“ fault tolerant energy aware data dissemination protocol in sensor networks ” , in dsn ’ 04

“Fault Tolerant Energy Aware Data Dissemination Protocol in Sensor Networks”, in DSN

’04ICS280

Lee, Kyoungwoo

2

SPMS• [Khanna, DSN04] Fault Tolerant Energy Aw

are Data Dissemination Protocol – G. Khanna, S. Bagchi, and Y. Wu– Dependable Computing Systems Lab at Electri

cal and Computer Engineering in Purdue Univ.– “Fault Tolerant Energy Aware Data Disseminati

on Protocol in Sensor Networks”, in DSN ’04

3

SPMS - Overview• Motivation

– Battery-powered sensor nodes– Data implosion in sensor networks meta-data negotiation

from SPIN (Sensor Protocols for Information via Negotiation)– Prone to link and node failures

• Problem– How to disseminate data reducing energy-consumption and

end-to-end delay in the face of node and link failures • Solution

– Each node maintains routes in the zone– Data transfer in multiple hops using the lowest energy level

• Contribution– Resilient to node and link failures– Lower overall delay– Energy efficient data dissemination protocol

4

SPMS Protocol• Initial phase

– Zone Neighbors: nodes which lie within a node’s zone

• Zone: the region that a node can reach at the maximum power level

– Maintain a routing table for each of its zone neighbors

• Meta-data exchange phase– Broadcast ADV (Advertise) to zone

neighbors– Send REQ (Request) to the source

through the shortest path• Directly to the source in SPIN and

who is next hop neighbor in SPMS• Indirectly to the source through

multiple hops • Data dissemination phase

– Transmit DATA to the destination in exactly the same manner as the received REQ

• Direct form the source to the destination if they are next hop neighbors

• Otherwise through multi-hop communication

S

zone of S

C

FD

B

A

2

2

1

1

3

5

Zone radiuswith Max Tx Power

Dest Cost Next Hop Neighbor

A 2 YesB 2 YesC 1 YesD 3 NoF 2 No

1

Routing Table of S

5

SPIN vs. SPMS• SPIN

① S broadcasts ADV② C,D, & F sends REQ to S③ S sends DATA to C,D, & F

• SPMS① S broadcasts ADV

1. D & F check routing table2. D & F start TADV waiting for F

& C to send ADV of the same data

② C sends REQ to S③ S sends DATA to C④ C broadcasts ADV

1. F cancels TADV and starts TDAT2. D resets TADV

⑤ F sends REQ to C⑥ C sends DATA to F⑦ F broadcasts ADV

1. D cancels TADV and starts TDAT⑧ D sends REQ to F⑨ F sends DATA to D

S

zone of S

C

FD

B

A

2

2

1

11

3

5

DATA

REQ

ADV13

2

1ADV

2REQ3

DATA

4ADV 5REQ

6DATA

7ADV

8REQ

9DATA

C,D, & F are interested in DATA and no failure

[Claim] SPMS is better than SPIN in terms of energy & delay•Multi-hop communication with varying transmit power levels can reduce energy since Ed2

•Reducing power level of transmission can cause a smaller level of MAC layer contention

6

SPMS Protocol for Fault Tolerance

• Main Idea– Maintain alternate node in the routing table– PRONE: Primary Originator Node

• First choice node for requesting data from– SCONE: Secondary Originator Node

• Second choice to be used in case the PRONE is unreachable• Algorithm

– Update PRONE and SCONE if closer node broadcasts ADV– Send REQ to PRONE with timer– Send REQ to SCONE if timer expires

• SPMS can tolerate– Failure of the source node after its data has been received

by any of its neighbor nodes– Failure of any intermediate node

7

SPMS – Fault Tolerance• CASE 1: F fails before

broadcasting ADV① S broadcasts ADV

1. D & F check routing table and update PRONE and SCONE

2. D & F start TADV waiting for F & C to send ADV of the same data

② C sends REQ to S③ S sends DATA to C④ C broadcasts ADV

1. F cancels TADV and starts TDAT2. D resets TADV and updates

PRONE and SCONE⑤ F may send or not REQ to C⑥ C may send or not DATA to F⑦ F can not broadcast ADV

1. TADV of D expires2. TDAT of D starts

⑧ D sends REQ to C through F1. TDAT of D expires since F fails

⑨ D sends REQ to C (PRONE) directly using higher Transmit Power

⑩ C sends DATA to D directly

Dest Cost Next Hop Neighbor

S 3 NoC 2 NoF 1 Yes

S

zone of S

C

FD

B

A

2

2

1

11

3

51ADV

2REQ3

DATA

4ADV 5REQ

6DATA

7ADV

8REQ

10 DATA

C,F,& D interested in DATA and F fails

PRONE SCONE① S S

9 REQ

PRONE SCONE① S S④ C S

Routing Table of D

8

SPMS – Fault Tolerance (cont’)

• CASE 2: F fails after broadcasting ADV

① S broadcasts ADV1. D & F check routing table and

update PRONE and SCONE2. D & F start TADV waiting for F & C

to send ADV of the same data② C sends REQ to S③ S sends DATA to C④ C broadcasts ADV

1. F cancels TADV and starts TDAT2. D resets TADV and updates

PRONE and SCONE⑤ F sends REQ to C⑥ C sends DATA to F⑦ F broadcasts ADV and fails

1. TDAT of D starts⑧ D sends REQ to F

1. TDAT of D expires since F fails2. D regards F dead and TDAT for C

starts⑨ D sends REQ to C (SCONE)

directly using higher Transmit Power

⑩ C sends DATA to D directly

S

zone of S

C

FD

B

A

2

2

1

11

3

51ADV

2REQ3

DATA

4ADV 5REQ

6DATA

8REQ

10 DATA

C,F,& D interested in DATA and F fails

Dest Cost Next Hop Neighbor

S 3 NoC 2 NoF 1 YesPRONE SCONE① S S④ C S

9 REQ

7ADV

PRONE SCONE① S S④ C S⑦ F C

Routing Table of D

9

SPMS – Fault Tolerance (cont’)

• CASE 3: F and C fail① S broadcasts ADV

1. D & F check routing table and update PRONE and SCONE

2. D & F start TADV waiting for F & C to send ADV of the same data

② C sends REQ to S③ S sends DATA to C④ C broadcasts ADV

1. F cancels TADV and starts TDAT2. D resets TADV and updates PRONE

and SCONE⑤ F sends REQ to C⑥ C sends DATA to F⑦ F broadcasts ADV and F & C fail

1. TDAT of D starts⑧ D sends REQ to F

1. TDAT of D expires since F fails2. D regards F dead and TDAT for C

starts⑨ D sends REQ to C (SCONE) directly

using higher Transmit Power What if C and F all fail?

S

zone of S

C

FD

B

A

2

2

1

11

3

51ADV

2REQ3

DATA

4ADV 5REQ

6DATA

8REQ

C,F,& D interested in DATA and C&F fail

Dest Cost Next Hop Neighbor

S 3 NoC 2 NoF 1 YesPRONE SCONE⑦ F C

9 REQ

7ADV

•SPMS keeps a pair of PRONE and SCONE Multiple SCONE can increase fault tolerance

Routing Table of D

10

SPMS – Energy Analysis• ESPIN = (A+D+R)*E1 + (A+D+R)*Er

• ESPMS = k*A*E1 + k*(D+R)*Em + k*(A+D+R)*Er

-(k-1) relay nodes from the source to the destination -A: size of ADV-D: size of DATA-R: size of REQ-E1, E2, …, Em where Ei>Ei+1: energy consumed per transmitted bit corresponding to the different transmission power levels-Er: energy required to receive the packet

Src Dest1 2 3 k-1D*E1

ADVA*E1 DATA

REQ R*E1

D*Er

A*Er

R*Er

Src Dest1 2 3 k-1

A*E1

DATA

REQ

ADV

D*Em, R*Em

ADV ADV ADV

A*Er, D*Er, R*Er

DATA

REQ

ESPIN

ESPMS

11

SPMS – Energy Analysis (cont’)

• ESPIN = (A+D+R)*E1 + (A+D+R)*Er

• ESPMS = k*A*E1 + k*(D+R)*Em + k*(A+D+R)*Er

• Ratio of Energy (SPIN/SPMS) = ESPIN/ESPMS

(Observation) •Higher radius of transmission indicates higher distance from the source to the destination•In SPIN, the energy overhead increases exponentially since Ed2 but it increases linearly in SPMS

12

SPMS – Energy Experiments• Assumptions

– Sensor field with uniform density of nodes– Power level: 3.1622, 0.7943, 0.1995, 0.05, and 0.0125 mW taken from Berkeley MICA2– Distance: 91.44, 45.72, 22.86, 11.28, and 5.48 m– The maximum number of the zone is six– Size of DATA: 40 bytes– Size of REQ and ADV: 2 bytes– All-to-all communication: each node generates 10 new packets and every node is interested in them

(Observation) •SPMS saves 26 ~ 42 % of energy compared to SPIN•SPMS outperforms SPIN with increases of the number of nodes and radius of transmission

13

SPMS - Conclusion• Fault-Tolerance

– Maintain PRONE and SCONE at routing table

– Select alternate if node fails• Energy-Efficiency

– Multi-hop communication since Ed2

• End-to-end delay– Less delay due to less contention of MAC

14

SPMS - Discussion• Single-hop vs. multi-hop transmission

– Depends on application• Increase fault-tolerance using multiple SCO

NEs– Increase sleep time and decrease delay when

the multiple failures occur

[Szewczyk, SENSYS04] R. Szewczyk, A. Mainwaring, J. Polastre, and D. Culler, “An Analysis of a Large Scale Habitat Monitoring Application” in SenSys 04

Real measurement of lifetime of sensors in Habitat Application

Lifetime in the single hop network Lifetime in the multi-hop network

" Balancing Energy Efficiency and Quality of Aggregate

Data in Sensor Networks", To Appear in the VLDB Journal

Special Issue on Data Stream Processing, 2005

16

GaNC and TiNA

• [Sharaf, VLDB05] GaNC and TiNA – M. A. Sharaf, J. Beaver, A. Labrinidis, and P. K. Chrysa

nthis– Dependable Computing Systems Lab at Computer Sci

ence in Univ. of Pittsburgh– “Balancing Energy Efficiency and Quality of Aggregate

Data in Sensor Networks”, in VLDB ’05• Propose group-aware network configuration meth

od (GaNC) and a framework to use temporal coherency tolerances (TiNA) to provide significant energy savings and a negligible drop in quality of data

17

GaNC & TiNA - Overview• Motivation

– Further energy savings in the context of In-network aggregation• Goal

– Reduce the size of transmitted data– Minimize the number of transmitted messages– Without significant QoD

• Solution– GaNC can reduce the size of transmitted data– TiNA can minimize the number of transmitted messages as well

• Contribution– Propose enhanced network configuration scheme– Provide a framework on top of existing in-network configuration

18

GaNC• Group-Aware Network Configuration

method• Observation

– The length of messages sent depends on the number of groups in the routing subtree

• Idea– Reduce the number of groups to red

uce the length of messagesGroup-aware Network Configuration

• Cluster along the same path sensor nodes that belong to the same group

• Consider semantics of the query and properties of sensor nodes

• Reduce the size of transmitted data

A

B C

A

B C

Group1 Group2

Group2 Group2

19

TiNA• Temporal coherency-aware In-

Network Aggregation• Goal

– Reduce the size of transmitted data

– Minimize the number of transmitted message

• Idea– Exploit temporal correlation in

streams of sensor readings• Suppress insignificant readings• Potentially allow nodes to switch

to sleep mode• Work on top of existing in-

network aggregation• Introduce TOLERANCE for

temporal coherency tolerance

A

C

D •Old = 20•New = 21•If TOLERANCE = 10%,don’t Transmit Newbecause (21-20)/21 < 0.1

BGroup1 Group2

•Old = 20•New = 25•If TOLERANCE = 10%,Transmit Newbecause (25-20)/21 > 0.1

Reduce the numberof messages

Reduce the size of data

20

Synchronization in TAG• TAG

– Divide a given DURATION into Communication Slots

• Duration of each Communication Slot = DURATION/d

• where d = number of slots = maximum depth of routing tree

– Provide a query result every Epoch DURATION

• During a given Communication Slot, one level (A) sending and another level (B) listening

• At the next Communication Slot, A goes to sleep mode and B sending (C may be listening)

A

C

D

B

d (depth) = 3

1ABCD

2 3

Listening

Sending Sleep

Listening

Sending

Listening

Sending

Sleep

Sleep

Sleep

Sleep Sleep

21

Synchronization in Cougar• Cougar

– Pragmatic approach– Algorithm

• In a certain round, A adds C to its waiting list if A receives data from C

• In the following rounds, A waits to hear from all nodes in the waiting list

• To prevent indefinite waiting, each node transmits reading or notification

– Reduce response time for uncongested network

A

CB

waiting_list={B} {B,C}

22

Network Configuration Method

• First-Head-From Network Configuration

– Based on network proximity

– Algorithm1) Root sensor prepares query

msg with query spec. & Ls and broadcasts

2) Sensor i receives msg & sets Li=Ls+1

3) Sensor i sets Pi=Ids, then sets Ls=Li & Ids=Idi

4) Steps 2) & 3) repeated

• Group-aware Network Configuration

– Keep members of the same group within same path

– Algorithm1) Root prepares query msg wi

th query spec. Ls, & Gs and broadcasts

2) i receives msg & sets Li=Ls+1

3) i sets Pi=Ids & PGi=Gs, then sets Ls=Li , Ids=Id, & Gs =Gi

4) i continues to listen5) Tie-breaker conditions to se

lect better parent 6) Steps 2) to 5) repeated

Main Difference: GaNC can switch to a better parent• First tie-breaker: the same group ID preferred

same group can reduce size of msg• Second tie-breaker: the lower distance preferred

closer parent saves tx energy

23

TiNA - algorithm• Main Idea

– Use temporal correlation in a sequence of sensor readings by suppressing insignificant readings

• TOLERANCE clause in SQLTOLERANCE tct (eg: tct=10%)– Specify the temporal coherency toleran

ce for the query– Output filter– Only report readings differing from the l

ast reported readings by more than 10%

• Information to utilize TOLERANCE– Leaf node: keep the last reported readi

ng– Internal: last reported data from each

child as well

• Algorithm– Leaf node

• Report VNEW if VNEW violates tct s.t. |Vold-Vnew|/Vnew > tct

– Internal node1) Collect the data from children2) Compute the partial result3) Take its own reading which can

be aggregated within a group already exists in the partial result regardless of tct

4) If a new group, the reading is only added when violating tct

5) Compares an OLD partial result with the NEW partial with tct = 0

24

TiNA on top of TAG• Use the predefined

communication slots for sending and listening

• When communication slot expires, parent checks and takes the last reported data for each child it didn’t heard from

• Representation– Circles: nodes– Arrows: the flow of data– Boxes: current state

• Old – last reported reading• New – current reading

– Table: previously reported partial result

– Cost: the size of table•Every reading is sent from child to parent

25

TiNA on top of TAG (cont’)④ New = 6, Old = 5 and |6-5|/6 > 0.1,

thus send New⑤ Just add 11 to 6 and compares 17

(New) with 15 (Old), and |17-15|/17 > 0.0, thus send New partial aggregate value

⑥ |4.1-4|/4.1 < 0.1 thus suppress⑦ Aggregate reading to partial

(17+4=21) and compares it with Old (21), it suppresses since no change

• Algorithm– Leaf node

• Report VNEW if VNEW violates tct s.t. |Vold-Vnew|/Vnew > tct

– Internal node1) Collect the data from children2) Compute the partial result3) Take its own reading which can be agg

regated within a group already exists in the partial result regardless of tct

4) If a new group, the reading is only added when violating tct

5) Compares an OLD partial result with the NEW partial with tct = 0

•Less number of sent messages

26

TiNA on top of Cougar

• In Cougar, parents wait to hear from all their children

• Send heartbeat message [notification] when it can tolerate the quality

notify

notify

•Energy saving by sending notificationPacket instead of data packet with Respect to size of message

27

TiNA with GaNC• Further energy saving

– Reduce total size of messages– Reduce total number of messages

•Presentation•Circles: nodes•Groups: Blue or not•Boxes: New data (violating tct)•Value: difference b/w New & Old•m: transmission of a messageof unit size

Totally 5 messages sent& total size of messages is 6

Totally 4 messages sent& total size of messages is 4

Complementary data(+5 & -5) canceleach other savetransmission

28

Evaluation by simulation• Energy, REM (Relative Error Metri

c), and Response Time• Energy

– 4 main activities• Txing, Listening, • Sampling, Processing

– Parameters for Txing & Rxing• Sensor operates 3 volts• Data rate: 40 Kbps• Tx current: 0.012 Amp• Rx current: 0.0018 Amp

– Tcost = 3 V * 0.012 A * 1/40,000 sec = 0.9 uJoules

• Energy consumption for one bit transmission

– Rcost = 3 V * 0.0018 A * 1 sec = 0.0054 Joules

• Listening for one second• Independent of number of messag

es

29

Experiments• Sensitivity to temporal coherency tolerance

– Measure Energy, REM, response time for TiNA vs. for Cougar and TAG varying tct– Tradeoff between Energy Saving and REM

TiNA with Cougar uses 56% of energy by CougarAt tct=30%, only 24% but REM increases to 3.3%

TiNA with TAG uses 86% of energy by TAGAt tct=30%, only 74.9% but REM increases to 3.7%TiNA on TAG must listen for entire assigned time slot

30

Experiments (cont’)• Sensitivity to temporal coh

erency tolerance (2) – Tradeoff between energy savi

ng and response time– The time to hear from all chil

dren decreases• TiNA can send Notification i

nstead of readings within tct– For tct=0% (30%), the respon

se time of TiNA on Cougar 60% (27%) of Cougar’s

– The response times for TiNA on TAG are always same as the duration

31

Experiments – Energy vs. Duration &

Scalability

The amount of energy increases with an increaseof Duration in TAGmore sensors can send readings as Dur increases

Energy consumption increases with an increaseof number of sensors in CougarEnergy saving increases as network increases

32

Experiments –Energy Effect of GaNC

1) GaNC can save energy in sensor network for the most part (positive effect)GaNC can reduce the size of transmitted message

2) The energy savings of GaNC over FHF decreasesas tct increases (negative effect)Some nodes switch to parents in the same group(switching parents can cause more messages sent)

For larger network, positive effects outweigh negativeEffects. As tct increases, less nodes transmitting

33

Experiments – TiNA with GaNC vs. number of groups

AT the small group (eg. 5), GaNC consumes significantly (41%,38%,37%) children nodes can select parents in the same group as themAt the large group (eg. 50), not reduce dramatically (12%, 10%,9%)less chances that children can find parents in the same group

34

Conclusion• GaNC (Group-aware Network Configuration)

– Consider semantics of the query and properties of sensor nodes– Reduce the size of transmitted data

• TiNA (Temporal coherency-aware in-Network Aggregation)– Temporal correlation in conjunction with in-network aggregation– Minimize the number of transmitted messages– Decrease the size of transmitted data

• Significant energy saving while negligible drop in Quality of Data– TiNA can reduce power consumption for communication by up to

60% and extend the life by up to 270%– Additional 33% of energy can be saved by incorporating the GaNC

with TiNA