Opportunistic Routing Algorithms in Delay T olerant Networks - - PowerPoint PPT Presentation

▶
opportunistic routing algorithms in delay t olerant
SMART_READER_LITE
LIVE PREVIEW

Opportunistic Routing Algorithms in Delay T olerant Networks - - PowerPoint PPT Presentation

Nov 18, 2023 246 likes •749 views

Opportunistic Routing Algorithms in Delay T olerant Networks Eyuphan Bulut Rensselaer Polytechnic Institute Department of Computer Science and Network Science and Technology (NeST) Center PhD Thesis Defense Feb 4 th , 2011 Outline

slide-1
SLIDE 1

Opportunistic Routing Algorithms in Delay T

  • lerant Networks

Eyuphan Bulut

Rensselaer Polytechnic Institute Department of Computer Science and Network Science and Technology (NeST) Center PhD Thesis Defense Feb 4th, 2011

slide-2
SLIDE 2

Outline

  • Introduction to DTNs

– Challenges of routing

  • Proposed Algorithms

1) M ulti-period Spray and Wait routing 2) M ulti-period erasure coding based routing 3) Efficient single-copy routing utilizing correlation between node meetings 4) Social relation based routing

  • Summary of Contributions

2/ 4/ 2011 Bulut: PhD Defense (RPI) 2

Simulation Results Based on Real and S ynthetic Traces

slide-3
SLIDE 3

Outline

  • Introduction to DTNs

– Challenges of routing

  • Proposed Algorithms

1) M ulti-period Spray and Wait routing 2) M ulti-period erasure coding based routing 3) Efficient single-copy routing utilizing correlation between node meetings 4) Social relation based routing

  • Summary of Contributions

2/ 4/ 2011 Bulut: PhD Defense (RPI) 3

slide-4
SLIDE 4

Delay Tolerant Networks

  • Intermittently connected mobile networks

– Sparse mobile networks

  • M ain difference from M ANETs

– Lack of continuous end-to-end connectivity – Utilizes “store-carry-and-forward” paradigm in

routing

2/ 4/ 2011 Bulut: PhD Defense (RPI) 4

slide-5
SLIDE 5

Applications of DTNs

  • Space networks
  • Satellites and planets
  • M ilitary Networks
  • Soldiers, aircrafts
  • Social Networks
  • People, base stations
  • Vehicular Networks
  • Underwater networks

2/ 4/ 2011 Bulut: PhD Defense (RPI) 5

Mars Jupiter Earth

slide-6
SLIDE 6

Routing in DTNs

  • Challenges:

– Dynamic and sparse topology – Low probability of end-to-end connectivity – How to locate destination with local knowledge? – Opportunistic message exchanges

  • When nodes come to the range of each other
  • How to decide whom to forward/copy a message?

2/ 4/ 2011 Bulut: PhD Defense (RPI) 6

A B C D E B C

slide-7
SLIDE 7

Routing in DTNs

2/ 4/ 2011 Bulut: PhD Defense (RPI) 7

Knowledge Number of carriers Multiple Replication-based Full knowledge History based Single Erasure coding- based Flooding Quota based

  • Future meetings
  • Position information

Opportunistic (Whenever they are in the range of each other)

Selective (Prediction-based

  • r Probabilistic)

Random

i.e. first node met

Use information extracted from encounter history to predict future meetings

slide-8
SLIDE 8

Our Research Path

2/ 4/ 2011 Bulut: PhD Defense (RPI) 8

Random M obility M odels

(Random walk, waypoint, direction)

Analysis of Real DTN Traces

(E ffects of pair-wise relations in routing)

Social Behavior

(Human carried wireless devices)

Reliability Correlated node mobility

(Repetitive behavior, Importance of past)

M ulti-copy based Erasure coding based Single-copy based Single-copy based

slide-9
SLIDE 9

Outline

  • Introduction to DTNs

– Challenges of routing

  • Proposed Algorithms

1) M ulti-period Spray and Wait routing 2) M ulti-period erasure coding based routing 3) Efficient single-copy routing utilizing correlation between node meetings 4) Social relation based routing

  • Summary of Contributions

2/ 4/ 2011 Bulut: PhD Defense (RPI) 9

slide-10
SLIDE 10

Spray and Wait *

  • Random mobility model

– Exp. dist. intermeeting times

2/ 4/ 2011 Bulut: PhD Defense (RPI) 10

M essage delivered M essage copied M essage copied

time Cdf of delivery probability Destination Node with message copy Node without message copy

L1 < L2 < L3

* Spyropoulos et. al. Transactions on Networking, 08

L1 L2 L3

slide-11
SLIDE 11

Two Period Spray & Wait

  • Spray L1 copies at the beginning
  • Spray additional L2 - L1 copies at time xd (start of

second period)

2/ 4/ 2011 Bulut: PhD Defense (RPI) 11

L L2 L1

1) M aintain the same delivery rate by deadline (td) 2) Lower the average cost GOALS:

Delivery probability in first period

L1(P)+ L2(1-P) < L

slide-12
SLIDE 12

Three Period Spray & Wait

2/ 4/ 2011 Bulut: PhD Defense (RPI) 12

L1 L L2 L3

1st Period 2nd Period 3rd Period

slide-13
SLIDE 13

M ultiple Period Spray & Wait

  • If we currently have k spray and wait periods, to obtain k+1

periods:

– Partition each period into two sub-periods optimally – Take the one which makes the overall cost minimum

2/ 4/ 2011 Bulut: PhD Defense (RPI) 13

L1 L2 L5 L6 L3 L4 L

2 periods: L1 and L2

3 periods: Select either a) L1, L5 and L6 b) L3, L4 and L2

slide-14
SLIDE 14

Acknowledgment of delivery

  • Two types:

– Type I: Acknowledgment by flooding

  • Pros: Acks are small, lower cost
  • Cons: Takes time to reach all nodes, thus extra copying

may occur

– Type II: Single broadcast with powerful radio

  • Pros: Immediate acknowledgment
  • Cons: Cost of powerful radio

2/ 4/ 2011 Bulut: PhD Defense (RPI) 14

slide-15
SLIDE 15

Optimum Li’s from Analysis

2/ 4/ 2011 Bulut: PhD Defense (RPI) 15

Xd=285s

Cost=4.64 Cost=4.28 Cost=5.87

slide-16
SLIDE 16

Simulation Results

  • Percentage of Saving

– While achieving the same delivery ratio by deadline

2/ 4/ 2011 Bulut: PhD Defense (RPI) 16

slide-17
SLIDE 17

Extensive Simulations

  • Theoretical results are matching with simulation

results

  • Effect of different number of nodes, different desired

delivery ratios etc.

  • Results on Real Traces

– Demonstrates benefit, but still needs careful analysis due

to heterogeneous meeting behavior of different nodes

[IEEE/ACM Transactions on Networking’10], [Globecom’08], [ACIT A’08]

2/ 4/ 2011 Bulut: PhD Defense (RPI) 17

slide-18
SLIDE 18

Outline

  • Introduction to DTNs

– Challenges of routing

  • Proposed Algorithms

1) M ulti-period Spray and Wait routing 2) M ulti-period erasure coding based routing 3) Efficient single-copy routing utilizing correlation between node meetings 4) Social relation based routing

  • Summary of Contributions

2/ 4/ 2011 Bulut: PhD Defense (RPI) 18

slide-19
SLIDE 19

Replication vs. Erasure Coding

2/ 4/ 2011 Bulut: PhD Defense (RPI) 19

  • Replication

A message (M bytes) L copies Source Destination In total L relay nodes Wait for 1 of them to reach destination M*L bytes of data is transmitted to the network A message (M bytes) Divided into k small parts (M/k bytes of each) Source Destination In total R*k relay nodes Wait for k of them to reach destination ~M*R bytes of data is transmitted to the network (independent from k) Encoded into R*k blocks R: Replication factor If L=R, then the cost (transmitted bytes over the radio) becomes equal.

  • Erasure Coding (EC)
slide-20
SLIDE 20

Replication vs. Erasure Coding

  • Which one is better?
  • 1. Spraying L messages and waiting for 1 (to reach destination)?

– Spraying duration takes less time than the second one

  • 2. Spraying Φ=R* k messages and waiting for k?

2/ 4/ 2011 Bulut: PhD Defense (RPI) 20

Costs are the same when R=L

“EC” also provides more reliable routing:

In a failure of one packet, the performance of “ replication” routing is affected more than the performance of “erasure coding” based routing.

slide-21
SLIDE 21

Replication vs. Erasure Coding

  • If desired delivery rate is higher, we can achieve more cost

saving with erasure coding based routing compared to replication routing.

  • Optimum Single-period erasure coding based routing:

– Try to minimize Replication factor (R) since cost is proportional to R. – M aintain delivery rate by deadline

2/ 4/ 2011 Bulut: PhD Defense (RPI) 21

(R2,k) (R1,k) R1 > R2

slide-22
SLIDE 22

M ulti-period Erasure Coding-based Routing

  • In 1st period:

–

Create Φ2=kR* coded blocks, where R*>R

  • pt (optimum R in single period)
  • Linear time complexity of creating these packets (Tornado codes)

–

Spray Φ1= αkR* of them and try delivery with them

  • If delivery doesn’t happen in 1st period (by xd)

–

Spray remaining Φ2-Φ1 of them in 2nd period

  • Same goals:

–

M aintain delivery rate by deadline

–

Achieve lower cost on average

2/ 4/ 2011 Bulut: PhD Defense (RPI) 22

Start of second period

slide-23
SLIDE 23

Simulations

  • M essage size 100Kb
  • Costs at the delivery (Type II) and after all nodes are

acknowledged (Type I):

2/ 4/ 2011 Bulut: PhD Defense (RPI) 23

1 period 2 periods [ICC’10]

1 period Replication Based Cost=578 1 period Replication Based Cost=587 2 period Replication Based Cost=464 2 period Replication Based Cost=478

Lower cost than Replication based Routing Lower cost than Replication based Routing

slide-24
SLIDE 24

Outline

  • Introduction to DTNs

– Challenges of routing

  • Proposed Algorithms

1) M ulti-period Spray and Wait routing 2) M ulti-period erasure coding based routing 3) Efficient single-copy routing utilizing correlation between node meetings 4) Social relation based routing

  • Summary of Contributions

2/ 4/ 2011 Bulut: PhD Defense (RPI) 24

slide-25
SLIDE 25

Analysis of Real DTN Traces

  • Haggle Project (people, conference, Imote)
  • M IT Reality Project (campus, phone)
  • UM ass Diesel-Net Project (bus meetings)
  • RollerNet Traces (roller skate tour, Imote)
  • Others:

– Zebra, taxi etc.

  • Extracted information:

– Pair-wise and aggregate inter-meeting and contact

duration

2/ 4/ 2011 Bulut: PhD Defense (RPI) 25

Inter-meeting M eeting (contact) duration

slide-26
SLIDE 26

Single-copy DTN Routing

  • Shortest-path based routing

– DTN Graph M odel:

  • Vertices are nodes, edges are links between nodes
  • Weights of edges are average inter-meeting times

– Ex: M ED, M EED etc.

  • M etric-based (utility) routing

– When two nodes meet, one forwards its message to other if the

  • ther’s metric suggests more delivery chance with destination

– Ex: Prophet, Fresh, SimBet etc.

2/ 4/ 2011 Bulut: PhD Defense (RPI) 26

C D F E G I A B H

slide-27
SLIDE 27

M otivation

  • Conclusions from Analysis of Real-traces (human-based):

– Pair-wise intermeeting times follow log-normal distribution

  • NOT memory-less (as opposed to exp. dist.)

– Non-deterministic but cyclic mobility is frequent

  • Periodic meetings of same nodes
  • Conclusion from Current M etric-based Algorithms

– Forwarding decision based on only individual relations of nodes with

destination

  • “M eeting with each other” is not used.
  • M eetings of nodes are assumed independent (uncorrelated) from each
  • ther
  • BUT node meetings may be correlated

– Ex: Family (home), security guard (gate), office friends (work)

2/ 4/ 2011 Bulut: PhD Defense (RPI) 27

slide-28
SLIDE 28

Proposed M etric

  • Conditional Intermeeting Time:

– τA(C| B) = Average time it takes for node A to meet node C

after the time node A meets B (condition).

  • Can be computed from contact history
  • τA(C| C) standard intermeeting time

– If the meetings of a node with other nodes are correlated

(if the identity of B matters):

  • Residual time to a node’s next meeting with other nodes can be

predicted more accurately.

2/ 4/ 2011 Bulut: PhD Defense (RPI) 28

slide-29
SLIDE 29

2/ 4/ 2011 Bulut: PhD Defense (RPI) 29 29

Statistics from Real Traces ( )

29

M eetings of a node with other nodes are correlated How can we use this for efficient single copy based routing?

slide-30
SLIDE 30

M odification in Shortest-path based routing

  • Conditional Shortest Path Routing (CSPR)

– Finds path with minimum CSP and sends message over that path:

2/ 4/ 2011 Bulut: PhD Defense (RPI) 30

First hop weight Conditional intermeeting time

Updated DTN Graph M odel:

Node A that has last met node B

slide-31
SLIDE 31

M odification in M etric based Routing Algorithms

  • Current Design:

– When node A meets node B, A forwards its message

(which is destined to D) to B:

  • In Prophet: If A’s delivery probability to D is smaller than B’s

delivery probability.

  • In Fresh: If B has a more recent meeting with D than A.
  • M odification:
  • In C-Prophet and C-Fresh, we add one more condition for

forwarding:

2/ 4/ 2011 Bulut: PhD Defense (RPI) 31

slide-32
SLIDE 32

Simulations

  • Data Sets:

– Real DTN Traces

  • Cambridge Traces
  • RollerNet Traces
  • Haggle Project Traces

– Synthetic Traces based on Community M odels

  • Algorithms in Comparison:

– Group 1: Shortest path based routing

  • 1) SPR (M EED, M ED) 2) CSPR

– Group 2: M etric based routing

  • 1) Fresh 2) Prophet 3) C-Fresh 4) C-Prophet
  • Epidemic (Optimal delivery ratio, delay)

2/ 4/ 2011 Bulut: PhD Defense (RPI) 32

slide-33
SLIDE 33

Simulation Results (Group 1)

  • Higher delivery ratio, lower delay

2/ 4/ 2011 Bulut: PhD Defense (RPI) 33

slide-34
SLIDE 34

Simulation Results (Group 1)

  • Close Average Cost:

– Number of forwardings per message

  • Better Routing Efficiency (10%-23% improvement):

– Delivery Ratio/Average Cost

2/ 4/ 2011 Bulut: PhD Defense (RPI) 34

slide-35
SLIDE 35

Simulation Results (Group 2)

  • M odified versions are better in all performance

metrics:

2/ 4/ 2011 Bulut: PhD Defense (RPI) 35 35

RollerNet Traces

35

Synthetic Traces

slide-36
SLIDE 36

Simulation Results (Group 2)

  • M odified versions provide smaller cost and better routing efficiency

but same delivery ratio (due to not so clear repetitive meetings between nodes):

2/ 4/ 2011 Bulut: PhD Defense (RPI) 36

Cambridge Traces Haggle Traces

slide-37
SLIDE 37

Outline

  • Introduction to DTNs

– Challenges of routing

  • Proposed Algorithms

1) M ulti-period Spray and Wait routing 2) M ulti-period erasure coding based routing 3) Efficient single-copy routing utilizing correlation between node meetings 4) Social relation based routing

  • Summary of Contributions

2/ 4/ 2011 Bulut: PhD Defense (RPI) 37

slide-38
SLIDE 38

Social relation based routing

  • Social network metrics are used to understand contact

relations between nodes

  • Recent work:

– Similarity and betweenness: SimBet [M obihoc 07] – Community Detection: BubbleRap [M obiHoc 08], LocalCom [Secon 09]

  • Deficiencies:

– Lack of a metric to detect the node relations (also the opportunities

for message exchanges) accurately:

– Improper handling of indirect relations – Improper handling of temporal variations in node relations

2/ 4/ 2011 Bulut: PhD Defense (RPI) 38

slide-39
SLIDE 39

Analysis of “ M essage Exchange Opportunity” using Current M etrics

  • Average Contact frequency

– Can not differentiate cases a & b

  • Average Contact duration

– Can not differentiate cases b & c

  • Average Separation Period

– Can differentiate cases c & d utilizing variance (irregularity) between

different separation periods

– Can not differentiate cases b & e (also cases b & f)

2/ 4/ 2011 Bulut: PhD Defense (RPI) 39

slide-40
SLIDE 40

Proposed M etric

  • Social Pressure M etric (SPM )

– M easures “ how actual is the knowledge of a person

about his/ her friend?”

– What would be the average message forwarding delay to

node j if node i has a new message for node j at each time unit?

2/ 4/ 2011 Bulut: PhD Defense (RPI) 40

Ex: SPM = [(t 1)2 +(t 2)2 +(t 3)2]/ 2T

Node i’s delivery metric (weight) to j

slide-41
SLIDE 41

Indirect Relations

  • Relative-SPM (RSPM )

– What would be the average delivery delay of node i’s continuously

generated messages (for k) if they follow the path <i,j,k>?

  • Better than (SPM i,j + SPM j,k) due to a possible correlation between node j’s

meetings with i and k

2/ 4/ 2011 Bulut: PhD Defense (RPI) 41

Node i’s indirect delivery metric (weight) to k through j

slide-42
SLIDE 42

Friendship Community Formation

  • Each node i:

– computes its SPM i,j values with each j

– receives its RSPM i,j,k from j periodically

  • Node i forms its friendship community as the nodes having

direct or indirect weight larger than a threshold (τ):

2/ 4/ 2011 Bulut: PhD Defense (RPI) 42

slide-43
SLIDE 43

T emporal Variations in Node Relations

  • Aging of weights may cause wrong decisions
  • Proposed Solution: Generating different communities at different

times of the day –

Ex: 3 hour ranges

2/ 4/ 2011 Bulut: PhD Defense (RPI) 43

node 28 meets node 38 between 9am-7pm

slide-44
SLIDE 44

Forwarding Algorithm

  • When two nodes meet, message is forwarded

if the following case occurs:

2/ 4/ 2011 Bulut: PhD Defense (RPI) 44

A

M essage for D

B Y es Y es

Is D in your friendship community? Is your friendship with D better than mine? Then, I’ll forward the message to you

slide-45
SLIDE 45

Simulation Results (M IT traces)

  • Higher delivery ratio, lower delay
  • Lowest cost, best routing efficiency (delivery

rate/cost)

2/ 4/ 2011 Bulut: PhD Defense (RPI) 45

[Globecom 2010], [TPDS Journal in submission]

slide-46
SLIDE 46

2/ 4/ 2011 Bulut: PhD Defense (RPI) 46

Simulation Results (Haggle Traces) Simulation Results (S ynthetic Traces)

slide-47
SLIDE 47

Outline

  • Introduction to DTNs

– Challenges of routing

  • Proposed Algorithms

1) M ulti-period Spray and Wait routing 2) M ulti-period erasure coding based routing 3) Efficient single-copy routing utilizing correlation between node meetings 4) Social relation based routing

  • Summary of Contributions

2/ 4/ 2011 Bulut: PhD Defense (RPI) 47

slide-48
SLIDE 48

Summary of Contributions

  • M ulti-period spray and wait algorithm

– Fewer average copies while keeping the same delivery rate by the

deadline

  • Cost efficient routing with erasure coded messages

– Single and multi-period (reliability)

  • Efficient single-copy routing

– Conditional shortest path routing – M odifying metric based algorithms using conditional intermeeting

time (correlated node mobility)

  • Social relation based routing

– Friendship based routing – (Relative) Social-pressure metric

2/ 4/ 2011 Bulut: PhD Defense (RPI) 48

slide-49
SLIDE 49

Acknowledgment

  • M y advisor:

– Prof. Boleslaw Szymanski

  • M y colleagues:

– Zijian Wang Sahin Cem Geyik

  • Grants and Research Centers:

– –

2/ 4/ 2011 Bulut: PhD Defense (RPI) 49

slide-50
SLIDE 50

THANK YOU

2/ 4/ 2011 Bulut: PhD Defense (RPI) 50