Adaptive and resource-efficient rural area networks

Adaptive and Resource-Efficient
Rural Area Networks
Slides at: www.cs.bham.ac.uk/~pejovicv/cambridge

Veljko Pejovic
Research Fellow
University of Birmingham

Digital Divide
●

●

●

A division between those who do and those who do
not have access to and the capability to use modern
information and communication technologies (ICTs)
The digital divide is tightly connected with the living
standard, health care, economy, education, political
freedoms
Observed from different aspects: gender, age,
affluence

Digital Divide – A Broad View

Internet users per region
Source: ITU 2013

Digital Divide – Causes
Differences between regions that
impact ICT adoption:
●

●

●

●

Purchasing power – ICTs cost
Existing infrastructure – ICTs
need reliable power supply
Level of urbanization – ICTs are
designed for cities
Different cultures – the same ICTs
might not be suitable for all
societies

Digital Divide – Urbanisation Levels
100
80
60
40
20
0

Least dev.
UK
Zambia
OECD
South Korea
South Africa

Urbanisation level [%]
Source: World Bank

Digital Divide – Rural vs Urban
100
Rural

80

Urban

60
40
20
0

Belarus

Brazil

Egypt

S. Africa

S. Korea

Population that uses the Internet [%]
Source: ITU

US

Existing Problems
Technical
●

Poor signal propagation due to vast distances,
terrain configuration, vegetation

●

Wireless interference, especially in
the case of unlicensed solutions

●

Lack of reliable electrical energy supply

Socio-economic
●

Economic infeasibility of wide area coverage

●

Lack of locally relevant online content

●

Inability to engage a wider community into the network

●

Micro digital divides: castes, genders

Our view on why rural area
connectivity fails
In rural areas a unique set of
technical and social
challenges are obstacles to
Internet penetration.
The essence of the problem
lies in a general lack of
understanding of rural area
dwellers’ needs, and in the
development of
communication technologies
without consideration of
unique nuances of rural areas.

Holistic approach
Investigate existing solutions
identify obstacles and
true needs of our users

Develop technical solutions
with experts from target areas

Investigating Technical and Social
Challenges in Rural Areas
Analysis of existing rural wireless networks in Africa
(Macha, Zambia and Dwesa, South Africa):

Why Macha and Dwesa?
●
●
●
●

Real rural Africa
Community wireless networks
Different social settings
Strong collaboration links
through our partners

Challenges in Rural Areas
Analysis of existing rural wireless networks in Africa
(Macha, Zambia and Dwesa, South Africa):
●

Lightweight traffic monitoring system:
–

–
●

Packet headers on the satellite
gateway
Squid proxy logs

Social surveys
–

Go beyond just anecdotal evidence - quantifiable
data

–

Examine Internet usage, legacy communication
practices, social aspects of computer networking,
quality of service issues

Challenges – Key Findings
●

The location of Internet access (home/work/internet
café) impacts the type of applications used online:
–

●

There is a strong locality of interest:
–

●

Only at-home access allows full-fledged online
experience, including active OSN usage, content
generation; otherwise deliberate interaction model

The majority of voice-over-IP (VoIP) calls and instant
messages (IM) are exchanged within the village

Network performance and user behavior are tightly
intertwined:
–

people share files via USB drives when the network is
More about rural area network analysis in our WWW'11 paper
congested

Develop Technical Solutions
Guidelines:
●

Provide at-home Internet access to all

●

Support local communication

●

Facilitate content generation

●

Be resource (electrical energy, satellite bandwidth,
wireless spectrum) efficient

VillageNet
VillageCell
Enables free local
mobile phone calls via
off-the-shelf
Evolved into Kwiizya
open-source solutions.
Itdeployed no Macha, Zambia.
requires in modification
to the existing GSM
handsets and SIM cards.
Check: M. Zheleva et al.
●
●
●

"Kwiizya...", MobiSys'13
GNUradio
OpenBTS
Asterisk

VillageNet

VillageShare
Evolved into Kwaabana
Improves content generation
and sharingin Macha, Zambia
deployed in a village via
a local file sharing application.
and rural Eastern Cape
Enables extra upload capacity
Check: D. Johnson et al.
via time-delayed uploads.
"Kwaabana...", ACM DEV'13

VillageNet
VillageLink
Connect distant locations
through outdoor, non line
of sight wireless links
operating on unlicensed
frequencies.

Bring connectivity to every individual!

Wide-Area Wireless

Low population density:
●

●

Cell phone towers are not economically viable
for low income under-populated areas
WiFi networks have a limited range and
require a line of sight

New opportunities for rural area
connectivity
White spaces:
●

●

●

Frequency band from roughly
50MHz to 800MHz
Vacant after TV went digital;
potentially unlicensed spectrum
Excellent propagation properties:
–

Long range (path loss ~ f2)

–

Not absorbed by vegetation

–

Signal can bend around obstacles

White Spaces – Issues
●

●

White spaces encompass a few hundreds of MHz of
spectrum
Dynamic range in white spaces:
Technology

fL (MHz)

fU (MHz)

D (dB)

802.11 (2.4GHz)

2412

2484

0.26

802.11 (5 GHz)

5170

5700

0.85

GSM 900

935

960

0.23

White spaces

43.25

797.25

25.31

●

●

spectrum
Results from a 3 km long outdoor link in South Africa

Signal strength can be
tens of dB different
across the band.

●

●

spectrum

Why is this a problem?

Performance across the frequency band
cannot be described
solely by the propagation theory

Antenna properties and the environment
determine signal strength at different channels

White Spaces – Channel Allocation
●

●

We have a limited pool of vacant white space
channels
Network capacity depends on the useful signal
strength and the interference (plus noise) strength

How to allocate wireless channels
to network nodes so that
the network capacity is maximized?

Conventional Network – Channel
Allocation
●

Signal strength is equal at all frequencies. Channels
allocation strives to minimize interference.
Access points with
associated clients

Allocation
●

●

Graph coloring – assign colors (channels) so that no two
nodes that share a link in the interference graph are
assigned the same color.

Interference graph
among access points

Allocation
●

●

Graph coloring – assign colors (channels) so that no two
nodes that share a link in the interference graph are
assigned the same color.

White Space Network – Channel
Allocation
●

In white spaces signal strength varies over channels,
moreover the variation may be different for different pairs
of nodes.

Allocation
●

●

In white spaces signal strength varies over channels,
moreover the variation may be different for different pairs
of nodes.
Minimizing interference with graph coloring does not work
anymore. The graph depends on channel selection.

Allocation

Propagation diversity over a wide white space band
is highly varying and unpredictable

Even if we were to know propagation
over all frequencies for all links,
the problem would be intractable

Channel Probing and Medium
Access
●

●

Consider a network of base stations (BSs) with multiple
associated clients (CPEs)
BSs select their operating channels and CPS switch to a
channel selected by the BS they are associated with
●

CPE

BS

Selection of the operating channel
impacts the signal strength from a
BS to a CPE and the interference
from one BS to another.

CPE
CPE
CPE

BS
CPE

Channel Probing and Medium
Access
●

●

We extend the 802.22 protocol with inter-BS and BS-CPE
probing.
A probe is a packet whose content is known to the
receiver. By comparing the received probe with the sent
one, we can estimate the channel quality. A probe is sent
at each available channel.
●

CPE

BS
CPE

After the probing is completed each
BS knows channel quality between
itself and each of its CPEs and the
interference level between itself and
each of the neighboring Bss.

BS
●

The information is also propagated
to the neighboring BSs.

Channel Allocation Method
●

●

●

Gibbs sampling – obtain samples from a hard-to-sample
multivariate distribution
Draws samples from a multivariate
probability distribution: p( x1,..., x N )
Sample each of the variables (xi) in turn from a
j

j

j

j

conditional probability distribution: p( x i | x1 ,..., x i−1 , x i +1,..., x N )
Do this for each sample j = 1..k
●

In the end we have k samples from the joint distribution

How is Gibbs sampling
connected with
channel allocation?

1) Probability distribution is related to overall network
performance
2) Probability distribution depends on channels allocated
to BSs
3) Probability distribution favors states that lead to
maximum performance
4) Conditional probability distribution isolates the impact
of each of the nodes on the total optimization function
5) Conditional probability distribution can be calculated
independently at each of the base stations
If the above conditions hold,
Gibbs sampling of the performance distribution
(over channels at different BSs)
will lead to the optimal channel allocation

●

Network performance metric:
–

Total network capacity C, under a certain channel
allocation c independently at each of the base
stations: C(c) = C (c ) = W log(1+ SINR (c ))

∑

i

i

∑

i

i

i

Sum of the capacity of
each BS-CPE

●

SINR (signal to interference-plus-noise ratio)
is different at different channels
for different BS-CPEs due to high variability of
propagation in white spaces

Remember one of the conditions to use Gibbs sampling:
Conditional probability distribution isolates the impact of
each of the nodes on the total optimization function
SINR: a single BSs decision on the operating channel
affects interference at other BSs, yet we cannot isolate
the effect as it is “hidden” behind a log function
●

●

Network performance metric, take two:
–

Cumulative interference-plus-noise to signal
ratio (CINSR):
Noise Interference
CINSR(c) = ∑
i

1
=∑
SINRi (c i )
i

N 0W + ∑ ch(i, j)PH ji (c i )
j ≠i

PHi (c i )
Useful signal

ch(i,j) is 1 if BS i and BS j
operate on the same channel

●

Network performance metric, take two:
–

Cumulative interference-plus-noise to signal
ratio (CINSR)
●

Easy to isolate the impact of a single decision on
the total metric
 PH (c ) PH (c ) 
N 0W
ji
i
ij
i
CINSRi (c) =
+ ∑ ch(i, j)
+
 PH (c ) PH (c ) 

PHi (c i ) j ≠i

i
i
j
i 
Impact of a local
decision on own
CINSR

Impact of a local
decision on others
CINSR

All the information can be calculated locally!

●

Connect the network performance metric with a
probability distribution:
–

The Gibbs distribution:

π (c) =

Favors low CINSR states
especially at a low temperature (T)

–

e

CINSR (c )
T

∑e

−

CINSR ( c ′ )
T

c ′ ∈c N

Local Gibbs distribution:

π i (c) =

−

e

−

CINSR i (c i ,(c j ) j ≠i )

∑e

c ′ ∈c N

T
−

CINSR i (c i ,(c j ) j ≠i )
T

●

Each BS samples its local Gibbs distribution to obtain a
preferred channel selection

2
1
3
4
Available channels:

●

Each BS samples its local Gibbs distribution to obtain a
preferred channel selection

2
1
Is this channel allocation
optimal?

Available channels:

3
4

●

●

●

●

The distribution might be such that many states have low
energy, and the sampler might get stuck in a channel
selection which is good, but not optimal
Annealed sampler – change the temperature (T) as the
process progresses allows the exploration of a wider
solution space:
Depending on the temperature change schedule we get
different results
Inspiration from annealing
in metallurgy

VillageLink Algorithm
●

Distributed channel allocation algorithm (at each node):
–

While time t < tend
●

Calculate temperature T at time t (temperature
decreases over time)

●

Calculate local CINSRi for each possible channel decision

●

Calculate and sample local Gibbs distribution

●

●

–

π i (c)

Pick a channel according to the channel sampled from
the Gibbs distribution and disseminate that information
to neighbors
Listen to information about the channel selection of
neighbors

Switch the wireless interface to the last selected channel

Channel switching does not occur in the loop!

VillageLink Algorithm
●

Properties of the algorithm
–

Distributed algorithm - uses only local computations.

–

Uses propagation profiling results from channel probing.

–

Only the information on the channel that resulted from the
sampling process is used in each iteration. True channel
switching happens only once at the end of the process.

–

For certain cooling schedules converges towards the
globally minimal CINSR. However, there is no guarantee on
the number of iterations needed.

VillageLink – Evaluation
●

Simulation setup
–

Propagation in white spaces is influenced by the free space
loss, antenna patterns and the environment
●

●

Propagation calculation takes into account transmission
power, antenna gain and the distance between the
nodes
We closely model antenna irradiation patterns, frequency
selectivity and antenna orientation

–

We experiment with a varying number of base stations and
available channels

–

We model a wide area with a few TV stations that create
varying spectrum availability over the area

●

Is CINSR a good metric?
–

Comparison to the minimal interference metric

Over-provisioned channels
a lot of vacant channels, few BSs

Channel allocation in a rural network
is important even when
interference is not a problem

Under provisioned channels
a lot of BSs, few channels

Minimizing interference is
a good approach if the interference
limits the capacity

●

Alternatives to VillageLink
–

Least congested channel search (LCCS) – selects the
least used channel locally

–

Preferred intra-cell channel allocation (PICA) – selects
the channel for which the BS experiences the highest
channel gain towards its clients

–

VillageLink minimizes CINSR (cumulative interference
plus noise to signal ratio), thus taking into account
both preferred channels and interference

●

Total network capacity:
Ten available channels

Fifteen available channels

Twenty available channels

●

Fairness (Jain fairness index,
the closer the value is to 1 the better)
Ten available channels

Fifteen available channels

Twenty available channels

VillageLink – Conclusion
●

●

●

White space channel
allocation algorithm that
jointly minimizes interference
and maximizes BS-CPE
capacity
A practical solution that
requires the minimal number
of channel switching events
VillageLink is an integral part
of VillageNet, a set of
networking solutions we
developed for rural areas that
includes our previous work
VillageCell and VillageShare

VillageLink – Future
●

System Implementation
–

●

Deployment
–

●

From simulator to Software Defined Radio
Use case for VillageLink

Licensing
–

White spaces are still a grey zone when it comes to
licensing, especially in our target areas

Collaborators
Mariya Zheleva, UCSB

Albert Lysko, CSIR,
South Africa

David Johnson, CSIR, South Africa
Elizabeth Belding, UCSB

Gertjan van Stam, SIRDC,
Zimbabwe

Also:
● Meraka Institute, South
Africa
● LinkNet, Macha, Zambia

Thank you!
Veljko Pejovic
v.pejovic@cs.bham.ac.uk
More at: http://www.cs.bham.ac.uk/~pejovicv/publications.php

Digital Divide – Rethinking the
definition
●

A gap between those who
do and those who do not
have access to ICTs

Digital Divide – Rethinking the
definition
●

A variety of inequalities among people’s access to
ICTs, ability to use ICTs and benefits from using ICTs.

Measuring Success
●

This is an over simplified view of the divide

Measuring Success
●

A complex metric is necessary
–

Conventional metrics of access do not capture
differences in access quality

Measuring Success
●

Example – Connectivity Speed

Source: ITU

International Internet bandwidth (bit/s per user), by region

Measuring Success
●

Example – Connectivity

The average web page size grew
Speed about 50 times in 15 years

Source: ITU

International Internet bandwidth by region

Is the access in the developing world
effectively getting worse?

Measuring Success
●

Example – Location of Access

Source: ITU

Measuring Success
●

Example – Location of Access
–

Location of access is important:
●

Distance:
At home or a long walk to a terminal
Availability hours:
–

●

Any time or business hours only
Pre-determining online behavior:
–

●

Browse the web or prepare emails before the
access happens
Types of applications used:
–

●

–

Some applications are more suited for leisurely
at-home access - Facebook

Measuring Success
●

Example – Cost of Access

Source: ITU

Measuring Success
●

Cultural and socio-economic affordances of
connectivity:
–

Content in local languages

–

Availability of e-Government services

–

e-Commerce

–

Supporting infrastructure: roads, banking (credit
cards)

–

Social affordances: connectivity with local and global
population; online social networks; networked
individualism

Holistic approach
Evaluate success!

Investigate existing solutions
identify obstacles and
true needs of our users

Develop technical solutions
with experts from target areas

Adaptive and resource-efficient rural area networks

More Related Content

What's hot (20)

Viewers also liked (19)

Similar to Adaptive and resource-efficient rural area networks (20)

Recently uploaded (20)

Adaptive and resource-efficient rural area networks