1. DATA AGGREGATION Siddhartha Sarkar Roll no: 13000111128 CSE-4 th Year-7 th semester Sensor Networks (CS 704D) Assignment

2. OVERVIEW  Introduction  Basic Idea  Why Data Aggregation in WSN?  What is Aggregation?  System Model  Data Aggregation Process  Tiny Aggregation  System Diffusion  Energy consumption  Aggregation Queries  Secure Aggregation  Conclusion

3. WSN nodes perform sensing of a physical environment. The sensed data from multiple sources is collectively used to make inferences.  Large amount of raw data  Correlated data Communication in the network is significantly reduced by:  Elimination of redundant data  Accumulation and processing at intermediate nodes INTRODUCTIONINTRODUCTION

4. To exploit the data redundancy Packets from different nodes, are combined in – network. Implementation Who carries the data with redundancy Data-centric routing Differences Based on contents of the packets. Address-centric routing Routing based on an end-to-end manner.

5. What is Aggregation?

6.  Communication is the most energy consuming functionality  Energy consumed in transmitting one bit over 100m 1000× Energy consumed per instruction execution  Efficient bandwidth utilization Why Data Aggregation in WSN? – I

7. Cost of computing an aggregate such as AVERAGE on a binary tree: Message count per query response without aggregation in a binary tree of depth d: S = 2(1 20 + 2 21 + 3 22 + : : : + d 2d-1) S = 2(1 20 + 2 21 + 3 22 + : : : + d 2d-1) = (d - 1)2d+1 + 2 N lgN; = (d - 1)2d+1 + 2 N lgN; N is the number of nodes. In general, S grows as N logb N, b is the branching factor. Why Data Aggregation in WSN? - II

8. Message count if aggregation is used : N Message count if aggregation is not used : Nlogb N Why Data Aggregation in WSN? - II

9. System Model Tree with N nodes and a sink. Tree with N nodes and a sink. Time-slotted and synchronized network. Time-slotted and synchronized network. Aggregated event data needs to reach sink within a deadline. Aggregated event data needs to reach sink within a deadline. Arbitrary set of source nodes. Arbitrary set of source nodes. Sink requires aggregated form of data: Sink requires aggregated form of data: Symmetric Functions – f(x, y) = f(y, x). Symmetric Functions – f(x, y) = f(y, x). Function value does not depend on sensor Function value does not depend on sensor identity. identity. Aggregation functions supported – MIN, MAX, Aggregation functions supported – MIN, MAX, Sum, Mean, Variance, Higher order statistics etc. Sum, Mean, Variance, Higher order statistics etc.

10. There are several aggregation techniques followed in Wireless Sensor Network, Such as  Tree based Aggregation  In network Aggregation

11. Data Aggregation Process Sensor nodes are organized into a tree hierarchy rooted at the Base Station Non-leaf nodes act as the aggregators

12. Tiny Aggregation Distribution phase Aggregate queries are pushed down into the network Collection phase Aggregate values are continuously routed up from children to parents

13. Routing: via a tree rooted at the sink. Routing tree formation Routing: via a tree rooted at the sink. Routing tree formation algorithm (invoked periodically): algorithm (invoked periodically): Do upon receiving message M(n; l) from node n at Do upon receiving message M(n; l) from node n at level l level l if this node's level > l + 1 this node's level = l + 1 this node's level = l + 1 this node's parent = n broadcast M(this node's id, l + 1) broadcast M(this node's id, l + 1) Detection of a leaf node n0: n0 does not hear any message of the Detection of a leaf node n0: n0 does not hear any message of the form M(n’,.) form M(n’,.) Tiny Aggregation (TAG)

14. Query Model:  Single table called sensors  Append-only relational tabl e  One attribute per sensing functionality Form of queries: SELECT {agg(expr), attrs} FROM sensors WHERE {selPreds} WHERE {selPreds} GROUP BY {attrs} GROUP BY {attrs} HAVING {havingPreds} HAVING {havingPreds} EPOCH DURATION i EPOCH DURATION iExample: SELECT {MAX(temperature),building} FROM sensors WHERE block = ALL WHERE block = ALL GROUP BY building GROUP BY building HAVING MAX(temperature) > 100 HAVING MAX(temperature) > 100 EPOCH DURATION 60s EPOCH DURATION 60s Tiny Aggregation (TAG)-II

15. The aggregation clause:  An initializer i  A merging function f  An evaluator e The aggregated record = f(, ), where and are partial records. Example: AVERAGE A partial record is the tuple A partial record is the tuple i SUM) = i SUM) = f (, ) =, ) = <SUM1+SUM2,COUNT1+COUNT2> e(SUM, COUNT) = SUM/COUNT. Tiny Aggregation (TAG)-III

16. A classification of aggregates A classification of aggregates : Tiny Aggregation (TAG)-III

17. Synopsis Diffusion Motivation: If trees are used for data aggregation, such as in TAG, a link failure leads to loss of data from an entire sub-tree. However, if aggregation of duplicate sensitive aggregates, such as COUNT, is done on a graph, one needs to solve the problem of making the process insensitive to duplicate messages.

18.  Synopsis is a digest of data  Partial aggregates are represented by synopses  Order and duplicate insensitive (ODI) synopses The aggregation process: Let i denote sensor data, s denote synopsis and a denote the desired aggregate.  Synopsis generation function SG : i s  Synopsis fusion function SF : (s1; s2) s  Synopsis evaluation function SE : s a Synopsis Diffusion

19. Finding ODI synopses and their fusion functions is the main difficulty. We look at two cases, one trivial and the other not so trivial! Example 1: ODI synopsis for MAX. Let X be the variable. Synopsis : X (the number itself) SG() = Xi SF(Xi; Xj) = The larger of Xi; Xj SE(Xi) = Xi Synopsis Diffusion-II

20. ODI Correctness Test: A synopsis diffusion algorithm is ODI-correct if SF and SG are order and duplicate-insensitive functions. Synopsis Diffusion-III

21. ODI Correctness Test A synopsis diffusion algorithm is ODI-correct if SF and SG are order and duplicate-insensitive functions. Define a projection operator q : Multiset of sensor readings ordered set of values.  SG preserves duplicates: 8r1; r2 2 R : q(fr1g) = q(fr2g) ) SG(r1) = SG(r2). The same 8r1; r2 2 R : q(fr1g) = q(fr2g) ) SG(r1) = SG(r2). The same synopsis is generated for all duplicates. synopsis is generated for all duplicates.  SF is commutative: 8s1; s2 2 S : SF(s1; s2) = SF(s2; s1).  SF is associative: 8s1; s2; s3 2 S : SF(s1; SF(s2; s3)) = SF(SF(s1; s2); s3). 8s1; s2; s3 2 S : SF(s1; SF(s2; s3)) = SF(SF(s1; s2); s3).  SF is idempotent: 8s 2 S : SF(s; s) = s. Synopsis Diffusion-III

22. Energy Consumption

23. Declarative Queries for Sensor Networks Examples: SELECT Nodeid, light FROM sensors WHERE light > 400 EPOCH DURATION 1s EpochNodeidLightTempAccelSound01455xxx 02389xxx 11422xxx 12405xxx Sensors Time is partitioned into epochs of duration i A single aggregate value is produced to combine the readings of all Time is partitioned into epochs of duration i A single aggregate value is produced to combine the readings of all devices during the epoch devices during the epoch 1

24. Aggregation Queries SELECT roomNo, AVG(sound) FROM sensors GROUP BY roomNo HAVING AVG(sound) > 200 EPOCH DURATION 10s Rooms w/ sound > 200 3 2 SELECT AVG(sound) FROM sensors EPOCH DURATION 10s EpochAVG(sound)0440 1445 EpochroomNoAVG(sound)01360 02520 11370 12520

25. Topology Maintenance and Recovery How to address the unreliable nature of WSNs in TAG? Each node maintains a fixed size of its neighbors – Select a better parent node If a node does not hear from its parent for some time, it assumes that its parent has failed

26. Secure Aggregation

27. It is challenging to design suitable security mechanisms It is challenging to design suitable security mechanisms for Wireless Sensor Networks (WSNs) for Wireless Sensor Networks (WSNs) ― Stringent resource constraints on energy, processing power, memory, bandwidth, etc. WSNs need lightweight secure mechanisms WSNs need lightweight secure mechanisms We introduce an LCG-based secure aggregation scheme We introduce an LCG-based secure aggregation scheme ― Efficiency and simplicity Secure Aggregation

28. Wireless Controller Area Network (CAN) Based Prioritized MAC using Bit-wise arbitration o Used in Bosch’s CAN 2.0 o Extended to wireless channel 0110 1010 1100  The highest priority packet gets transmitted first  Compute MAX by using data as the priority  Compute MIN by using complement of data as the priority  Excellent time-complexity for MAX/ MIN in single broadcast domains domains  Limitations for other aggregates

29. Open Problems Joint aggregation and scheduling problem  Joint aggregation and scheduling problem - Spatio-temporal optimization - Spatio-temporal optimization  Multi-query optimization  Correlated source coding, compressed sensing

30. Conclusion Data-aggregation leads to bandwidth and energy  Data-aggregation leads to bandwidth and energy efficiency efficiency  Pure flooding is wasteful  Multicast tree with large node degrees is not optimal