You are on page 1of 6

A Routing Strategy for Vehicular Ad Hoc Networks in City

Environments1
Christian Lochert

Hannes Hartenstein

Jing Tian

Network Laboratories
NEC Europe Ltd.
Heidelberg, Germany
lochert@ccrle.nec.de

Network Laboratories
NEC Europe Ltd.
Heidelberg, Germany
hartenst@ccrle.nec.de

Distributed Systems Department


University of Stuttgart
Stuttgart, Germany
jing.tian@informatik.uni-stuttgart.de

Holger Fler

Dagmar Hermann

Martin Mauve

Department of Computer Science


University of Mannheim
Mannheim, Germany
fuessler@informatik.uni-mannheim.de

Traffic Assistance Systems


DaimlerChrysler AG
Stuttgart, Germany
dagmar.hermann@daimlerchryler.com

Department of Computer Science


University of Mannheim
Mannheim, Germany
mauve@informatik.uni-mannheim.de

Abstract
Routing of data in a vehicular ad hoc network is a challenging task due to the high dynamics of such a network.
Recently, it was shown for the case of highway traffic
that position-based routing approaches can very well deal
with the high mobility of network nodes. However, baseline position-based routing has difficulties to handle twodimensional scenarios with obstacles (buildings) and voids
as it is the case for city scenarios. In this paper we analyze a position-based routing approach that makes use of
the navigational systems of vehicles. By means of simulation we compare this approach with non-position-based ad
hoc routing strategies (Dynamic Source Routing and AdHoc On-Demand Distance Vector Routing). The simulation makes use of highly realistic vehicle movement patterns derived from DaimlerChryslers Videlio traffic simulator. While DSRs performance is limited due to problems
with scalability and handling mobility, both AODV and the
position-based approach show good performances with the
position-based approach outperforming AODV.

1 Introduction
Communication between vehicles by means of wireless
technology has a large potential to improve traffic safety
and travel comfort of drivers and passengers. Current advances in the field of wireless ad hoc networks show that
inter-vehicle communication based on vehicular ad hoc networks is a feasible approach that has a competitive edge over
cellular network-based telematics with respect to several aspects: low data transport times for emergency warnings, robustness due to the networks mesh structure, and low costs
for usage due to the use of unlicensed frequency bands. The
1 This work has been carried out within the framework of the FleetNet
project as part of BMBF contract no. 01AK025D. J. Tian acknowledges
support from EU IST Project CarTalk 2000 (IST-2000-28185).

FleetNet project is currently developing a communication


platform for inter-vehicle communication following the ad
hoc networking paradigm [6].
Several potential applications in the area of inter-vehicle
communications require data routing algorithms for the underlying ad hoc network: when communication endpoints
are not within their respective radio transmission range, unicast routing is required to establish communication between
two vehicles or between a vehicle and a fixed gateway.
Communication partners are either selected based on their
identity, e.g., IP address, or based on their geographic position. The latter case refers to applications where a person
in a vehicle requests some information, e.g., on traffic flow
or road conditions, from a specific geographic region. To
support such application, geocast routing [7] should also be
provided by the underlying routing protocol.
Traditional ad hoc routing protocols have difficulties in dealing with the high mobility specific to vehicular ad hoc networks. In a recent paper [5] we have shown for highway
scenarios that routing approaches using position information, e.g., obtained from on-board GPS receivers, can very
well deal with the mobility of the nodes. A general survey
on position-based routing approaches is presented in [12].
Related to our work with its focus on vehicular networks
is the work described in [14] that proposes to add geocast
abilities to the Ad Hoc On-demand Distance Vector protocol (AODV).
In this paper we discuss aspects of position-based routing
for the case of vehicular ad hoc networks in a city environment. In contrast to highway scenarios, we now have
to deal with problems associated to a truly two-dimensional
system area as well as with problems like radio obstacles
due to buildings. We present a simulation study that compares a position-based routing approach with classical ad
hoc routing methods (Dynamic Source Routing and Ad Hoc
On-Demand Distance Vector). To the best of our knowl-

edge, our study is the first one that evaluates ad hoc routing
protocols over a realistic vehicle movement pattern for a city
scenario. The vehicular traffic simulation was done using a
simulation tool of DaimlerChrysler and is based on actual
traffic measurements taken in the city of Berlin, Germany.
The paper is structured as follows: In Section 2 we first discuss the challenges of position-based routing in a city scenario, show the shortcomings of existing approaches and
then outline our algorithm that is based on ideas from the
Terminodes project [2] and on a proposal presented in [15].
Section 4 gives details on the generation of the vehicle
movement patterns used in our simulation study. The results of the routing protocol simulations are presented and
discussed in Section 5. Section 6 summarizes our analysis
and outlines open issues.

The perimeter mode of GPSR consists of two elements: i) a


distributed planarization algorithm and ii) an online routing
algorithm for planar graphs. The planarization algorithm
transfers (locally) the connectivity graph into a planar graph
by eliminating redundant edges. The elimination criteria
used for the perimeter mode approach are illustrated in Figure 1. The online routing algorithm then forwards a packet
along the faces2 of the planar graph towards the destination. Intuitively, as depicted in Figure 4(a) the packet will
be routed according to the Right Hand Rule from node u
where perimeter mode is entered via v, w, x, and y to destination node D. In case the packet traverses an inner face
and the destination is not part of the this face, the algorithm switches towards the face closer to the destination
node whenever it encounters an edge that crosses the imaginary line from the perimeter mode starting point (here: u)
to the destination node D. By referring the reader to [9] for
details on this algorithm, we now study the deficiencies of
the perimeter mode approach for the case of city scenarios.

2 Challenges of Position-Based Routing in a City


Environment
Position-based routing bases forwarding decisions on position information. Thus, there are several requirements on
the availability of position information: first of all, positionbased routing requires position-awareness of all participating nodes, e.g., through a GPS receiver on each node. Furthermore, it is assumed that each node is aware of the positions of its direct neighbors: each node periodically sends
out beacon messages that indicate the current position of
the node1 . In order to send a packet to a destination node,
a sending node also requires information on the current geographic position of the destination in order to include it in
the packet header and to make the routing decision. This information is gained via a so-called location service. While
several proposals for location services have recently been
made in the literature [12], in this paper we make use of
a very simple reactive location service (RLS) described in
Section 3.
With the above-mentioned information at hand a node forwards the packet to the direct neighbor that is closest to
the destination position. This strategy is called greedy
position-based routing. Since this is a purely local decision,
no route setup or maintenance is required. Instead, forwarding hops are determined on the fly. Greedy forwarding,
however, can fail when there is no neighbor available that is
closer to the destination than the current forwarding hop. In
such a case, position information points in the right direction
but is not correlated with available paths to the destination.
Several recovery strategies like Perimeter Mode in Greedy
Perimeter Stateless Routing (GPSR) [9] or face-2 [3] were
proposed in the literature. For the remainder of this paper,
we stick to the notations and definitions of [9].
1 The beacon messages are not forwarded, they are only intended to inform the direct neighbors.

w
w
u

(a) GG

(b) RNG

Figure 1: Gabriel Graph and Relative Neighborhood Graph criteria for planarization: the edge between u and v is eliminated when there is another node in the grey-shaded
region.

Network disconnection. City scenarios with almost all


area between streets covered with buildings significantly
limit the applicability of purely greedy position-based routing and of corresponding recovery strategies. Due to these
obstacles, nodes that would have seen each other in a freespace model might not be aware of each other anymore.
Since the planarization methods outlined above assume that
the connectivity between nodes only depends on the node
distances, planarization of the neighborhood in the presence
of obstacles can lead to network disconnection as illustrated
in Figure 2. Both criteria, GG and RNG, would lead to elimination of the edge between u and v under the misconception
that nodes w1 and w2 can reach node v.
Too many hops. A planarized connectivity graph for vehicles along a single street essentially leads to a graph where
a vehicle no longer sends packts to the neighbor with the
largest forward progress (Fig. 3). Thus, compared to greedy
routing, many more nodes have to be traversed when routed
in perimeter mode. This leads to increased delays and, frequently, to elapsed hop counts.
2 A planar graph partitions the plane in an unbounded outer face and
several bounded inner faces.

w
x
b

w_2

v
u

w_1

(a) Working
Perimeter
Mode

u
a

a
S

(b) Routing Loop

Figure 2: Planarization problems.


Figure 4: Problems while Routing in Perimeter Mode.




(a) traversing a planar graph












(b) greedy routing

method that is supported by a map of the city. It is called Geographic Source Routing (GSR). The presence of a map is a
valid assumption, e.g., when vehicles are equipped with onboard navigation systems. As outlined in the introduction to
position-based routing above, we make use of our reactive
location service in order to learn the current position of a
desired communication partner.

Figure 3: Traversing a planar graph versus greedy routing.

Routing Loops. As indicated in [9, 8] mobility can induce


routing-loops for packets being routed in perimeter mode.
Figure 4 sketches a corresponding scenario. Node S wants
to send a packet to node D. In node u, greedy forwarding
fails because there is reachable node closer to the destination node. So the forwarding mode is switched to perimetermode. The initial face will be set to uv. In a static network,
the packet would reach D according to the Right Hand
Rule as expected, but with movement of node x reaching
node vs radio range while v already has sent the packet,
a routing loop is created between vwx by the Right Hand
Rule. The traversal of the initial face would be used to determine a face-loop but since it is never traversed again, the
packet is circled until the max hop count is reached.
Wrong Direction. As one can see in Fig. 4(a) the perimeter
mode following the Right Hand Rule is biased towards a
specific direction when selecting a next hop. When there exist more than one routing alternative this can lead to routes
longer than necessary. These long routes lead again to the
problem of Too many hops and also is prone to error because of mobility.

3 Geographic Source Routing


As a strategy to deal with the high mobility of nodes on the
one hand and with the specific topological structure of a city
on the other hand, we have chosen a position-based routing

RLS is a direct translation of the route discovery procedure


used in reactive non-position-based ad hoc routing protocols
to the position discovery of position-based routing. Essentially, the querying node floods the network with a position request for some specific node identifier. When the
node that corresponds to the requested identifier receives
the position request, it sends a position reply back to the
querying node. In order to avoid extensive flooding as well
as the well-known broadcast storm problem, several optimizations have been implemented for RLS. An extensive
analysis of RLS is given in [11].
With this information, the sending node can compute a path
to the destination by using the underlying map of the streets.
In other words, the sender computes a sequence of junctions
the packet has to traverse in order to reach the destination.
The sequence of junctions can either be put into the packet
header as it would be done in a geographic source routing
approach [2] or it can be computed by each forwarding
node. Clearly, there is a trade-off between bandwidth consumption and required processing performance when choosing between these two options. But one can see that there
are only differences in the result if a node uses information about its neighborhood, i.e., to look for connectivity
breaks of a route which was chosen before and therefore
recomputes another route to the destination. Forwarding a
packet between two successive junctions is done on the basis of greedy forwarding since here no obstacles should
block the way. In the current implementation for our simulations, the path between source and destination is determined by a Dijkstra shortest path calculation based on the
street map. Since we do not take into account whether there

are enough vehicles on a street to provide connectivity between the two involved junctions, there is a large potential
to improve our results by other path-finding strategies. Simulations results for the outlined methods in comparison to
non-position-based routing strategies are given in Section 5.
The following section provides some insight into the modeling of the vehicles in a city environment.

time: 2.00245

4 City Traffic Simulations


City traffic simulation itself is a complex challenge because
the traffic flow in conurbation deeply depends on rules at
its intersections and on the capacity of the roads and intersections. The traffic flow simulator Videlio [10], developed
by DaimlerChrysler AG, is based on microscopic simulation
using time depending origin destination matrices. Core elements are the Optimal Velocity Model [1], a special lane
changing model [10] and the C-logit model [16] to calculate
the traffic assignments. Videlio uses a detailed description
of the road network with information about, e.g., lane numbers, traffic regulations, and time tables of the traffic lights.

Figure 6: Graph of streets of our vehicular movement simula-

For the vehicular movement pattern generation, a small part


(6.25 km 3.45 km) of the city of Berlin was modeled as a
graph of streets with 28 vertices and 67 edges as depicted in
Figures 5 and 6. In total, movements of 955 vehicles have
been simulated.

We assume that nodes transmit according to the IEEE


802.11 wireless LAN standard (2 Mbps). The transmission
range was originally set to 250 m but an analysis of the connectivity graph of the network shows that with this range
too often there is no connectivity along the street between
two junctions. Real-world tests with vehicles equipped
with IEEE 802.11b cards also have shown that transmission
ranges in the order of 500 to 800 meters are feasible with
external antennas. We thus have set the transmission range
in our simulations to 500 m.

tions.

the CMU-extensions [4]. The simulator was extended with a


basic form of obstacle modeling: spaces between streets are
assumed to be buildings and, therefore, radio waves cannot
propagate through them. Thus, two nodes can only communicate directly with each other when they are in their
respective transmission ranges and also obey the line-ofsight criterion.

We selected several pairs of nodes randomly; the reactive


location service is used to determine the current position
of the communication partner. Then, each pair of nodes
exchanges 20 packets over 5 seconds. We measured the
achieved packet delivery rate (Fig. 7) versus the distance
between the two communication partners (at the beginning
of the communication) as well as the associated bandwidth
consumption (Fig. 8), latency for the first packet (Fig. 9)
and number of hops (Fig. 10). Each point in the above mentioned graphs is based on at least 10 000 packets exchanged.
Figure 5: Part of the city of Berlin modeled for our simulation.

5 Simulations and Results


The simulation study presented in this section compares
the map-based geographic source routing approach (GSR)
with Dynamic Source Routing (DSR) and Ad Hoc OnDemand Distance Vector Routing (AODV), two classical
non-position-based ad hoc routing strategies. Simulations
were done with the network simulator NS-2 [13] in its version ns-2.1b8a (with some bug fixes from ns-2.1b9a) with

The study on achievable packet delivery rate (Fig. 7) shows


good results for the position-based approach compared to
DSR that shows some performance problems. In comparison to AODV there is still some advantage for the positionbased approach. A detailed analysis of the causes for not
delivering packets indicates that the position-based routing
approach fails when the connectivity on a street selected by
the path-finding algorithm is broken. This is also confirmed
by Fig. 11 that shows the effect of reducing the transmission
range from 500 m to 250 m. Thus, with improved and more
adaptive path selection procedures one can expect to improve the obtained results for the position-based approach.

Avg. Latency of First Delivered Packet


1.8

0.9

1.6

0.8

1.4

0.7

1.2

0.6

time [sec]

packet delivery rate

Avg. Delivery Rate


1

0.5
0.4

1
0.8
0.6

0.3

0.4

0.2

GSR 500
AODV 500
DSR 500

0.1

0
500

1000

AODV 500
GSR 500
DSR 500

0.2
1500

2000
2500
distance [m]

3000

3500

0
500

4000

1000

1500

2000
2500
distance [m]

3000

3500

4000

Figure 9: Latency of first packet per connection.

Figure 7: Packet delivery ratio versus distance of communication


partners.
Avg. Total Bandwidth

Avg. Number of Hops

16384

16

8192
8

2048
1024

# hops

bandwidth [kbps]

4096

512

256
2

128
AODV 500
GSR 500
DSR 500

64
32
500

1000

1500

2000
2500
distance [m]

3000

3500

4000

Figure 8: Average Total Bandwidth consumption per second.


The main problem of DSR appears to be the network load
it generates with its signaling traffic (Fig. 8). A study on
the performance of DSR that makes use of an ideal MAC
scheme with (almost) unlimited bandwidth available shows
significantly better results for DSR. The achieved packet delivery rate grows significantly, e.g., for communication partners with a distance up to 4 km from about 20 % up to the
range of 60 % to 70 % (Fig. 12). But still DSR creates large
packets because of the source route inscribed in the headers,
especially during route discovery phase, which leads to a
significant bandwidth overload. The second cause of failure
for DSR is mobility causing frequent route breaks. But this
is less dramatic than for highway scenarios since mobility
rate of city scenarios is much lower than of highway scenarios. On the positive side DSR has less problems when a
given street does not provide connectivity, since it can very
well find other non-direct routes.
In contrast to the significant bandwidth consumption of
DSR, simulations of AODV show delivery rates similar to
GSR. By extending the communication distances bandwidth consumption of AODV is increasing up to the regions
of GSR because AODV uses an expanding ring searchtechnique for route discovery. With the position-based ap-

1
500

GSR 500
AODV 500
DSR 500
1000

1500

2000
2500
distance [m]

3000

3500

4000

Figure 10: Average number of hops depending on distance.

proach, the packets flooded during the reactive location request period are of constant size.
The observed latency Fig. 9 for the first packet of a connection is similar for DSR and GSR approaches with a small
advantage for DSR. This was to be expected since the route
establishment in DSR and the location discovery in positionbased routing are very similar. The usage of expanding ring
search technique of AODV is responsible for the higher latency since it is a trade-off between bandwidth consumption
and latency.
In contrast to DSR where a node drops a packet when route
breaks occur GSR uses some recovery strategies (fall back
on greedy mode) to by-pass this particular node. As can be
seen in Fig. 10 this strategy implies a slightly longer route to
the destination node. AODV using routing tables instead of
source routes shows similar results to GSR. One can assume
that DSR is much more aggressive in routing to a destination, i.e., uses the node with the largest progress which also
could go out of radio range shortly. This would lead to more
route breaks involving more packet drops which we have
seen in our simulations.

vehicles. Third, we study approaches that do not need a


navigational system but where local maps, particularly for
junctions, are inferred from observing transmitted packets
and vehicle movement patterns.

Avg. Delivery Rate


1
0.9

packet delivery rate

0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

GSR 500
GSR 250
0
500
1000

1500

2000
2500
distance [m]

3000

3500

4000

Figure 11: Packet delivery ratio of GSR for case 250 m and 500 m
transmission range.
Avg. Delivery Rate
1
0.9

packet delivery rate

0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
500

DSR 500 - EMU


DSR 500
1000

1500

2000
2500
distance [m]

3000

3500

4000

Figure 12: Comparison of DSR with 802.11 and with an idealized MAC scheme with (almost) unlimited bandwidth
available.

In general, the bumpiness of the curves have their reason


in the restricted scenario and is not due to an insufficient
number of packet transmission.

6 Conclusions and Future Work


In this paper we proposed geographic source routing
(GSR), which combines position-based routing with topological knowledge, as a promising routing strategy for vehicular ad hoc networks in city environments. We demonstrated by a simulations study that made use of realistic vehicular traffic in a city environment that GSR outperforms
topology-based approaches like DSR and AODV with respect to delivery rate and latency.
We are currently extending the work into the following three
directions. First, we are going to perform large scale simulations in order to get results independently from the structure of a given scenario. Second, we work on strategies to
avoid running into streets with no or not sufficiently many

References
[1]
M. Bando, K. Hasebe, A. Nakayama, A. Shibata, and
Y. Sugiyama. Dynamical model of traffic congestion and numerical simulation. Physical Review, E 51:10351042, 1995.
[2]
L. Blazevic, S. Giordano, and J. Y. Le Boudec. Selforganizing wide-area routing. In In Proc. of SCI 2000/ISAS 2000,
Orlando, 2000.
[3]
P. Bose, P. Morin, I. Stojmenovic, and J. Urrutia. Routing
with guaranteed delivery in ad hoc wireless networks. In Proc. of
3rd ACM Intl. Workshop on Discrete Algorithms and Methods for
Mobile Computing and Communications DIAL M99, pages 4855,
1999.
[4]
The cmu monarch wireless and mobility extensions to ns-2.
http://www.monarch.cs.cmu.edu/cmu-ns.html.
[5]
H. Fler, M. Mauve, H. Hartenstein, M. Ksemann, and
D. Vollmer. Location-based routing for vehicular ad hoc networks.
student poster, ACM Mobicom 02, Atlanta, Georgia, USA, 2002.
[6]
H. Hartenstein, B. Bochow, A. Ebner, M. Lott,
M. Radimirisch, and D. Vollmer. Position-aware ad hoc wireless
networks for inter-vehicle communications: the FleetNet project.
In Proceedings of the second ACM international symposium on
Mobile and ad hoc networking and computing (MobiHoc 01),
2001.
[7]
X. Jiang and T. Camp. Review of geocasting protocols for
a mobile ad hoc network. In Proceedings of the Grace Hopper
Celebration (GHC), 2002.
[8]
B. Karp. Challenges in Geographic Routing: Sparse Networks, Obstacles, and Traffic Provisioning. Talk at DIMACS
Workshop on Pervasive Networking, May 2001.
[9]
B. Karp and H. T. Kung. GPSR: Greedy Perimeter Stateless Routing for Wireless Networks. In Proceedings of the sixth
annual ACM/IEEE International Conference on Mobile computing and networking (MobiCom 00), pages 243254, Boston, Massachusetts, August 2000.
[10] W. Kronjger and D. Hermann. Travel time estimation
on the base of microscopic traffic flow simulation. ITS World
Congress, 1999.
[11] M. Ksemann, H. Fler, H. Hartenstein, and M. Mauve. A
Reactive Location Service for Mobile Ad Hoc Networks. Technical Report TR-02-014, Department of Computer Science, University of Mannheim, Nov. 2002.
[12] M. Mauve, J. Widmer, and H. Hartenstein. A Survey on
Position-Based Routing in Mobile Ad-Hoc Networks. IEEE Network, 15(6):3039, November/December 2001.
[13] The ns-2 network simulator. http://www.isi.edu/nsnam/ns/.
[14] C. Schwingenschlgl and T. Kosch. Geocast enhancements
of AODV for vehicular networks. ACM SIGMOBILE Mobile Computing and Communications Review, 6(3):9697, 2002.
[15] J. Tian, I. Stepanov, and K. Rothermel. Spatial aware geographic forwarding for mobile ad hoc networks. Technical Report
TR2002/01, University of Stuttgart, 2002.
[16] F. Zurheide. Dynamische Verkehrsumlegung in einer
mikroskopischen Simulation. Masters thesis, Fachhochschule
Braunschweig/Wolfenbttel, 1999.