1. 2 Enterprise WLAN setting 2 Vivek Shrivastava Wireless controller Access Point Clients Internet NSDI 2011.

1 1Vivek Shrivastava

PIE in the Sky : Online Passive Interference Estimation for Enterprise WLANs

WiNGS Labs

Vivek Shrivastava* Nokia Research Center, Palo Alto

NSDI 2011

Shravan Rayanchu, Suman Banerjee University of Wisconsin-Madison

Konstantina Papagiannaki Intel Labs, Pittsburgh

*[email protected]

2

Enterprise WLAN setting

2Vivek Shrivastava

Wireless controller

Access Point

Clients

Internet

NSDI 2011

3


3Vivek Shrivastava

Wireless controller

Access Point

Clients

Internet

Functionalities implemented at controller• Intrusion detection system• Interference management (channel assignment, power control)

NSDI 2011

4


4Vivek Shrivastava

Wireless controller

Access Point

Clients

Internet

Functionalities implemented at controller• Intrusion detection system• Interference management (channel assignment, power control)

NSDI 2011


Problems with wireless*

Support

“Flaky how ?”

“We will take a look.”

“The wireless is being flaky.”

“Well my connection dropped earlier and now it seems to be slow”

“Wait, now it seems fine.”User

NSDI 2011

*Slide borrowed from Cheng et. al (Jigsaw, Sigcomm ’06)

6 6

Vivek Shrivastava

Problems with wireless

NSDI 2011 6

Hidden terminals


Mismatch in data rates, slows down fast links


Rate anomaly

6 Mbps

54 Mbps

NSDI 2011

8 8

Vivek Shrivastava


NSDI 2011 8

Increasing client density and mobility ….

9 9

Vivek Shrivastava


NSDI 2011 9

Increasing client density and mobility….

10

10

Vivek Shrivastava


NSDI 2011 10

…. changing interference patterns

11

11

Vivek Shrivastava

Interference management in WLANs

NSDI 2011 11

12

12

Vivek Shrivastava


Estimate Interference dynamically

Manage Interference (data scheduling, transmit power control, channel

assignment)

NSDI 2011 12

13

13

Vivek Shrivastava



assignment)

NSDI 2011 13

Estimate Interference dynamically

14

14Vivek Shrivastava

How to estimate interference ?

Use bandwidth tests

NSDI 2011

15

15Vivek Shrivastava


Use bandwidth tests

Interferer

AP-Client pair

NSDI 2011

16

16Vivek Shrivastava


NSDI 2011

1) Measure AP-Client delivery in isolation

17

17Vivek Shrivastava


Isolation delivery= 0.95

NSDI 2011


18

18Vivek Shrivastava


NSDI 2011


2) Activate interferer and measure delivery

19

19Vivek Shrivastava


Interference delivery= 0.66

NSDI 2011



20

20Vivek Shrivastava




NSDI 2011

Link Interference Ratio (LIR) = del Interference / del isolation

21

21Vivek Shrivastava


NSDI 2011



Link Interference Ratio (LIR) = 0.66 / 0.95 = 0.69

22

22Vivek Shrivastava


NSDI 2011



LIR0 1

Strong Weak

23

But are bandwidth tests practical ?

•Can we use bandwidth tests in live settings

• Good accuracy –

• Network downtime required - X

• Not scalable (~ 1 hr for 20 AP-Client pair network) - X

• Not based on realistic rates and packet sizes – X

• Inefficient in dynamic scenario (client mobility) – X

23Vivek Shrivastava NSDI 2011

24

But are bandwidth tests practical ?

•Can we use bandwidth tests in live settings

• Good accuracy –

• Network downtime required - X

• Not scalable (~ 1 hr for 20 AP-Client pair network) - X

• Not based on realistic rates and packet sizes – X

• Inefficient in dynamic scenario (client mobility) – X


Can we estimate interference in a passive, real-time way ?

25

PIE Outline

•Motivation

•Conventional bandwidth tests not sufficient

•Passive Interference Estimation (PIE)

•Polling period of PIE

•Accuracy of PIE

•Realistic trace replay with PIE

•Applications of PIE

•Summary


26

26Vivek Shrivastava

Estimating interference passively

Sniffer

Sniffer

NSDI 2011

27

27Vivek Shrivastava


Sniffer

Sniffer

NSDI 2011

• Sniffer could be a dedicated wireless radio• Clocks synchronized using wired backplane

28

28Vivek Shrivastava


Sniffer reports

NSDI 2011

29

29Vivek Shrivastava


Sniffer reports

Timestamp, duration, rate, success..

NSDI 2011

30

30Vivek Shrivastava


Hidden terminals

NSDI 2011

31

31Vivek Shrivastava

Hidden terminals


NSDI 2011

32

32Vivek Shrivastava

Hidden terminals


NSDI 2011

33

33Vivek Shrivastava

1. Carrier sense

Hidden terminals


NSDI 2011

34

34Vivek Shrivastava

2. Channel free, transmit

Hidden terminals


1) Note timestamp, rate, duration

NSDI 2011

35

35Vivek Shrivastava

3. Collision !


1) Note timestamp, rate, duration

2) Note if transmission is a success (ack received ?)

NSDI 2011

http://images.google.com/imgres?imgurl=http://blogs.suntimes.com/ebert/1kapow.gif&imgrefurl=http://blogs.suntimes.com/ebert/1kapow.html&usg=__YfQGGlIAWY21t6vlLYPSwYsbbyk=&h=401&w=460&sz=21&hl=en&start=1&um=1&tbnid=Gdr-BKajmtLarM:&tbnh=112&tbnw=128&prev=/images?q=kapow&hl=en&rls=com.microsoft:en-us:IE-SearchBox&rlz=1I7ADBR_en&um=1

36

36Vivek Shrivastava

Estimating interference passivelyTimestamp: T0Rate: 6MbpsLength: 1400 bytesSuccess: False

Timestamp: T0 + δRate: 12MbpsLength: 600 bytesSuccess: False

NSDI 2011

37

37Vivek Shrivastava


T06Mbps

1400 bytesFalse

T0 + δ12Mbps

600 bytesFalse

Interference estimation

NSDI 2011

38

38Vivek Shrivastava


T06Mbps

1400 bytesFalse

T0 + δ 12Mbps

600 bytesFalse

NSDI 2011

39

39Vivek Shrivastava


T06Mbps

1400 bytesFalse

T0 + δ 12Mbps

600 bytesFalse

NSDI 2011

40

40Vivek Shrivastava


T06Mbps

1400 bytesFalse

T0 + δ12Mbps

600 bytesFalse

&

NSDI 2011

41

41Vivek Shrivastava


T06Mbps

1400 bytesFalse

T0 + δ12Mbps

600 bytesFalse

&

Red & Green clients

interfere

NSDI 2011

42


Sniffer reports

Infer interference

Scenarios

Reception

XX

XX

Vivek Shrivastava NSDI 2011

NSDI 201143


Sniffer reports

Infer interference

Scenarios

Reception

XX

XX

Red and Green packets overlaps

=> both lostVivek Shrivastava

44


Sniffer reports

Infer interference

Scenarios

Reception

XX

XX

No overlap, no problem !



Sniffer reports

Infer interference

Scenarios

Reception

XX

XX Both way

hidden terminals

Vivek Shrivastava 45NSDI 2011

46


Sniffer reports

Infer interference

Scenarios

Reception

X X


47


Sniffer reports

Infer interference

Scenarios

Reception

X X

Red and Green packets overlaps=> Green is lostVivek Shrivastava

48


Sniffer reports

Infer interference

Scenarios

Reception

X X One way hidden

terminalsVivek Shrivastava NSDI 2011

49

Computing interference measure in PIE

•Compute Isolation loss rate

•Fraction of non-overlapping packets lost

•Compute Interference loss rate

•Fraction of overlapping packets lost

•Interference measure (LIR):

(1 – Interference loss) / (1 – Isolation loss )


50

How quickly can PIE converge ?

•Time taken by PIE to converge depends on two key properties

• Periodicity with which sniffer reports are collected by the controller

•Traffic patterns for the links which dictate the number of interference events captured in a time interval


51

How quickly can PIE converge ?

•Time taken by PIE to converge depends on two key properties

• Periodicity with which sniffer reports are collected by the controller

•What is the minimum polling period ?

•Traffic patterns for the links which dictate the number of interference events captured in a time interval

•How much time does PIE take under realistic access patterns ?


52

PIE Outline

•Motivation




•Accuracy of PIE



•Summary


53

53Vivek Shrivastava

What is the minimum polling period ?

TimeI1 I2time interval

P P P

NSDI 2011

54

54Vivek Shrivastava


P P P

R(I1) R(I1)

NSDI 2011


55

55Vivek Shrivastava



P P P

LIR (I1) R(I1) R(I1)

NSDI 2011

56

56Vivek Shrivastava



P P P

R(I2) R(I2)

LIR (I1)

NSDI 2011

57

57Vivek Shrivastava



P P P

LIR (I2).

R(I2) R(I2)

LIR (I1)

NSDI 2011

58


58Vivek Shrivastava

LIR

Polling period (ms)Stability of interference

measure for saturated trafficNSDI 2011

59


59Vivek Shrivastava

Measure stabilizes after ~85 ms (at least 20 overlap samples)LI

R


measure per polling periodNSDI 2011

60


60Vivek Shrivastava

LIR


measure per polling periodNSDI 2011

We use a polling period of 100ms

61

PIE Outline

•Motivation




•Accuracy of PIE



•Summary


62

How accurate is PIE ?

62Vivek Shrivastava

% o

f lin

k-in

terf

ere

r p

air

s

Mean Error in LIR estimation

63


63Vivek Shrivastava

% o

f lin

k-in

terf

ere

r p

air

s


64


64Vivek Shrivastava

% o

f lin

k-in

terf

ere

r p

air

s


95% of link-interferer pairs, LIR computed by PIE is within +/- 0.1 of the value reported by BW test

65

PIE Outline

•Motivation




•Accuracy of PIE



•Summary


66

66Vivek Shrivastava

PIE with realistic access patterns

NSDI 2011

67



• Evaluate PIE using realist traffic patterns on a 15 node topology (7 AP – 8 laptops)

• Each client laptop replays the traffic patterns of an actual client from a real wireless trace

• Three activity periods: heavy (> 40 % medium busy), medium (40 – 20% busy), light (< 20% busy)

68


68Vivek Shrivastava

Tim

e to

est

imat

e (m

s)

NSDI 2011

Traffic period

Heavy Medium Light 0

200

400

600

800

1000

1200

69


69Vivek Shrivastava

Tim

e to

est

imat

e (m

s)

NSDI 2011

Traffic period

Heavy Medium Light 0

200

400

600

800

1000

1200

• Convergence is faster for higher client activity

• Even for light activity, median time of estimate LIR is less than 650 ms

70

PIE Outline

•Motivation




•Accuracy of PIE



•Summary


71

71Vivek Shrivastava

What is the impact on WLAN applications?

NSDI 2011

AP-Client pairs

72

72Vivek Shrivastava


NSDI 2011

AP-Client pairs

Evaluate usefulness of PIE for an interference mitigation mechanism (data scheduling using CENTAUR – Mobicom ‘09)

73

73Vivek Shrivastava


NSDI 2011

AP-Client pairs

1. Estimate interference using PIE

74

74Vivek Shrivastava


NSDI 2011

AP-Client pairs

1. Estimate interference using PIE2. Input estimate to a centralized data scheduler

75

75Vivek Shrivastava


NSDI 2011

AP-Client pairs

1. Estimate interference using PIE2. Input estimate to a centralized data scheduler3. Evaluate performance under dynamic scenarios

76

What is the impact on end users ?

76Vivek Shrivastava

Sys

tem

T

hrou

ghpu

t (M

bps)

Static scenario

Distributed

Cent. Sched (BW)

Cent. Sched (PIE)

0

2

4

6

8

10

12

14

Static

NSDI 2011

Distributed

Cent. Sched (BW)

Cent. Sched (PIE)

0

2

4

6

8

10

12

14

Static

77


77Vivek Shrivastava

Sys

tem

T

hrou

ghpu

t (M

bps)

Static scenario

NSDI 2011

Distributed

Cent. Sched (BW)

Cent. Sched (PIE)

0

2

4

6

8

10

12

14

Static

78


78Vivek Shrivastava

Sys

tem

T

hrou

ghpu

t (M

bps)

Static scenario

Static scenarios, PIE is comparable to BW test

NSDI 2011

79


79Vivek Shrivastava

Sys

tem

T

hrou

ghpu

t (M

bps)

Mobile scenario

Distributed

Cent. Sched. (BW)

Cent. Sched. (PIE)

0

2

4

6

8

10

12

14

Series2

NSDI 2011

Distributed

Cent. Sched. (BW)

Cent. Sched. (PIE)

0

2

4

6

8

10

12

14

Series2

80


80Vivek Shrivastava

Sys

tem

T

hrou

ghpu

t (M

bps)

Mobile scenario

NSDI 2011

Distributed

Cent. Sched. (BW)

Cent. Sched. (PIE)

0

2

4

6

8

10

12

14

Series2

81


81Vivek Shrivastava

Sys

tem

T

hrou

ghpu

t (M

bps)

Mobile scenario

Mobility scenarios, PIE outperforms BW test

NSDI 2011

82


•PIE can also be used to monitor production systems (like Jigsaw)

•We monitored two production WLANs

•Use testbed nodes in proximity of production APs as sniffers

•Identify hidden terminals and rate anomaly problems


83



WLANs

WLAN1

WLAN2

Hidden terminal cases (LIR < 0.7)

Rate anomaly cases (Ratio of rates < 0.2)

8%

11%

21%

22%

84



WLANs

WLAN1

WLAN2

Hidden terminal cases (LIR < 0.7)

Rate anomaly cases (Ratio of rates < 0.2)

8%

11%

21%

22%

•Hidden terminals are rare, but can become pain points for clients

•Rate anomaly is more frequent, but do not cause drastic performance issues

85

PIE Outline

•Motivation




•Accuracy of PIE



•Summary


86

Related Work

•PIE leverages techniques from Jigsaw, WIT (Sigcomm 2006) and builds on their ideas

• Focus of Jigsaw, WIT was to understand interference, ours is to compute it in real-time

• CMAP also infers interference to harness exposed terminals, but requires physical layer change

•Active techniques like Microprobing (CoNext 2008) still require downtime and do not use realistic traffic


87

PIE Limitations

•Does not handle non-WiFi interferer like microwaves.

•Can miss external interferers if none of the enterprise APs can listen to the interferer

• May miss client conflicts, can use client participation in PIE to enhance the system

•Interference detection techniques at the physical layer may be more accurate in some scenarios where diversity is too low for PIE to function


88

PIE Summary

•Online interference estimation important for interference mitigation

•BW test incurs high overhead, requires downtime

•PIE is a passive mechanism, generates interference estimates in real time

•Leverages centralized infrastructure to collect real time reports from APs

•Non-intrusive with good accuracy


Vivek Shrivastava 89

Thank you !

[email protected] www.cs.wisc.edu/~viveks

NSDI 2011

http://www.cs.wisc.edu/~viveks

http://www.cs.wisc.edu/~viveks

90

Is there sufficient diversity ?

90Vivek Shrivastava

% Overlap in transmit time

% o

f tr

ansm

it

pair

s

NSDI 2011

91


91Vivek Shrivastava


% o

f tr

ansm

it

pair

s

Overlap in transmit times for transmitter pairs in real WLAN traces

NSDI 2011

92


92Vivek Shrivastava


% o

f tr

ansm

it

pair

s

80% of transmit pairs have less then 10%

overlap


NSDI 2011

93

How quick is PIE ?

93Vivek Shrivastava

Convergence time of PIE under varying loads

Traffic load on link and interferer

Con

verg

en

ce

tim

e

NSDI 2011

94

How quick is PIE ?

94Vivek Shrivastava



Con

verg

en

ce

tim

e

NSDI 2011

95

How quick is PIE ?

95Vivek Shrivastava



Con

verg

en

ce

tim

e

Converges within 650 ms for light traffic loads as well

NSDI 2011

96

Synchronization error in PIE ?

96Vivek Shrivastava

Synchronization error using TSF beacon synchronizationNSDI 2011

97

97Vivek Shrivastava


InterfererAP-Client pair

Non Interferer

1. Activate link, interferer and non-interferer

NSDI 2011

98

98Vivek Shrivastava


InterfererAP-Client pair

Non Interferer

1. Activate link, interferer and non-interferer2. Measure LIR for interferer and non-interferer

NSDI 2011

99

99Vivek Shrivastava


NSDI 2011

AP-Client pairs

100

100Vivek Shrivastava

What is the impact on end users ?AP-Client

pairs

NSDI 2011

101



pairs

Evaluate PIE on a 7 AP - 8 Client topology

NSDI 2011

102



pairs

Evaluate usefulness of PIE for an interference mitigation mechanism

NSDI 2011

103



pairs

1. Estimate interference using PIE

NSDI 2011

104



pairs

1. Estimate interference using PIE2. Input estimate to a centralized data scheduler

NSDI 2011

105



pairs

1. Estimate interference using PIE2. Input estimate to a centralized data scheduler3. Evaluate performance under dynamic scenarios

NSDI 2011

106


Carrier sensing (conservative)


Exposed terminals

NSDI 2011

107

107

Vivek Shrivastava



assignment)

NSDI 2011 107

Estimate Interference

108


Key Challenge

Q. Is there a way to dynamically identify the precise set of nodes in an enterprise WLAN that interfere with each other and the degree

to which they do so ?

NSDI 2011

109

PIE Outline

•Motivation



•Accuracy of PIE

•Convergence of PIE

•Agility of PIE

•Applications of PIE (data scheduling)

•Summary


110


How agile is PIE ?

Client moves from AP towards interferer

Interferer

AP-Client pair

NSDI 2011

111


How agile is PIE ?


Interferer

AP-Client pair

NSDI 2011

112


How agile is PIE ?


Interferer

AP-Client pair

NSDI 2011

113

How agile is PIE ?


InterfererAP

Tracking client mobility.

SNR

SN

R (

dB)

NSDI 2011

114

How agile is PIE ?


InterfererAP


SNR

Tput

SN

R (

dB)

Tpu

t (M

bps)

NSDI 2011

115

How agile is PIE ?



SNR

Tput

LIRS

NR

(dB

)T

put (

Mbp

s)LI

R (

PIE

)

NSDI 2011

116



LIR

NSDI 2011

Can PIE classify interferers accurately ?

117




LIR

NSDI 2011

118




LIR

When overlap is less than 80%, PIE is able to classify accurately

NSDI 2011

119

Convergence of PIE•Test stability of interference measure with different

polling periods

•Polling period of 85 ms is sufficient for stable interference measure in presence of saturated traffic

•Test stability of interference measure with different traffic rates

•Higher convergence times for low traffic rates, as number of events per polling period is low

•Still converges within 600ms for very lightly loaded links (when load is 2-3% of link capacity)


120

Estimating carrier sense passively

Sniffer reports

Infer carrier sense


121


Sniffer reports

Infer carrier sense


Determine CS by observing direction of packet overlaps for contending

transmitters

122


Sniffer reports

Infer carrier sense

Scenarios


Observe packet overlaps in both directions

123


Sniffer reports

Infer carrier sense

Scenarios


Both Red and Green APs do not sense each

otherObserve packet overlaps

in both directions

124


Sniffer reports

Infer carrier sense

Scenarios


Red senses Green, but

not vice versa

Observe packet overlaps in only one direction

125


Sniffer reports

Infer carrier sense

Scenarios


Green senses Red, but not vice

versa

Observe packet overlaps in only one direction

126


Sniffer reports

Infer carrier sense

Scenarios


No contending packets overlap

Green and Red sense each other

127

Computing carrier sense measure in PIE

•Compute overlap probability for contending packets ( Poverlap )

•If (1 – Poverlap ) > δt , report no carrier sensing

•else if Poverlap > δt

•If Pone-way > δt, report one-way CS

•else report mutual carrier sensing


128



Interferer

AP-Client pair

NSDI 2011

129



Interferer

AP-Client pair

1. Activate both link and interferer

NSDI 2011

130



Interferer

AP-Client pair

1. Activate both link and interferer2. Measure LIR using PIE and BW test

131

How accurate is PIE with multiple interferers ?

Scenarios

Reception

X X

Vivek Shrivastava

Red and Green packets overlap=> Green is lost

One way hidden

terminals

NSDI 2011

132

How accurate is PIE with multiple interferers ?

Scenarios

Reception

Vivek Shrivastava

Blue and Green packets overlap=> None is lost

No interference

NSDI 2011

133




% o

f tr

ansm

it

pair

s

NSDI 2011

NSDI 2011134




% o

f tr

ansm

it

pair

s


NSDI 2011135




% o

f tr

ansm

it

pair

s

80% of transmit pairs have less then 5%

overlap


136

What about multiple interferers?

X

Vivek Shrivastava

X

Ambiguous of loss source

(Red/Blue)

NSDI 2011

137


X

Vivek Shrivastava

X

Overlap with Blue

=>No Loss

NSDI 2011

138


X

Vivek Shrivastava

X

Overlap with Red => Loss

NSDI 2011

139


X

Vivek Shrivastava

X

Isolation =>No Loss

NSDI 2011

140


X

Vivek Shrivastava

X

Isolation =>No Loss

Overlap with Blue

=>No Loss

NSDI 2011

141


X

Vivek Shrivastava

X

Isolation =>No Loss

Overlap with Blue

=>No Loss

Overlap with Red => Loss

NSDI 2011

142


X

Vivek Shrivastava

X

PIE needs transmission diversity to identify interferers accurately

NSDI 2011

143



% o

f lin

k -

inte

rfer

er

pairs

NSDI 2011

Convergence time (ms)

144



% o

f lin

k -

inte

rfer

er

pairs

NSDI 2011


145



% o

f lin

k -

inte

rfer

er

pairs

NSDI 2011


• Convergence is faster for higher client activity

146



% o

f lin

k -

inte

rfer

er

pairs

NSDI 2011

Convergence time (ms)• Convergence is faster for higher client activity

• Even for light client activity, 90% of link-interferer pairs converge within ~ 10 seconds

147

Synchronization error


Synchronization error for a 19 node topology

NSDI 2011

CD

F

Synchronization error (microsec)

148

Synchronization error


Synchronization error as a function of beacon interval

NSDI 2011

Clo

ck s

kew

(u

s)

Time (ms)

149

1. 2 Enterprise WLAN setting 2 Vivek Shrivastava Wireless controller Access Point Clients Internet NSDI 2011.

Documents

interference delivery

estimate interference

vivek shrivastava problems

isolation slide

wireless nsdi

measure delivery slide

measure delivery nsdi

vivek shrivastava mismatch