1 REED: Robust, Efficient Filtering and Event Detection in Sensor Networks Daniel Abadi, Samuel Madden, Wolfgang Lindner MIT United States VLDB 2005

2 What Problem Are We Trying To Solve? Complex data filtering in sensor networks

3 Example Filter Query TimestampTemp 3:05PM 74 MinTSMaxTSMinTempMaxTemp 2:00PM2:30PM7075 2:30PM3:00PM7378 3:00PM3:30PM7580 3:30PM4:00PM8388 4:00PM4:30PM8590 4:30PM5:00PM7075 5:00PM5:30PM7277 5:30PM6:00PM7580 Join Predicate: TS > MinTS && TS MaxTemp) X Sensor Data Predicate Table

4 Constraints: Sensor Networks Sensor nodes are small, battery-powered devices Power conservation is important –Sensing and transmitting data typically dominate power usage Berkeley Mote 4Mhz uProc 900Mhz Radio ( ft. range) 4 K RAM, 128 K Program Flash, 512 K Data Flash

5 Sensor Database Motivation Programming Apps is Hard –Limited power budget –Lossy, low bandwidth communication –Require long-lived, zero admin deployments –Distributed algorithms –Limited tools, debugging interfaces Solution: database style interface (e.g. TinyDB [Madden 2002], Cougar [Yao 2003])

6 TinyDB Root 0 Main PC Controller How TinyDB Works: 1.Form a routing tree 2.Distribute query to nodes 3.Every time node produces data tuple, filter by expression and pass result up tree, aggregating if necessary

7 Naïve Join Algorithm Send all tuples from data table to root; perform join at root Root 0 Main PC Controller B C D A X B X X Predicate Table

8 Ideal Join Algorithm Root 0 Main PC Controller A B C D A B C D A B C D A B C D A B C D A B C D A B C D X X BX X X Send join table to each node At node, perform join Problem: Severe Node Memory Constraints

9 X REED Algorithm 1 Cluster nodes into groups Store portion of predicate table in each group member Send sensor data tuples to every member of group Root X D 8 X A B C D A B C D A B C D X XXX

10 Group Formation 1 43 Neighbor list: {1, 2, 3, 4, 6} Broadcast: Want to make group Choose Me! {1, 3, 4, 6} Space: 4 CurrList: {1} Potential: {1, 2, 3, 4, 6} Space: 8 CurrList: {1, 4} Potential: {1, 3, 4, 6} Choose Me! {1, 3, 4} Space: 2 Space: 10 CurrList: {1, 3, 4} Potential: {1, 3, 4} Group Accepted: {1, 3, 4} 6 Neighbor list: {1, 3, 4} Neighbor list: {1, 3, 4, 6} Neighbor list: {1, 4, 6}

11 Table Distribution Group members figure out amongst themselves how the table will be divided across group Table flooded to network

12 Bloom Filter Optimization Temp: 20 Temp: hash Bloom Filter Step 1: Hash domain of sensor values onto Bloom Filter Step 2: Send Bloom Filter to Each Sensor Node Root Might produce false positives but never false negatives Can be used in conjunction with previous REED algorithm X X

13 Cache Diffusion Root Cache non-joining ranges on a per node basis Also will produce false positives but no false negatives 21

14 Results: Experimental Setup Ran experiments both in simulation and on real motes For simulation, 40 sensor nodes arranged in a grid Use TinyOS Packet Level Simulation Models CSMA backoff Carrier sense packet delivery model Overlap between 2 receptions leads to both being corrupted Use TinyOS MintRoute for MultiHop Routing Layer

15 REED Performs Well at Most Selectivities

16 REED Algorithm Overhead is Negligible

17 Simulated Results Match Real Results from Motes Ran REED algorithm on a simple 5 node sensor network

18 Conclusion Contributions: –Complex filters  table of expressions  join –REED algorithms capable of Running with limited amounts of RAM Robustness in the face of message loss and node failure –Experiments show benefits of doing complex join-based filters in the sensor network

19 Backup Slides Selectivity Number of Transmissions (1000s) Selectivity Number of Transmissions (1000s)

20 REED Performs Well even at low AVG node Depths

21 Cache Diffusion Takes Advantage of Data Locality

22 Distributed Join Group Formation Root Process: 1.Every node maintains list of nodes it can hear by listening in on packets 2.After a random interval, a node P which is not in a group broadcasts a form group request 3.Every node N which hears that request and is not currently in a group replies to P with a list of neighbors and amount free space 4.Node P collects the replies, and determines who should be in the group. For every node N which replied, P sends either a group reject or a group accept message. 5.Group accept message contains a list of nodes in the group A Group is a set of nodes where every node is in broadcast range of every other node. {1,2,3,4} {3,1,4} {4,1,3, 6} {1,2,5} {5,2,6,7} {6,5,7, 4} {7,5,6}

23 Distributed Join Join Table Distribution Root Process: 1.When a node enters a group, it sends a request to the root for join table data 2.Per group, the root gives out non- overlapping segments of the join table to every member 3.Once all the nodes in a group have received join tuples, they begin processing data tuples as a group Get me some tuples! (3) Get me some tuples! (2) Get me some tuples! (4)

24 Distributed Join Operation Root For nodes not in group: 1.When generating a data tuple or receiving data tuple from child, pass on to parent 2.When receiving a result from child, pass on to parent For nodes in group: 1.When generating a data tuple or receiving data tuple from child, broadcast to group (including self). 2.Upon receiving data tuple broadcast from group, join with stored subset of join table and pass result up to parent. 3.When receiving a result from child, pass on to parent. a a a a1

25 Related Work Gamma[8] and R* [15] systems both looked at ways to horizontally partitioning a table to perform a distributed join –Different optimization goals TinyDB [19,20,21] and Cougar [31] both present a range of distributed query processing techniques –No joins Bonfils and Bonnet [6] propose a scheme for join-operator placement within sensor networks –Look at joins of sensor data, not an external table

26 Motivating Applications Industrial Process control –Distributed sensors measure environmental variables –Want to know if exceptional condition is reached Failure and Outlier Detection –Look for de-correlated sensor readings Power scheduling –Minimize power consumption by distributing work across sensors

27 Results Experimental Setup Sensor Nodes in a 2 x 20 grid Use TinyOS Packet Level Simulation Models CMSA backoff Carrier sense packet delivery model Overlap between 2 receptions leads to both being corrupted Use TinyOS MintRoute for MultiHop Routing Layer root 5 feet