1

Pradeep Kumar Gunda(Thanks to Jigar Doshi and Shivnath Babu for some slides)

TAG: a Tiny Aggregation Service for Ad-Hoc Sensor

NetworksSamuel Madden, Michael J Franklin, Joseph M Hellerstein, Wei Hong

2

TAG - Motivation

Sensor Networks used for monitoring in various fieldsCivil engineers to monitor buildings during earthquakes

Biologists for habitat monitoring

People prefer summary reports not individual values

Aggregation common to all these applications!

Must be a core service and easy to use.

TAG fills this void

3

Before TAG

Centralized approachTransfer everything to base station

No suppression – high energy usage, traffic

Directed DiffusionViewed aggregation as a application-specific operation

Aggregation API in routing layer

No declarative query language like TAG

Not for any generic aggregation operators

4

What is TAG

Tiny Aggregation for Sensor Networks

SQL – like interface eg. Min, Max, Count

Sensitive to constraints of ad-hoc sensor networks

Query inserted into network over an existing routing protocol

Aggregation done along the reverse path

Combines the research in networking community with database community

5

DBMS in a nutshell

Select max(wins), team fromBasketballwins

Where year=‘2002’

Group by tournament

Team Year Wins Tournament

Duke 2002 25 ACC

Duke 2003 35 NCAA

UNC 2002 20 ACC

UNC 2003 25 NCAA

6

DataBase Management System

High-levelHigh-levelQuery QQuery Q

DBMS

Data

Answer

Translates Q intobest execution plan

for current conditions,runs plan

Keeps data safe and correct

despite failures, concurrent

updates, online processing, etc.

7

Data Streams

User/ApplicationUser/Application

Register Register Continuous QueryContinuous Query(Standing Query)(Standing Query)

Stream QueryProcessorInput streams

ResultResult

8

What can DBMS offer for Sensor Networks??

Express aggregation as SQLSpecify what you want. Not how to getUsers need not write low level programming language code!!!

Less bugsDon’t worry about optimization

Techniques from parallel/distributed dbSensor network is a stream of sensor readings to base station

9

Query ModelOne Table : sensorsSELECT AVG(volume), room FROM sensorsWHERE floor = 6GROUP BY roomHAVING AVG(VOLUME) > thresholdEPOCH DURATION 30s

In generalSELECT {agg(expr), attrs} FROM sensorsWHERE {selPreds}GROUP BY {attrs}HAVING{havingPreds}EPOCH DURATION IDifference between TAG & SQL : Continuous Output

10

Aggregate Structure

Standard SQL supports “the basic 5”:MIN, MAX, SUM, AVERAGE, and COUNT

TAG supports any function conforming toInitializer i: Instantiates a record for a single sensor valueMerging function f. Merges two partial state recordsEvaluator e: Computes the actual value of the aggregate from a partial state record

Example - averagei{v} <v,1>f{<S1, C1>, <S2, C2>} < S1 + S2 , C1 + C2>e{<S1, C1>} S1/C1

TAG supports MEDIAN, HISTOGRAM and COUNT DISTINCT also

11

Classifying AggregatesDuplicate Sensitive (yes/no)Exemplary/SummaryMonotonics’ = f(a,b) e(s’) >= MAX(e(s1),e(s2)) OR e(s’) <= MIN(e(s1),e(s2))

Decides whether predicate can be applied in network

Partial StateDistributive (partial state’s size same as final aggregate)Algebraic (partial states are not themselves aggregate)Holistic (No useful partial aggregation)UniqueContent Sensitive

12

Aggregate Taxonomy

Distributive Algebraic Holistic Unique Content-sensitive

Partial statesize

Same asfinalaggregate

Constant Dataproportional

Distinctvalueproportional

Proportionalto somedataproperty

13

Requirements of the Routing Algorithm

Deliver Query requests to all nodes

Route from every node to root

No Duplicates ! (Affects some aggregates like count, avg)

Does it violate the end-to-end principle?

A simple example proposed : tree based routing

14

Tree Based Routing

One rootAny interior node sets sender as parent and sets its level to that of parent + 1RebroadcastsMessage sent by node to its parent eventually reaches rootReselect parent after k silent epochs

Query

P:0, L:1

2

1

5

3

4

6

P:1, L:2

P:1, L:2

P:3, L:3

P:2, L:3

P:4, L:4

15

The TAG Algorithm

2 PhasesDistribution: Queries are pushed down the network.

Parents broadcast queries to their children

Collection: Aggregate values continuously sent from children to parents

Reply from all children required before forwarding an aggregate value

TDMA like partitioningChildren must deliver records during a parent-specified time interval

Parent collects all values (including its own) and sends the aggregate up the tree

16

Flow of partial State

Parent reception interval must be chosen carefullyAll children must be able to reportCannot exceed end of epoch

However we can always make the algorithm pipelined

17

Pipelined Aggregation

122132

111111

123143

54321

1

2 3

4

5

1

2

31

4

Sensor #

Ep

och

#

Epoch 3SELECT COUNT(*) FROM sensors

18

Pipelined Aggregation

123143

122132

111111

123154

54321

1

2 3

4

5

1

2

31

5

Sensor #

Ep

och

#

Epoch 4SELECT COUNT(*) FROM sensors

19

Pipelined Aggregation

123154

123143

122132

111111

123155

54321

1

2 3

4

5

1

2

31

5

Sensor #

Ep

och

#

Epoch 5SELECT COUNT(*) FROM sensors

20

Grouping

Simple aggregation mechanismComplicated by HAVING clauseGroup eviction to solve storage problemEvicted tenant sent to parent

21

Simulation Environment

Java-based simulation & visualization for validating algorithms, collecting data.

Sensors arranged on a grid, radio connectivity by Euclidian distance

Communication modelLossless: All neighbors hear all messages

Symmetric linksNo collisions, hidden terminals, etc.

RealisticNumber of hops related to distance but not proportional

22

Simulation Results

2500 nodes, d=50TAG outperforms centralized approach by an order of magnitude in most cases

Does equally well in the worst caseActual benefit depends on the topology

23

Optimizations

SnoopingOverhear packets – can initiate aggregation if missed

Can also be used for suppression!

Hypothesis testing Guess the value of aggregate & suppress

Send only if current val > MAX

Can be applied to a variety of aggregates such as MAX

24

Experiment: Hypothesis Testing

Uniform Value Distribution, MAX Query

Messages/ Epoch vs. Network Diameter

0

500

1000

1500

2000

2500

3000

10 20 30 40 50

Network Diameter

Messages / Epoch

No GuessGuess = 50

Guess = 90Snooping

25

TAG Loss Tolerance

Maintain List of the link signal quality to the neighbors and if better shift parent

Pick a new parent if no hello for a .You can pick a node below you in the tree so child may have to reselect their parent

26

Experiment: Effects of Loss

Percent Error From Single Loss vs. Network Diameter

0

0.5

1

1.5

2

2.5

3

3.5

10 20 30 40 50

Network Diameter

Percent Error From Single Loss

AVERAGECOUNT

MAXMEDIAN

27

Experiment: Benefit of Cache

Percentage of Network I nvolved vs. Network Diameter

0

0.2

0.4

0.6

0.8

1

1.2

10 20 30 40 50

Network Diameter

% Network

No Cache

5 Rounds Cache9 Rounds Cache

15 Rounds Cache

28

Summary

TAG is based on a declarative interfaceMakes network tasking easier for the end user

TAG outperforms centralized approaches in most cases

Relies heavily on underlying routing layerPlacement of query tree constrained by characteristics of routing tree

29

Critique

Fault Tolerance – too simplisticHow high is the failed node on the tree

MobilityMultiple Queries ?Generic?Multiple sinks / Sink mobility?More hypothesis testing - more problems!Compression + aggregation ?Energy budgeting ?

30

Questions ?