  Aims to improve flooding efficiency  Reducing the constant implicitly used in O(N), where N is the number of nodes in the network. In contrast, the comb-needle scheme achieves O(sqrt(N)) in the best case by balancing push and pull

6. Related Work Reducing the discovery/query cost by taking into account application semantics  Discovery and dissemination protocols  The scaling laws for structured and unstructured information queries are studied under storage and energy constraints

7. Related Work Reducing the cost of search via efficient distributed indexing schemes  Distributed indexing scheme called Semantic Routing Tree (SRT), supporting queries from a fixed node  Distributed index for multidimensional data (DIM), allowing queries to be issued from any node

8. Related Work Trajectory-based routing  Develop cross-shaped trajectories to disseminate service information in the network  Focus : one alternative attempt to develop cross-shaped trajectories

9. Balancing ‘Push and Pull’ Sensor nodes continuously gather information and report to one or more sink nodes Sensor network is considered as a distributed database, where information can be extracted only when needed  Communications of information that is not needed result in a waste of resources and should be minimized when possible

10. Balancing ‘Push and Pull’ Support both mobile and stationary query nodes Entry point can be anywhere in the network and occurs at any time Assume that the speed of the mobile node is much smaller than that of communication in the sense that disconnection does not happen during a query process Assume all nodes in the network have information on their own locations

11. Balancing ‘Push and Pull’ Assume that the network is ad hoc and uniform in the sense that all nodes are equivalent and the network does not have a built-in hierarchy Assume sensor nodes are stationary Use packet-hop as a metric to measure communication efficiency and, thus, as an indication of energy consumption Assume that the size of a query packet is the same as that of a data-duplication packet

12. Balancing ‘Push and Pull’

13. EXAMPLE : L = 5 & s = 3 n nodes, located at (i,j), where 0 < i,j < n Update, on the event, to (L – 1) of its vertical neighbors and thus builds a vertical needle of length L  Updates to nodes (i,j + 1), (i,j + 2) … (i,j + L/2) and (i,j – 1), (i,j – 2) … (i,j – L/2 +1) Query is sent vertically from (i,j) to (n,j) and to (0,j)  Query is fanned out horizontally from nodes (i,j), (i +- s,j), (i +- 2s,j) …, where s is the interspike spacing or combing degree

14. Balancing ‘Push and Pull’ A lifetime parameter, τ ( tau ), can be included in the query message to indicate its time-window of interest C L = L – 1, the communication cost for each query

15. Balancing ‘Push and Pull’ Average query reply cost for each relevant event is αn (alpha*n), where 0.5 <= α <= 1 is the distance factor reflecting the positions of the query node Query dissemination cost (cost to build one vertical query line and multiple horizontal query lines) :

16. Balancing ‘Push and Pull’ Average distance (averaging over all locations of events) to node (0,0) : Total expected reply cost :

17. Balancing ‘Push and Pull’ Total communication cost per query : L = s is required to guarantee that a query meets all relevant event :

18. Balancing ‘Push and Pull’ Minimum communication cost of comb- needle :

19. Balancing ‘Push and Pull’

20. For Reverse Comb, the per-query communication cost : The average distance is (s’ – 1)/2 :

21. Balancing ‘Push and Pull’ Adaptive Comb-Needle Strategy  A query node can estimate the value of f d based on the number of replies it obtains. The query node calculates s.  A data node uses its most recent information on s to synchronize the needle length L with the comb width s.

22. Balancing ‘Push and Pull’ Fixed-Node Query  Communication cost per unit time : Binary Query  Average query cost per query :

23. Balancing ‘Push and Pull’ Compare comb with fixed-node queryCompare the width of regular comb and sequential comb

24. Simulation Environment Transmission model :  P transmit : signal strength at the transmitter  P rec,ideal (d) : ideal received signal strength at distance d  α (alpha) and β (beta) are random variables with normal distributions to σ (sigma) of N(0, σ α ) & N(0,σ β )

25. Simulation Environment Routing protocol in pseudocode :

26. Research Results Two separate tests are performed  To discover what is the best spacing for the comb for a grid network with small random offset  To show the robustness of the protocol with a varying comb width w for a grid network with large random offset There exists a trade-off between the delivery ratio and the communication cost

27. Research Results

28. Analysis of the comb-needle structure in a regular grid network  The trajectory of query and event duplications should cross each other to guarantee event discovery  The structure should adapt based on query and event frequencies

29. Research Results f q = 0.1, f e = 1 f q = 0.1, f e = 0.1

30. Research Results Full pull versus optimal comb-needle :

31. Research Results

32. In a network with a random topology, build approximations of the combs and needles using a Constrained Geographical Flooding Use trajectory-based routing schemes to develop trajectories for query and data duplication Query reliability is an important issue and is inherent in all wireless sensor systems

33. Conclusion The comb-needle is not the only possible structure Cannot prove that comb-needle is an optimal structure Further work needed in discovering the optimal structure for information dissemination and discovery, in particular in random networks Semi-conclusion : the shape of the optimal structure may be determined by the particular network topology

34. Conclusion The comb-needle strategy :  a simple yet efficient data discovery scheme for supporting queries in large-scale sensor networks  used as a substrate to study the benefit of balancing push and pull in data gathering and dissemination in large-scale wireless networks  (including the reverse one) covers a spectrum of push and pull strategies, with the pure push-based and pure pull-based schemes in two extremes

35. Conclusion Data aggregation and compression can be integrated into the comb-needle strategy to reduce the communication cost Optimal structures for information dissemination and discovery, in particular in random networks, are unknown

36. Questions or comments?