1

Designing High Performance Enterprise Wi-Fi Networks

Rohan Murty Harvard University

Jitendra Padhye, Ranveer Chandra, Alec Wolman, and Brian Zill Microsoft Research

2

Trends in Enterprise Wi-Fi Networks

Increased adoption and usage [Forrester]

Culture of mobility: Users tend to use Wi-Fi even when wired connections are available [Gartner, Forrester, Economist]

Move towards an all wireless officeUsers want wire-like performance from wireless networks

3

Capacity of Conventional Corporate WLANs

Corporate WLAN Study: 12 users < 1 Mbps each

4

Characteristics of Conventional Corporate WLANs

Focus on coverage Fewer APs than clients Clients talk to APs far away; worsens rate anomaly

Clients pick APs to associate with Use RSSI of beacon packets Agnostic to channel load at APs

Lack adaptive behavior No load balancing; fixed channel assignments Congestion and hotspots worsen

5

DenseAP

Focus on capacity Lots of APs; densely deployed Clients can talk to APs near by; mitigates rate

anomaly

Infrastructure picks client-AP associations Global view of network conditions (channel load,

interference, etc.)

Adaptability Load balance associations; Dynamic channel

assignment Redistributes load away from local hotspots

6

DenseAP is Practical

No client modifications Works with legacy clients Changes limited to the infrastructure Easy to deploy

Self-managing

7

DenseAP Central Controller (DC)

DenseAP System Architecture

AssociationsChannel AssignmentsLoad Balancing

DenseAP Nodes (DAPs)

Commands to DenseAP

nodes

Summarized Data from DenseAP nodes

Summarized

Data Wired NetworkCommand

s

Interface with clientsSend summaries to DC

8

Key Challenges

Controlling Associations Mechanisms Policy

Dynamic Channel Assignment Mechanism Policy

Load Balancing Mechanism Policy

9

Probe Request

Probe Request

Probe Request

ACL

ACL

ACL00:09:5B:5A:1F:4F

Association Control in DenseAP

10

ACL

ACL

ACL00:09:5B:5A:1F:4F

Probe Request MAC = 00:09:5B:5A:1F:4F

RSSI = 30

Probe RequestMAC = 00:09:5B:5A:1F:4F

RSSI = 42

Probe Request MAC = 00:09:5B:5A:1F:4F

RSSI = 40

Association Control in DenseAP

11

Accept Client

ACL00:09:5B:5A:1

F:4F

ACL

ACL00:09:5B:5A:1F:4F

Association Control in DenseAP

Client only sees one DAP at any given time

Probe Response

12

Association Policy

What is the quality of a connection between a client and a DAP? (rate)

How busy is the medium around each DAP?

Overall goal: Associate client with a DAP

that will yield good throughput

13

A Metric for DAP Selection

Expected Transmission-

Rate (Mbps)

Available Capacity (AC)

(Mbps)

Free Air Time(%)

X=

Probe Request

Pro

be R

eq

uest

Prob

e Req

uest

Free air time = 0.35

DAP2

DAP1

DAP3

RSSI = 20

RSSI = 10

RSSI = 30

Free air time = 0.45

Free air time = 0.22

DAP Free Air-

Time

RSSI

DAP1 0.35 20

DAP2 0.22 10

DAP3 0.45 30

Accept Client

Probe Response

DAP Free Air-Time

RSSI

Ex. Tx-Rate

AC

DAP1

0.35 20 18 6.3

DAP2

0.22 10 6 1.32

DAP3

0.45 30 48 21.6

15

RateMap: Estimating Expected Transmission-Rate

Correlation between RSSI of Probe Request

packets Avg. throughput between

a DAP-client pair

Rough approximation - ordering of DAPs

Online profiling method that builds RSSI to data-rate estimates Upload and RSSI correlation

= 0.71Download and RSSI correlation = 0.61

0-9 10 - 19

20 - 29

30 - 39

40+0

2

4

6

8

10

12

14

16

18Upload Download

RSSI Values

Avera

ge T

hro

ug

hp

ut

(Mb

ps)

16

Estimating Free Air Time

Estimate how busy is the medium around at a DAP

Technique similar to ProbeGap* Measure time taken to

finish a packet transmission

Estimates match up closely with offered traffic load

*Lakshminarayan et al., 2004

*Vasudevan et al., 2005

0 20 40 60 80 100

0

20

40

60

80

100

Actual Load (%)

Esti

mate

d L

oad

(%

)

17

Channel Assignment

Integrated into the association process

DAPs not discovered by clients don’t need channels

A DAP is assigned a channel only when it goes from being passive (no clients) to active (services at least one client) Central controller assigns channel with

least load

18

Re-evaluating Associations

So far, associations when a new client joins the network

No association is perfect Client traffic demands change Local hotspots created

19

Load Balancing

Central controller monitors load on every DAP

When channel load on a DAP crosses a certain threshold Client causing most load is determined Moved to less loaded DAP nearby Ensure client continues to get at least as much

available capacity at the new DAP

Load balancing achieved via handoffs Use association control; manipulate ACLs on DAPs

20

Results

21

UP

DN

DN

UP

30003000

30013001

30023002

30043004

30053005

30063006

3012SHUOCHEN

3012

3014ARUNKU3014

3016CHADV3016

3018BRADDAN

3018

3019EMREK3019

30203020

30213021

30223022

3024YMWANG

3024

30263026

30283028

30303030

3032MFORNEY

3032 30333033

3034RAMANC3034

3035V-DOAVELV-JONAHAV-TEDKAO

3035

30373037

30393039

30403040

3042V-JWING

3042

30443044

3046DESNEY3046

3048DCR30483052

3052

31003100

3102MARYCZ3102

310331033104

GGR3104

3106GREGSMI

3106

3108BRIANME

3108

31103110

3112SZELISKI

3112

3114NURIA3114

3116MATTU3116

3118BROWN3118

3120SWINDER

3120

31233123

31253125

31303130

31313131

3134STEEDLY

3134

3136SBKANG3136

31383138

3139T-LUCASK

3139

31403140

31413141

31433143

31453145

31473147

31493149

31513151

31533153

3154MCOHEN

3154

3156ERUDOLPH

3156

31593159

31603160

31613161

31633163

31653165

31673167

3172LARRYZ3172

3174A-HGOOD

3174

3176PETERJ3176

3178CRAIGVI

3178

31793179

31803180

31813181

32003200

32013201

32033203

32043204

32053205

3206KENTOY

3206

32073207

3208ANANDAN

3208320932093210

3210

32113211

3212LORIMAC

3212

32133213

3214CLOOP3214

32153215

3216MARCEL3216

3217BLINN3217

32183218

32193219

32203220

32223222

3224HHOPPE3224

3226JOHNSNY

3226

3228BGUENTER

3228

32303230

32323232

32333233

3234V-AVD3234

3235BODHIP3235

3237CYL

3237

32393239

32403240

32423242

3244JFAY3244

3246GEORGP3246

3248SDRUCKER

3248

3252WONG3252

33003300

3302JGEMMELL

3302

33033303

3304JDUNAGAN

3304

3305MZH3305

3306BZILL3306

3307DMALTZ3307

3308DANSIMON

330833093309

33103310

3312RICHDR3312

3314HOWELL

3314

3316SAGARWAL

3316

3318HELENW

3318

3320PADHYE3320

332133213323

3323

33253325

33293329

33303330

33313331 3333

3333

3334PADMANAB

3334

3336LIUJ3336

3338JELSON3338

3339SUMANN

3339

33403340

33413341

3343RANVEER

3343

3347T-ANDRES

3347

33513351

3354ZHAO3354

33563356

3359JOHNDO

3359

33603360

3363RATUL3363

33713371

3372LORCH3372

3374BAHL3374

3376ALECW3376

33773377

3378BOLOSKY

3378

33793379

33803380

CH31CH31

CH32CH32

CH33CH33

CI32CI32

EL 31EL 31

EL 32EL 32

STR31STR31

STR32STR32

(0,0)

(98,32)

Testbed

1 Corp AP

24 DAPs

24 Clients

802.11 a/bg

22

Results: Roadmap

Performance Density Channels Intelligent Association

Load Balancing

23

2 4 6 8 10 1202468

101214161820

DenseAP Corporate WLAN

Number of Clients

Per-

Clien

t Th

rou

gp

ut

(Mb

ps)

2 4 6 8 10 1202468

101214161820

Corporate WLAN

Number of Clients

Per-

Clien

t Th

rou

gp

ut

(Mb

ps)

Overall DenseAP Performance: 802.11a

Gains due to• More channels• DAP density• Intelligent

associations

1250% gain

Why?

24

Exploring the impact of density

Put all DAPs on the same channel

Factors out Channels Intelligent Associations: same load on all DAPs

Single out impact of Density

25

Impact of Density: Using only 1 channel

Higher density provides better performance

1 2 3 4 5 60

5

10

15

20

Corporate WLAN

Number of Clients

Avg

. P

er-

Clien

t Th

rou

gh

pu

t (M

bp

s)

26

Is intelligent association control necessary?

27

Why does intelligent association matter?

Client-Driven Disable intelligent association control Let clients pick DAP to associate with (conventional

WLANs)

Compare with DenseAP

Factors out Channels Density

Single out impact of Intelligent association

28

2 4 6 8 10 1202468

101214161820

Client-Driven DenseAP

Number of Clients

Per-

Clien

t Th

rou

gh

pu

t (M

bp

s)

Necessity of the Association Policy

Intelligent association policy is necessary

160% gain

29

Load Balancing

30

0 50 100 150 200 250 3000

5

10

15

20

25

Client 3

Client 1

Client 2

Time (s)

Th

rou

gh

pu

t (M

bp

s)

Load Balancing

Client 1

moved

Client 1

improves

Clients 2 & 3

improve

Client 2

moved

31

Other Details and Results in the Paper

Load balancing algorithm and mechanism

Mobility

Performance Fewer DAPs Fewer channels 802.11g …..

Scalability

32

Related Work

Plenty of prior work on static channel assignment, power control and associations Each studied each aspect in isolation Require client modifications [Ramani and Savage,

Infocom 2005] SMARTA [Ahmed et al., CoNext 2006]

Examines channel and power control Increase overall network capacity Does not consider associations, load balancing

MDG [Broustis et al., MOBICOM 2007] Identified tuning channel, power and associations Studies the order in which these knobs must be tuned Requires client modifications

33

Overall Contributions

Practical system How do density, intelligent association,

and more channels affect capacity? Adaptive system

Future directions Impact of hidden terminals Heterogeneous mix of client traffic patterns Other backhauls: e.g. Wireless, powerline