Вы находитесь на странице: 1из 8

2012 Second International Workshop on Vehicular Communications and Networking

Multi-objective OLSR optimization for VANETs


Jamal Toutouh, Enrique Alba
Dept. de Lenguajes y Ciencias de la Computacion
University of Malaga, ETSI Informatica
Campus de Teatinos, Malaga - 29071, Spain
{jamal, eat}@lcc.uma.es

AbstractVehicular ad hoc networks (VANETs) are protocol. We have chosen OLSR as a use case, however the
infrastructure-less and self-organized networks deployed among methodology presented in this article can be directly applied
vehicles and other road users. Due to the limitations of the over any other protocol. OLSR has already been used in
wireless technologies used and the rapid topology changes,
designing efficient routing protocols for VANETs is becoming VANETs research [6], [7], [8], [9], since it offers a series of
a major concern. In this study, we applied a multi-objective features that make it well-suited for VANETs: it exhibits short
optimization metaheuristic, in order to find efficient OLSR delays in the transmission of data packets, it adapts well to
parameterizations that improve the QoS of the OLSR RFC continuous topology changes, and it allows a simple operation,
and a previous optimized configurations. Our optimized allowing an easy integration into different kinds of devices.
configuration significantly reduces OLSR scalability problems
keeping competitive packet delivery rates. The OLSR routing The main drawback of OLSR is that the excessive routing
overhead is reduced between 47% and 76% and the delivery load can cause network congestion, then limiting the data
times are between 32% and 38% shorter when using our packets exchange. However, as shown in this study and in [10],
optimized settings. [11], it admits a wide range of improvement by changing the
Index TermsVANETs; OLSR; NSGA-II; optimization configuration parameters.
The methodology used here consists in exploring the search
I. I NTRODUCTION space defined by all possible feasible OLSR configurations
Vehicular ad hoc networks (VANETs) are infrastructure- using the non-dominated sorting genetic algorithm, version II
less and self-organized networks deployed in vehicular envi- (NSGA-II) [12] to find optimized parameter settings. In this
ronments for the communication between vehicles, elements context, we decide that an efficient OLSR configuration should
of roadside infrastructure, and pedestrian personal devices. maximize the packet delivery ratio and minimize the packet
Such networks present the opportunity to develop cooperative delivery delays. In turn, this configuration should minimize
applications capable of gathering, processing, and distributing the routing workload to mitigate network congestion problems.
road traffic information. IEEE 802.11p is one of the most The evaluation of the solutions (tentative configurations) is car-
promising standards to allow wireless access in VANETs. ried out by analyzing the communications in realistic VANET
The limitation in coverage, the high mobility of the nodes, scenarios by using the ns-2 network simulator [13]. Finally,
and the presence of obstacles generate continuous link failures the best OLSR configuration found, the best one presented
among VANET nodes. As a matter of fact, the communica- in [11], and the OLSR RFC 3626 settings are analyzed in
tions suffer from frequent topology changes, packet loss, and different VANET scenarios to validate the results.
network fragmentation. In order to face these problems, the In summary, the main contributions of our work are: i)
research community dedicates a great deal of effort to design to propose an optimization strategy in which the NSGA-II
efficient MAC strategies [1] and routing protocols [2], [3]. is coupled with a VANET simulator to find efficient OLSR
Some VANET routing methods have been defined based on configurations, that can be generalized to be applied over any
prior ad hoc network architectures by targeting the specific other protocol; ii) to suggest new OLSR configurations whose
needs of vehicular environments [3]. QoS are better than the RFC 3626 defined configuration;
In order to provide VANETs with efficient communications, and iii) to validate the results by comparing a set of OLSR
an optimal routing strategy that makes better use of resources configurations in realistic VANET scenarios taken from the
and offers the best quality-of-service (QoS) is needed. A metropolitan area of Malaga (Spain).
proper parameterization of these protocols can improve the The remainder of this paper is organized as follows. In
performance of the whole network, even making the difference the next section, we introduce some approaches on routing
between a network that does or does not work, e.g., the optimization. Section III describes the OLSR QoS optimiza-
networks with high routing information exchange are more tion problem. Section IV presents the NSGA-II algorithm.
likely to suffer from network congestion and cannot ensure Section V describes the optimization methodology followed
timely delivery of messages [4]. to tackle the problem. Section VI presents the experiments
Thus, in the present study, we aim to solve an off-line carried out, as well as the results, comparisons, and analyses.
multi-objective optimization problem (MOP) to efficiently and Finally, conclusions and further work are presented in Sec-
automatically tune OLSR [5], a widely used ad hoc routing tion VII.

978-1-4673-1430-5/12/$31.00 2012 IEEE 571


II. R ELATED W ORK III. E FFICIENT OLSR FOR VANET S
The volatility of mobile ad hoc networks (MANETs) Exchanging real-time information among vehicles is the
and VANETs critically limits their performance in terms of most salient feature of VANETs. In this sense, data packets
throughput and coverage. For this reason, there is a great deal travel through the network nodes to reach the destination
of effort being made to provide MANET and VANET with node. Finding efficient routing paths is a complex task due
efficient communications by means of protocol optimization. to the volatility of the network and the absence of any central
manager. Distributed routing protocols operate in the core of
In this area, most of studies have focused on obtaining
VANETs, they compute updated paths among the nodes to
improvements in terms of both the QoS offered (maximizing
permit an effective information exchange.
packet delivery rates, minimizing delivery delays, etc.) and in
The methodology presented here can be applied over any
resource consumption, e.g. minimizing power requirements.
other VANET protocol. In our case study, we have chosen
Most of the defined protocol optimization problems are multi-
OLSR [5] for a use case, as it presents a series of features that
objective with objectives conflicting in nature. These kinds of
make it suitable for VANETs: i) it adapts well to the frequent
optimization problems usually show high complexity due to
and rapid topological changes because it follows a proactive
the huge number of possible solutions. Thus, the use of auto-
strategy, ii) its simple operation allows an easy integration
matic intelligent tools is a mandatory requirement when facing
into existing devices and operating systems, and iii) it is
them. Along these lines, metaheuristic algorithms [14] have
appropriate for most VANET warning applications due to its
emerged as efficient techniques able to solve such problems.
short transmission delays. The main drawback of OLSR is the
Multi-objective evolutionary algorithms (MOEAs), as
excessive routing workload generated to maintain the routing
NSGA-II, CellDE, PAES, etc., have emerged as metaheuristics
table updated, such traffic may provoke network congestion.
that are able to find a set of solutions, which optimize two
or more conflicting objectives given a problem [15]. The use A. Optimized Link State Routing Protocol
of MOEAs to solve protocol optimization problems provides Optimized Link State Routing (OLSR) is a proactive link-
solutions that optimize to varying degrees different QoS or state routing protocol specifically designed for ad hoc net-
resources consumption metrics. Thus, the network designer is works with low bandwidth and high mobility. OLSR uses a
able to select a solution (protocol configuration) depending on subset of special nodes of the network that acts as multipoint
the specific needs of the network. relays (MPRs) to periodically broadcast the routing control
Several studies have applied MOEAs to solving MANET information. This way, it reduces the number of required
optimization problems. Alba et al. [17] applied a specialized transmissions, trying to limit the network workload [5].
cellular multi-objective GA for finding an optimal configura- Two processes neighborhood discovery and topology dis-
tion for the Delayed Flooding with Cumulative Neighborhood semination mainly govern the core functionality of this
broadcasting strategy. Ruiz et al. [18], [19] applied a hybrid protocol. Such processes exchange three types of messages [5]:
multi-objective optimization algorithm (CellDE) to maximize
HELLO messages, exchanged between neighbor nodes
the coverage and minimize the power consumption and broad-
for link sensing, neighborhood detection, and MPR se-
cast time of the EDB protocol, and Abdou et al. [20] employed
lection. Periodically, they are generated containing infor-
different multi-objective optimization algorithms to optimize
mation about the links between the network interfaces.
a broadcasting strategy in mobile networks.
TC (topology control) messages, generated by MPRs to
In the literature, however, there is a lack of studies applying
indicate which other nodes have selected it as MPR. They
MOEAs to solve protocol optimization problems in vehicular
are forwarded to the nodes through the entire network.
networks. The multi-objective protocol optimization problem
The information is used for routing table calculations.
has usually been transformed into a single objective problem
MID (multiple interface declaration) messages are sent to
by defining a weighted factor function of each objective.
report information about the network interfaces employed
This means the problem is solved by using a mono-objective
to participate in the network, since a node can have
metaheuristic, resulting in a single optimized solution based on
multiple interfaces with distinct addresses participating
the value of combined objective function. Unfortunately, the
in the communications.
solution obtained by this process depends largely on the values
assigned to the weighting factors. Various studies have used The OLSR processes are regulated by a set of parameters
this strategy to solve such problems: a set of mono-objective predefined in the OLSR RFC 3626 [5] (see Table I). These
metaheuristics were used to find efficient parameter settings parameters are: the timeouts before resending HELLO, MID,
for VDTP [21], AODV [22], and OLSR [11], [23] protocols. and TC messages (HELLO INTERVAL, MID INTERVAL, and
TC INTERVAL, respectively); the validity time of the in-
Here, we study the application of the NSGA-II multi-
formation received via these three message types, which
objective algorithm to solve the off-line optimization problem
are: NEIGHB HOLD TIME (HELLO), MID HOLD TIME
of finding QoS efficient OLSR settings. In previous com-
(MID), and TOP HOLD TIME (TC); the WILLINGNESS of
prehensive analysis, the off-line mono-objectively optimized
a node to act as an MPR; and DUP HOLD TIME, that
protocols significantly improved the standard ones in several
represents the time during which the MPRs record information
VANET scenarios that reflect different traffic situations [16].
about the forwarded packets.

572
TABLE I
M AIN OLSR PARAMETERS AND OLSR RFC 3626 SPECIFIED VALUES . A MOP solution is a vector of decision variables x
Parameter Standard value Range (solution vector) which satisfies the constraints formulated by
HELLO INTERVAL 2.0 s R [2.0, 15.0] the functions G(x), offering adequate trade-off values for the
MID INTERVAL 2.0 s R [5.0, 15.0]
TC INTERVAL 5.0 s R [4.0, 35.0] functions fi (x). Considering a minimization MOP, a solution
WILLINGNESS 3 Z [0, 7] w is said to dominate the other solution v (it is denoted
NEIGHB HOLD TIME 3 HELLO INTERVAL R [5.5, 45.0]
MID HOLD TIME 3 TC MID R [10.5, 90.0] w v), if fi (x) fi (w) j/fj (x) > fj (w). The set
TOP HOLD TIME 3 TC INTERVAL R [10.5, 90.0] of optimal solutions for a MOP is composed by the non-
DUP HOLD TIME 30.0 s R [10.5, 90.0]
dominated feasible vectors, named Pareto optimal set. It is
B. OLSR Quality of Service Optimization in VANETs defined by P = {x /x , F(x )  F(x)}. The
The performance of the communication protocols depends region of points defined by the optimal Pareto set in the
on the selection of the parameter settings that determine objective function space is known as Pareto front, formally
their behavior, e.g., in OLSR, the detection of topological defined by P F = {(f1 (x), . . . , fk (x)), x P }.
changes can be adjusted by changing the HELLO INTERVAL NSGA-II is a multi-objective version of the well-known
parameter. Thus, optimized configurations could decisively genetic algorithm (GA) [12] and it is one of the reference
improve the protocols QoS and resources consumption. algorithms to solve MOPs. Its pseudocode is presented in
The standard configuration of OLSR offers a moderate QoS Algorithm 1. NSGA-II makes use of a population (P ) of
when is used in VANETs [9]. Hence, taking into account solution vectors (known as individuals). In each generation, it
the impact of the parameters configuration in the network works by creating new individuals after applying the genetic
performance, we tackled the OLSR QoS MOP to discover operators to P (recombination and mutation), in order to create
efficient protocol configurations. We have computed the ranges a new population Q (lines 5 to 8). Then, both the current
for each parameter by following OLSR RFC restrictions for (P ) and the new population (Q) are joined; the resulting
taking into account just feasible configurations (see Table I). population, R, is ordered according to a ranking procedure
Following this idea, we have defined a solution vector of real and a density estimator known as crowding distance (line 13).
variables that can be fine tuned by using an optimization For further details, please see [12]. Finally, the population P
technique with the aim of obtaining QoS efficient OLSR is updated with the best individuals in R (line 14). These steps
configurations for VANETs. are repeated until the termination condition is fulfilled.
We have employed three of the most commonly used QoS Algorithm 1 Pseudocode of NSGA-II.
metrics in this area [7] to evaluate the computed OLSR con- 1: P initialize population() // P = population
figurations (tentative solutions) and guide the search process: 2: Q // Q = auxiliary population
3: while not termination condition() do
Packet Delivery Ratio (PDR): the fraction of the data 4: for i 1 to (nsga-II.popSize / 2) do
packets originated by an application that are delivered. 5: parentsselection(P)
6: offspringrecombination(nsga-II.Pc,parents)
Normalized Routing Load (NRL): the ratio of routing 7: offspringmutation(nsga-II.Pm,offspring)
packet transmissions to data packets delivered. 8: evaluate fitness(offspring)
9: insert(offspring,Q)
End-to-End Delay (E2ED): the time spent from the packet 10: end for
is originated until is received at its destination. 11: RPQ
12: ranking and crowding(nsga-II, R)
Therefore, an efficient OLSR parameter setting for VANETs 13: P select best individuals(nsga-II, R)
maximizes PDR and minimizes NRL and E2ED. 14: end while
The search of possible combinations of OLSR parameter V. NSGA-II FOR OLSR O PTIMIZATION
values is not an easy problem. The dimension of the search In this work, the search for efficient OLSR parameter set-
space increases exponentially with both the number and the tings is carried out by the NSGA-II algorithm. The evaluation
range of possible parameter values. Thus, exact search meth- of tentative solutions involves realistic ns-2 VANET simula-
ods are not useful for efficiently solving the problem. In this tions in which the nodes use the tentative OLSR configuration
context, we have employed NSGA-II to compute accurate (see Fig. 1). After the simulation, the ns-2 output is analyzed
efficient configurations in reasonable times. to obtain the PDR, NRL, and E2ED to evaluate the solution.
We use the NSGA-II implementation available in the jMet-
IV. N ON - DOMINATED S ORTING G ENETIC A LGORITHM -II alCpp library [24]. In order to better adapt NSGA-II to
A MOP means to optimize a group of objective functions, the search for efficient OLSR parameter settings, we decide
that are generally in conflict with each other. Unlike in a to introduce some variations to the canonical initialization,
single-objective problem, there is not a unique optimized recombination, and mutation operators. We present the details
solution, but a set of solutions that represent different trade- of these operators bellow.
offs among the optimized functions. A generic formulation of A. Problem Encoding
a minimization MOP is [12]: The OLSR protocol is governed by eight different configu-
min F(x) = (f1 (x), f2 (x), . . . , fk (x)) ration parameters, as presented in Table I. As shown in Fig. 2,
subject to G(x) = (g1 (x), g2 (x), . . . , gj (x)) 0 the NSGA-II solutions are represented by individuals encoded
x as a vector with eight genes, one for each parameter.

573
C. Recombination
The genetic information exchange among the individuals
(solutions) is carried out by a modified version of N-point
crossover operator for real-valued problem encodings. It de-
fines a linear combination of two chromosomes Parent A and
Parent B to generate two new individuals Offspring A and
Offspring B. We include problem-related information in the
operator to avoid pointless configurations.
The concept of related-genes is defined to refer to
those genes that represent OLSR parameters which
have relation between their values described in the
Fig. 1. Optimization methodology followed in this work. OLSR RFC 3626. There are three pairs of related-
genes: HELLO INTERVAL and NEIGHB HOLD TIME;
The first three genes are real valued, and they represent
MID INTERVAL and MID HOLD TIME; and
the timeout timers before resending control messages (HE-
TC INTERVAL and TOP HOLD TIME. Thus, we
LLO INTERVAL, MID INTERVAL, and TC INTERVAL,
have defined five different possible gene exchanges (see
respectively). The forth one encodes the WILLINGNESS,
Fig. 3): the related-genes (Crossover-1 to Crossover-3), the
then taking an integer value from zero to seven. Finally,
WILLINGNESS (Crossover-4), and the DUP HOLD TIME
the last four are real valued, and they denote the timeout
(Crossover-5). Each time that the recombination operator
hold timers (NEIGHB HOLD TIME, MID HOLD TIME,
is called three of these five genes exchanges are randomly
TOP HOLD TIME, and DUP HOLD TIME, respectively).
applied to create the final offspring.
The valid ranges for each gene are presented in Table I.

Fig. 2. Solution encoding for the energy-aware OLSR tunning problem.

B. Initialization
The population initialization should distribute the individ-
uals uniformly in the search space as much as possible.
However, this uniform pattern is not easy to obtain when
using random operators and small populations. Therefore,
we propose using a uniform initialization to ensure that the
initial population contains individuals from different areas of
the search space. The initialization operator splits the search
space into pop diagonal subspaces (where pop is size of the Fig. 3. Solution encoding for the energy-aware OLSR tunning problem.
population), and it forcibly ensures that there is an individual
D. Mutation
located in each diagonal subspace [25]. The following equa-
The mutation operator introduces new genetic information,
tions summarize the initialization procedure.
and therefore, diversity to the population of the NSGA-
(0) 0
, high0p,i

xp,i = rand lowp,i i [0, 7], p [0, pop 1] (1) II. After analyzing the algorithm of the OLSR protocol,
0
we decide to introduce some problem-related information in
lowp,i = (p z sizei,pop ) + z(i,MIN ) (2) the mutation operator (see Equation 5). Thereby, the new
(0)
high0p,i = lowi + z sizei,pop (3) genetic information is randomly generated, always ensuring
the feasibility of the resulting mutated individual. More-
z(i,MAX) z(i,MIN )
z sizei,pop = (4) over, when a gene is modified it is checked whether the
pop related-genes continue fulfilling the restrictions defined in
the OLSR RFC. If there is a restriction that is not satisfied
For each gene i that encodes the i-th OLSR parameter,
then the related gene is also mutated taking into account
we divide the range of values into pop equal segments (see
the information about the previously mutated gene. For ex-
Equation 4, where z(i,MAX) and z(i,MIN ) are the upper
ample, when HELLO INTERVAL is modified it is checked
and lower values for the i-th parameter following the ranges
whether the NEIGHB HOLD TIME fulfills the restriction
defined in Table I). So, according to Equation 1, the initial
(NEIGHB HOLD TIME > HELLO INTERVAL).
value for each gene i in the solution vector that encodes the
(0)
p-th individual (xp,i ) is set to a randomly distributed value in
0 (g+1)
and high0p,i ).

the p-th segment (bounded between lowp,i xp,i = rand z(i,MIN ) , z(i,MAX) (5)

574
VI. E XPERIMENTAL A NALYSIS network workloads. They are grouped in low-rates (33, 66,
This section presents the experimentation carried out in and 100 kbps) and high-rates (333, 666, and 1000 kbps).
the multi-objective optimization of OLSR by using NSGA- The vehicles network devices employ the evaluated OLSR
II algorithm and ns-2 simulator, as the validation process. The parameterization. Table III summarizes the main features of
details of the NSGA-II implementation are described in the the network used in our VANETs simulations.
previous section. The simulations are performed by using the TABLE III
VANET COMMUNICATIONS SPECIFICATION .
ns-2.34 extended with UM-OLSR [26]. The study is carried
out by running 30 independent runs of NSGA-II in a Gigabit parameter value/protocol
Ethernet cluster of Pentium IV 2.4 GHz cores, 1 GB RAM, Propagation model Nakagami
and O.S. Linux Fedora core 6. Carrier Frequency 5.89 GHz
Medium Capacity 6 Mbps
A. Urban VANET Scenarios PHY/MAC Layer IEEE 802.11p
The simulation testbed is composed of 18 urban VANET Routing Layer OLSR
scenarios covering a real urban area of Malaga (Spain) by Transport Layer UDP
taking into account different road traffic densities and network CBR Packet Size 512 bytes
workloads. Tables II and III summarize the main features of 500 kbps
the scenarios used in our experimentation. CBR Data Rate 33/66/100 kbps
333/666/1000 kbps
CBR Time 60 s

Ns-2 is used to evaluate the VANET communications.


With the aim of reflecting the network interactions in a
trustworthy manner, the simulated vehicles are equipped with
Unex DCMA-86P2, a real transceiver specifically designed to
support IEEE 802.11p. In turn, the fading Nakagami radio
propagation model representing the WAVE radio propagation
in urban scenarios [28] is included in the simulations .

B. Optimization Performance
Fig. 4. Malaga urban areas taken into account in each VANET scenario.
This section presents the quality of results and the com-
putational efficiency of the optimization method employed.
The vehicular environment is generated by using the
We perform 30 independent runs of the steady state NSGA-II
SUMO [27] traffic simulator, taking into account a 240,000m2
that is configured with eight individuals that evolve during 65
area from a suburb of Malaga (see Fig. 4). Different road
generations, a crossover probability Cr=0.9, and a mutation
traffic densities named L (low), M (medium), and H (high)
factor =0.125. The evaluation of the tentative solutions is
are also defined. Thus, we study how the compared OLSR
carried out by invoking the simulation of such a configuration
configurations perform when the network size and road traffic
over the low traffic density scenario (20 vehicles) in which the
density change. In these geographical areas, there are different
there are ten data transfers of 60 seconds with a data rate of
number of vehicles moving through the roads following real
500 kbps. This scenario is named L-500kbps.
traffic rules during three minutes (see Table II).
After the experimentation, we computed the global Pareto
TABLE II front by gathering the non-dominated solutions found in the
VANET SCENARIOS DETAILS . 30 independent runs, which is composed by 44 non-dominated
solutions (OLSR configurations). These solutions present dif-
scenario name # of vehicles CBR sources ferent trade-offs among the optimized QoS metrics.
L (low) 20 10 The distance of the solutions to the ideal vector [15] is
M (medium) 30 15 employed to compare the obtained configurations. The ideal
H (high) 40 20 vector contains the optimum for each separately considered
objective achieved at the same point in the search space. Thus,
In each scenario, there is a number of data exchanges this vector presents the best theoretical solution assuming
between pairs of vehicles. The data is generated by a constant equally weighted objectives. For this reason, we have mea-
bit rate (CBR) generator for 60 seconds. The number of sured the quality of the solutions giving that the QoS metrics
vehicles that generate the information (CBR sources) to be are equally important by using their distance to the ideal
sent to the other nodes depends on the VANET scenario (see vector. In this work, the ideal vector has three components
Table II). In turn, we have defined different scenarios by that are the obtained maximum PDR and minimum NRL and
using six different traffic data rates to experiment with several E2ED during the optimization process (99.96, 0.38, 1.69).

575
We focus on the solutions that produce the best values for
each QoS metric maximum PDR (max-PDR), minimum NRL
(min-NRL), and minimum E2ED (min-E2ED) and the closest
one to the ideal vector (min-distance). Table IV presents their
QoS and distance to the ideal vector. In turn, it shows the QoS
achieved by the OLSR RFC 3626 (OLSR-RFC) configuration.
Finally, the last row contains the median value for each metric
for the 44 non-dominated solutions found.
TABLE IV
Q O S RESULTED SIMULATING THE OLSR RFC 3626 AND FOUR
OPTIMIZED CONFIGURATIONS OVER L-500kbps VANET SCENARIO , AND
MEDIAN VALUES FOR THE WHOLE PARETO FRONT.

configuration PDR NRL E2ED distance Fig. 5. Normalized QoS metrics of the Pareto front solutions.
min-distance 99.93% 1.02% 2.45ms 0.120
Finally, in terms of the computational effort required to
max-PDR 99.96% 1.01% 2.52ms 0.124
automatically obtain QoS efficient OLSR settings, the exe-
min-NRL 79.10% 0.38% 3.72ms 0.238
cution time for each NSGA-II execution is measured. The
min-E2ED 24.54% 2.02% 1.69ms 1.024
mean execution time is 2.1765E+4 seconds (about 6 hours)
OLSR-RFC 78.33% 5.16% 66.04ms 1.198 with a deviation of 8.24%. This relative low effort in the
NSGA-IImed 75.60% 0.74% 2.43ms 0.192 protocol design is completely justified by the subsequent
benefits obtained in the quality of the communications, as we
The max-PDR configuration delivers almost all the data show in the following analysis.
packets (PDR=99.96), it does not generate an excessive routing
overhead and the packets reach the destination nodes in C. Experimental Validation
competitive times. The min-NRL settings generates 13 times A set of VANET simulations involving the 18 urban sce-
lesser control messages than the RFC 3626 configuration narios presented in Section VI-A are carried out to compare a
(min-NRL NRL=0.38% and OLSR-RFC NRL=5.16%). How- MO optimized configuration with the OLSR RFC 3626 (RFC)
ever, the routing tables are updated less often, producing a and with PSO-OLSR, another optimized version of OLSR
reduction in the performance in terms of PDR and E2ED. computed with mono-objective PSO algorithm [11].
The configuration that delivers the data packets in shortest As we do not focus on any specific VANET application, we
times (E2ED=1.69ms), the min-E2ED, is the optimized con- have taken into account the min-distance OLSR configuration
figuration that generated the largest NRL. In turn, this is the in the validation experiments because it is the most
least balanced configuration of the studied ones because it balanced configuration of the Pareto front, and therefore,
provokes a severe drop of delivered packets (PDR=24.51%). it presents the most competitive trade-off among the QoS
The closest configuration to the ideal vector (min-distance) metrics. The min-distance parameterization is HELLO INT=
presents the most balanced QoS, since it offers the second 4.368, REFRESH INT=34.222, TC INT=14.852, WILL-
best values for each metric being these values close to the INGNESS=1, NEIGHB H T=44.808, MID H T=14.852,
best ones (PDR=99.93%, NRL=1.01%, and E2ED=2.45ms). TOP H T=73.920, and DUP H T=10.935.
Globally, the optimized configurations drastically reduce the The experimental validation evaluates the three QoS metrics
delivery times (E2ED=2.43ms) and the routing workload used during the optimization process and also the size of the
(NRL=0.74%) keeping similar packet delivery rates concern- routing paths regarding the number of hops required to reach
ing to the OLSR RFC 3626. the destination, named Routing Path Length (RPL). Table V
Fig. 5 shows the normalized values of the three optimized shows the average results grouped by road traffic density.
objectives for the Pareto front solutions ranked according to TABLE V
the distance to the ideal vector. Thus, the solution number one VALIDATION EXPERIMENTS RESULTS GROUPED BY TRAFFIC DENSITY.
is the closest solution to the ideal vector and number 44 is the config.
QoS metrics
PDR NRL E2ED RPL
furthest from it. The solutions are grouped in three sets (Set A, urban L scenario
Set B, and Set C) according to their QoS in comparison with min-distance 76.05% 3.40% 27.67ms 1.16
PSO-OLSR 80.74% 6.23% 32.14ms 1.18
the OLSR RFC. Set A contains the first 19 solutions, which RFC 85.96% 16.61% 63.14ms 1.10
significantly improve the studied QoS metrics. Set B groups urban M scenario
min-distance 62.21% 4.10% 88.36ms 1.40
the solutions which reduce NRL and E2ED with a slight loss PSO-OLSR 58.46% 9.35% 132.51ms 1.52
of PDR (<15%). Finally, Set C presents the less competitive RFC 69.94% 18.70% 137.68ms 1.19
urban H scenario
solutions, which reduce both, NRL and E2ED, but suffer a min-distance 49.86% 8.18% 151.14ms 1.93
PDR degradation greater than 15%. This figure confirms the PSO-OLSR 49.49% 14.05% 216.19ms 2.02
RFC 52.64% 29.67% 231.16ms 1.41
results presented in Table IV, since most of solutions improve average
or keep the PDR while reducing significantly the NRL and the min-distance 62.71% 5.23% 89.06ms 1.50
PSO-OLSR 62.90% N 9.88% N 126.94ms 1.58
E2ED regarding OLSR-RFC. RFC N 69.51% N 21.70% N 143.99ms N 1.23

576
With the aim of providing the comparison with statistical
confidence, we apply Wilcoxon signed-rank statistical test [29]
to compare the distribution of the results of PSO-OLSR and
RFC against the MO optimized configurations (min-distance).
The confidence level is set to 95% (p-value<0.05). The results
of this test are presented in the two last rows of Table V by
using triangles. On the one hand, the black triangle (N) means
that there exists a statistical difference against min-distance.
On the other hand, the white one () means that there is not
statistical difference.
Summarizing the results presented in Table V and figures 6,
7, and 8, we observe that the QoS offered by these three Fig. 8. E2ED results grouped by road traffic density.
protocols gets worse the PDR decreases and the NRL, E2ED,
Finally, the routing paths generated by the standard OLSR
and RPL increases with the road traffic density, and therefore,
configuration are shorter than the ones generated by the
with the number of nodes connected in the network.
By examining the PDR indicator (see Fig. 6), we can check optimized configurations. However, on average, the routes
that both optimized routing protocols behaved similarly. The computed by the three studied configurations require lesser
than 1.6 hops to reach the destination nodes, which is a
average difference between their PDR is lower than 0.20%.
However, the RFC configuration delivers a larger number of competitive length in highly dynamic networks as VANETs.
packets, even its performance gets close to the optimized Globally, we observe that the RFC configuration delivers
configurations in the high road traffic density scenario (H). more data packets through the shortest routing paths. However,
the differences with the optimized configurations decrease in
the highest road traffic density. This is because the standard
configuration exchanges more routing messages in order to
maintain the routing tables up to date. The min-distance con-
figuration is the most robust in terms of resources consumption
and scalability. It produces statistically the least routing load
with a degradation lower than 7% in the PDR. In turn, the
reduction of routing overhead mitigates the OLSR congestion
and medium access problems, and therefore, even when the
routing paths are longer the packets reach the destination nodes
in significantly shorter times.
Fig. 6. PDR results grouped by road traffic density.
The most important improvement of the configuration found VII. C ONCLUSION
by NSGA-II appears in the routing traffic workload (NRL). As
The design of efficient protocol communications is an
Fig. 7 shows, the RFC configuration generates between three
important issue in VANETs deployment. In this study, we have
and five times more routing traffic than our computed min-
analyzed the use of an automatic intelligent method to obtain
distance configuration. In turn, the multi-bjective optimized
QoS optimized OLSR configurations by coupling a MOEA,
configuration of OLSR reduces between 41% and 56% the
NSGA-II, and the ns-2 simulator. So, the main contribution
routing load generated by PSO configuration.
of this study is the application of a MOEA to a VANET
protocol optimization, avoiding the problems of using mono-
objective optimization metaheuristics and the weighted factors
functions. Moreover, the NSGA-II optimized configuration
found is validated comparing it against the mono-objective
PSO optimized settings and the standard one in RFC 3626,
studying their performance in terms of QoS in 18 urban
VANET scenarios.
Regarding the VANET communications, the experimental
analysis demonstrates that significant improvements are ob-
tained in terms of routing overhead when using the QoS opti-
Fig. 7. NRL results grouped by road traffic density. mized configurations by NSGA-II (average routing workload
Concerning E2ED (see Fig. 8), the times transferring data reduction of 47% and 76% regarding PSO and OLSR RFC
packets are significantly shorter when the min-distance OLSR configurations, respectively), mitigating the OLSR congestion
configuration is used. On average, the packets require 89.06 ms problem. In turn, this feature makes this routing the most
to reach the destination node, which is a 32% and 38% less robust in terms of resources consumption and scalability of
time than when the PSO and RFC ones are used, respectively. the compared configurations. In addition, the reduction of

577
the routing protocol medium accesses statistically improves [8] A. Laouiti, P. Muhlethaler, F. Sayah, and Y. Toor, Quantitative evalua-
the packet delivery times even when the computed paths are tion of the cost of routing protocol OLSR in a Vehicle Ad Hoc NETwork
(VANET), in Proceedings of the 67th IEEE Vehicular Technology
slightly longer with regard to the OLSR RFC 3626 config- Conference, VTC Spring 2008. IEEE, 2008, pp. 29862990.
uration (E2ED is up to 38% lower). All these features are [9] J. Santa, M. Tsukada, T. Ernst, O. Mehani, and A. F. Gomez-Skarmeta,
achieved with a PDR degradation lower than 7% as regards Assessment of VANET multi-hop routing over an experimental plat-
form, Int. J. Internet Protoc. Technol., vol. 4, no. 3, pp. 158172, 2009.
RFC. [10] C. Gomez, D. Garca, and J. Paradells, Improving performance of a real
The optimization methodology used in this study (coupling ad hoc network by tuning OLSR parameters, in ISCC 05: Proceedings
NSGA-II multi-objective algorithm and VANET simulations of the 10th IEEE Symposium on Computers and Communications.
Washington, DC, USA: IEEE Computer Society, 2005, pp. 1621.
in the evaluation process analysis) offers the possibility of au- [11] J. Toutouh, J. Garcia-Nieto, and E. Alba, Intelligent olsr routing proto-
tomatically and efficiently customizing any VANET protocol, col optimization for vanets, Vehicular Technology, IEEE Transactions
not just OLSR. In turn, this approach can be used to improve on, vol. 61, no. 4, pp. 18841894, 2012.
[12] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, A fast and elitist
other existing strategies applied over OLSR that appear in the multiobjective genetic algorithm: Nsga-ii, Evolutionary Computation,
literature. This is an added value of this line of research. IEEE Transactions on, vol. 6, no. 2, pp. 182197, 2002.
The main lines of future research are related to the issue [13] The Network Simulator Project - Ns-2, [online]
http://www.isi.edu/nsnam/ns/, 2011.
of improving the method used in the automatic search to [14] C. Blum and A. Roli, Metaheuristics in combinatorial optimiza-
find optimized QoS OLSR configurations for VANETs. The tion: Overview and conceptual comparison, ACM Computing Surveys,
application of other MOEAs should be considered, including vol. 35, no. 3, pp. 268308, 2003.
[15] C. Coello, G. Lamont, and D. Van Veldhuizen, Evolutionary algorithms
new QoS metrics, such as the RPL. In turn, the parallelization for solving multi-objective problems. Springer-Verlag New York Inc,
of the population evaluation by multithreading techniques 2007, vol. 5.
could allow the use of larger populations and number of [16] J. Toutouh and E. Alba, Optimizing OLSR in VANETs with Differ-
ential Evolution: A Comprehensive Study, in First ACM International
generations used in these optimization methods with the conse- Symposium on Design and Analysis of Intelligent Vehicular Networks
quent improvement of the search of new solutions. Finally, the and Applications (DIVANet 11), Miami, Florida, USA, 2011.
approach proposed in this paper could be extended by using [17] E. Alba, B. Dorronsoro, F. Luna, A. Nebro, P. Bouvry, and L. Hogie,
A Cellular MOGA for Optimal Broadcasting Strategy in Metropolitan
several VANET scenarios to evaluate each OLSR tentative MANETs, Computer Communications, vol. 30, no. 4, pp. 685697,
configuration to obtain even more accurate results. 2007.
[18] P. Ruiz, B. Dorronsoro, and P. Bouvry, Optimization and performance
ACKNOWLEDGMENT analysis of the AEDB broadcasting algorithm, in Computer Communi-
cations and Networks (ICCCN), 2011 Proceedings of 20th International
Authors acknowledge funds from the CICE, Junta An- Conference on, 31 2011-aug. 4 2011, pp. 1 6.
[19] P. Ruiz, B. Dorronsoro, G. Valentini, F. Pinel, and P. Bouvry, Optimisa-
dalucia, under contract P07-TIC-03044 (DIRICOM project) tion of the enhanced distance based broadcasting protocol for manets,
and Spanish Ministry of Sciences and Innovation (MICINN) The Journal of Supercomputing, pp. 128, 2011.
and FEDER under contracts TIN2011-28194 (roadME) and [20] W. Abdou, A. Henriet, C. Bloch, D. Dhoutaut, D. Charlet, and F. Spies,
Using an evolutionary algorithm to optimize the broadcasting methods
TIN2008-06491-C04-01 (M*). Jamal Toutouh is supported by in mobile ad hoc networks, J. Netw. Comput. Appl., vol. 34, no. 6, pp.
grant AP2010-3108 from the Spanish Government. 17941804, Nov. 2011.
[21] J. Garca-Nieto, J. Toutouh, and E. Alba, Automatic tuning of commu-
R EFERENCES nication protocols for vehicular ad hoc networks using metaheuristics,
Eng. Appl. Artif. Intell., vol. 23, no. 5, pp. 795805, 2010.
[1] N. Lu, Y. Ji, F. Liu, and X. Wang, A dedicated multi-channel MAC [22] J. Garca-Nieto and E. Alba, Automatic parameter tuning with meta-
protocol design for VANET with adaptive broadcasting, in Wireless heuristics of the AODV routing protocol for vehicular ad-hoc networks,
Communications and Networking Conference (WCNC), 2010 IEEE, in EvoApplications (2), ser. Lecture Notes in Computer Science, C. D.
April 2010, pp. 1 6. Chio, A. Brabazon, G. A. D. Caro, M. Ebner, M. Farooq, A. Fink,
[2] T. Taleb, E. Sakhaee, A. Jamalipour, K. Hashimoto, N. Kato, and J. Grahl, G. Greenfield, P. Machado, M. ONeill, E. Tarantino, and
Y. Nemoto, A stable routing protocol to support ITS services in VANET N. Urquhart, Eds., vol. 6025. Springer, 2010, pp. 2130.
networks, Vehicular Technology, IEEE Transactions on, vol. 56, no. 6, [23] J. Toutouh, S. Nesmachnow, and E. Alba, Fast energy-aware olsr
pp. 3337 3347, November 2007. routing in vanets by means of a parallel evolutionary algorithm, Cluster
[3] F. Li and Y. Wang, Routing in vehicular ad hoc networks: A survey, Computing, pp. 116, 2012.
Vehicular Technology Magazine, IEEE, vol. 2, no. 2, Jun. 2007. [24] J. J. Durillo and A. J. Nebro, jmetal: A java framework for multi-
[4] W. Zhang, A. Festag, R. Baldessari, and L. Le, Congestion control objective optimization, Advances in Engineering Software, vol. 42, pp.
for safety messages in VANETs: Concepts and framework, in ITS 760771, 2011.
Telecommunications, 2008. ITST 2008. 8th International Conference on, [25] C. Chou and J. Chen, Genetic algorithms: initialization schemes and
October 2008, pp. 199 203. genes extraction, in The Ninth IEEE International Conference on Fuzzy
[5] T. Clausen and P. Jacquet, Optimized Link State Routing Pro- Systems, vol. 2, 2000, pp. 965968.
tocol (OLSR), IETF RFC 3626, [online] Available in URL [26] F. J. Ros, Um-olsr, http://masimum.dif.um.es/?Software:UM-OLSR.
http://www.ietf.org/rfc/rfc3626.txt, United States, 2003. [27] D. Krajzewicz, M. Bonert, and P. Wagner, The open source traffic
[6] T. Chen, O. Mehani, and R. Boreli, Trusted routing for VANET, simulation package SUMO, in RoboCup06, Bremen, Germany, 2006,
in ITST 2009, 9th International Conference on Intelligent Transport pp. 110.
Systems Telecommunications, M. Berbineau, M. Itami, and G. Wen, Eds. [28] V. Taliwal, D. Jiang, H. Mangold, C. Chen, and R. Sengupta, Empirical
Piscataway, NJ, USA: IEEE Computer Society, October 2009, pp. 647 determination of channel characteristics for DSRC vehicle-to-vehicle
652. communication, in Proceedings of the 1st ACM international workshop
[7] J. Harri, F. Filali, and C. Bonnet, Performance comparison of AODV on Vehicular ad hoc networks, ser. VANET 04. New York, NY, USA:
and OLSR in VANETs urban environments under realistic mobility ACM, 2004, pp. 8888.
patterns, in Med-Hoc-Net 2006, 5th Annual Mediterranean Ad Hoc Net- [29] R. Wilcox, New statistical procedures for the social sciences. Hillsdale,
working Workshop, S. Basagni, A. Capone, L. Fratta, and G. Morabito, 1987.
Eds. IFIP, June 2006.

578

Вам также может понравиться