Predictable 802.11 Packet Delivery from Wireless Channel...

Predictable 802.11 Packet Delivery fromWireless Channel Measurements

Daniel HalperinWenjun Hu, Anmol Sheth, David Wetherall

Daniel Halperin, SIGCOMM 2010, [email protected]

802.11 Wi-Fi technology

• Fast - 600 Mbps in 802.11n represents a 300x speedup in 12 years

• Reliable - vehicular speeds, extended range, stable hardware and software

• Ubiquitous - few dollars per chip allows integration everywhere

2


New, exciting apps on the horizon

802.11 Wi-Fi technology

• Fast - 600 Mbps in 802.11n represents a 300x speedup in 12 years

• Reliable - vehicular speeds, extended range, stable hardware and software

• Ubiquitous - few dollars per chip allows integration everywhere

2


New apps stress network

3

WirelessDisplay

WirelessInput

MobileWireless



3

All-wirelessHome



3

All-wirelessHome

Performance really matters


Performance – in theory

39Mbps

4

Channel Measurements

Textbook Algorithms

Rate Selection


Performance – in theory

39Mbps

4


Textbook Algorithms

Rate Selection

In practice, this has never worked!


Performance – In practice

65 Mbps?Nope!

65 Mbps?Nope!

65 Mbps?Nope!

52 Mbps?Nope!

13 Mbps?Okay!

5

Statistics-based Adaptation

Problem:Convergence time

• Dynamic environments

• Large search spaces– >300 tx configs in 802.11n– Combined rate & power

Both are trends


Goals: BridgingTheory and Practice

• Accurately predict performance over real channels

• Agile response to changing channels

• Leverage measurements available in real NICs

• Extend to 802.11n and more applications

6

Key: an accurate channel metric


Today’s talk

• Why it’s hard to predict performance withRF measurements today

• Our solution: an accurate channel metric using Effective SNR

• Evaluation of Effective SNR in Wi-Fi Networks


SNR based on RSSI

• Received Signal Strength Indicator– Measures total power received in packet– With Noise, gives SNR for packet

• Treated as if directly reflects performanceE.g., NIC manufacturers list per-rate ‘sensitivity’

0 5 10 15 20 25 300

20

40

60

80

100

Packet−level SNR (dB)

PRR

6.51319.526395258.565

9

• In practice, SNR at which a rate starts to work can vary more than 10 dB for real links


802.11: OFDM and MIMOOrthogonal FrequencyDivision Multiplexing

Multiple-InputMultiple-Output

Frequency-selective fading Spatial diversity

10

Power

Frequency


802.11: OFDM and MIMOOrthogonal FrequencyDivision Multiplexing

Multiple-InputMultiple-Output

Frequency-selective fading Spatial diversity

10

Key: Different subchannelshave different SNRs

Power

Frequency


5

15

25

35

-28 -14 0 14 28

SNR

(dB)

Subcarrier index

Packet SNR for 4 faded links

11


5

15

25

35

-28 -14 0 14 28

SNR

(dB)

Subcarrier index


11

30 dB


5

15

25

35

-28 -14 0 14 28

SNR

(dB)

Subcarrier index


11

30 dB

17 dB


Packet SNR

5

15

25

35

-28 -14 0 14 28

SNR

(dB)

Subcarrier index


11

30 dB

17 dB


Packet SNR

5

15

25

35

-28 -14 0 14 28

SNR

(dB)

Subcarrier index


11

30 dB

17 dB

Errors


Packet SNR

5

15

25

35

-28 -14 0 14 28

SNR

(dB)

Subcarrier index


11

30 dB

17 dB

Errors

Fundamental SNR mismatch


An 802.11n opportunity

• 802.11n provides detailed channel measurementsUsed for advanced MIMO techniques

• Channel State Information (CSI) measuresMIMO and OFDM!– Matrix captures per-antenna paths– One matrix per subcarrier

• Can we use it to predict packet delivery?In theory? In practice?

12


Today’s talk

• Why it’s hard to predict performance withRF measurements today

• Our solution: an accurate channel metric using Effective SNR



Effective SNR• Introduced by Nanda and Rege in 1998

• Packet SNR: total power in the link

• Effective SNR: useful power in the link

14

5

15

25

35

-28 -14 0 14 28

SNR

(dB)

Subcarrier index


Effective SNR• Introduced by Nanda and Rege in 1998

• Packet SNR: total power in the link

• Effective SNR: useful power in the link

14

5

15

25

35

-28 -14 0 14 28

SNR

(dB)

Subcarrier index

Effective SNR


Using Effective SNR

39Mbps

15


Textbook Algorithms

Rate Selection


Using Effective SNR

39Mbps

16

Textbook Algorithms

Rate Selection

Channel State Information(MIMO & OFDM)


Using Effective SNR

17

39Mbps

Rate Selection

Channel State Information(MIMO & OFDM) Effective

SNR Model


Using Effective SNR

18


SNR Model

Working Configurations;Application Decision

1x65 ✘1x52 ✘2x26 ✔3x13 ✔


Using Effective SNR

19


SNR Model


1x65 ✘1x52 ✘2x26 ✔3x13 ✔


• RX measures CSI from packet preambleNICs do this for MIMO/OFDM operation

• For every received frameMeasures all antennas + subcarriers used

Obtaining CSI

20

3-antenna Link

✹ ✹ ✹

✹ ✹ ✹

✹ ✹ ✹

3x3 Matrix

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

✹ ✹ ✹

✹ ✹ ✹

✹ ✹ ✹

One matrixper Subcarrier


Using Effective SNR

21


SNR Model


1x65 ✘1x52 ✘2x26 ✔3x13 ✔


Computing Effective SNR

• Single antenna link (1x1)CSI gives the per-symbol SNR

• Multiple RX antennas (1xN)Maximal-ratio combining

• MIMO link (MxN)Minimum mean-square error (MMSE)

A B

D E

A B

D E

A B

D E

A B

D E

A B

D E

A B

D E

A B

D E

A B

D E

✹ ✹

✹ ✹

22

Compute SNRs

per symbolCSI

SNRs


Computing Effective SNRA B

D E

A B

D E

A B

D E

A B

D E

A B

D E

A B

D E

A B

D E

A B

D E

✹ ✹

✹ ✹

23

ComputeBERs

per symbol

SNRs

BERs

CSI

Modulation BER(ρ)

BPSK

QPSK

QAM-16

QAM-64

Q��

2ρ�

Q (√ρ)

Q��

ρ/5�

Q��

ρ/21�

Textbookformulas


BEReff


D E

A B

D E

A B

D E

A B

D E

A B

D E

A B

D E

A B

D E

A B

D E

✹ ✹

✹ ✹

24

Average:Effective BER

SNRs

BERs

CSI



D E

A B

D E

A B

D E

A B

D E

A B

D E

A B

D E

A B

D E

A B

D E

✹ ✹

✹ ✹

25

SNRs

BERs

CSI

SNReffConvert back to SNR

BEReff


Using Effective SNR

26


SNR Model


1x65 ✘1x52 ✘2x26 ✔3x13 ✔


Predicting Packet Delivery

27

• Effective SNR thresholds for each rate

• Threshold per NIC implementation,not per NIC or per channel

• Adds flexibility to handle real NICs

• Hard vs soft decoding

• Other special techniquese.g., use optimal Maximum Likelihood receiveronly for small modulations


Example Applications

• Rate/MIMO/Channel width selection:What is the fastest configuration for this link?

28

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

✹ ✹ ✹

✹ ✹ ✹

✹ ✹ ✹

3x3, 40 MHz




28

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

✹ ✹ ✹

✹ ✹ ✹

✹ ✹ ✹

3x3, 40 MHz

1x3




28

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

✹ ✹ ✹

✹ ✹ ✹

✹ ✹ ✹

3x3, 40 MHz

2x3




28

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

✹ ✹ ✹

✹ ✹ ✹

✹ ✹ ✹

3x3, 40 MHz

3x3




28

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

✹ ✹ ✹

✹ ✹ ✹

✹ ✹ ✹

3x3, 40 MHz

20 MHz40 MHz



• Power Consumption:Which receive antenna is best to disable to save power?

29

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

✹ ✹ ✹

✹ ✹ ✹

✹ ✹ ✹

3x3, 40 MHz

RX AntennaSelection



• Spatial Reuse:What is the lowest transmit power at which I can support 100 Mbps bitrate?

30

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

A B C

D E F

G H I

✹ ✹ ✹

✹ ✹ ✹

✹ ✹ ✹

3x3, 40 MHz

Power ×


Today’s talk

• Why it’s hard to predict performance withRF measurements

• Our solution building a better metric using Effective SNR



Implemented in Intel NIC

• Intel Wi-Fi Link 5300 NIC (3x3, 450 Mbps)

• Two testbeds with > 200 widely varying links

• Linux (2.6.35-rc3) open source iwlwifi driver

• Firmware debug mode: send CSI to RX host

• Real-time computation: ~4 µs per 3x3 CSI32


Evaluation Questions

• Does Effective SNR accurately predict packet delivery?

• Does an Effective SNR rate selection algorithm perform well?

• More results in the paper

• Wireless link transition region• Transmit power control• Collisions

33


Predicting Optimal 3x3 Rate

34



34

0

13

26

52

65

0 10 20 30 40 50 60

Rat

e / s

tream

(Mbp

s)

SNR (dB)Packet SNR (dB)



35

0

13

26

52

65

0 10 20 30 40 50 60

Rat

e / s

tream

(Mbp

s)

SNR (dB)Effective SNR (dB)


Rate control evaluation

• 802.11a: Does Effective SNR match related work?ESNR versus SampleRate, SoftRate, OPT

• 802.11n: Does Effective SNR extend to 802.11n?ESNR versus OPT

• Channel simulation over mobile traceto compare against related work & vary speed

• MATLAB simulation + SoftRate GNU Radio

• Effective SNR algorithm gets corrupted CSI

36

Daniel Halperin, SIGCOMM 2010, [email protected] 37

Effective SNR for 802.11a

• Matches or beats 802.11a algorithms

• All within 15% of OPT

0

10

20

30

40

50

60

70

0 50 100 150 200 250

Avg

. deliv

ere

d r

ate

(M

bps)

Packet trace index (x100)

ESNRSampleRate

SoftRateSampleRate fixedSampleRate fixed retry

Daniel Halperin, SIGCOMM 2010, [email protected] 37

Effective SNR for 802.11a

• Matches or beats 802.11a algorithms

• All within 15% of OPT

0

10

20

30

40

50

60

70

0 50 100 150 200 250

Avg

. deliv

ere

d r

ate

(M

bps)


ESNRSampleRate

SoftRateSampleRate fixedSampleRate fixed retry

No rate fallback on retries:50% performance gap


ESNR extends to MIMO

• 80% accuracy, 10% overselection

• 24 rates vs 8, larger gap vs Previous-OPT38

0

50

100

150

200

0 50 100 150 200

Avg

. d

eliv

ere

d r

ate

(M

bp

s)


OPTPrevious-OPT

ESNR


Related work

39


Related work802.11a MIMO &

Ant Sel.TX

PowerChannelWidth

RealNICs

SoftRate (2009)

AccuRate (2010)

Error Estim. Codes (2010)Effective

SNR

✔

✔ ✔ ✔

✔ ✔

✔ ✔ ✔ ✔ ✔

39


Conclusions

• For the first time, we can usemeasurements available in real NICs topredict packet delivery over real channels

• Matches good performance of existing rate adaptation algorithms and extends to 802.11n

• Applies to a broad problem space and provides a simple, practical API for protocols

• Lots more in the paper!

40

Thanks! [email protected]

Predictable 802.11 Packet Delivery from Wireless Channel...

Documents

Transcript of Predictable 802.11 Packet Delivery from Wireless Channel...