Auto-configuration of 802.11n WLANs Mustafa Y. Arslan - Konstantinos - - PowerPoint PPT Presentation

auto configuration of 802 11n wlans
SMART_READER_LITE
LIVE PREVIEW

Auto-configuration of 802.11n WLANs Mustafa Y. Arslan - Konstantinos - - PowerPoint PPT Presentation

Auto-configuration of 802.11n WLANs Mustafa Y. Arslan - Konstantinos Pelechrinis - Ioannis Broustis UC Riverside University of Pittsburgh UC Riverside Srikanth Krishnamurthy - Sateesh Addepalli - Konstantina Papagiannaki UC


slide-1
SLIDE 1

Auto-configuration of 802.11n WLANs

Mustafa Y. Arslan - Konstantinos Pelechrinis - Ioannis Broustis UC Riverside University of Pittsburgh UC Riverside Srikanth Krishnamurthy - Sateesh Addepalli - Konstantina Papagiannaki UC Riverside Cisco Inc. Intel Labs, Pittsburgh ACM CoNEXT 2010

slide-2
SLIDE 2

Channel Bonding (CB)

  • Goal of CB is to combine two adjacent 20 MHz channels to double the

bandwidth (raw transmission rate)

slide-3
SLIDE 3

Channel Bonding (CB)

  • Goal of CB is to combine two adjacent 20 MHz channels to double the

bandwidth (raw transmission rate)

Spectral Mask Channel 1 (20 MHz) Channel 2 (20 MHz)

slide-4
SLIDE 4

Channel Bonding (CB)

  • Goal of CB is to combine two adjacent 20 MHz channels to double the

bandwidth (raw transmission rate)

Spectral Mask Channel 1 (20 MHz) Channel 2 (20 MHz)

Spectral Mask Bonded Channel (40 MHz)

slide-5
SLIDE 5

Channel Bonding (CB)

  • Goal of CB is to combine two adjacent 20 MHz channels to double the

bandwidth (raw transmission rate)

Spectral Mask Channel 1 (20 MHz) Channel 2 (20 MHz)

Spectral Mask Bonded Channel (40 MHz)

  • Fact: CB also increases interference

✓ Pelechrinis et. al, Shrivastava et. al.

slide-6
SLIDE 6

Channel Bonding (CB)

  • Goal of CB is to combine two adjacent 20 MHz channels to double the

bandwidth (raw transmission rate)

Spectral Mask Channel 1 (20 MHz) Channel 2 (20 MHz)

Spectral Mask Bonded Channel (40 MHz)

  • Fact: CB also increases interference

✓ Pelechrinis et. al, Shrivastava et. al.

  • Public belief: CB always gives throughput benefits in isolation
slide-7
SLIDE 7

Contributions

  • Public belief: CB always gives throughput benefits.
slide-8
SLIDE 8

Contributions

  • Public belief: CB always gives throughput benefits.
slide-9
SLIDE 9

Contributions

  • CB, when blindly applied, hurts throughput!

✓ Extensive measurements with WARP and off-the-shelf 802.11n ✓ PHY and MAC observations

  • User association + frequency selection
  • Public belief: CB always gives throughput benefits.
slide-10
SLIDE 10

Contributions

  • CB, when blindly applied, hurts throughput!

✓ Extensive measurements with WARP and off-the-shelf 802.11n ✓ PHY and MAC observations

  • User association + frequency selection
  • Auto-COnfiguRation of 802.11N WLANs

✓ First system custom built for 802.11n ✓ 1.5x - 6x throughput gain per AP

  • Public belief: CB always gives throughput benefits.
slide-11
SLIDE 11

Roadmap

  • CB - why and when does it fail?

✓ Effect on the PHY ✓ MAC and application layer observations

  • Designing ACORN

✓ User association, channel selection

  • Evaluation
slide-12
SLIDE 12

CB at the PHY

slide-13
SLIDE 13

CB at the PHY

  • 20 MHz vs 40 MHz (twice OFDM subcarriers in a symbol with CB)
slide-14
SLIDE 14

CB at the PHY

  • 20 MHz vs 40 MHz (twice OFDM subcarriers in a symbol with CB)
slide-15
SLIDE 15

CB at the PHY

  • 20 MHz vs 40 MHz (twice OFDM subcarriers in a symbol with CB)
  • Thermal Noise

✓ N (dBm) = -174 + 10log(B) ✓ 3 dB higher (twice) noise - noise per subcarrier is the same

slide-16
SLIDE 16

CB at the PHY

  • 20 MHz vs 40 MHz (twice OFDM subcarriers in a symbol with CB)
  • Thermal Noise

✓ N (dBm) = -174 + 10log(B) ✓ 3 dB higher (twice) noise - noise per subcarrier is the same

  • Subcarrier energy

✓ For a given TX power, energy per subcarrier is halved (3 dB loss)

slide-17
SLIDE 17

CB at the PHY

  • 20 MHz vs 40 MHz (twice OFDM subcarriers in a symbol with CB)
  • Thermal Noise

✓ N (dBm) = -174 + 10log(B) ✓ 3 dB higher (twice) noise - noise per subcarrier is the same

  • Subcarrier energy

✓ For a given TX power, energy per subcarrier is halved (3 dB loss)

  • SNR per subcarrier is 3 dB less with CB
slide-18
SLIDE 18

CB at the PHY

150

  • 140
  • 130
  • 120
  • 110
  • 100
  • 95
  • 92
  • 80
  • 20
  • 10

Fc1020

Power / frequency (dB / Hz) Frequency (MHz) 20 MHz 40 MHz

slide-19
SLIDE 19

CB at the PHY

a) without CB b) with CB

slide-20
SLIDE 20

CB at the PHY

a) without CB b) with CB

slide-21
SLIDE 21

CB at the PHY

a) without CB b) with CB

slide-22
SLIDE 22

CB at the PHY

a) without CB b) with CB

slide-23
SLIDE 23

CB at the PHY

CB increases baud error rate increase in BER

a) without CB b) with CB

slide-24
SLIDE 24

CB at the PHY

1e-07 1e-06 1e-05 0.0001 0.001 0.01 0.1 3 6 9 12 Bit Error Ratio SNR (dB) BER-20Mhz BER-40Mhz Theory 1e-07 1e-06 1e-05 0.0001 0.001 0.01 0.1 5 10 15 20 25 Bit Error Ratio Transmit Power [0:63] BER-20Mhz BER-40Mhz

slide-25
SLIDE 25

CB at the PHY

  • For a given TX power, BER is higher when CB is employed

1e-07 1e-06 1e-05 0.0001 0.001 0.01 0.1 3 6 9 12 Bit Error Ratio SNR (dB) BER-20Mhz BER-40Mhz Theory 1e-07 1e-06 1e-05 0.0001 0.001 0.01 0.1 5 10 15 20 25 Bit Error Ratio Transmit Power [0:63] BER-20Mhz BER-40Mhz

slide-26
SLIDE 26

Roadmap

  • CB - why / when does it fail?

✓ Effect on the PHY ✓ MAC and application layer observations

  • Designing ACORN

✓ User association, channel selection

  • Evaluation
slide-27
SLIDE 27

CB at the MAC

  • PHY observations with CB may not be exported to MAC

✓ Coding (FEC) ✓ What is the impact on PDR?

  • Throughput (T) = Rate (R) * PDR

✓ T20 = R20 * PDR20 ✓ T40 = R40 * PDR40 = 2 * R20 * PDR40

slide-28
SLIDE 28

CB at the MAC

  • PHY observations with CB may not be exported to MAC

✓ Coding (FEC) ✓ What is the impact on PDR?

  • Throughput (T) = Rate (R) * PDR

✓ T20 = R20 * PDR20 ✓ T40 = R40 * PDR40 = 2 * R20 * PDR40

slide-29
SLIDE 29

CB at the MAC

  • PHY observations with CB may not be exported to MAC

✓ Coding (FEC) ✓ What is the impact on PDR?

  • Throughput (T) = Rate (R) * PDR

✓ T20 = R20 * PDR20 ✓ T40 = R40 * PDR40 = 2 * R20 * PDR40

σ = PDR20 PDR40

slide-30
SLIDE 30

CB at the MAC

  • PHY observations with CB may not be exported to MAC

✓ Coding (FEC) ✓ What is the impact on PDR?

  • Throughput (T) = Rate (R) * PDR

✓ T20 = R20 * PDR20 ✓ T40 = R40 * PDR40 = 2 * R20 * PDR40

  • T20 > T40 if

σ > 2

σ = PDR20 PDR40

slide-31
SLIDE 31

CB at the MAC

  • T20 > T40 if

σ > 2

σ = PDR20 PDR40

2 - 3 dB of critical region

slide-32
SLIDE 32

CB at the end-user

10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80 Throughput-40Mhz (Mbits/s) Throughput-20Mhz (Mbits/s) UDP TCP

slide-33
SLIDE 33

CB at the end-user

10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80 Throughput-40Mhz (Mbits/s) Throughput-20Mhz (Mbits/s) UDP TCP

slide-34
SLIDE 34

CB at the end-user

CB hurts for poor links!

10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80 Throughput-40Mhz (Mbits/s) Throughput-20Mhz (Mbits/s) UDP TCP

slide-35
SLIDE 35

Summary

  • CB does not always benefit

✓ SNR decrease ✓ Increased BER ✓ Increased PER

  • Culprit for poor links
slide-36
SLIDE 36

Roadmap

  • CB - why / when does it fail?

✓ Effect on the PHY ✓ MAC and application layer observations

  • Designing ACORN

✓ User association, channel selection

  • Evaluation
slide-37
SLIDE 37

ACORN

  • User association

✓ Group similar quality clients in a cell

AP Poor Client Good Client

slide-38
SLIDE 38

ACORN

  • User association

✓ Group similar quality clients in a cell

AP Poor Client Good Client

slide-39
SLIDE 39

ACORN

  • User association

✓ Group similar quality clients in a cell

AP Poor Client Good Client

slide-40
SLIDE 40

ACORN

  • User association

✓ Group similar quality clients in a cell

AP Poor Client Good Client

slide-41
SLIDE 41

ACORN

  • User association

✓ Group similar quality clients in a cell

AP Poor Client Good Client

slide-42
SLIDE 42

ACORN

  • User association

✓ Group similar quality clients in a cell

CB CB AP Poor Client Good Client

slide-43
SLIDE 43

User Association

: aggregate transmission delay of AP i : channel access time of AP i ( = 1 with no contention, saturated traffic) : long term per-client throughput of AP i : number of clients of AP i (including u)

Mi ATDi

Mi

ATDi

Ki

slide-44
SLIDE 44

User Association

: aggregate transmission delay of AP i : channel access time of AP i ( = 1 with no contention, saturated traffic) : long term per-client throughput of AP i : number of clients of AP i (including u)

max.

Mi ATDi

Mi

ATDi

Ki

slide-45
SLIDE 45

User Association

: aggregate transmission delay of AP i : channel access time of AP i ( = 1 with no contention, saturated traffic) : long term per-client throughput of AP i : number of clients of AP i (including u)

max.

Mi ATDi

Mi

ATDi

Ki

aggregate throughput of AP i

slide-46
SLIDE 46

User Association

: aggregate transmission delay of AP i : channel access time of AP i ( = 1 with no contention, saturated traffic) : long term per-client throughput of AP i : number of clients of AP i (including u)

max.

Mi ATDi

Mi

ATDi

Ki

aggregate throughput of AP i aggregate throughput of other APs

slide-47
SLIDE 47

Channel Selection

slide-48
SLIDE 48

Channel Selection

The problem reduces to graph coloring and is NP-complete

  • In every iteration:

✓ AP with the max. increase in aggregate throughput picks a new

channel

  • When there is no improvement, terminate
slide-49
SLIDE 49

Channel Selection

20 MHz 40 MHz

slide-50
SLIDE 50

Channel Selection

20 MHz 40 MHz

  • 3 dB
slide-51
SLIDE 51

Channel Selection

20 MHz 40 MHz

  • 3 dB

+3 dB

slide-52
SLIDE 52

Channel Selection

20 MHz 40 MHz

  • 3 dB

+3 dB

BER

Theoretical BER

slide-53
SLIDE 53

Channel Selection

20 MHz 40 MHz

  • 3 dB

+3 dB

BER

Theoretical BER

PER

1 - (1 - BER)L

slide-54
SLIDE 54

Channel Selection

20 MHz 40 MHz

  • 3 dB

+3 dB

BER

Theoretical BER

PER

1 - (1 - BER)L

Set of Interferers

slide-55
SLIDE 55

Channel Selection

20 MHz 40 MHz

  • 3 dB

+3 dB

BER

Theoretical BER

PER

1 - (1 - BER)L

Set of Interferers

Scale down channel access ratio by (# Interferers + 1)

slide-56
SLIDE 56

Roadmap

  • CB - why / when does it fail?

✓ Effect on the PHY ✓ MAC and application layer observations

  • Designing ACORN

✓ User association, channel selection

  • Evaluation
slide-57
SLIDE 57

Evaluation

  • 18 node 802.11n testbed - Ralink chipset
  • Comparison with a legacy auto-configuration system

✓ Kauffmann et. al. - Infocom’07

  • Legacy user association

✓ Minimize total ATD of all users

  • Legacy channel selection

✓ Minimize total interference between APs ✓ Modified to aggressively pick 40 MHz channels

slide-58
SLIDE 58

Evaluation

AP2 AP1 AP5 AP4 AP3

Pictorial representation of actual testbed deployment

slide-59
SLIDE 59

Evaluation

AP2 AP1 AP5 AP4 AP3

Pictorial representation of actual testbed deployment

slide-60
SLIDE 60

Evaluation

AP2 AP1 AP5 AP4 AP3

Pictorial representation of actual testbed deployment

5 10 15 20 AP4 AP5 Throughput (Mbps)

Legacy ACORN

slide-61
SLIDE 61

Evaluation

AP2 AP1 AP5 AP4 AP3

slide-62
SLIDE 62

Evaluation

AP2 AP1 AP5 AP4 AP3

slide-63
SLIDE 63

Evaluation

AP2 AP1 AP5 AP4 AP3

Mid-quality client group - AP3 serves one good client

slide-64
SLIDE 64

Evaluation

AP2 AP1 AP5 AP4 AP3 15 30 45 60 AP1 AP3 Throughput (Mbps)

Legacy ACORN

slide-65
SLIDE 65

Evaluation

AP2 AP1 AP5 AP4 AP3

With ACORN, higher congestion at AP1 Aggregate throughput does not change!

15 30 45 60 AP1 AP3 Throughput (Mbps)

Legacy ACORN

slide-66
SLIDE 66

Conclusion

  • CB can hurt throughput even in isolation

✓ User association becomes critical

  • CB increases interference

✓ Addressing channel selection

  • ACORN performs both functions in tandem

✓ Trade off fairness for aggregate throughput

  • Implementation on a testbed and evaluations show:

✓ ACORN outperforms legacy approaches agnostic to CB

slide-67
SLIDE 67

THANK YOU!