32
Addressing Burstiness for Reliable Communication and Latency Bound Generation in Wireless Sensor Networks Department of Computer Science University of Virginia Sirajum Munir, Shan Lin, Enamul Hoque, S. M. Shahriar Nirjon, John A. Stankovic, and Kamin Whitehouse

Department of Computer Science University of Virginia

  • Upload
    celina

  • View
    24

  • Download
    0

Embed Size (px)

DESCRIPTION

Addressing Burstiness for Reliable Communication and Latency Bound Generation in Wireless Sensor Networks. Department of Computer Science University of Virginia. Sirajum Munir , Shan Lin, Enamul Hoque, S. M. Shahriar Nirjon, John A. Stankovic, and Kamin Whitehouse. Problem Definition. - PowerPoint PPT Presentation

Citation preview

Page 1: Department of Computer Science University  of Virginia

Addressing Burstiness for Reliable Communication and Latency Bound

Generation in Wireless Sensor Networks

Department of Computer ScienceUniversity of Virginia

Sirajum Munir, Shan Lin, Enamul Hoque, S. M. Shahriar Nirjon, John A. Stankovic, and Kamin Whitehouse

Page 2: Department of Computer Science University  of Virginia

Problem Definition

• Reliable delivery– Retry on each link until the

packet is receivedN1 N2 N3

N4 N5N6 N7

N8 N9N10

Source

Destination

23 3

5

Page 3: Department of Computer Science University  of Virginia

Problem Definition

• Reliable delivery– Retry on each link until the

packet is received

• Problem:– Unbounded E2E latency– Not acceptable for real-time

applications

N1 N2 N3

N4 N5N6 N7

N8 N9N10

Source

Destination

Page 4: Department of Computer Science University  of Virginia

Overview

• Basic Approach– Estimate maximum “burst”

length for each link• #consecutive failures

N1 N2 N3

N4 N5N6 N7

N8 N9N10

Source

23

3 200

335

3

5

100

5

4 562

2

Destination

Page 5: Department of Computer Science University  of Virginia

Overview

• Basic Approach– Estimate maximum “burst”

length for each link

– Key insight:• Burstiness caused by physical

world dynamics• Some links are relatively

insulated from these dynamics

N1 N2 N3

N4 N5N6 N7

N8 N9N10

Source

23

3 200

335

3

5

100

5

4 562

2

Destination

Page 6: Department of Computer Science University  of Virginia

Overview

• Basic Approach– Estimate maximum “burst”

length for each link– Choose routes that only use

non-bursty links

N1 N2 N3

N4 N5N6 N7

N8 N9N10

Source

23

3 200

335

3

5

100

5

4 562

2

Destination

Page 7: Department of Computer Science University  of Virginia

Overview

• Basic Approach– Estimate maximum “burst”

length for each link– Choose routes that only use

non-bursty links– Schedule packet transmission

to avoid interference between links

N1 N2 N3

N4 N5N6 N7

N8 N9N10

Source1 Destination2

Source2 Destination1

Page 8: Department of Computer Science University  of Virginia

Outline

• Modeling link burstiness• E2E latency bounds• Evaluation

Page 9: Department of Computer Science University  of Virginia

Modeling Burstiness

• Modeling link Bursts– Bmax, B’min per link

• Bmax = Maximum No. of time slots where transmission can fail

• B’min = Minimum No. of time slots available for transmission

• W = Bmax + B’min

– Example• B’min = 1• W = 2

1 0 0 1 1 0 1 …

Page 10: Department of Computer Science University  of Virginia

Modeling Burstiness

• Modeling link bursts– Bmax, B’min per link

• Bmax = Maximum No. of time slots where transmission can fail

• B’min = Minimum No. of time slots available for transmission

• W = Bmax + B’min

– Example• B’min = 1• W = 3• Bmax = 2

1 0 0 1 1 0 1 …

Page 11: Department of Computer Science University  of Virginia

Modeling Burstiness

• Different from existing models– The β Factor [Srinivasan et al.]

• Models burstiness based on the distribution of burst lengths

– Our model only cares about the maximum burst length

X X X X X X X X X …

Page 12: Department of Computer Science University  of Virginia

Empirical Study

• 21 Days-long • Indoor testbed• 48 Tmote Sky nodes • 3.6 M packets/link• 200 packets/sec• Compute Bmax, B’min, PRR of every link

Page 13: Department of Computer Science University  of Virginia

Empirical Study

• Verified:– Some links are very bursty– Some links are not bursty

• Bmax is not predicted by PRR– Some highly reliable links

(PRR>0.99) still have very large bursts

Page 14: Department of Computer Science University  of Virginia

Outline

• Modeling link burstiness• E2E latency bounds• Evaluation

Page 15: Department of Computer Science University  of Virginia

E2E Latency Bound

• Min Latency Bound: NP-Hard• Greedy solution: Principles

– Routing : Least burst routing

N1 N2 N3

N4 N5N6 N7

N8 N9N10

Source

Destination

N1 N2 N3

N4 N5N6 N7

N8 N9N10

Source

Destination

Bmax

BmaxBmax Bmax

Bmax

Bmax

Bmax

Bmax

Bmax

Bmax

Bmax

BmaxBmax

Bmax

Page 16: Department of Computer Science University  of Virginia

E2E Latency Bound

• Greedy solution : Principles– Routing: Least burst route– Schedule packet transmission

• Allocating time slots:– How many time slots to allocate per link?

» Allocate Bmaxi+1 contiguous time slots, for i-th link

– Can we do even better?» Yes ! Overlap some streams’ time slot allocation

N1 N2 N3

N4 N5N6 N7

N8 N9N10

Source

Destination

Page 17: Department of Computer Science University  of Virginia

E2E Latency Bound

• What is overlapping?– Assume Link L(1,2) has Bmax=2, B’min=4

– 2 Streams: S1, S2

• Why do we need overlapping ?– W/O overlapping: Avg LB = (3 + 6)/2 = 4.5– W/ overlapping: Avg LB = (3 + 4) /2 = 3.5

N1 N2

1 2 3 4 5 6

S1 S1 S1

S2 S2 S2

Schedule w/o overlapping

1 2 3 4

S1 S1 S1

S2 S2 S2

Schedule w/ overlapping

Prioritizing rule

Time Slots

Page 18: Department of Computer Science University  of Virginia

E2E Latency Bound

• How much to overlap?– Assume Link L(1,2) has Bmax=2, B’min=4

– 2 Streams: S1, S2

– Overlap at most B’min number of streams

N1 N2

1 2 3

S1 S1 S1

S2 S2 S2

Complete overlapping: Doesn’t work !

1 2 3 4 5 6

S1 S1 S1

S2 S2 S2

S3 S3 S3

S4 S4 S4

Time Slots

Time Slots

Page 19: Department of Computer Science University  of Virginia

E2E Latency Bound

• How much to overlap?– Assume Link L(1,2) has Bmax=2, B’min=4

– 2 Streams: S1, S2

– Overlap at most B’min number of streams

N1 N2

1 2 3

S1 S1 S1

S2 S2 S2

Complete overlapping: Doesn’t work !

1 2 3 4 5 6

S1 S1 S1

S2 S2 S2

S3 S3 S3

S4 S4 S4

Time Slots

Time Slots

Page 20: Department of Computer Science University  of Virginia

E2E Latency Bound

• How much to overlap?– Assume Link L(1,2) has Bmax=2, B’min=4

– 2 Streams: S1, S2

– Overlap at most B’min number of streams

N1 N2

1 2 3

S1 S1 S1

S2 S2 S2

Complete overlapping: Doesn’t work !

1 2 3 4 5 6

S1 S1 S1

S2 S2 S2

S3 S3 S3

S4 S4 S4

Time Slots

Time Slots

Page 21: Department of Computer Science University  of Virginia

E2E Latency Bound Summary

• Greedy solution : Principles– Routing: Least burst routing– Allocating time slots:

• How many time slots to allocate per link?– Bmaxi+1 contiguous time slots

– Without complete overlapping

• How much to overlap?– Overlap at most B’mini streams’ time slot allocation

• How to handle interference?– Use IM to avoid interference

Page 22: Department of Computer Science University  of Virginia

Outline

• Modeling link burstiness• E2E latency bounds• Evaluation

Page 23: Department of Computer Science University  of Virginia

Evaluation

• Experimental Setup– Same testbed as empirical study

• 48 Tmote Sky nodes

– Same packet transmission rate• 200 packets/sec

– RBS style time-synchronization

Page 24: Department of Computer Science University  of Virginia

• Effect of Bmax– B’min = 1– Multiplying factor: K

• Allocate Bmaxi*K + 1 time slots, for i-th link

– As K increases• Average LB increase linearly • E2E DMR becomes 0

at K = 0.6 !• Allocate Bmaxi*0.6 + 1

time slots -> save 12.4% latency

– K allows us to control E2E DMR and LB !

• Avg. LB increases linearly

Evaluation

12.4%

Page 25: Department of Computer Science University  of Virginia

Evaluation

• Effect of B’min– As B’min increases

• LB decreases• Then starts to increases again !

– Minimize average LB by an intelligent selection of B’min.

Page 26: Department of Computer Science University  of Virginia

Contributions

• New model of link burstiness– Estimates maximum consecutive packet loss – Not captured by β factor or PRR

• New scheduling algorithms for E2E latency bounds

• Empirical evaluation– 21 day link characterization– Testbed evaluation of LB miss ratio with 10

simultaneous streams

Page 27: Department of Computer Science University  of Virginia

Conclusions

• Can provide reasonable estimate of latency bounds– Not a guarantee– The “K” parameter helps control the trade-off between

miss ratio and latency

• One important step to combine wireless networking with real-time control

Page 28: Department of Computer Science University  of Virginia

Questions?

Page 29: Department of Computer Science University  of Virginia

Backup Slides

Page 30: Department of Computer Science University  of Virginia

One Final Issue…

• Change in burst behavior?– Packet Recovery

• Each node queues un-transmitted packets.• Transmits later if free slot available.

– Link Adaptation• Each node keeps a record when it fails to transmit• Sends this report to B.S. periodically• B.S. reschedules by doubling/ halving the allocated

time slots• LB expands/shrinks dynamically

Page 31: Department of Computer Science University  of Virginia

Stationarity

• Can we assume that Bmax is stationary?– Can classify links:

• Bursty links had highly variable Bmax• Non-bursty links were more consistent

– Why? Due to physical dynamics

– Must ensure that measurement period captures all physical dynamics

• No stronger requirement than any model

Page 32: Department of Computer Science University  of Virginia

IM

• Characterizing Interference:– Define an Interference Matrix, IM

– Measurement based on PRR

otherwise 0,

range ceinterferen in are Lj link and Li linkif 1,j)IM(i,

Ni1 Ni2

Nj1 Nj2

Li

Lj

L1 L3