Enabling Efficient and Accurate Large-Scale Simulations
of VANETs for Vehicular Traffic Management
Moritz Killat
∗
Felix Schmidt-Eisenlohr
∗
Hannes Hartenstein
∗
Christian Rössel
†
Peter Vortisch
†
Silja Assenmacher
‡
Fritz Busch
‡
∗
Institute of Telematics
University of Karlsruhe, Germany
{killat, fschmidt,
hartenstein}@tm.uni-
karlsruhe.de
†
PTV AG
Karlsruhe, Germany
{christian.roessel,
peter.vortisch}@ptv.de
‡
Chair of Traffic Engineering and Control
Technical University of Munich,
Germany
{silja.assenmacher,
fritz.busch}@vt.bv.tum.de
ABSTRACT
To study the impact of inter-vehicle communications on (ve-
hicular) transport efficiency, e.g., for traffic management
purposes, there is a need for efficient and accurate large-
scale simulations that jointly consider both, the vehicular
traffic and the communication system. To overcome the
scalability limitations of current discrete event-based net-
work simulators like NS-2, we propose a hybrid simulation
approach that can significantly reduce the number of sched-
uled events by making use of statistical models. Basically,
we treat some data traffic, which is not the primary con-
cern of the simulation study, as ‘noise’ (e.g., beaconing of
nodes). While accurately modeling this background traffic
we only need to simulate via discrete event-based simulation
the actual application we are interested in (e.g., a data dis-
semination protocol). We outline how the characterization
of the background traffic is gained, statistically validated
and used. The achievable speed-up is demonstrated in a
first application study where a speed funnel is built using
inter-vehicle communications. In this scenario, the conser-
vatively estimated speed-up factor is about 500 compared to
a pure discrete event-based simulation.
Categories and Subject Descriptors
C.2.1 [Network Architecture and Design]: Wireless com-
munication; I.6.3 [Simulation and Modeling]: Applica-
tions
General Terms
Design, Performance
Keywords
Vehicular networks, simulation, modeling, traffic manage-
ment
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
VANET’07, September 10, 2007, Montréal, Québec, Canada.
Copyright 2007 ACM 978-1-59593-739-1/07/0009 ...$5.00.
1. INTRODUCTION
Inter-vehicle communication (IVC), the direct wireless ex-
change of information between, to and from vehicles, is as-
sumed to have a beneficial impact on traffic efficiency and
on road safety. Recently, many papers in this research field
have been published that analyze the challenges of vehicu-
lar ad hoc networks (VANETs) and propose communication
strategies. In most cases their performance, however, is ex-
pressed in terms of ‘increased packet delivery ratio’ or other
communication-centric figures of merit. However, eventually
it is required to prove the impact of inter-vehicular commu-
nications based on metrics directly related to traffic safety
and efficiency. Clearly, simulations will be the first method
of choice.
This paper is based on a project funded by the traffic
management authority of the state of Hessen, Germany, in
order to investigate the impact of car-to-x communications
on transport efficiency. The expectations of a traffic man-
agement center’s employee on a tool for simulation of car-
to-x-enabled vehicular scenarios are pretty high:
• Scalability: typical scenarios that are interesting from
a vehicular traffic management point of view easily
comprise a few ten thousands of vehicles over tens or
hundreds of kilometers.
• Performance: simulation results should be available
within a few minutes at least for a variety of configu-
rations.
• Ease of use: configuration of simulation scenarios have
to be done ‘on the fly’.
It is hard to imagine that vehicular traffic experts will make
direct use of simulators like NS-2 [1] or comparable choices
since the requirements mentioned above are hard to match
with such a packet-level network simulator. Instead, for
them it looks more natural to make use of their vehicular
traffic modeling suites that already contain powerful mod-
ules on traffic sign/light information etc., and to extend
these simulators by modules that mimic the information dis-
semination as achieved by vehicular networks.
Significant speedups with respect to the time required for
a simulation cannot be simply done by tuning a current net-
29
work simulator or by porting it to a powerful machine. In-
stead, accuracy has to be traded off for the sake of simulation
speed. Of course, a ‘graceful’ or controlled degradation of
accuracy has to be guaranteed.
To achieve the goal of efficient large-scale simulations we
make use of the well-known concept of hybrid simulations:
“By combining, in a hybrid model, discrete-event simula-
tion and mathematical modeling, we are able to achieve a
high level of agreement with the results of an equivalent
simulation-only model, at a significant reduction in compu-
tational costs” [2]. Hybrid simulations do not represent an
‘exotic’ idea as they are part of current packet-level net-
work simulators regarding, e.g., radio wave propagations:
instead of simulating the various paths a signal can take or
the various bits in the data packet in a microscopic way, a
macroscopic model is used and ‘executed’. Our idea now is
to replace also ‘background data traffic’, which is typically
simulated in a discrete event-based fashion, by a suitable
macroscopic model, reducing computational costs by two to
three orders of magnitude.
For example, in active safety applications every vehicle
sends out periodic beacon messages to provide location,
speed and direction to neighboring vehicles. The resulting
load on the channel can be very high and will affect other
data transmissions like data dissemination protocols that in-
form drivers about a traffic jam or a construction site. While
the beaconing activity provides for the majority of events in
an event-driven simulator, the actual object of interest might
be the data dissemination process of a warning message, i.e.,
the probability of reception and corresponding latency of the
warning message, and the resulting driver and traffic behav-
ior. Therefore, we propose to put the ‘background data traf-
fic’ like the beaconing activity into a statistical model that
can be executed analytically but still provides realistic wire-
less channel conditions for the observed application. Thus,
not only fading and shadowing effects are coded into the
statistical model but also the packet collision probabilities.
In this paper we proceed as follows:
• We quantify the amount of events that are generated
by beaconing. Thus, we show the potential gain w.r.t.
the reduction in events.
• We derive a macroscopic statistical model for the prob-
ability of successful packet reception at a given dis-
tance, dependent on the vehicular traffic density and
the amount of ‘background data traffic’.
• We show for tractable scenarios that the hybrid ap-
proach and a pure NS-2-based approach do not show
relevant statistical differences.
• We outline an architecture how to couple a vehicu-
lar communication module that implements our hybrid
approach to a well-known vehicular traffic simulator,
VISSIM [3]. The architecture also considers the mod-
eling of the driver behavior since without proper mod-
eling of driver behavior the effect of access to informa-
tion cannot be seen. Our work can be interpreted in
a way that we add to the simulation of vehicular traf-
fic governed by driver behavior an additional source of
information, namely the one obtained by inter-vehicle
communications (see Figure 1).
• We present first results on our implementation for a
large-scale scenario comprising 2.500 to 3.000 vehicles.
Figure 1: Modeling the driving behavior: result of a
process influenced by many sources. Car-to-X commu-
nication enhances the information set available to the
driver or vehicle.
The conservatively estimated speedup factor is about
500 compared to the NS-2 simulation.
The structure of the paper will follow this list of items after
a discussion of previous and related work. Before continuing
we like to emphasize that the paper cannot be understood
to ‘replace’ packet-level network simulators by our hybrid
approach since the standard network simulators are required
for building the correct statistical models we make use of
in the hybrid approach. Also, we want to emphasize that
in the following the term ‘model’ refers to a mathematical
description of a behavior that is derived from empirical data
extracted from simulation results.
2. RELATED WORK
During the last years several attempts to combine in-
ter vehicle networking and vehicular traffic simulation were
presented. These proposals either suggested to incorporate
both disciplines, vehicular traffic and networking, into a sin-
gle simulation engine (cf. e.g., [4]) or to couple and synchro-
nize two simulators of the respective area (cf. e.g., [5], [6]).
The majority of the works following the latter approach used
the network simulator NS-2 and interlinked it with diverse
traffic simulators (e.g., CARISMA, VISSIM). The studies
present results on traffic performance (e.g., average speed)
and on network characteristics (e.g., latency), however, their
focus is not primarily on scalability.
As already described, to allow large-scale simulations we
follow the strategy of hybrid simulation. A historic view on
hybrid simulations development can be found in [7]. This
concept for simulation was first described and analyzed by
means of accuracy and efficiency for a queuing system [2].
Later, [8] clarified and unified the definitions used for hybrid
simulations and analytical models, and introduced a catego-
rization scheme for hybrid simulation systems. Depending
on how the problem solution is achieved and how simulation
and analytical model depend on each other each system falls
in one of four classes, ours being in “Class IV”, in which “a
simulation model is used as an overall model of the total sys-
tem, and it requires values from the solution procedure of
an analytic model representing a portion of the system” [8].
30
0
2e+07
4e+07
6e+07
8e+07
1e+08
1.2e+08
1.4e+08
1.6e+08
1.8e+08
0 100 200 300 400 500 600 700 800 900 1000
number of events
number of vehicles
500 veh/km
400 veh/km
200 veh/km
100 veh/km
50 veh/km
20 veh/km
10 veh/km
Figure 2: Number of scheduled events in NS-2 in depen-
dency of the number of vehicles and of the traffic density.
All scenarios ran 100s and each node broadcasted a single
500-byte packet each second.
The applicability and usefulness of a hybrid simulation ap-
proach can, for instance, be seen in [9]. The authors showed
that having the flexibility to choose the time granularity of
simulations allows to run large-scale simulations, and still
being efficient and accurate. Again, a crucial condition is
the availability of an appropriate model.
Several publications by R. Sargent et.al., e.g. [10], tackle
methods to validate and verify the simulation models used
in simulations. Our work makes use of the modeling pro-
cess and these methods to achieve high credibility of the
models developed. Sophisticated models and frameworks
for detailed vehicular packet level simulations with NS-2 are
proposed, e.g, in [11], [12] and [13].
3. HYBRID SIMULATION APPROACH
As indicated in Section 1 the dominating number of simu-
lation events results from ‘background data traffic’ like bea-
con messages. Simulation events belonging to the actual
object of interest of an application designer or civil engineer
are comparatively negligible. Accordingly, most of the sim-
ulation runtime is ‘wasted’ for simulating the environment.
We argue in this section that this runtime can be saved by
introducing a statistical model that mimics realistic condi-
tions for an application.
The remainder of this section is structured as follows:
firstly, we analyze the costs of the ‘background data traffic’
by making use of the discrete-event network simulator NS-2.
The results of this process justify the proposed hybrid simu-
lation approach for which a statistical model is required and
derived in the following. Finally, we will show in tractable
scenarios that the hybrid simulation approach statistically
agrees with a discrete-event simulation.
3.1 Beaconing costs in discrete-event
simulations
Previous approaches to simulate vehicular ad hoc net-
works coupled a traffic simulator with a communication sim-
ulator (cf. Section 2). When neglecting efforts required to
synchronize both simulators, the runtime costs belonging to
the communication part depend on the number of scheduled
Figure 3: Average runtime required by NS-2.31 to simu-
late 100s on an Intel Pentium 4 CPU 3.4 GHz, 3.5 GByte
main memory. Each vehicle broadcasts a 500-byte packet
once a second in average.
simulation events. Regarding beacon messages, the number
of events is mainly determined by the amount of vehicles, the
beaconing interval and by the traffic density. The reason for
the latter is the necessity to schedule events at each node
within a certain distance whenever a packet is broadcasted.
We set up scenarios for a varying number of wireless nodes
and for changing distances between them in the simulator
NS-2. In order to mimic different traffic conditions we make
use of a sequence of static snapshots with a defined number
of vehicles and a varying traffic density on a straight lane,
each. All scenarios ran 100s and each node broadcasted a
single 500-byte beacon message once a second in average.
The number of counted events in the global event scheduler
of NS-2 is depicted in Figure 2. Additionally, we ran the
same scenarios for more than 50 seed values and measured
the average runtime NS-2 required for each simulation af-
ter the initialization process was carried out (cf. Figure 3).
The simulations were conducted using an extended version
of NS-2.31 [14] on a Linux 2.6 system equipped with an Intel
Pentium 4 CPU having 3.4GHz and 3.5GByte main memory.
The results show time-consuming simulations for dense
traffic scenarios even for a relatively small amount of vehi-
cles. Furthermore, note that the scenarios assumed a quite
relaxed beaconing interval of one second. Accordingly, the
runtime performance becomes additionally stressed by more
frequent transmissions.
3.2 Modeling beacon traffic analytically
The previous Subsection 3.1 has shown that the discrete-
event simulation of the channel load, represented by the bea-
con traffic, turns out to be time-consuming. This subsection
now aims to save this effort by deriving a statistical model
giving the probability of a successful packet reception in de-
pendency of a varying environment. Once the model is built
it can be used to reduce the number of simulation events.
We start with an analytical derivation of the probability of
reception for a scenario only consisting of a single sender.
This result will serve as a starting point to build models for
many senders in varying traffic densities.
31
Figure 4: Probability of reception in dependency of the
distance for a Nakagami m-distribution fast fading model
in combination with a Friis/TRG path loss model. Com-
parison of analytical approach and average probability
gained from 1.000.000 simulated samples for each dis-
tance. The crossover distance at 556m can clearly be
seen.
Single sender:
Experiments on real roads have shown that the Nakagami
fast fading model, in combination with a Friis/Two-Ray-
Ground path loss model, is a suitable representation of the
radio propagation in vehicular ad hoc networks [11]. A
quadratic radio attenuation over distance according to the
Friis path loss model is assumed for smaller distances. As
suggested by Rappaport [15] at farther distances the Two-
Ray-Ground path loss model is applied that assumes a stronger
signal attenuation. The distance for switching from the Friis
to the Two-Ray-Ground model is the so called crossover
distance that depends on used antenna heights and on wave
length. On top of path loss calculation, the Nakagami model
considers disturbances due to fast fading effects. The Nak-
agami m fading parameter identifies the intensity of such
effects, stronger fading being indicated by a lower value of
m. Consequently, the radio reception strength at a certain
distance is probabilistically distributed. Likewise, the prob-
abilistic manner gives reason for a non-deterministic packet
reception behavior.
In case of a setup containing a single sender, expressions
that give the probability of successful packet reception at
each distance d from the sender can be derived analytically.
A detailed description of the derivation is given in the ap-
pendix. For d ≤ crossover distance we obtain
P
R
(d, CR)=
e
−3
(
d
CR
)
2
“
1+3
`
d
CR
´
2
+
9
2
`
d
CR
´
4
”
(1)
and, likewise, for d > crossover distance we get
P
R
(d, CR; γ)=
e
−3γ
„
d
2
CR
«
2
„
1+3γ
“
d
2
CR
”
2
+
9
2
γ
2
“
d
2
CR
”
4
«
. (2)
In this study the Nakagami fading parameter m is set to 3.
CR denotes the ‘intended’ communication range depending
0
0.2
0.4
0.6
0.8
1
0 100 200 300 400 500 600 700 800
probability of reception
distance sender/receiver [m]
single sender
25 sender per km
100 sender per km
250 sender per km
400 sender per km
Figure 5: Comparison of probabilities of packet recep-
tion depending on the distance between sender and re-
ceiver and varying number of transmitters per km. Data
are gained via simulation runs in NS-2.
on the configured transmission power at the sender. CR is
the maximum achievable communication distance when only
assuming path loss according to Friis/Two-Ray-Ground and
neglecting fast fading effects. Equation (2) includes an addi-
tional parameter γ, that depends on antenna heights (1.5m
in this study) and on the radio wavelength (5.08cm); see the
appendix for further details. According to our settings the
crossover distance equals 556m.
Figure 4 illustrates the average reception probabilities re-
sulting from 1.000.000 samples simulated with NS-2 and the
probabilities due to Equation (1) and (2) showing a perfect
match.
Multiple senders:
Clearly, with an increasing number of senders the load on the
communication channel grows and affects ongoing transmis-
sions. Regarding vehicular ad hoc networks, interferences
and resulting ‘disturbances’ intensify with an increasing traf-
fic density as more vehicles might interfere packet transmis-
sions. This issue is expressed in Figure 5 that exemplarily
compares chosen traffic densities with the previously intro-
duced single sender scenario.
Illustrated is the average probability of reception in de-
pendency of the distance between sender and receiver de-
termined via simulations in NS-2. For each density 1.000
simulation runs comprising 100 simulation seconds were con-
ducted. According to the investigated traffic density in each
run nodes were randomly positioned linearly beside a specific
sender under observation. While all nodes in the scenario
broadcasted a single 500-byte message once per second, we
put our focus on the successful receptions of the packets sent
out by the sender under observation. Figure 5 obviously
agrees with the intuitive expectation of decreasing reception
probabilities along with increasing traffic densities. Exem-
plarily, at a distance of 350m the divergence in the prob-
ability of reception reaches 40%. Clearly, this discrepancy
may have an impact on the performance (e.g., delay) of an
application protocol and might be the focus of a simulation
study.
32
-15
-10
-5
0
5
10
0 50 100 150 200 250 300 350 400
parameter value
number of vehicles per km
parameter x
1
parameter x
2
parameter x
3
parameter x
4
Figure 6: Parameters of the chosen approximation func-
tion show linear behavior w.r.t. the traffic density.
3.3 Derivation of statistical model
In the previous subsection we derived an analytical model
for scenarios consisting of a single sender. To meet the re-
quirements of multiple senders a more flexible model is re-
quired. Therefore, we seek a mathematical expression that
fits to the previously mentioned simulation results (cf. Fig-
ure 5). We describe our proceeding for distances smaller
than the crossover distance in the following; the same pro-
cedure, of course, is applied separately for distances larger
than the crossover distance.
Our methodology comprises nonlinear curve fitting tech-
niques for each simulated traffic density. We applied the
Levenberg-Marquardt method, the “standard of nonlinear
least-squares routines” [16], given a slight modification of
Equation 1 as starting point. The modification consists of a
parameterization (variables x
1
through x
4
) and of the addi-
tional consideration of the first and third monomial resulting
in
P
R,approx
(d, CR)=e
−3
(
d
CR
)
2
1+
4
X
i=1
x
i
„
d
CR
«
i
!
. (3)
Note, that in our settings the ‘intended’ fixed communica-
tion range, CR, matches 500m.
The optimization algorithm determined the fitting param-
eters x
1
through x
4
for all investigated traffic densities. As
shown in Figure 6 the parameters indicate a linear behavior
w.r.t. to the traffic density. By combining these approxi-
mations with Equation 3 we gained a statistical model. A
comparison of the model and the simulation results (cf. Sec-
tion 3.2) showed an average error of 0.25%, never exceeding
a value of 2.5%. Exemplarily, the differences are visualized
in Figure 7 and 8 for four chosen traffic densities.
3.4 Evaluation of proposed model
In the following we substantiate the promising results of
the previous Subsection 3.3 by evaluating a case study. We
set up and analyzed a scenario using the proposed hybrid
simulation approach and the corresponding implementation
(for the implementation, see Section 4). We compare the
obtained results with corresponding NS-2 simulations.
In the scenario vehicles were approaching and passing an
infrastructure point, or road side unit (RSU),fromthege-
10%
5%
0%
-5%
-10%
0 100 200 300 400 500 600 700 800
differences in the probability of reception
distance sender/receiver [m]
25 veh per km
100 veh per km
Figure 7: Difference of determined approximation and
average probability of reception for traffic densities of 25
and 100 vehicles per kilometer.
10%
5%
0%
-5%
-10%
0 100 200 300 400 500 600 700 800
differences in the probability of reception
distance sender/receiver [m]
250 veh per km
400 veh per km
Figure 8: Difference of determined approximation and
average probability of reception for traffic densities of
250 and 400 vehicles per kilometer.
ographical position x
−
= −1000m to x
+
= 1000m with an
average speed of 128km/h in a traffic density of 23 vehicles
per kilometer. The infrastructure point located at x
I
=0m
was broadcasting 500-byte data messages once a second. A
simulation ran for 10.000 seconds.
First we evaluated at which distances to the infrastructure
point packets were successfully received by the vehicles. Fig-
ure 9 illustrates the probability mass function (pmf) derived
from the experiment as well as the pmf determined on ba-
sis of NS-2 simulations. The consensus of both curves is
substantiated by examining the actual number of received
packets at certain distances for both approaches (cf. Fig-
ure 10). Outliers close to the infrastructure point (distances
nearby 0m in Figure 10) are naturally explained by the set
up of the scenario having installed the infrastructure point
a few meters apart from the roadside. Statistically we em-
phasized the validity of the model for the presented scenario
by performing a χ
2
test. Therefore, we discretized the dis-
tances between RSU and vehicles into 1001 bins of 1m and
hence obtained 1000 degrees of freedom (note, that distance
33
