EnergyEfficient Data Management for Sensor Networks

1
Energy-Efficient Data Management for Sensor
Networks

• Niki Trigoni, Cornell University
• niki_at_cs.cornell.edu
• Al Demers, Cornell University
• Johannes Gehrke, Cornell University
• Rajmohan Rajaraman, Northeastern University
• Yong Yao, Cornell University
• The Cougar Project
• http//cougar.cs.cornell.edu

2
Outline

• Our Model
• Sensor Network
• Data and Queries
• 2. First approach to saving energy
• Motivation
• Multi-Query Optimization
• Experiments
• 3. Second approach to saving energy
• Motivation
• Wave Scheduling
• Experiments
• 4. Conclusions and Future Work

3
Sensor Network

• Small stationary sensor nodes connected through
  a multi-hop wireless network
• Communication Limited bandwidth, variable
  latency, packet drops
• Computation Limited processing power and
  memory sizes
• Power Limited supply of energy

4
The DB View of Sensor Networks

• Traditional
• Procedural addressing of individual sensor
  nodes user specifies how task executes data is
  processed centrally.
• DB Approach
• Declarative querying and tasking user isolated
  from how the network works in-network
  distributed processing.

SensId
Loc
Time
Type
Value
temperature
1
(2,5)
3
60
pressure
1
(2,5)
6
62
User IF
SensId
Loc
Time
Type
Value

light
2
(4,2)
3
55
pressure
2
(4,2)
5
30
SensId
Value
Loc
Time
Type
humidity
3
(3,1)
3
70
5
Example Queries

• Snapshot queries
• What is the concentration of chemical X in the
  northeast quadrant?SELECT AVG(R.concentration)
  FROM Sensordata RWHERE R.loc in
  (50,50,100,100)
• Long-running queries
• Notify me over the next hour whenever the
  concentration of chemical X in an area is higher
  than my security threshold.SELECT
  R.area,AVG(R.concentration)FROM Sensordata
  RWHERE R.loc in rectangleGROUP BY
  R.areaHAVING AVG(R.concentration)gtTDURATION
  (now,now3600)EVERY 10

6
Saving energy Two distinct approaches

• Reduce the number of messages
• Turn off radios whenever not needed

7
Outline

• Our Model
• Sensor Network
• Data and Queries
• 2. First approach to saving energy
• Motivation
• Multi-Query Optimization
• Experiments
• 3. Second approach to saving energy
• Motivation
• Wave Scheduling
• Experiments
• 4. Conclusions and Future Work

8
First approach Previous work

• Two choices
• Centralized processing
• In-network processing
• Why in-network processing?
• Sensor network is power
• (and bandwidth) constrained
• Local computation is much
• cheaper than communication

9
Motivating Example

• Scenario Parking lots
• City with several attractions
• Parking lots around the attractions
• Sensor network covers parking lots
• Web service allows users to send queries

Gateway
10
Motivating Example

• Scenario Parking lots
• City with several attractions
• Parking lots around the attractions
• Sensor network covers parking lots
• Web service allows users to send queries

Gateway
11
Motivating Example

• Scenario Parking lots
• City with several attractions
• Parking lots around the attractions
• Sensor network covers parking lots
• Web service allows users to send queries

Gateway
12
Extensions to the standard SensorDBMS

• Multiple aggregate queries executed regularly
• Probabilistic sensor updates
• Ad-hoc aggregate queries

13
Problem Definition

• Time is divided into periods.
• Given
• Aggregate Query Workload QWltQ1,p1gt,
• Sensor Update Workload DWlts1,u1gt,
• Tree that connects sensors to the query point
• Cost of communicating n bits on a link is
  ??n
• Find a query execution plan that reduces the
  total communication cost.

14
Extension 1 multiple queries

• Scenario Queries and sensor updates are
  deterministic
• Intuition Identify values of sub-aggregates that
  can be shared among queries
• Query Workload Example
• Q1 cd
• Q2 e
• Q3 cde

a
b
c
e
d
15
Extension 1 multiple queries

• Queries cd, cde, e are not linearly
  independent

Gauss-Jordan
Elimination
Rank2
16
Extension 1 multiple queries

• Theorem 1 When queries and sensor updates are
  deterministic, the reduction of projected queries
  to their basis minimizes the communication cost.

17
Extension 2 probabilistic sensor updates

• Scenario Queries are deterministic, but updates
  are probabilistic
• Intuition Propagate only the changes
• Example only sensors c and d are updated

a
b
c
e
d
18
Extension 2 probabilistic sensor updates

• Query-based encoding
• Result Code b tells a which queries are updated
  
• Result Data b sends to a only the results of the
  updated queries
• (in fact only the basis)
• gt 2 data values

a
b
c
e
d
19
Extension 2 probabilistic sensor updates
Extension 2 probabilistic sensor updates
EC-based encoding Result Code b tells a which
equivalence classes are updated Result Data b
sends to a only the new values of the updated
equivalence classes (in fact only the basis) gt
1 data value
a
b
c
e
d
c and d are indistinguishable. They belong to the
same equivalence class
20
Extension 2 probabilistic sensor updates
EC-based encoding Result Code b tells a which
equivalence classes are updated Result Data b
sends to a only the new values of the updated
equivalence classes (in fact only the basis) gt
1 data value
a
b
c
e
d
21
Extension 2 probabilistic sensor updates

• Theorem 2 When queries are deterministic and
  sensor updates are probabilistic, the size of the
  Reduced Result Data component in EC-based
  encoding is a lower bound on the size of the
  optimal result message.

22
Extension 3 ad-hoc aggregate queries
Push Model
Pull Model
Hybrid Model
View Node
23
Extension 3 To pull or to push? (Ex1)
a
p probability of Q q query message cost r
result message cost
Q
b
model
push
pull
edge
r
qp r
a-b
a
a
pull if q lt (1-p) r
push otherwise
q
p r
r
b
b
24
Extension 3 To pull or to push? (Ex2)
Q1 cd Q2 de pr(Q1) pr(Q2) p q query
message cost r result message cost (qltr)
a
b
Q1
Q2
c
d
e
model
push
pull
edge
a-b
b-c
b-d
b-e
25
Extension 3 To pull or to push? (Ex2)
pr(Q1) p pr(Q2) p
a-b
b-c
b-d
b-e
26
Putting it all together

• Simulation Phase it monitors local edge costs
  for a number of periods
• EC-based vs Query-based result costs
• push vs pull costs
• Dynamic Programming Phase It makes compatible
  decisions that optimize the global network cost.

27
Experimental evaluation

• Grid topology (1010 network)
• Gateway (root of the tree) at the top left corner
• Minimum spanning tree
• Query Workload
• Rectangles of randomly selected dimensions.
• Query probabilities are the same for all queries.
• We vary the no of queries and their
  probabilities.
• Sensor Workload
• Sensor update probabilities are the same for all
  sensors.
• We vary sensor probabilities in different
  experiments.
• 2-phase algorithm Simulation (50 periods) and
  dynamic programming

28
Extension 1 Reducing the data values
queryProb1, updateProb1
NonReduced
Reduced
total cost
number of queries
29
Extension 2 Result encoding techniques
queryProb1, updateProb0.01
Query-based

total cost
EC-based
number of queries
30
Extension 3 Hybrid pull-push model
queries50, updateProb1, a64, b1
pull
push
hybrid pull-push

total cost
query probability
31
Overall benefits
queries80, updateProb0.05, a64, b1
naive - push
naive- pull

total cost
our algorithm
lower bound
query probability
32
Contributions of first approach

• Extended the usage model of sensor networks
• Defined formally the problem of minimizing the
  communication cost under such usage model
• Introduced three simple techniques that improve
  the communication cost significantly
• Reduction of results to the basis of the query
  space
• Hybrid (Query-based and EC-based) result encoding
• Hybrid pull-push model
• Proposed a two phase algorithm that makes optimal
  decisions at the entire tree
• Experiments showed significant cost savings

33
Outline

• Our Model
• Sensor Network
• Data and Queries
• 2. First approach to saving energy
• Motivation
• Multi-Query Optimization
• Experiments
• 3. Second approach to saving energy
• Motivation
• Wave Scheduling
• Experiments
• 4. Conclusions and Future Work

34
Second Approach Assumptions

• Nodes are location-aware
• Neighboring nodes in the network are synchronized
  (GPS, distributed time synchronization
  algorithms)
• Recent studies about radio energy consumption
• idle receive transmit
• 1 1.2 1.7
• 1 2 2.5
• 1 1.05 1.4
• We have a few gateway nodes

35
Data dissemination problem

• A data dissemination protocol in a sensor network
  has two components
• a scheduling algorithm and a routing algorithm.
• Problem
• Find an energy-efficient scheduling-routing
  pair.

36
Scheduling and Routing Results

• Let ??n be the cost of sending n bits on a
  link.
• Theorem 1 For any ?gt0 and ?gt0, finding an
  optimal routing-scheduling pair to minimize
  energy is NP-hard.
• Theorem 2 For any ?gt0 and ?gt0, given a set of
  source-destination routes, the problem of finding
  an activation schedule that minimizes energy is
  NP-hard.

37
Previous work Tiny Aggregation trees
a
b
c
Sending
Receiving
d
e
Idle
f
g
38
Scheduling multiple TAG trees

• If scheduled consecutively delay grows linearly
  with number of destinations
• If scheduled concurrently energy grows more than
  linearly in the number of destinations
  (contention)
• TAG scheduling is good for only a small number of
  trees

39
Our approach Wave Scheduling

• Goals
• Minimize collisions at the MAC layer
• Manage the radio in a power efficient manner
• Select efficient routes wrt energy or latency

40
GAF Topology Control
GAF Geographic Adaptive Fidelity
Periodically re-elect leader in each cell
41
Scheduling Trees embedded on a grid
42
Our approach Wave Scheduling
43
Our approach Wave Scheduling
EAST WAVE
Sending
Receiving
Idle
44
Our approach Wave Scheduling
EAST WAVE
Sending
Receiving
Idle
45
Our approach Wave Scheduling
EAST WAVE
Sending
Receiving
Idle
46
Our approach Wave Scheduling
EAST WAVE
Sending
Receiving
Idle
47
Our approach Wave Scheduling
EAST WAVE
Sending
Receiving
Idle
48
Wave Scheduling Properties

• Repeats in all 4 directions
• (North, East, South, West)
• Non-interfering edges are scheduled concurrently
• Simple or Interleaved
• General-purpose schedule. Every edge of the
• network is activated exactly once per period
• Can think of it as a real wave

49
Wave Scheduling Intuition
Sending
Receiving
Idle
50
Wave Scheduling Intuition
Sending
Receiving
Idle
51
Wave Scheduling Intuition
Sending
Receiving
Idle
52
Wave Scheduling Intuition
Sending
Receiving
Idle
53
Wave Scheduling Routing

• Schedules have handed-ness
• e.g. The (N,E,S,W) schedule favors
• paths that take right-hand turns
• Given a schedule, we can select
• Min-Delay Routes or
• Min-Energy Routes
• What about holes in the grid?

E
N
S
W
54
Wave Scheduling Routing
Long right turn path has smaller latency
than short left turn path
DESTINATION
SOURCE
gt
55
Wave Scheduling Routing Anomalies

• Solution
• Proactive distance-
• vector routing scheme
• Get min-feasible-hops
• with min delay or
• min-feasible-delay
• with min hops
• Small routing tables
• 1 bit per edge per
• destination node

c
d
a
b
DESTINATION
SOURCE
gt
56
Wave Scheduling Routing Anomalies
Problem Infinite Loop (because of heavy
traffic) Solution Decide the next hop for a
message at the time it is received by a node
A
B
C
SOURCE
DESTINATION
57
Comparison with optimal

• ? ?n, where ? is the start-up cost
• Energy-based routing minimizes the second
  component of the cost (?n).
• If ?ltlt ?n then the wave schedules are
  asymptotically as efficient (wrt energy) as the
  optimal schedule
• Otherwise, for networks without holes and traffic
  under capacity, the wave algorithm is at most 4
  times more expensive than the optimal.

58
Scheduling Experimental Setup

• NS-2 Network Simulator
• Mac layer IEEE 802.11
• communication range 250m
• interference range 550m
• transmitreceiveidle 1.61.21.0
• grid cell has size (100m 100m)
• grid 100 (10 by 10) cells
• We assume that we have one node per grid cell
  (GAF)
• Experiments lasted 5000 seconds simulation
  time
• Sources and destinations of messages are
  randomly selected.

59
Scheduling Experiments
Wave ENERGY Varying the number of destination
nodes
Average Energy Consumption
Energy (Joules)
1,2,3,5,10 trees (gateways)
Messages
60
Scheduling Experiments
Tree ENERGY Varying the number of destination
nodes
Average Energy Consumption
10 trees
5 trees
Energy (Joules)
3 trees
1 tree
Messages
61
Scheduling Experiments
ENERGY Energy-based vs Delay-based Routing
Average Energy Consumption
Delay-based
Energy (Joules)
Energy-based
Messages
62
Contributions of the second approach

• Wave scheduling a class of algorithms for
  scheduling sensor nodes.
• Energy-based vs Delay-based routing
• Routing algorithms interact symbiotically with
  the scheduling decisions.
• Energy is saved by
• avoiding collisions at the MAC layer,
• turning off the radios whenever not needed.
• Experimental results show significant energy
  savings at the cost of higher latency.

63
Conclusions

• Energy is a limited resource in sensor networks
• Two approaches to saving energy
• By reducing the number of messages.
• Multi-query optimization in a hybrid pull-push
  model
• By turning off radios whenever possible
• Sensor node coordination with (global) wave
  schedules

64
Future Work

• Multi-query optimization
• Selecting trees given query and update workloads
• Multi-query optimization of a larger class of
  queries
• Scheduling
• Explore space between tree and wave scheduling
• Wave scheduling in the presence of dynamic
  failures
• Interaction between the two approaches
• Efficient wave schedules given specific
• query execution plans

65
Thank you!