High Density Emulation Platform for Wi-Fi Performance Testing

Germán Capdehourata,∗, Germán Álvareza , Martı́n Álvareza , Pedro Porteiroa , Fernando Bagalciaguea
a Plan Ceibal, Avda. Italia 6201, Edificio Los Ceibos, 11500, Montevideo, Uruguay

Abstract
The IEEE 802.11 standard has become the basis of one of the most successful wireless communication technologies of all time.
Originally created to provide wireless connectivity for a few devices, a couple of decades later it may support thousands of users in a
single wireless LAN. This fact has made 802.11 a relevant research topic, and as it happens with other wireless technologies, many
of the work carried out is based on simulations. In particular, studies for scenarios with high user density are usually performed
this way, in many cases leading to conclusions which do not apply to real world situations. This mismatch can be due to multiple
factors, such as the specific protocol implementations or the hardware and drivers used. In this article we present a novel 802.11-
based testing platform, which aims to bridge the gap between simulations and the real world, in order to carry out research work
for typical high density scenarios. The platform is compatible with standard 802.11-based wireless cards on the market and it was
tested with two different radios, The validation metrics considered were the TCP throughput, the airtime utilization and the effective
data rate, with relative errors ranging from 0 up to 15%. The potential of the tool is illustrated with real world measurements from
two example use cases in education facilities: a school classroom and a conference room. The results indicate this might be the first
step towards an open platform to enable active Wi-Fi performance testing for large scale scenarios. Further emulation capabilities
are shown with different application tests already integrated to the platform, such as QoE tests for YouTube video playback or
e-learning platforms.
Keywords: IEEE 802.11, client emulation, performance testing

1. Introduction In this research area, as it happens with other wireless tech-

In the past decade, the great success of the IEEE 802.11 stan-
dard, popularly known as Wi-Fi, has been one of the key fac-
tors that has driven the massive deployment of wireless Inter-
net access worldwide. This technology, originally created to
provide a wireless solution for small-size LAN environments,
now supports thousands of clients in a single WLAN (e.g. large
conferences or stadiums [1]), becoming one of the most suc-
cessful wireless communication technologies of all time [2].
This evolution was obviously supported by much research on
the subject, which has also made 802.11-based systems one of
the most relevant research topics in wireless networks.
One of the major challenges for the 802.11 evolution is the
exponential increase in the number of devices and traffic that
it will have to support in the near future. As it was the case
in the past, with the explosion of smartphones and tablets in
the mass consumer market, the novel Internet of Things (IoT)
forecasts a new wave of significant growth in the number of
connected devices [3]. Every day new products arise, either
wearables or home sensors, and many of them are choosing
Wi-Fi as their preferred wireless connectivity technology [4].
This is a tough technical challenge for 802.11-based networks,
and many research efforts would be necessary to find efficient
solutions to support the larger amount of devices and higher
densities expected.
∗ Corresponding author: gcapdehourat@ceibal.edu.uy.
Preprint submitted to Ad Hoc Networks
nologies, much of the work carried out is based on simula-
tions. In particular, studies for large scale scenarios and high
device density are usually performed this way, majorly due to
the greater simplicity and scalability against the difficulty of
mounting suitable real world testbeds. In many cases this op-
tion is suitable, choosing carefully the proper simulator for the
particular purpose [5, 6, 7, 8]. However, there are also many
situations where the results obtained do not match with the per-
formance seen in practice, with real world implementations [9].
This mismatch can be explained by multiple factors such as
inaccurate radio channel modeling due to wrong assumptions
[10], improper parameters settings for the protocols implemen-
tation, or specific hardware and drivers issues which are not
considered in typical simulators [11].
Concerning high-density in 802.11-based networks, it is
clear that simulation provides a simpler option because of the
large amount of nodes necessary to implement a typical WLAN
network. Hence the research carried out on real hardware is
usually performed with few devices, but it is very uncommon
for large scale scenarios due to the inconveniences involved in
managing a large number of devices. We strongly believe that
this typical research practice may lead in some cases to quite
different results from what happens in real world environments.
To illustrate the problem an example is given below.
One of the most studied topics in 802.11 is the performance
of the MAC layer, which is based on the CSMA/CA medium
access algorithm. In this area, one of the typical questions
March 17, 2022
posed is to find out the protocol scalability as the client count in-
creases [12]. It is well known that when traffic is UDP and there
is no RTS/CTS, the transport layer capacity is equivalent to the
one at the MAC layer (neglecting the headings overhead), with
a continuous drop as the number of client rises. The situation is
quite different for TCP traffic, as we have two dynamics which
run simultaneously: the CSMA/CA algorithm at the MAC layer
and the TCP congestion control at the transport layer. This in-
teraction has a positive effect for high density scenarios, as the
TCP congestion control avoids the large number of stations to
contend simultaneously, which provides a nice scalability prop-
erty [13]. This effect seems to be ideal looking at several simu-
lation based research papers [14, 15, 16], where the aggregated
capacity hardly changes as the client count increases.
However, the real world results show something different.
In spite of the good scalability, there is still a non-negligible
capacity loss as the number of stations rises [17, 18]. In this
empirical observations coming from the Wi-Fi industry, the ex-
planation for the drop in the aggregate capacity seems to be the
rise of control packets when the number of devices increases.
Those control packets are related to energy saving issues and it
remains unclear if the effect has to do with the particular devices
involved in the tests, or if it is something which is possible to
generalize for any other Wi-Fi devices. This gap between sim-
ulations (and the analytical models derived from them) against
what reality shows, has a more significant impact with the evo-
lution of the IEEE 802.11 standard, due to the higher data rates
available and the larger densities supported by today WLANs.
An error in the capacity estimation of the WLAN cells may lead
to wrong sizing decisions and underperforming networks.
The previous is just an example of the possible wrong conclu-
sions induced by working with simulators. Other real world is-
sues such as capture effect or the carrier sense algorithm imple-
mentation, typically omitted by simulators or analytical models,
are also found in the literature [19]. This problem becomes ev-
ident looking at the Wi-Fi industry, where large-scale testing is
required to verify the networks operation [20], because it is not
possible to trust in any simulation to guarantee that things will
work in the same way in practice. This kind of stress tests serve
to analyze the network performance in a high-density high-load
situation, in order to determine how many clients and which
applications the network would be able to support. It is a suit-
able way to address extensive real world testing, except for one
point: managing hundreds of devices each time you perform
a test. This fact implies several inconvenient aspects: the lo-
gistics to transport the devices and other additional equipment
(e.g. spectrum analyzer), tens of technicians and several hours
of field work.
In this work we present a novel platform for effective and
reproducible performance analysis of 802.11-based systems,
which enables to easily generate a high number of clients in
the network. We believe that the proposed solution is a contri-
bution in order to bridge the gap between simulations and the
real world, in order to carry out research work for typical high
density scenarios. The low-cost portable solution presented of-
fers a flexible setup that can emulate dynamic topologies which
classical testbeds cannot due to their fixed design. The platform
2
proposed could be useful not only as a research tool, but also as
a performance testing solution for real world Wi-Fi infrastruc-
ture. The tool enables to conduct stress tests in a much easier
way and also helps to reduce costs and simplify the work of the
technicians involved. Moreover, the flexibility to integrate test
automation for different applications just developing the proper
new software module, allows to adapt the tool for other specific
requirements.
The platform performance, considered as the ability to em-
ulate the real devices behavior, was evaluated in a laboratory
environment with two different models of 802.11 radio cards.
The validation was carried out considering transport layer met-
rics such as the TCP throughput, and lower layers metrics such
as the airtime usage and the effective data rate. The results show
an accurately emulation of the real devices, with relative errors
between 0 up to 15%, which indicate this might be the first step
towards an open platform to enable active Wi-Fi performance
testing for large scale scenarios. The potential of the tool is il-
lustrated with real world field measurements from two example
use cases in the educational context: a school classroom and a
conference room. In the field of Wi-Fi networks deployment,
the last years have shown a switch between coverage-based de-
sign into a capacity-based approach, particularly focused on the
quality of experience (QoE) [21, 22, 23] of users. The tool
not only proved to be a useful solution to perform high density
Wi-Fi tests, but also enables future customization for different
purposes because of the system modular design. An example
of this possibility is shown below, with the integration of an
open-source software to estimate QoE for playback of YouTube
videos [24].
The rest of this article is structured as follows. In the next
section we present other works related to testbed development
and a short review of 802.11 emulation solutions. In section 3
the problem addressed is presented, defining the main require-
ments for the platform development. Next, section 4 illustrates
the design and development of the solution proposed, with one
subsection dedicated to the platform hardware and the other to
the software architecture and its main modules. In section 5 we
present the procedure and test results carried out to validate the
accurate performance of the platform. Two example use cases
are presented in section 6 which illustrate the potential of the
tool. Finally, section 7 concludes the article and presents sev-
eral new research lines.
2. Related Work
Several approaches are typically used to carry out research
in wireless networks, ranging from analytical models and sim-
ulation, to testbed developments, emulation and real world field
tests. For example, a lot of academic research in 802.11-based
networks is based on the popular open source simulator ns-31 ,
with a large community involved and available implementations
for almost any relevant wireless standard. There is a clear ad-
vantage of using simulation, which is the possibility to generate
1 ns-3: https://www.nsnam.org/.
hundreds of nodes in a network with only several lines of code
and execute tests in a short period of time. However, simu-
lators are typically developed for specific purposes with many
assumptions and simplifications of real world scenarios [25].
Depending on the particular research purposes, simulation may
become an unsuitable option because of its limitations [26, 27],
or in other cases still very difficult to develop the proper simu-
lator [28, 29].
One of the alternatives that has gained relevance in the area
is the development of real world testbeds (e.g. Roofnet [30]
and ORBIT [31], among others). The advantage for emula-
tion testbeds over simulation, is that tests are implemented with
real nodes, concerning firmware as well as hardware and real
protocols implementation. In this line, many efforts have been
made during the last decade in order to generate emulation so-
lutions for mobile ad hoc networks (MANETs) [32], vehicular
ad hoc networks (VANETs) [33] and wireless sensor networks
(WSNs) [34, 35, 36]. In other cases the efforts are particularly
focused on channel emulation [37], attending to the problems
related to the variability of the real world wireless channel. An-
other approach is to use directly end user devices in the tests,
but the academic research found based on this approach is typ-
ically carried out with very few devices [38, 39] or with fixed
deployments [40, 41].
Granelli et al. [42] follow a similar approach to our work
but focused on wireless mesh networks (WMNs), presenting an
open experimental toolkit for testbeds. The hardware is based
on off-the-shelf 802.11 routers (ALIX boards from PCEngines2
and Cambria boards from Gateworks3 ) supporting two or four
radio cards per node. The software stands on the popular open-
source operating system OpenWRT4 . Another testbed based on
OpenWRT is presented in [43], also made with off-the-shelf
hardware (Atheros boards), with several software generated
clients but using only one radio interface. Most of the works
found in the literature which operate with off-the-shelf 802.11
hardware use radio cards mostly coming from backhaul solu-
tions, but none of them includes the typical radio cards found
in end user devices such as commercial laptops.
If we look at commercial testing tools in the industry, there
are also some existing platforms related to our work. For exam-
ple, products like IxVeriWave5 or Landslide 6 , provide a com-
plete system to properly test a Wi-Fi network. However, these
are expensive solutions and developed with specific high quality
hardware, different from the common 802.11 radios included
in common end user devices. Thus, the results obtained with
these tools may present great differences from the real perfor-
mance that users will get in the same real world environment.
Moreover, this kind of solutions typically emulate the different
RF channel between transmitter and receiver by simply using
different power levels for the different clients, but still using
the same RF chain and a single antenna for all of them. This
2 PCEngines ALIX boards: https://pcengines.ch/alix.htm.
3 Gateworks Cambria boards: http://www.gateworks.com/product#cambria.
4 OpenWRT: https://openwrt.org/.
5 IxVeriWave from Ixia: https://www.ixiacom.com/es/products/ixveriwave.
6 Landslide from Spirent: https://www.spirent.com/Products/Landslide.
3
way, the actual physical channel, determined by the locations
of clients and APs, antennas position, orientation and polariza-
tion, is not properly emulated. As Wi-Fi beamforming tech-
nology evolves, this variables seem to be too important to be
ignored. While these products are suitable to verify compliance
with a certain standard and perform laboratory testing of dif-
ferent hardware, they are not the right approach if we want to
find out the real world performance of end user devices under
typical Wi-Fi deployed infrastructure. This fact is also reflected
in the approaches seen in the industry, where large scale stress
tests with real end user devices are a common way to carry out
performance testing [17, 18, 20].
To the best of our knowledge, this work is the first one aimed
in constructing a portable multi-client emulation platform for
performance testing of infrastructure based 802.11 networks,
particularly focused in large scale scenarios with high device
density. There are several highlights which differentiate the
proposal from previous works. While many previous testbeds
have worked with off-the-shelf hardware, this one integrates
common 802.11 radios included in typical end user devices.
Unlike other proposals which only support to emulate a few
devices, the platform proposed includes up to 32 devices in a
single emulation unit. Each of them has its own dedicated WiFi
hardware, which implies that each radio runs its own 802.11
state machine, avoiding the virtual emulation via software of
the MAC layer. In addition, the proposed flexible setup is able
to emulate different scenarios, which classical testbeds cannot
due to their fixed design. For example, it is possible to carry out
tests using several emulation units, easily reaching more than a
hundred active devices in the same test. Finally, the system is
based on a standard Linux kernel, which allows to easily inte-
grate other developments for different tests. An example of this
possibility is shown in section 4.2.
3. Problem Addressed and Platform Requirements
The main goal was to develop a platform which provides
an effective solution to accurately emulate a high number of
802.11 end user devices. In order to design and develop an ap-
propriate solution several requirements were established for the
system implementation, which were then taken into account to
select the particular hardware and software to build the plat-
form. The main points are listed below:
1. The system should be a cost-effective and portable solu-
tion for Wi-Fi performance testing.
2. The solution should emulate as accurate as possible the
hardware and software of typical end user devices such as
commercial laptops.
3. Each emulation unit should emulate at least 20 end user
devices.
4. The system should enable proper client MAC layer and RF
emulation, with one complete Tx/Rx chain per device.
5. Validation tests should consider different metrics, not only
from transport layer measurements but also from air traffic
captures.
 6. The platform should be flexible enough to enable the inte-
gration of new modules for the future evolution.

The development was carried out in the educational context,

in particular for the nationwide one-to-one program in Uruguay,
Plan Ceibal [44]. Thus, some of the points considered have to
do with this fact, while others are just generic requirements de-
fined. For example, the minimum of 20 end user devices corre-
sponds to the ability to represent at least one typical classroom
with a single emulation unit. This way we are able to perform a
complete WiFi infrastructure test in a school using several units,
one per classroom. In addition, although it was not specified as
a requirement, in order to validate the platform developed two
of the laptop models from Plan Ceibal were used.
The third requirement forced to use one 802.11 card per
client, with the corresponding cables, connectors and antenna,
per each emulated device. This ensures a far more realistic
lower layers emulation of end user devices than virtualization-
based approaches. Concerning the validation procedure, al-
though the main goal was to have accurate results for the
throughput measurements at the transport layer, we also con-
sidered that the observed behavior at the lower layers should be Figure 1: Scheme of the platform hardware.
compared.

3.1. Plan Ceibal laptop models
As mentioned before, two different laptop models used by
Plan Ceibal [45] were considered for the platform implemen-
tation, Magallanes 3 (MG3) [46] and Positivo BGH 11cle2
(11cle2) [47], which are the two most used devices in the edu-
cational program.
The first laptop has an Intel Atom microprocessor, 1 Gbyte of
RAM memory and 8 GB of flash SSD storage. Its wireless net-
work interface controller (WNIC) is the Intel Wireless-N 100
mini PCI express radio card [48]. The most important details
about the WNIC are:
The operating system of both laptop models provided by
Plan Ceibal to the students is Ubuntu7 , the popular open source
Linux distribution.
4. Design and Development of the Emulation Platform
The proposed system is composed by two main parts, the
hardware platform which supports a large number of 802.11
radio cards, and the software layer which integrates the different
modules for each particular application performance test.

• IEEE 802.11b/g/n Wi-Fi wireless adapter (only supports 4.1. Hardware

2.4 GHz).
• MIMO 1x2:1, so it only supports one spatial stream (1SS).
• The maximum datarate at the 802.11 physical layer is
72.2 Mbps for a channel width of 20 MHz.
The second model considered has an Intel Celeron micropro-
cessor, 2 Gbyte of RAM memory and 16 GB of flash SSD stor-
age. The WNIC is an Intel Wireless-N 7260 mini PCI express
radio card [49]. The most important details about the WNIC
are:
• IEEE 802.11a/b/g/n Wi-Fi wireless adapter (dual-band
2.4 GHz and 5 GHz).
• MIMO 2x2:2, so it supports up to two spatial streams
(2SS).
An scheme of the emulation system hardware is depicted in
Figure 1. Basically, it consists of a main computer with several
802.11 interfaces which are connected to their respective anten-
nas. The 802.11 WNICs selected were exactly the same as the
ones integrated in each laptop model considered, Intel Wireless-
N 100 and Intel Wireless-N 7260 respectively. The final setup
for one main box (see Figure 2) has 32 radios with their respec-
tive antennas, enabling the emulation of up to 32 clients per
unit, with a full dedicated RF chain for each of them.
In order to integrate a large number of 802.11 WNICs, a
motherboard with several PCI ports was selected. Then, each
PCI port was expanded to several miniPCI slots, which are the
standard connectors for 802.11 radio cards. For this purpose
we used x8 splitters8 , which provide 8 miniPCI slots for each
PCI port. Due to the number of PCI ports available in typical
off-the-shelf available motherboards, we reached a maximum

• The maximum datarate at the 802.11 physical layer is 7 Ubuntu OS: https://www.ubuntu.com/.
144.4 Mbps for a channel width of 20 MHz. 8 Amfeltec: http://amfeltec.com/splitters/.

4
of 32 radios supported in a single emulator unit. For novel ra-
dio cards which have different connectors (e.g. M.2 instead of
miniPCI), it is possible to adapt the solution just using slightly
different splitters.
Flexible arms for the antennas setup were designed, in order
to make the tool more portable for on-site measurements. Fig-
ure 2 shows the resulting customized case for the main box and
the folding arms, composed by several parts printed with a 3D
printer. The solution achieved consists of two 2.4 GHz/5 GHz
antennas, each of them with a 3-meter-length coaxial cable
RG174/U, ending in two QMA connectors9 . This setup sup-
ports any miniPCI 802.11 radio card with up to two Tx/Rx
chains (i.e. MIMO 2x2). For novel radio cards supporting more
spatial streams (e.g. MIMO 3x3) it is necessary to extend the
solution with extra connectors, cables and antennas, which is a
relatively easy adaptation. The long cables make it possible to
emulate the actual location of the clients, for example distribut-
ing them appropriately around the main box.
The connection between the mini PCI express 802.11 WNICs
and the antennas is solved with U.FL to QMA pigtails. The
U.FL connector is an standard RF connection included in
typical 802.11 radio cards, while the QMA allows to easily
plug/unplug each of the RF cables attached to each antenna.
As mentioned before, the final setup supports up to two spatial
streams, so the RF chains are used in a different way depending
on the WNIC considered. On the one hand the Intel Wireless-N
100 is MIMO 1x2:1, which means it only uses one RF chain
for transmission and both RF chains are used for reception di-
versity. On the other hand the Intel Wireless-N 7260 is MIMO
2x2:2, so it uses the two RF chains for both, transmission and
reception multiplexing respectively.
To complete the main computer, the rest of the components
are just standard computer elements, such as a microprocessor,
RAM memory, a hard disk and a power source. The compo-
nents selected for the final setup are detailed below:
• Motherboard: ASUS MAXIMUS VI EXTREME.
• Microprocessor: Intel Core i7-4790.
• RAM memory: CORSAIR Vengeance 16GB.
• Hard disk: SanDisk Solid State Drive 128G.
• Power source: CORSAIR CX600 600W ATX12V.
4.2. Software
The platform operating system is Ubuntu, in order to assure
that the wireless drivers are the same to the ones used by Plan
Ceibal devices. In addition, using a popular Linux distribu-
tion facilitates the appropriation of the tool by the community.
Moreover, Ubuntu already has drivers for most of the main Wi-
Fi radio manufacturers (e.g. Intel, Broadcom, Atheros, Realtek
and Marvell). In particular the driver used in our case was iwl-
wifi, which is the one that corresponds to Intel WNICs.
9 The QMA connector is a quick connect/disconnect version of the typical
RP-SMA connector, which is quite suitable for testing and lab equipment.
5
Figure 2: Customized case for the main box of the emulator and folding arms
for the antennas setup.
Several open source tools available, such as iPerf 10 , wget11 ,
PhantomJS12 and Pytomo13 , were used for the implementation
of different application tests. We believe this is not an extensive
list, and in fact many others available tools could be used as
basic bricks to construct new applications tests. This is another
advantage of the operating system selected, as it allows easy
integration of new modules. Next, each of the application tests
modules developed are detailed.
Stress Test. The first software module developed corresponds
to the automation of stress tests within a WLAN, which was
also the performance test used to validate the platform. Un-
like the other cases, this is the only one that requires additional
hardware to perform the test. Since the goal is to measure only
the WLAN capacity, it is necessary to have a local server in
the same LAN than the Access Points (APs), which serves as
traffic source. This module is based on iPerf, an open source
tool for active measurements of the maximum achievable band-
width on IP networks. The traffic type considered was TCP
downlink flows from the server to each of the wireless clients.
This can be thought of as the situation ocurring when multiple
stations are downloading large files simultaneously, but in this
case measuring the throughput at the transport layer. As we
have two hardware components for the stress test, the emulator
and the local server, two different software modules were de-
veloped for the test, one for each component. On the one side,
the emulator software module handles all the Wi-Fi related is-
sues in the test, such as the automatic association to the desired
wireless network, defined by the SSID and the WPA pre-shared
key. In addition, the proper routing is configured, as well as
enabling iPerf servers for each interface, which will receive the
traffic from each Wi-Fi radio card. On the other side, the local
server software module is in charge the test execution, gener-
ating one iPerf client for each wireless interface. Finally, the
10 iPerf:https://iperf.fr/.
11 GNU Wget: https://www.gnu.org/software/wget/.
12 PhantomJS: http://phantomjs.org/.
13 Pytomo: https://pypi.python.org/pypi/Pytomo.
corresponding test logs are processed in order to generate the
final test results, such as the average total throughput and the
bandwidth sharing between clients.
Web Browsing and File Downloading. The next software mod-
ule developed and integrated to the platform was based on wget,
a GNU free software package for retrieving files using the most
widely-used Internet protocols such as HTTP, HTTPS and FTP.
In this case the goal was to emulate two different common ap-
plications: web browsing and file downloading. A simple com-
mand line interface (CLI) was developed, which allows first to
select between both options, and then to choose from a prede-
fined set of web pages or file repositories a particular test. Due
to Plan Ceibal’s needs, the web lists implemented consisted of
the main web portal and other relevant pages with educational
content, while the file sources included were a digital library
and a repository used for software updates for Plan Ceibal lap-
tops and tablets. The module also allows to use customized
lists of pages or files repositories for the tests. In both cases,
web browsing and file sharing, the resulting test measurement
consist of the time elapsed for each client to get the desired con-
tent. The test results are summarized in a report which includes
several overall statistics such as the average time between all
clients and the worst case registered.
e-Learning Platform. In this case the goal was to develop an-
other module which enables performance evaluation of an on-
line educational platform for mathematics. While working with
the previous module a wget limitation was detected, which
is the inability to support JavaScript. As many other online
platforms, the Intelligent Tutoring System (ITS) used by Plan
Ceibal14 , aimed for the adaptive learning of mathematics, is
based on this language. The solution found is based on Phan-
tomJS, a tool for web testing which supports JavaScript and im-
plements the full web stack, so no browser is required. In a sim-
ilar way than in the previous case, a CLI was developed, which
allows to select different options from a list of typical tours in
the ITS, which always include the following steps: select the
main topic, select a book of exercises, and access to a particular
exercise. In both cases, the resulting test measurement con-
sist of the time elapsed for each client to do the complete set
of steps of the selected tour, up to the complete loading of the
selected exercise. It is well studied that for the case of online
platforms, this type of metric based on the user waiting times,
is directly related to the quality of experience (QoE) perceived
by the user [22]. As in the previous case, a summary with the
main statistics is included in the test report.
YouTube Video Playback. The last module integrated to the
platform is based on a third party development, which was
adapted to run in the emulator. The credit corresponds to Juluri
et al. [24] which developed a tool called Pytomo for analyzing
the playback quality of YouTube videos. This tool allows to es-
timate the QoE for YouTube providing three main performance
indicators: start time, number of interruptions and buffering
14 The platform is provided by http://bettermarks.com/.
6
time. The first one corresponds to the time elapsed between
selecting a video and the moment when the playback finally
starts. The second one is the number of times that the video
freezes during the playback, while the last one is the amount of
time that the video remains stalled.
While Pytomo provides the complete solution for YouTube
videos playback performance testing, the tool was intended for
a single client, so it only supports a single interface with In-
ternet access in the computer which runs the test. In order to
integrate the tool to the emulator, we have introduced minor
modifications which enable a multi-client test. The resulting
new module allows to select a list of YouTube videos which are
assigned to each wireless interface. Thus, it is possible to test
the simultaneous YouTube video playback for all the wireless
emulator interfaces, based on the same core of the original Py-
tomo code. Finally, the three original Pytomo metrics for each
client are taken into account to register the test results. We also
generated a new metric, which integrates the previous three, in
order to have a single value to evaluate the QoE, in the same
way that typical mean opinion score (MOS) metrics.
In order to estimate the QoE we defined a proper exponential
function combining the three indicators provided by Pytomo,
which are the typical perceptual models found in the literature
[22]. To avoid the hard work involved to carry out the tests for
learning the proper function to compute a mean opinion score
(MOS), we opted for a simpler option in this case. We con-
sidered each of the Pytomo indicators as an independent fac-
tor affecting the QoE, and included an exponential term to take
into account its influence in the global QoE estimation. It is
important to notice that this function should vary between 1 to
5, which is the scale used for MOS metrics. Then, we forced
certain values for particular situations, such as best case (no de-
lay to start, no interruptions) and worst case (where maximum
tolerable values were defined for each parameters). The test du-
ration was predefined to 30 seconds (Pytomo default value) and
a fixed video quality of 720p, because Pytomo does not support
DASH15 and we assumed this quality is enough for education
purposes.
Finally, the resulting QoE model was:
 
QoE = f (S T, NI, DI) = max 1, QoE f
     
QoE f = 5/8 · 1 + e−S T/C1 · 1 + e−NI/C2 · 1 + e−DI/C3
where ST stands for start time, NI for number of interruptions
and DI for duration of interruptions respectively. In order to
adjust the model, extreme worst cases were predefined for the
three indicators. When each of the parameters takes the worst
case value, the resulting QoE was enforced to be 1, no matter
what the value of the other two was. The selected thresholds
were: a starting time of 10 seconds, 6 interruptions and a total
buffering time of 10 seconds. Imposing this condition leads
to the following values for the model constants: C1 = 5.64,
C2 = 3.39 and C3 = 5.64.
15 DASH: Dynamic Adaptive Streaming over HTTP.

Figure 3: Setup for the validation test.
We highlight this last case, showing the integration of a tool
such as Pytomo into the proposed emulation platform, as a clear
example of the flexibility and versatility which the tool can of-
fer. We believe this is a key feature which enables to easily
integrate many other developments, with a larger exploitation
of their potential, using this tools for taking measurements in
high density environments. To better illustrate this module and
the potential of this type of integrations, a use case scenario is
presented in Section 6, where the test results for a particular
network are presented.
5. Platform Validation Tests and Results
The validation test to measure the emulation accuracy of the
developed tool was done with the setup shown in Figure 3. The
application considered was an stress test, with simultaneous
TCP downlink traffic flows generated with iPerf. The tests were
conducted in the same way for both laptop models, comparing
the results between the real and the emulated versions in both
cases. The test details are presented in Table 1.
The tests were done in a classroom located in an office en-
vironment, with controlled interference conditions. The chan-
nel was selected as the least occupied from the standard Wi-Fi
channels in 2.4 GHz (1, 6 or 11). For this purpose we used the
WiSpy spectrum analyzer16 , controlling that the channel utiliza-
tion due to external interference was always below 10%. The
antennas were disposed over the desktops in the same places
as the laptops throughout the room, thus emulating the same
client distribution for both cases. This way we ensure similar
conditions in both tests.
5.1. Validation Metrics
Two type of measurements were considered to compare the
behavior of real clients versus emulated clients:
• Higher layer
TCP throughput - Total amount of transferred data
per time unit at the transport layer.
16 WiSpy: http://www.metageek.com/products/wi-spy/.
7
Validation test details
Access Point Cisco Aironet 2702
Local server miniPC Giada17
Laptops models MG3 and 11cle2
802.11 radios Intel N100 and Intel N7260
Frequency / Channel width 2.4 GHz / 20 MHz
Traffic type TCP downlink
Test duration 5 minutes
Table 1: Configuration details for the validation test.
• Lower layer
Airtime utilization - Time percentage during which
the channel is occupied by the AP or the clients.
Effective data rate - Average transmission data rates
over the air weighted by the corresponding packet sizes.
The first one was computed with the iPerf logs for each client,
while the latter were obtained from air packets traffic captures.
For this purpose we used a specific device (in this case an Apple
Macbook Pro18 ) running the standard capture tool Wireshark19 ,
because standard USB 802.11 adapters only support one spatial
stream and are not able to capture packets transmitted using
multiple spatial streams.
In order to avoid the TCP transient and synchronization ef-
fects, the first and last minute from the 5-minute-length tests
were discarded for computing performance metrics. This way
the simultaneous stationary operation of all the clients was en-
sured during the interval considered. Thus, all airtime con-
sumption measures are expressed as a percentage of the 3-
minute-length subpart of the test. In addition, this long run
guarantees the statistical validity of the test results, as they can
be seen as an average between several shorter tests.
In addition to the aforementioned metrics, other indicators
were computed from test measurements. The distribution be-
tween clients was studied, analyzing both levels, on the one
hand the bandwidth sharing and on the other hand the airtime
consumption per client. Finally, the time consumed by retries
and the transmission data rates histograms were also compared
for all tests.
5.2. Test Results for laptops MG3 - Intel N100 cards
The validation tests for the laptop model MG3 were con-
ducted with 23 clients, as one of the laptops failed at the mo-
ment of carrying out the tests. In order to have the same number
of clients for the comparison, one of the emulator WNICs was
turned off, to operate only with 23 clients also. The higher layer
results, which correspond to iPerf logs, are analyzed first. The
total average results are summarized in Table 2. The aggregated
TCP traffic was the same in both cases, while the variations be-
tween clients were larger for laptops than for the emulator. This
result is ratified by Figure 4, which shows the average TCP
18 Macbook Pro 13: http://www.apple.com/macbook-pro/.
19 Wireshark: https://www.wireshark.org/.
 Total TCP
Average per Std. Dev.
Throughput
client (Mbps) (Mbps)
(Mbps)
Laptops 16 0.7 0.50
Emulator 16 0.7 0.35 Airtime Std. Dev. Retries Eff. data
(%) (%) (%) rate (Mbps)
Table 2: Throughput results for emulation of laptops MG3. Laptops 67.7 1.9 9 43.5
Emulator 70.1 1.1 14 41.8

Table 3: Airtime consumption results for emulation of laptops MG3.

Figure 4: Throughput TCP comparison for emulation of laptops MG3.

throughput per client in ascending order for both laptops and

emulator.
While a similar general behavior is observed for both cases,
it can be noticed that the laptops with larger throughputs are
above the corresponding emulator clients, and the opposite hap-
pens with the laptops with smaller throughputs, which are be-
low the emulator clients. There is no clear explanation of this
bandwidth sharing difference, but one possible hypothesis is Figure 5: Airtime per client comparison for emulation of laptops MG3.
that the hardware variations among different laptops (even be-
ing all of the same model) could produce this effect, while the
emulator clients are all over the same base hardware, so there is
no possible hardware variability in that case.
Concerning air packet captures, Table 3 summarizes the main
results. Again, a larger variation between clients is observed for
the laptops, which is more clear looking at the airtime consump-
tion per client shown in Figure 5. Most parameters are similar
for both cases, with the exception of the airtime consumed by
retries, which is larger in the emulator test. On the other hand,
the airtime consumed by control frames is quite similar in both
tests, with 24% for the laptops and 26% for the emulator.
In Figure 6 the data rates histograms are compared, showing
an intersection of 84.3%, which means that most of the time
the operation is similar in both cases. These results indicate
that the behavior is quite similar, not only looking at the higher
layer considering throughput results, but also taking into ac-
count what happens over the air.We believe this is important to
validate the emulator behavior and ensure that the results for
different test scenarios are similar to the ones that would be ob-
tained with the laptops. Figure 6: Data rate histogram comparison for emulation of laptops MG3.

5.3. Test Results for laptops 11cle2 - Intel N-7260 cards

The validation tests for the laptop model Positivo 11cle2
were conducted with 24 clients, with the emulator operating
with the full setup for 3 PCI to mini PCI splitters, each of them
8
 Total TCP
Average per Std. Dev.
Throughput
client (Mbps) (Mbps)
(Mbps)
Laptops 44.2 1.8 0.82
Emulator 52.0 2.2 0.46

Table 4: Throughput results for emulation of laptops 11cle2.

Airtime Std. Dev. Retries Eff. data

(%) (%) (%) rate (Mbps)
Laptops 65.6 1.4 5 111
Emulator 73.6 0.7 4 118
Figure 8: Airtime per client comparison for emulation of laptops 11cle2.
Table 5: Airtime consumption for emulation of laptops 11cle2.

with 8 Wi-Fi WNICs. The first analysis corresponds to higher

layer performance. In Figure 7 the average TCP throughput per
client for both laptops and emulator is shown, while Table 4
summarizes the aggregated averages results. In this case the to-
tal aggregated throughput was superior for the emulated clients,
with a relative error of 15%. As for the previous laptop model,
the variations between laptops throughput is larger than for the
emulator clients. In particular, the difference is more noticeable
for the laptops which get smaller throughputs, which are clearly
below the corresponding emulated clients.
Looking at the behavior over the air, the results are summa-
rized in Table 5. The laptops present again a higher variability,
which becomes clear with the airtime consumption per client
shown in Figure 8, and is consistent with higher layer results.
In this case the airtime consumed by retries is almost the same
in both cases, while the airtime corresponding to control frames
is also quite similar in both tests, with 24% for the laptops and
22% for the emulator.
Concerning the packet data rates, Figure 9 shows the two
histograms, for both laptops and emulated clients. In this case
they present an intersection of 81.7%, a little bit less than for
the previous laptop model, but still with a high degree of co-
incidence. As in the previous case for the MG3 laptops, now
for the Positivo 11cle2 model the measurements taken from air
captures confirm a similar behavior for both laptops and emu-
lated clients.
Figure 7: Throughput TCP comparison for emulation of laptops 11cle2.
9
Figure 9: Data rate histogram comparison for emulation of laptops 11cle2.
5.4. Interference Validation Test
The previous validation tests prove that the behavior of lap-
tops and the emulator is quite similar, which is enough to con-
sider the emulator as a suitable tool in order to make perfor-
mance measurements in a single cell scenario. An additional
test is necessary to validate the operation of the emulator in
multiple simultaneous channels. In a single cell scenario, with
only one 20 MHz channel, there is no possible interference
as only one radio can transmit at a time, due to the 802.11
medium access mechanism. However, if we have the emula-
tor connected to two different APs in different channels, some
radios could transmit simultaneously and eventually generate
interference between them. In order to validate that there is no
interference between the emulator radios, we did a simple test
with only the dual band Intel N7260 radios.
Two APs were placed two meters apart, operating in two dif-
ferent 20 MHz non-overlapping channels in the 5 GHz band.
We repeated the same stress test than before but for two differ-
ent situations. First, we measured the capacity with only one
AP turned on and 16 active radios in the emulator. Then, we
measured the capacity with both APs turned on and the emula-
tor with 32 active radios, connecting half of them to each AP.
For the first case the average aggregated capacity measured was
66.8 Mbps, while the second test result was 130.1 Mbps. Basi-
cally the results show that the capacity measured for two cells
is almost twice the one for a single cell. This way we prove that
the interference between wireless cards in the emulator does not
affect the measurements for multiple cell scenarios. Notice that
the results for a single cell are well above the ones obtained in
the previous validation test. This is explained because the test
took place in other date and location, in a more clean channel
in the 5 GHz band with almost no external channel utilization
and also with less amount of devices.
5.5. Comparison with analytical models and simulations
Finally, we compare the validation tests results with the ca-
pacity calculated with analytical models [16] and with ns-3 sim-
ulations [50]. For both the analytical models and the simu-
lations, it is necessary to define the 802.11 operation param-
eters, as well as other variables not specified in the standard
and which depends on the equipment manufacturers implemen-
tations. As an example of the latter we can list the rate adapta-
tion mechanisms, the 802.11n aggregation policy, the usage of
block ACKs and the fairness policy of the AP, among others.
That said, we will compare the estimates provided by both the
theoretical models and the simulations for the tested scenarios.
For the 1SS capable radios the maximum rate at the phys-
ical layer is 72.2 Mbps, for which analytical models for TCP
estimates a maximum capacity ranging from 28 Mbps up to
45 Mbps, depending on the aforementioned defined variables.
The same test scenario simulated in ns-3 results in a TCP ag-
gregated capacity of 20 Mbps for the worst case, up to more
than 50 Mbps for the most efficient configuration possible (i.e.
with maximum packet aggregation). The case for the 2SS ca-
pable radios is analogous, but now with a maximum rate at the
physical layer of 144 Mbps. This results in estimations that are
twice the ones for the previous case, but again with a great pos-
sible range depending on the particular configuration selected.
In both cases the validation test results were below the maxi-
mum capacity estimates. A possible explanation is the perfect
scalability that most models assume for TCP over 802.11 [13],
which seems not valid in real world scenarios [17, 18].
Although it is possible to adjust the previous estimates, it
is still quite difficult to select all the proper parameters to ac-
curately represent the test scenario. As mentioned above, the
specific aggregation policy used by the AP is a proprietary al-
gorithm, so it can only be analyzed by some kind of reverse
engineering. The same happens with the rate adaptation and
the fairness policy introduced by the AP to reach client airtime
fairness. All these real world variations are not easy to include
in analytical models or simulators. There are so many 802.11
chipsets and proprietary algorithms from different vendors, that
we believe the proposed emulation framework is more suitable
as a tool for real world measurements. Our proposal assures
accurate results for the particular equipment that we are dealing
with, in a much simpler way than fitting a complex model to a
specific scenario and equipment.
6. Use Case Examples
In this section two example use cases in the Plan Ceibal ed-
ucational context are presented. In the first case, one emulation
unit is used to take measurements in a typical school classroom.
In the second example, several emulation units are used to mea-
sure the Wi-Fi capacity in a conference room, generating more
10
Num. of Avg. Tput Std. Dev. Agg. Tput
Clients (Mbps) (Mbps) (Mbps)
1SS-T1 24 1.8 1.2 43.5
1SS-T2 29 1.4 1.2 41.4
2SS-T1 32 6.2 5.7 199
2SS-T2 32 5.4 5.0 173
Table 6: School classroom test for 1SS and 2SS capable cards. T1 was during
winter holidays and T2 in a regular school day.
than one hundred active devices. In both cases the Wi-Fi in-
frastructure corresponds to Cisco APs Aironet 2702 and run-
ning with a centralized wireless LAN controller (WLC) from
the 7500 series.
6.1. School classroom
In order to test the platform in the field, and also attending
to specific Plan Ceibal needs in a primary school, several tests
were carried out to measure the Wi-Fi capacity in a particu-
lar classroom. The measurements were taken in two different
dates, the first one during the winter holidays, and the second
one in a regular day of class. This way, the results enable
to compare the impact of the interference generated by other
devices in neighboring classrooms, and other Wi-Fi networks
from buildings next to the school. The traffic was generated
from a local server located in the same LAN as the APs, in the
same way that in the validation tests presented in Section 5.
The classroom tested does not have APs installed in the
room, so the Wi-Fi infrastructure corresponds to the APs in-
stalled in the neighbouring rooms. The emulator association
process is exactly the same as the one of the real end user de-
vices. Typically the AP selection is not controlled by the user,
who only selects the network ESSID20 . Thus, the selected AP
ends up depending on the signal strength and the particular card
and driver implementation. It is important to clarify that this
case does not correspond to a single WLAN cell test, since
the school has several APs installed, reason why the clients are
distributed among them during the association process. Thus,
this test reproduces the real classroom scenario with the same
amount of laptops than emulated clients, evenly distributed in
the room, since they associate in the same fashion to the net-
work.
The tests results are summarized in Table 6, comparing the
results between a winter holidays day and a regular school day.
Two different emulators were used in the tests, one with the sin-
gle band 1SS capable radios Intel N100 and the other with the
dual band 2SS capable radios Intel N7260. Each result corre-
sponds to one stress test, which lasts five minutes as described
in Section 5. As we can see, the results for the single band 1SS
radios were quite similar for both tests, while some aggregated
capacity decrease is observable for the case of the dual band
2SS radios. Thus, the results indicated there was no significant
impact of the interference generated on a regular school day,
compared with the situation during holidays.
20 ESSID: Extended Service Set Identifier, which defines the network name.
 Figure 10: Throughput per client in classroom tests for 2SS capable radios.
Looking more closely at the test results for the 2SS capa-
ble radios, Figure 10 shows the TCP throughput achieved by
each client, sorted in ascending order. The multi-cell scenario
with several APs may generate imbalanced client distribution
between them, which is reflected in the bandwidth sharing in-
equalities between clients observed, and could also explain the
difference between the tests. This fact leads to some clients
getting a capacity well above others, as they are sharing the cell
with fewer devices.
The tests results also illustrate the significant difference be-
tween devices which support both frequency bands (2.4 GHz
and 5 GHz), compared to the ones that only operate in 2.4 GHz.
In addition, the wireless network at the school uses 40 MHz
channels for the 5 GHz band, while 20 MHz is used for
2.4 GHz, where less channels are available, which makes more
difficult to maintain a low CCI21 . This fact evidences the prob-
lem of dealing with single band devices in high density sce-
narios, where not having dual band devices can make almost
impossible to provide connectivity that meets the requirements
of typical applications.
6.2. Conference room
In this second example the measurements correspond to a
conference room with a maximum allocation between 100 and
200 people. Several emulation units were used simultaneously
in this case, in order to carry out a capacity test with more than
100 active devices, in a total approximate space of 150 m2. The
wireless infrastructure installed in the room consists of 5 dual
band APs, two of them operating only in 5 GHz, while the oth-
ers have active radios in both bands. The channel width config-
uration was 20 MHz in the 2.4 GHz band and 40 MHz in 5 GHz.
Figure 11 shows the antennas distribution throughout the room,
which were uniformly placed in all the available space.
Different stress tests were carried out, combining the emula-
tors unit equipped with single and dual band radio cards. The
results are summarized in Table 7 detailing how many clients
21 CCI: Co-Channel Interference.
11
Figure 11: Partial view of the conference room during the tests.
of each type were active in each test. Looking at the results of
the different tests it is possible to notice some useful insights.
For example, comparing the total capacity for tests T1 and T2
we found quite similar results, which illustrates the tremendous
difference between the traditional 2.4 GHz Wi-Fi band versus
the less congested 5 GHz band, with more available channels.
Another interesting result was found comparing tests T3 and
T4, where exactly the same type of clients were used in both
tests (dual band 2SS capable devices), but located in the op-
posite side of the room. In test T3 the antennas were located
in the northern part of the room, where the building ends and
no neighbouring buildings are present. On the other hand, in
test T4 the antennas were in the opposite side, where the build-
ing has other Wi-Fi networks operating and also neighbouring
buildings with their own networks. This fact has a large impact
in the total capacity obtained in each case, with almost doubling
the capacity in the sector of the room less affected by interfer-
ence.
In Figure 12, the TCP throughput achieved by each client is
shown for the case with 111 active client devices (test T1), 44
devices operating only in 2.4 GHz and 67 dual band devices
in 2.4 GHz and 5 GHz. Looking at the results it is possible
to identify the different groups of clients operating in the same
cell, which can be associated with each step in the curve, as
the bandwidth sharing within a cell tends to be equal to each
client due to the standard 802.11 operation. Thus, the unfair
sharing observed probably corresponds to the different client
capabilities (e.g. single band versus dual band, 1SS versus 2SS)
and the number of devices sharing the same cell, defined by one
radio from one access point (as each AP has two radios, one on
each frequency band).
Besides the stress tests, we also carried out the evaluation of
specific applications, using the tools introduced in Section 4.2.
In particular, the quality of experience (QoE) of YouTube
videos playback was estimated, with the module successfully
integrated to the platform based on Pytomo [24]. As it was
previously mentioned, three different indicators are reported by
 Figure 14: Platform integration example for YouTube QoE estimation with py-
Figure 12: Stress test results for more than 100 simultaneous active clients. tomo.

Single Band Dual Band Avg. Tput Agg. Tput

Devices Devices (Mbps) (Mbps) we can see the resulting estimation of the YouTube video play-
T1 44 67 6.5 718 back QoE, integrating the three different indicators reported by
T2 0 64 11.3 725 Pytomo. The results are consistent with the network user expe-
T3 0 32 18.3 586 rience, with most of the QoE values ranging from 4 to 5, which
T4 0 32 10.1 323 is an excellent QoE.
T5 28 0 2.2 61

Table 7: Conference room stress test results. 7. Conclusions and Future Work

In this work we present the design, development and vali-

Pytomo, start time, number of interruptions and buffering time,
which were combined in a suitable way to obtain a QoE estima-
tor, similar to the typical mean opinion score (MOS) metrics.
Figure 13 shows the resulting histogram for the start time
after 10 tests carried out with Pytomo integrated to the emu-
lator. During every test each of the 32 radio interfaces were
simultaneously doing YouTube video playback during 30 sec-
onds, with a fixed video quality of 720p for all of them. It is
important to note that the network under test has a suitable in-
ternet access with a 120 Mbps downlink capacity. This fact is
reflected in the results, which show quite acceptable values for
the start time, ranging from 0 up to 3 seconds. In Figure 14
Figure 13: Platform integration example with pytomo for YouTube testing:
Start Time histogram for 10 tests.
12
dation of a Wi-Fi multi-client emulation platform. The system
developed enables to carry out performance testing in high den-
sity Wi-Fi environments, with an accurate emulation of the real
world end user devices. The platform core is composed by a
main computer with several Wi-Fi WNICs and a dedicated RF
chain for each emulated client. The proposed solution is an
important contribution to bridge the gap between Wi-Fi simu-
lations and the real world, in particular for the research focused
on high density scenarios. We believe the platform is useful
for several scenarios, not only for research purposes, but also
as testing tool for real world Wi-Fi deployments. In addition,
the flexible setup of the solution presented enables the emula-
tion of different network topologies, which classical testbeds
cannot due to their fixed design.
The emulation platform was validated for two different Wi-
Fi radio cards. The validation tests were conducted with more
than 20 laptops in the 2.4 GHz band. Several validation metrics
were taken into account, such as TCP throughput, the airtime
consumption and the effective data rate, achieving an accept-
able accuracy for all of them, with relative errors ranging from
0 up to 15%. The validation results guarantee that performance
evaluation conducted with the emulator will be accurate enough
for many research and testing purposes, while avoiding the in-
convenience related to the large amount of devices involved in
the tests. Different use case examples were also presented, in-
cluding one with more than one hundred active clients. This
case scenario is typically tackled by the research community
with simulation tools, mainly because of the lack of an easy and
inexpensive way of carrying out this kind of real world exper-
iments. The proposed platform viabilizes these large scale ex-
periments, which are very important for analyzing the WLANs
behavior in high density scenarios.
Among other advantages, the developed system is based on
a open source Linux distribution that can be adapted for em-
ulating specific client devices. This characteristics make it a
flexible solution and its modular design allows other users to
customize it, for example by changing the WNICs or the cor-
responding antennas, and also developing their own software
for other tests. We believe this is an important property of the
platform developed. In addition, as all the system runs on a
standard Linux environment, where every wireless card is just
seen from the operating system as a standard network interface,
it is quite easy to integrate other developments. This fact en-
ables to conduct novel tests for specific purposes, that run also
over the same hardware, which allows to make profit of the high
density client emulation capability provided by the tool.
This is the case of the example presented with Pytomo, a
tool for analyzing playback quality of YouTube videos, which
was successfully integrated to the emulation platform with only
some minor changes in order to operate with several wireless
interfaces. Thus, the resulting tool enables to perform large
scale YouTube tests to estimate the QoE of different users, em-
ulating many clients watching videos simultaneously. A use
case example was presented, with results gathered from a real
world operating network. We are not aware of any other exist-
ing platform that allows to do this kind of tests. In addition, this
example shows how easy is to integrate other type of tests to the
platform, just selecting the proper software measurement tool.
Regarding future work, new research lines have already been
discussed and started to be explored. Several improvements
to the existing platform were identified, such as industrial de-
sign modifications for a better tool usage in the field. We also
analyzed the possibility to collect more data during the tests,
particularly from the wireless environment, which could be an-
other important source of information to understand the high
level metrics obtained in a particular experiment. Other possi-
ble use cases of the platform could be explored, such as per-
formance evaluation of different Wi-Fi WNICs and different
wireless drivers, in order to characterize the 802.11 standard
implementation typically found in off-the-shelf Wi-Fi devices.
Another research line analyzed is the addition of virtualiza-
tion capabilities to the platform. The final goal would be to
be able to emulate the operating system and applications of
the end user devices, rather than only the specific Wi-Fi hard-
ware. For this purpose the emulation platform should run a
virtual machine (VM) for each device, with the same operating
system than the real device. One problem to be addressed is
to find the best way to assign the wireless interfaces for each
VM, which affects directly how each VM detects the WNIC
and which drivers are used. Although a more expensive hard-
ware is necessary to support a large number of VMs (e.g. 32),
we believe it could worth it, because of the broad range of new
possibilities that this feature may offer. The first experiments
carried out indicate that it could be very useful to have this new
software layer, in order to achieve a more powerful high density
emulation platform.
13
References
[1] Extreme Networks, Super Bowl XLVIII Stats Infographic, ???? URL:
http://www.extremenetworks.com/super-bowl-stats/.
[2] W. Lemstra, V. Hayes, Unexpected innovation: The case of Wi-Fi, in:
Proceedings of EuroCPR2008, 2008, pp. 1–30.
[3] Cisco, Cisco Visual Networking Index: Global Mobile Data Traffic Fore-
cast Update, 20152020, 2016.
[4] A. Pyattaev, K. Johnsson, S. Andreev, Y. Koucheryavy, Communica-
tion challenges in high-density deployments of wearable wireless devices,
IEEE Wireless Communications 22 (2015) 12–18.
[5] M. Jevtic, N. Zogovic, G. Dimic, Evaluation of wireless sensor network
simulators, in: The 17th Telecommunications Forum, TELFOR 2009,
2009, pp. 1303–1306.
[6] S. Mehta, N. Ullah, M. H. Kabir, M. N. Sultana, K. S. Kwak, A case study
of networks simulation tools for wireless networks, in: 2009 Third Asia
International Conference on Modelling Simulation, 2009, pp. 661–666.
[7] B. Musznicki, P. Zwierzykowski, Survey of simulators for wireless sensor
networks, International Journal of Grid and Distributed Computing 5
(2012) 23–50.
[8] L. Carla, F. Chiti, R. Fantacci, T. Pecorella, Chapter 15 - current and
future trends in open source network simulators for wireless systems, in:
M. S. Obaidat, P. Nicopolitidis, F. Zarai (Eds.), Modeling and Simulation
of Computer Networks and Systems, Morgan Kaufmann, Boston, 2015,
pp. 451 – 464.
[9] R. Khattak, A. Chaltseva, L. Riliskis, U. Bodin, E. Osipov, Com-
parison of wireless network simulators with multihop wireless network
testbed in corridor environment, in: Proceedings of the 9th IFIP TC
6 International Conference on Wired/Wireless Internet Communications,
WWIC’11, Springer-Verlag, Berlin, Heidelberg, 2011, pp. 80–91.
[10] D. Kotz, C. Newport, R. S. Gray, J. Liu, Y. Yuan, C. Elliott, Experimental
evaluation of wireless simulation assumptions, in: Proceedings of the 7th
ACM International Symposium on Modeling, Analysis and Simulation of
Wireless and Mobile Systems, MSWiM ’04, ACM, New York, NY, USA,
2004, pp. 78–82.
[11] A. D. Stefano, G. Terrazzino, L. Scalia, I. Tinnirello, G. Bianchi, C. Gia-
conia, An experimental testbed and methodology for characterizing IEEE
802.11 network cards, in: 2006 International Symposium on a World of
Wireless, Mobile and Multimedia Networks(WoWMoM’06), 2006, pp. 6
pp.–518.
[12] G. Bianchi, Performance analysis of the IEEE 802.11 distributed coordi-
nation function, IEEE Journal on Selected Areas in Communications 18
(2000) 535–547.
[13] S. Choi, K. Park, C. kwon Kim, Performance impact of interlayer depen-
dence in infrastructure WLANs, IEEE Transactions on Mobile Comput-
ing 5 (2006) 829–845.
[14] G. Kuriakose, S. Harsha, A. Kumar, V. Sharma, Analytical Models for
Capacity Estimation of IEEE 802.11 WLANs Using DCF for Internet
Applications, Wirel. Netw. 15 (2009) 259–277.
[15] J. Yu, S. Choi, D. Qiao, Analytical modeling of TCP dynamics in
infrastructure-based IEEE 802.11 WLANs, Journal of Communications
and Networks 11 (2009) 518–528.
[16] F. Larroca, F. Rodrı́guez, An overview of WLAN performance, some im-
portant case-scenarios and their associated models, Wirel. Pers. Commun.
79 (2014) 131–184.
[17] C. Lukaszewski, L. Li, Very High-Density 802.11ac Net-
works Theory Guide - Validated Reference Design, 2015. URL:
http://www.arubanetworks.com/assets/vrd/Aruba_
VHD_VRD_Theory_Guide.pdf.
[18] No Strings Attached Show, Cisco AP 2700 Multi Client Test,
2015. URL: http://nsashow.com/AP2700/Cisco_AP2700_
Multi_Client_Test.pdf.
[19] J. Lee, J. Ryu, S.-J. Lee, T. T. Kwon, Improved modeling of IEEE 802.11a
PHY through fine-grained measurements, Computer Networks 54 (2010)
641 – 657. Advances in Wireless and Mobile Networks.
[20] Wireless LAN Professionals, Wi-Fi Stress Test, 2013. URL:
http://www.wlanpros.com/wp-content/uploads/2015/
06/Wi-Fi-Stress-Test-Report.pdf.
[21] B. Lewcio, S. Mller, Perception of quality changes in wireless networks,
in: Quality of Experience, Springer, 2014, pp. 395–410.
[22] R. Schatz, T. Hossfeld, L. Janowski, S. Egger, Datatraffic monitoring and
analysis, Springer-Verlag, Berlin, Heidelberg, 2013, pp. 219–263.
[23] B. Chihani, K. ur Rehman Laghari, E. Bertin, D. Collange, N. Crespi, [43] S. Dutt, D. Habibi, I. Ahmad, A low cost atheros system-on-chip and
T. Falk, User-centric quality of experience measurement, in: Mobile openwrt based testbed for 802.11 wlan research, in: TENCON 2012
Comput., Appl., and Serv., volume 130, Springer, 2014, pp. 33–46. IEEE Region 10 Conference, 2012, pp. 1–4.
[24] P. Juluri, L. Plissonneau, D. Medhi, Pytomo: A tool for analyzing play- [44] E. Moscatelli, The success of Plan Ceibal, Master’s thesis, University of
back quality of youtube videos, in: 2011 23rd International Teletraffic Texas at Austin, USA, 2016.
Congress (ITC), 2011, pp. 304–305. [45] Plan Ceibal, Plan Ceibal: One Laptop per Child implementation in
[25] R. Skehill, P. Scully, S. M. Grath, Characteristics, results and findings Uruguay., 2007-2015. URL: http://www.ceibal.org.uy/.
of IEEE 802.11 in an RF isolated testbed, in: 2007 IEEE 18th Interna- [46] JP Sá Couto, Laptop Magallanes 3, 2013. URL: http://www.jp-ik.
tional Symposium on Personal, Indoor and Mobile Radio Communica- com/.
tions, 2007, pp. 1–5. [47] Informatica Fueguina SA, Laptop Positivo BGH 11cle2, 2013. URL:
[26] A. Alvarez, R. Orea, S. Cabrero, X. G. Pañeda, R. Garcı́a, D. Melendi, http://www.positivobgh.com/hogar/spa/home.
Limitations of network emulation with single-machine and distributed [48] Intel Corp., Centrino Wireless-N 100, 2013. URL: http:
ns-3, in: Proceedings of the 3rd International ICST Conference on //www.intel.com/content/www/us/en/processors/
Simulation Tools and Techniques, SIMUTools ’10, ICST (Institute for centrino/centrino-wireless-n-100-brief.html.
Computer Sciences, Social-Informatics and Telecommunications Engi- [49] Intel Corp., Dual Band Wireless-N 7260, 2013. URL: http://www.
neering), ICST, Brussels, Belgium, Belgium, 2010, pp. 67:1–67:9. intel.la/content/www/xl/es/wireless-products/
[27] S. Rampfl, Network simulation and its limitations, in: Network Archi- dual-band-wireless-n-7260-brief.html.
tectures and Services, Seminars FI / IITM / ACN SS2013, 2013. [50] M. Zubelda, A. Ferragut, F. Paganini, Overcoming performance pitfalls
[28] M. Nekovee, R. S. Saksena, Simulations of large-scale wifi-based wire- in rate-diverse high speed wlans, Computer Networks 57 (2013) 3673 –
less networks: Interdisciplinary challenges and applications, Future 3685.
Gener. Comput. Syst. 26 (2010) 514–520.
[29] S. Yoo, Y. Shin, S. Kim, S. Choi, Toward realistic WiFi simulation with
smartphone ”Physics”, in: Proceeding of IEEE International Symposium
on a World of Wireless, Mobile and Multimedia Networks 2014, 2014,
pp. 1–6.
[30] J. Bicket, D. Aguayo, S. Biswas, R. Morris, Architecture and evaluation
of an unplanned 802.11b mesh network, in: Proceedings of the 11th
Annual International Conference on Mobile Computing and Networking,
MobiCom ’05, ACM, New York, NY, USA, 2005, pp. 31–42.
[31] D. Raychaudhuri, I. Seskar, M. Ott, S. Ganu, K. Ramachandran,
H. Kremo, R. Siracusa, H. Liu, M. Singh, Overview of the orbit radio
grid testbed for evaluation of next-generation wireless network protocols,
in: IEEE Wireless Communications and Networking Conference, 2005,
volume 3, 2005, pp. 1664–1669 Vol. 3.
[32] K. N. Patel, R. h. Jhaveri, A survey on emulation testbeds for mobile
ad-hoc networks, Procedia Computer Science 45 (2015) 581 – 591. Inter-
national Conference on Advanced Computing Technologies and Applica-
tions (ICACTA).
[33] H. Ahmed, S. Pierre, A. Quintero, A flexible testbed architecture for
vanet, Vehicular Communications 9 (2017) 115 – 126.
[34] H. M. A. Fahmy, Simulators and Emulators for WSNs, Springer Singa-
pore, Singapore, 2016, pp. 381–491.
[35] H. M. A. Fahmy, Testbeds for WSNs, Springer Singapore, Singapore,
2016, pp. 251–379.
[36] S. Saginbekov, C. Shakenov, Testing wireless sensor networks with hy-
brid simulators, CoRR abs/1602.01567 (2016).
[37] J. Beshay, K. Subramani, N. Mahabeleshwar, E. Nourbakhsh,
B. McMillin, B. Banerjee, R. Prakash, Y. Du, P. Huang, T. Xi, Y. You,
J. Camp, P. Gui, D. Rajan, J. Chen, Wireless networking testbed and em-
ulator (winetester), Computer Communications 73, Part A (2016) 99 –
107.
[38] S. Choi, K. Park, C.-k. Kim, On the performance characteristics of
WLANs: Revisited, in: Proceedings of the 2005 ACM SIGMETRICS In-
ternational Conference on Measurement and Modeling of Computer Sys-
tems, SIGMETRICS ’05, ACM, New York, NY, USA, 2005, pp. 97–108.
[39] N. Khademi, M. Welzl, S. Gjessing, Experimental evaluation of TCP
performance in multi-rate 802.11 WLANs, in: 2012 IEEE International
Symposium on a World of Wireless, Mobile and Multimedia Networks
(WoWMoM), 2012, pp. 1–9.
[40] P. Serrano, C. J. Bernardos, A. de la Oliva, A. Banchs, I. Soto, M. Zink,
FloorNet: Deployment and Evaluation of a Multihop Wireless 802.11
Testbed, EURASIP J. Wirel. Commun. Netw. 2010 (2010) 8:1–8:12.
[41] Y. Yiakoumis, M. Bansal, A. Covington, J. van Reijendam, S. Katti,
N. McKeown, Behop: A testbed for dense wifi networks, in: Pro-
ceedings of the 9th ACM International Workshop on Wireless Network
Testbeds, Experimental Evaluation and Characterization, WiNTECH ’14,
ACM, New York, NY, USA, 2014, pp. 1–8.
[42] F. Granelli, R. Riggio, T. Rasheed, D. Miorandi, WING/WORLD: An
Open Experimental Toolkit for the Design and Deployment of IEEE
802.11-Based Wireless Mesh Networks Testbeds, EURASIP Journal on
Wireless Communications and Networking 2010 (2010) 730198.

14