a measurement study of a commercial-grade urban wifi meshpages.cs.wisc.edu/~shravan/mcb-talk.pdfa...

A Measurement Study of a Commercial-gradeUrban WiFi Mesh

Vladimir Brik Shravan Rayanchu Sharad SahaSayandeep Sen Vivek Shrivastava Suman Banerjee

Wisconsin Wireless and NetworkinG Systems (WiNGS) LabUniversity of Wisconsin Madison

IMC 2008

WiNGS Lab, UW-Madison Measurement Study of MadMesh

Introduction

What?

Measurement study of the performance and usagecharacteristics of an operational commercial urban wirelessmesh network

Why?

Better understanding of these networks

Previous measurement studies: Roofnet, TFA, DGP (all arecustom testbeds built for experimentation)

What is the state-of-the-art in the industry?

How much of prior work is applicable?


IntroductionMadMesh overview

MadMesh

More than 250 MAPs covering 10 sq. miles

Has been operational for about 2 years now

Provides service to more than 1000 residential customers,small businesses, public safety organizations


IntroductionMadMesh Architecture

Architecture

Cisco 1510 MAPs, RAPs,Mesh controller

Tree-based routing (vendorproprietary protocols)

Backbone interface(802.11a, 5GHz)

Access interface (802.11b/g, 2.4 GHz)


Study Goals & Data Sets

Categories of study:

Mesh planning and deployment

Mesh routing

User experience

Usage characterization

Data collection:

Period infrastructure logs: SNMP (every 3 min), tools at thecontroller

Passive measurements: Deployed indoor/outdoor nodes

Active measurements: coverage, throughput experiments

Two week period - 1.7 million SNMP records; 100 hrs ofactive measurements


Mesh Planning & Deployment

Deployment Characteristics:

What kind of connectivity does each MAP have? Is thenetwork robust against failures?

What are the link-level error rates and signal qualities on thebackbone and access links?

Does the network topology lend itself to new techniques likewireless network coding?


Mesh Planning & DeploymentAverage MAP Degree

Q. What is the average MAP degree?

Lower degree ⇒ less re-routing choices during losses

Higher degree ⇒ over-provisioning, self-interference


Mesh Planning & DeploymentAverage MAP Degree

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

1 2 3 4 5 6 7 8910 20 30 40

Fra

ctio

n

Average degree of MAPs (in Logscale)

MadMeshRoofnet Mesh

MAPs: 20% have degree ≤ 2 and 50% have degree > 3Much higher degree for Roofnet


Mesh Planning & DeploymentRobustness

Q. Is the deployment robust?

Min-cut: the minimum number of edges, whose removalwould disconnect the MAP from the graph


Mesh Planning & DeploymentRobustness

path to the root

mincut = 1

Connectivity for thehighlighted node

11

3

18

58 23

102

17

160

56

21

39 48

34

19

0

1

3

5

7

9

11

1 2 3 4 5 6 7 8 9 10 11

Min

cut o

f Nod

e

Average Degree of Node

8% of the MAPs have min-cut ≤ 2Some MAPs with degree as high as 7 have min-cut ≤ 2


Mesh Planning & DeploymentError Rates

Q. How good are the access and backbone links?


Mesh Planning & DeploymentError Rates

Backbone

0.20.40.60.81.0

.02 .04 .06 .08 .10 .12

Fra

ctio

n

MAC level Error Rate

# of MAPs 224

Access

0.2

0.4

0.6

0.8

1.0

.05 .1 .15 .2 .25 .3 .35

Fra

ctio

n

MAC level Error Rate

# of MAPs 15

Much higher loss rates on access links

Why is this the case?


Mesh Planning & DeploymentChannel Selection

Backbone

0.10.20.30.40.50.60.70.80.9

10 15 20 25 30 35 40 45 50

Fra

ctio

n

SNR (in dB)

# of MAPs 224

Access

Current

Best

# of MAPs 224

0

20

40

60

80

100

−100−90 −80 −70 −60 −50 −40

Fra

ctio

n

Interference (in dBm)

Good quality backbone links

PER on access: (1) low SNR (2) interference

≥ −70dB interference on 80% of the access links

Channel selection can help


Mesh Planning & DeploymentApplicability of Network Coding

Q. Are techniques like network coding applicable?



Q. Are techniques like network coding applicable?

COPE: XOR operations, opportunistic listening

Coding rule:

To transmit n packets, p1, ..., pn, to n nexthops,r1, ..., rn, a node can XOR the n packets togetheronly if each next-hop ri has all n− 1 packets pj forj 6= i.

We calculate the maximum coding gain at each MAP

Depends on a number of factors; Measure of overhearingsupported by the deployment



NodesLeaf

No Overhearing

24% (Overhearing)

1

2

3

4

5

6

0 20 40 60 80 100 120 140 160 180 200MAP Index

Max

imum

cod

ing

gain

For 66% of the MAPs, coding gain was 224% of MAPs had coding gain > 2 (Max. coding gain was 6)

Network coding can indeed improve the performance


User Experience

Characterizing the user experience:

How good is the client connectivity? Are coverage holesprevalent?

What are the observed client throughputs?


User ExperienceClient connectivity

Q. How prevalent are coverage holes?

Estimate using a path-loss model:

PdBm(d) = PdBm(d0)− 10αlog10(d

d0) + ε

Collected signal strength information at 25 locations for eachMAP, and then estimated α, ε



Path-loss modeling results:

α = 2.3 (Campus)

-100

-90

-80

-70

-60

-50

-40

-30

-20

0 50 100 150 200 250 300 350

Rec

eive

d S

igna

l Str

engt

h (d

Bm

)

Distance (m)

Pathloss Exp = 2.3082Shadowing Std. = 14

Mean+7 Stdev-7 Stdev

Measurements

α = 2.9 (Downtown)

-100

-90

-80

-70

-60

-50

-40

-30

-20

0 50 100 150 200 250 300 350R

ecei

ved

Sig

nal S

tren

gth

(dB

m)

Distance (m)

Pathloss Exp = 2.9082Shadowing Std. = 8.1

Mean+4 Stdev-4 Stdev

Measurements

Path-loss varies with each MAP (location)



Propagation model based radio map shows this area to be’covered’

Simple monitoring tool at the clients

More holes prevalent at vehicular speeds

Client feedback can really help


User ExperienceClient Throughput

Q. What are the throughputs achieved at different locations?



Q. What are the throughputs achieved at different locations?

Random sample of 100 locations

3 runs of TCP iperf tests, 100 seconds each



Results of throughput tests:

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

1

0.0 0.2 0.4 0.6 0.8 1.0 1.2

CD

F o

f sam

pled

loca

tions

TCP Throughput (Mbps)10% of the clients have less than 0.2 Mbps, while 80% haveless than 1 Mbps



Results of throughput tests:

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

1

0.0 0.2 0.4 0.6 0.8 1.0 1.2

CD

F o

f sam

pled

loca

tions

TCP Throughput (Mbps)

Factors:

RSSI

Hop-count

Channel Congestion

Shared Congestion

10% of the clients have less than 0.2 Mbps, while 80% haveless than 1 Mbps


Usage Characterization

Characterizing mesh usage:

How many clients are using the network? How does the usagevary with time?

How does client distribution vary with MAPs, across differenthops etc. ?


Usage CharacterizationDistribution across time

Q. How does the usage vary with time?


Usage CharacterizationDistribution across time

Q. How does the usage vary with time?

avg = 498

max = 627

0

100

200

300

400

500

600

700

3:00 6:00 9:00 12:00 15:00 18:00 21:00 24:00

Ave

rage

num

ber

of u

sers

Time of day

Most number of clients were connected at around 10 PM


Usage CharacterizationDistribution across MAPs

Q. How does the usage vary across MAPs?


Usage CharacterizationDistribution across MAPs

Q. How does the usage vary across MAPs?

50% (40)

25% (14)

75% (80)

0

1

2

3

4

5

6

7

0 50 100 150 200

Ave

rage

num

ber

of c

lient

s

MAP Index

Clearly, certain MAPs are more popular than others

50% of clients are connected to 40 MAPs (20%)


Summary

Main observations:

Robustness – ensure path diversity

Bottleneck – it is the access link; channel selection can help

Management – client feedback can really help

Techniques like network coding are applicable

User characteristics – night time peaks and uneven usage


Questions?


a measurement study of a commercial-grade urban wifi meshpages.cs.wisc.edu/~shravan/mcb-talk.pdfa...

Documents