Professional Documents
Culture Documents
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/277009005
CITATIONS READS
5 598
5 authors, including:
All in-text references underlined in blue are linked to publications on ResearchGate, Available from: Georgios Z. Papadopoulos
letting you access and read them immediately. Retrieved on: 01 November 2016
1
[koskri, peris]@it.teithe.gr
AbstractVerification of theoretical analysis is a vital step to evolve are often application-specific and too complex to be
the development of an application or a protocol for wireless reproduced precisely. More specifically, simulators allow users
networks. Most of proposals are evaluated through mathematical to implement some basic assumptions (e.g. link quality, radio
analysis followed by either simulation or experimental validation
campaigns. In this paper, we analyze a large set of statistics propagation, medium interferences, topologies) [3]. Although,
on articles published (i.e. 674 papers in total) in Ad-Hoc and the majority of the simulation models cannot capture real
Wireless Sensor Networks (WSNs) related top representative con- world complexity [4] [5], they are often utilized as a first step.
ferences (i.e. ACM/IEEE IPSN, ACM MobiCom, ACM MobiHoc Our purpose is to show that this step is not sufficient to show
and ACM SenSys) during the period 2008-2013. We focus on the the consistency of a solution as well as that low cost devices
evaluation methodologies provided by researchers. More specif-
ically, our goal is to explore the role of simulators and testbeds have steered researchers and engineers to enrich performance
in the theoretical analysis of a model throughout the protocol evaluation with testbeds.
development procedure. We show that there is a tendency for Experimental evaluation is performed either custom or over
more and more researchers to rely on custom or open testbeds in open testbeds, and exhibits potential unexpected failures and
order to evaluate the performance of their proposals. Simulators problems that the proposed solutions by researchers would
indeed fail to reproduce actual environment conditions of the
deployed systems. Experimentation with real hardware allows face during real deployments. Even though performing well
our research community to mind the gaps between simulation over testbeds, they remain in vitro deployments with more
and real deployment. Still, as experimental approach through or less controlled environment conditions. Such a proof of
custom testbeds results in a low reproducibility level (i.e., 16.5%), concept must then be transposed into the real world. Designing
we investigate to what extent such performance evaluation and setting up a complete Ad-Hoc or WSN system under
methods will be able to bridge those gaps. We finally discuss
experimental testbeds and their potential to replace simulators real conditions that can support robust applications is a very
as the cornerstone of performance evaluation procedures. complex task [6]. Researchers and production system archi-
tects, first need an appropriate plan of deployment and later
Index TermsAd-Hoc, wireless sensor networks, performance
evaluation, network simulation, experimental testbeds. number of tools, simulators/emulators and testing facilities for
real experiments, in order to initially validate their concept or
model and then to develop the appropriate infrastructure.
I. I NTRODUCTION Throughout this study, we compile a large set of statistics on
Ad-Hoc and Wireless Sensor Networks (WSNs) have en- literature review of 674 articles published in top conferences
abled a large variety of applications. Environmental and that are related with Ad-Hoc and WSNs over the 2008-2013
wildlife monitoring, clinical medical and home-care monitor- period. We focus on the evaluation provided by authors, and
ing, monitoring and control of industrial processes including especially to what extent experiments on testbeds have become
agriculture, smart houses and cities are just some of the a must for performance evaluation of new communication
examples of Ad-Hoc and WSN applications, where low-cost, algorithms and protocols. Hence, we exhibit the tendency
and easily deployed multi-functional sensor nodes is the ideal for performance evaluation procedures to rely on experiments
solution [1]. As a result, during the last years we experience with real hardware and environment, to the detriment of
the emergence of a new paradigm called Internet of Things simulations. The question of scientific results versus proofs
(IoT) in which smart and connected objects cooperatively of concepts therefore arises. Indeed, we discuss the meaning
construct a (wireless) network of things [2]. However, the of reproducibility and of a proof of concept as a prototype
unique features of Ad-Hoc and WSN technologies can pose being designed to determine feasibility. In this paper, we
significant challenges. Hence, envisioned solutions must be also analyze the selection of the evaluation methodology
verified before being deployed in a real-world WSN deploy- (e.g. simulator, testbed), and simplicity of the overall design
ment, either by utilizing simulators or emulators or through that should be provided for validation, understanding and
experimentations by employing testbeds. explanation. Finally, this work aims to investigate and gather
Simulation evaluation is an essential phase during the de- the pros both from simulation and experiments so that real-
sign and development of an Ad-Hoc or WSN infrastructure. world experiments could lead to reproducible scientific results
However, environments in which Ad-Hoc or sensor networks for our research community.
2
50
140 Total Ad-Hoc and WSN related (over 2008 - 2013) IPSN
Ad-Hoc & WSN related 45 MobiCom
100 MobiHoc
120 SenSys
40
Number of articles
Number of articles
100 80
Articles (%)
35
80 60 30
60 25
40
40 20
20
20 15
0 2008 2009 2010 2011 2012 2013 0 IPSN MobiCom MobiHoc SenSys 10
2008 2009 2010 2011 2012 2013
(a) Number of articles per year (all (b) Appropriateness of our conference (c) Publication flows over the 2008 - 2013
conferences are considered). sample. period.
Fig. 1: Published articles in ACM/IEEE IPSN, ACM MobiCom, ACM MobiHoc and ACM SenSys from 2008 to 2013.
II. P ERFORMANCE EVALUATION PROCEDURES results. As a result, during the simulation evaluation the
environmental conditions should not affect the behavior of the
In a typical research process cycle, once the modeling phase
nodes. Hence, it would be ideal if the authors first verify their
is done, network researchers and developers continue with
model by employing experimental tests in order to reflect the
the validation procedure in which they evaluate their concept
reality that their proposals would face during real deployment.
by using either a simulator or an emulator. Later, network
engineers and developers may proceed with experimentation
to further cross-verify their proposal [7]. Thus, once both the B. A Thorough Literature Study
simulation performance and the experimental measurements In this article, we carry out a thorough study over top
are satisfactory then real deployments can be initiated. representative conferences that are strongly related to Ad-Hoc
and WSN research fields. In particular, we have studied all ar-
ticles that have been published at the ACM/IEEE International
A. Simulating protocols or experimenting algorithms Conference on Information Processing in Sensor Networks
Since we face complex environments that are very difficult (IPSN), ACM Annual International Conference on Mobile
to be theoretically analyzed and we also take into account Computing and Networking (MobiCom), ACM International
the difficulties of setting up a real-world (e.g. large-scale) Symposium on Mobile Ad Hoc Networking and Computing
deployment, simulations are often considered as the optimal (MobiHoc) and ACM Conference on Embedded Networked
approach for studying the performance of Ad-Hoc and WSNs. Sensor Systems (SenSys) conferences in order to derive the
Many open source and freely available simulators allow users current tendency of the validation methodology that authors
to have a better control of the nodes by often employing a follow with respect to previously reported issues. Hence, we
Graphical User Interface (GUI), and to retain or simplify some go through and study 674 articles in total, published in the
assumptions in order to evaluate their solutions. Simulation conference proceedings for the last six years from 2008 to
evaluation is a provisioning procedure during the protocol 2013, out of which 596 are related to Ad-Hoc & WSN
development. However, even if the simulation performance (see Figure 1a). Indeed, we identified 78 articles that deal
presents coherent results with mathematical analysis, past real- with other wireless technologies such as WiFi and WiMAX,
world deployments show that it is not recommended to proceed that are studied in the context of cellular networks. All of
directly with real deployment since engineers may face unpre- these papers have been found in MobiCom (i.e. 140 out of
dictable phenomena such as node crashing or network discon- 185) and MobiHoc (i.e. 142 articles out of 175) conferences
nection [5], [8]. Intermediate experimentation platforms can (see Figure 1b), which are not entirely dedicated to Ad-
therefore be considered to bridge the gap between simulations Hoc and WSN however have a broad scope on mobility and
and real world deployments. Nevertheless, while simulations wireless communications. In the rest of the article, we further
can offer wider sets of assumptions to test and therefore emphasize our investigation over these 596 articles. During
potentially more complete evaluations, testbeds do impose our investigation, we obtain plethora of information for each
many characteristics (e.g. physical environment, hardware, work and we then categorize the considered articles based on
network topology). Such facilities offer the opportunity to their common features.
have their solutions facing real conditions, thus being more Figure 1c provides detailed information about the total
realistic than those modeled under software simulators. Yet, number as well as the Ad-Hoc & WSN related published
numerous parameters (e.g. radio dynamics, link stability and articles per proceeding year. We actually observe that, there is
symmetry, impact of the weather on communications [9]) a decreasing tendency of published articles in the proceedings,
appear so unpredictable that they may lead to results that can indeed we identified 43 articles less from 2008 to 2013. More
not be reproduced with sufficiently tight confidence intervals. specifically, MobiHoc and IPSN reduced the total accepted
The ambition of obtaining scientific results should then lead articles, from 44 to 24 (MobiHoc) and from 41 to 24 (IPSN)
researchers to allow for further repeatability of the presented respectively, while MobiCom and SenSys kept a steady flow.
3
S & E, 10
IPSN-Simulations IPSN-Experiments
MobiCom-Simulations MobiCom-Experiments 90
MobiHoc-Simulations MobiHoc-Experiments Simulation
SenSys-Simulations SenSys-Experiments M & S & E, 121 80 Experiment
100
70
Number of articles
E, 22
80 M & E, 239 60
M, 51
S, 3 50
Articles (%)
60
40
40 30
20
20
10
M & S, 150
0
2008 2009 2010 2011 2012 2013 0 2008 2009 2010 2011 2012 2013
(a) Mobility scenarios in performance (b) Use of Mathematics (M), Simulations (c) Total simulation versus experimentation
evaluation procedures. (S), Experiments (E) and their combinations evaluated articles.
in validation procedures of 596 Ad-Hoc and
WSN related articles.
Fig. 2: Published articles in ACM/IEEE IPSN, ACM MobiCom, ACM MobiHoc and ACM SenSys from 2008 to 2013.
Modern technologies introduced the feature of mobility. of simulations is decreasing every year (except in 2011) while
Consequently, the research community focuses on developing experimentations still remain present at a relatively stable rate.
and testing such aspects and scenarios. Our study results jus- Over the 2008-2013 period, 284 studies followed a simula-
tify this trend, owing to the 148 articles (57.7%) that simulated tion evaluation to test their proposal. We noted the simulator
mobile scenarios. Still, our statistical results for MobiHoc usage, the scales of simulated networks and the programming
and MobiCom, the mobile oriented conferences, show that languages used for custom simulators (see Figure 3). Only
not all of their articles implement mobility scenarios. For 43.3% are validated through a known simulator while 42.3%
instance, during the 2008 MobiHoc conference we determined of articles did not even provide any information about the tool
only 13 out of 28 simulation-based articles that introduced that their authors have utilized (see Figure 3a). Finally, 14.4%
mobility in their tests. As shown in Figure 2a, 57% of articles (with respect to 284 studied simulation-evaluated articles)
involving mobility are less induced by our conference sample developed a homemade simulator (Figure 3a), by utilizing
(half of the conferences, MobiCom and MobiHoc, being programming languages, such as Python and Java (see Figure
theoretically focused on mobility-related topics) than by the 3c for the distribution of the most popular programming
global enthusiasm for mobile scenarios. languages).
We are next interested in determining the usage of the
III. R ESULTS OF ANALYSIS simulators. As can be observed in Figure 3b, MATLAB is the
A. Evaluation procedures first choice in our community counting more than 35 articles,
In this subsection, we expose our analysis on the validation followed by TOSSIM which has been reported in almost 20
procedures that the authors followed. As a first step, we aimed articles. Furthermore, Network Simulator 2 (ns-2) comes third
to categorize the reviewed articles according to the employed with 13 articles.
evaluation method. In particular, we examine the proportion Nowadays, the research community is able to evaluate
of simulation, experimental and mathematical (i.e. modeling proposed protocols, models, even new technologies over open
or analysis) evaluated works. Our primary analysis exposes testbeds at a very large-scale [2]. Increasingly, network re-
interesting results. More specifically, our investigation shows searchers are using experimentations to enlarge the scope of
that the majority (i.e. 561) of the articles provide an analytical their performance evaluation, Figure 2c. Moreover, as it can
representation of their solution. The remaining 35 have only be observed in Figure 4a, our investigation shows that the
simulation or experimentation results. Furthermore, 284 verify majority of the researchers, 91.3%, choose to set up their
their proposal by employing simulation evaluation while on the own testbeds. Even though to the current day, there is a
other hand 392 of the articles include experimental evaluation number of open facilities providing to the developers the
for their validation. Finally, only one out of five (i.e. 20.3%) infrastructure needed for experimental Ad-Hoc, WSN or IoT
articles examines all three phases of the research process cycle studies, only 10.7% of the articles use open platforms. Our
(i.e. analysis, simulation and experimentation). The number of compiled statistics tend to show that researchers would rather
articles with the previously stated properties (with respect to favor their own setups for small scale deployments. In fact,
596 studied papers) is illustrated in Figure 2b. among the 392 articles exposing experimental results, 78%
We now present the characteristics of the articles that we of them do not exceed 40 nodes for their experimental setup
studied. The percentage of simulation versus experiment-based (see Figure 4a). Hence, the increased difficulty to apprehend
studies (with respect to 596 studied articles) is illustrated a remote open testbeds (e.g. specific hardware and software,
in Figure 2c. As can be observed, while simulations and network topology, booking procedure) may have induced
experiments used to be equally deployed until 2009, the usage researchers to set up their own relatively small scale networks.
4
90 40 9
Simulated networks
80 100 35 8
70 80 30 7
Number of articles
Number of articles
Number of articles
60 Articles (%) 60 6
25
50 40 5
20
40 20 4
15
30 0 3
ated ade ified
20 Indic Homem Unspec 10 2
10 5 1
0
<=20 21 - 40 41 - 60 61 - 80 81 - 100 >=101 0 0 Python C++ Java C C# Perl
LAB SIMns-2 COOJA ualnetns-3 Avrora NET+S+PsimrgyPluMs oSim
Number of nodes in simulations MAT TOS Q OM M Ene Glo
(a) Simulator usage and scales of simulated (b) Popularity of simulators. (c) Programming language popularity for
networks. custom simulators.
Fig. 3: Simulation evaluation methods (left) and popularity of simulators (center) and programming languages (right).
300
Homemade testbed 60
Open testbed 10
250
100 50
8
Number of articles
Number of articles
Number of articles
200
80 40
Articles (%)
60 6
150
30
40
100 4
20 20
50 0 Homemade 2 10
Open
0
<=20 21 - 40 41 - 60 61 - 80 81 - 100 >=101 0 0
ab ST iya ge et lab TN ab AB ley IT et AP ye SEI sB ky Az RP A2 RP te2 de ee IS PIC Fly ck ote
Number of nodes in experiments teL WI dr ira elN u D ARPL CKLBerk ORB utorn M NetE KAN Telo ote SMIC US MIC WA iMo inyNo ZigB IR E Fire Fle iM
Mo T In M Dies Em W FLO T Tm T
(a) Testbed usage and scales of (b) Popularity of open testbeds. (c) Motes popularity.
experimented networks.
Fig. 4: Experimental evaluation methods (left) and popularity of open testbeds (center) and motes (right).
Finally, we evaluated the popularity of the devices in home- provide some details about simulation setups. Among those,
made experiments. In Figures 4b and 4c the utility of the open 72.5% precise the number of involved nodes. Finally, we
testbeds and motes is presented. Even though a small number decided of non complete setups as soon as there was a lack
of articles experimented over open testbeds, we pointed out the of critical details regarding the tools used during simulations.
popular open platforms. As observed in Figure 4b, Harvards For instance, as discussed earlier, MATLAB stands as the
Motelab comes first (11 articles), followed by TWIST (10 most popular software for simulations. In order to use it as a
studies). That can be simply explained as those facilities were network simulator, researchers must import external libraries
the first to open up to the scientific community. Regarding (e.g. as developed by the WISLAB1 team). It is difficult, if not
the Indriya testbed, even though it was made available only impossible, to reproduce a simulation study when the version
since 2011, it was used in 8 articles. The fact that users can of a publicly available simulator is unknown, and only 21.5%
interact with the testbed through the same intuitive web-based provide us with the employed version or the utilized library of
interface as MoteLabs, could explain this success among the the simulator, which essentially concludes our outcome about
community. the reproducibility of the simulation-evaluated articles.
We followed similar methodology for the experimental-
based validations. Taking into account the nature of open
B. Reproducibility platforms, the 42 articles, we consider that these articles
We continue our study by investigating the feasibility of overall are reproducible. However, we counted 8 papers where
reproducing results that are presented in the reviewed articles, the authors tested their ideas over both custom and open
both for simulation and experimental campaigns. To proceed testbeds, with only 3 of them providing enough information to
so, we looked for some critical information (e.g. simulation be assumed reproducible. On the other hand, the experimental
setup, simulator indication, simulator details such as version results that are retrieved through homemade testbeds can be
or library, number of nodes), that should be provided by the considered as difficult or even impossible to reproduce. This is
studied articles. In order to reproduce the proposed solution, explained since most of them are deployed in offices, houses
we assumed that the authors should provide a complete or even outdoor installations where the environmental radio
simulation or experiment settings subsection. activity varies, due to the interpolation of external features
Regarding the simulation based evaluations, while only
43.3% of the articles indicate the simulator, 78.5% of those do 1 http://wislab.cz/
5
MAC, 81
such as mobile phones, wireless routers and access points
Data Link, 50
and so on. Nevertheless, owing to the nature (e.g. application
layer) of the tested solution, we detected 31 homemade-based Network, 52
studies that may be reproduced. Finally, by summarizing the Physical, 62
previous statements, we calculated that only 16.5% (65) of the Transport, 18
experimental-based papers present reproducible results.