Geographic Routing with Partial Position Information
Tony Ducrocq
1
, Micha
¨
el Hauspie
2
and Nathalie Mitton
1
1
Inria Lille - Nord Europe, 59650 Villeneuve d’Ascq, France
2
Universit
´
e Lille1, Cit
´
e Scientifique, 59100 Lille, France
Keywords:
Wireless Sensor Networks, Geographic Routing Algorithm, Position Based Routing.
Abstract:
Geographic routing protocols show good properties for Wireless Sensor Networks (WSN). They are stateless,
local and scalable. However they require that each node of the network is aware of its own position. While it
may be possible to equip each node with GPS receiver, even if it is costly, there are some issues and receiving
a usable GPS signal may be difficult in some situations. For these reasons, we propose a geographic routing
algorithm, called
HGA
, able to take advantages of position informations of nodes when available but also
able to continue the routing in a more traditional way if position information is not available. We show with
simulations that our algorithm offers an alternative solution to classical routing algorithm (non-geographic) and
offers better performances for network with a density above
25
and more than
5%
of nodes are aware of their
position.
1 INTRODUCTION
Wireless sensor networks consist of sets of mobile
wireless nodes without the support of any fixed infras-
tructure. Such wireless sensor networks offer great
application perspectives. Sensors are tiny devices with
hardware constraints (low memory storage, low com-
putational resources) that rely on battery. Sensor net-
works thus require energy-efficient algorithms to make
them work properly in a way that suits their hardware
features and application requirements.
Low power sensor nodes have limited transmission
power, thus they can communicate only to a limited
number of nodes. This set of nodes is called the neigh-
borhood of the node. In order to send messages at
longer range, nodes are using multi-hop communica-
tion. Multi-hop communication means that data will
need to be routed from source to destination by other
nodes. An efficient way to route messages is to use
nodes position information. A node uses a metric
based on its own position, its neighbors positions and
the destination position in order to choose the next hop
for the route. To use such a technique it is possible to
equip each node of the network with a GPS receiver or
to configure them at setup with their position if they
are static.
This work is partially supported by CPER Nord-Pas-
de-Calais/FEDER Campus Intelligence Ambiante and ANR
ECOTECH BinThatThinks.
On the other hand there exist non geographic rout-
ing algorithms. They do not require node position to
route data but only neighborhood knowledge.
In the Smart Cities context, it is likely that the net-
work is heterogeneous (Al-Hader et al., 2009). Some
nodes may be aware of their position while some other
are not. The reason to not equip nodes with GPS re-
ceiver may be to save money or energy. It is also not
possible to setup position of mobile nodes as it will
change over time. Furthermore, even if all nodes are
equipped with a GPS receiver, it is possible that some
of them do not receive the signal because of environ-
mental factors. For instance, two districts of a city may
be connected through a tunnel, in which GPS signal is
not received.
In such contexts, the availability of a routing al-
gorithm taking advantages of nodes positions when
possible but also functioning when the position is
not available is interesting. Thus, we introduce the
HGA
algorithm (
H
ybrid-
G
reedy-
A
ODV). The algo-
rithm works like the Greedy geographic routing while
position information allows it, and if position informa-
tion is not available, a route request message is sent in
the node neighborhood in order to find a path toward
the destination.
The rest of the paper is organized as follows: Sec-
tion 2 presents related work concerning geographic
and pseudo geographic routing algorithms. We intro-
duce some background works on which our algorithm
165
Ducrocq T., Hauspie M. and Mitton N..
Geographic Routing with Partial Position Information.
DOI: 10.5220/0004872901650172
In Proceedings of the 3rd International Conference on Sensor Networks (SENSORNETS-2014), pages 165-172
ISBN: 978-989-758-001-7
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
relies in Section 3. Section 4 describes the
HGA
algo-
rithm operation in details and Section 5 present sim-
ulation and results of its performances. Finally we
conclude in Section 6.
2 RELATED WORKS
To the best of our knowledge, there is no existing rout-
ing solution that considers that only a subset of nodes
is aware of its position. Classical routing algorithms
and those using virtual coordinates are the closest ap-
proaches to the one we propose. In this section we
describe related works as well as works on which our
proposal relies.
The main idea of geographic routing protocols is to
route data closer to the destination at each step of the
routing. The Greedy algorithm, on which part of our
work relies, chooses the closest node to the destination
in a node’s neighborhood as the next forwarder of the
data packet. The Greedy algorithm and some variants
from literature are shown on Figure 1.
S
D
a
b
c
e
Figure 1: Comparison of geographic algorithms. Node
a
is chosen by MFR algorithm, node
b
by Greedy, node
c
by
NFP and node e by Compass.
Some algorithms allow the deduction of position
for nodes by using their neighbors (
ˇ
Capkun et al.,
2002), (Ermel et al., 2005). Nevertheless, these so-
lutions exhibit poor performances due to the difficulty
to determine distance between two nodes without spe-
cialized hardware (Benkic et al., 2008). Moreover,
Seada et al. (Seada et al., 2004) showed that a small
error on position can generate several losses of packets
in geographic routing algorithms.
Routing algorithms over virtual coordinates allow
the benefits of geographic routing algorithms without
the need of real geographic position of nodes. The
main drawback of these solutions resides in the setup
of the virtual coordinates system. Indeed, it is needed
to flood all the network several times at the beginning
of the network lifetime.
The VCap solution (Caruso et al., 2005) relies on
virtual coordinates and a greedy routing. Some partic-
ular nodes are chosen for their interesting properties
to act as starting points for the different coordinates
of the system. These nodes are named anchors. The
coordinate system relies on the distance in number
of hops between a node and the different anchors. A
predefined sink node is used to initiate anchors elec-
tion (or any other node). This node, by broadcasting
a message in the whole network allows to elect a first
anchor. At its turn, this anchor broadcast a message
in the whole network to elect a second one and so on.
Anchors are chosen to be close to the network bound-
aries a far away to each other to limit duplicates in
coordinates. In VCap the network is flooded at least
5
times. Otherwise, because uniqueness of coordinates
is not guaranteed, the delivery can not be guaranteed.
The cost over progress concept idea (Stojmenovic,
2006) is to optimize the ratio between the cost of a
transmission by the progress made by the transmission.
Vcost (Elhafsi et al., 2007) relies on this idea and on
VCap coordinate system to reduce energy needs for
the routing.
The LTP algorithm (Ch
´
avez et al., 2007) uses a
root and a tree construction as the only coordinate
system instead of anchors. Thanks to the tree, this
routing algorithm guarantees delivery but the chosen
routes may be long. Above all, the tree construction is
costly and hard to maintain.
HECTOR (Mitton et al., 2012) combines both co-
ordinates systems of LTP and VCap. Its chooses the
best node in the VCap coordinate system at the same
level or that allow a progress in LTP coordinate system.
HECTOR is then able to guarantee delivery of packets
and to reduce the energy used for the routing.
Other virtual coordinates solutions have been pro-
posed, some are without guaranteed delivery (Ben-
badis et al., 2006), (Fang et al., 2005), (Niculescu
and Nath, 2001), some other with guaranteed deliv-
ery (Ch
´
avez et al., 2007), (Liu and Abu-Ghazaleh,
2008). However, algorithms using virtual coordinates
have the drawback to not support node mobility. Fur-
thermore, it is needed to flood the whole network at the
beginning, introducing high latency at startup when
the network is big.
3 BACKGROUND
Our proposal relies on route requests and route replies
similar to those used in AODV (Perkins and Royer,
1999). We introduce their behavior in the following.
Figure 2 shows a route search in AODV.
AODV is a non-geographic routing, totally reactive.
SENSORNETS2014-InternationalConferenceonSensorNetworks
166
S
D
(a) Reverse path setup.
S
D
(b) Forward path setup.
Figure 2: Route search example from S to D in AODV.
Indeed, routes are established only on-demand, when
they are needed. To route a data packet toward the
destination, a node broadcasts a RREQ (route request)
in its neighborhood. Each node receiving the RREQ
keeps track of it until expiration delay expires (defined
to
3000 ms
in AODV). Keeping track of RREQ allows
the creation of the reverse path as shown in Figure 2(a)
that will be used when destination is found. If a node
S does not know a route towards the destination D,
S is not the destination and it has not received this
RREQ earlier, it broadcasts a RREQ at its turn in its
neighborhood. When a node knows a route toward
the destination, or is itself the destination, it sends a
RREP packet (route reply) to the node that has sent
the RREQ. The RREP packet then follows the reverse
path and each node on the route saves the source of
the RREP in its neighborhood table in order to create
the reverse path. Figure 2(b) shows the creation of the
forward path and expiration of reverse paths.
4 PROPOSAL
In this paper, we propose
HGA
, a
H
ybrid
G
reedy-
A
ODV algorithm. The main idea of
HGA
is to apply
a greedy geographic routing whenever it is possible.
When a greedy routing is not possible, a RREQ is sent
in the
k
-neighborhood of the node. When a node needs
to route a data packet and none of its neighbors allows
a progress toward the destination, a route is searched
in AODV fashion (Perkins and Royer, 1999).
We first describe the notations and the model used
and then the algorithm details.
4.1 Notations and Model
The
k
parameter is the maximum depth for searching
a route and defines the variant of HGA : HGA-
k
. The
set of nodes aware of their position is
P
, then if
u
knows its position
u P
.
R
defines the routing table
of a node, then
R(u)
is the routing table of node
u
.
RREQ T IMEOU T
is the delay until a RREQ is ig-
nored if no RREP has been received. The Euclidean
distance between nodes
u
and
v
is noted
|uv|
.
N(u)
is the set of neighbors in communication range of
u
.
RREQs in the “waiting” state belong to the set W .
4.2 Algorithm Description
The header of a data packet contains the destination
position and the last known position where the packet
passed by.
To route a data packet, a node
u
starts by verifying
if one of its neighbors allows a progress toward the
destination. If
u
knows its own position, the progress
depends of the distance between
u
and the destina-
tion. If
u
is not aware of its position, the progress
depends on the last known position contained in the
data packet if available, otherwise, any node aware of
its position will be considered as making a progress.
If no neighbor of
u
allows a progress,
R(u)
is sought
to find a route allowing a progress in an AODV fash-
ion. If there is no such a route, a RREQ is sent
N(u)
and a new research in routing table is made after the
delay
RREQ T IMEOU T
. Algorithm 1 sums up route
search on a node u.
When u receives a RREQ, it can either transmit it,
answer with a RREP or discard it. Node
u
starts by
analyzing
R(u)
to verify, by comparing the sequence
number and the source node of RREQ to those in the
routing table entries, whether
u
has already received
this RREQ. If so, the RREQ is ignored and nothing
happens, otherwise
u
search in
N(u)
if a node is closer
to the destination than the source of the RREQ. If a
route is found among neighbors of
u
, a RREP is sent
to the nodes that transmitted the RREQ to
u
. The
RREP contains the source id and sequence number
from the RREQ, it also contains position of the closest
neighbor of
u
to destination. If no route is found, the
RREQ is forwarded if the maximum depth
k
for RREQ
(
MAX HOP RREQ
) has not been reached. When a
RREQ is transmitted, a new entry in the routing table
is created with a ‘waiting” flag allowing to know that a
RREQ has been sent but no route is known yet. Algo-
rithm 2 describes operations made at RREQ reception.
When a RREQ is received by a node
u
, the corre-
sponding entry with “waiting” flag in the routing table
is updated. The flag is removed and the neighbor allow-
ing the destination to be reached is written is the table
(this information is known thanks to the RREQ). Like-
wise, position of the destination is updated, indeed, the
RREP may point to a position allowing to approach
the destination but not necessarily the destination itself.
By saving position of the node contained in the RREQ,
routing table informations are more accurate.
GeographicRoutingwithPartialPositionInformation
167
Algorithm 1:
Data packet reception at node
u
towards
D
with
l
being last node with known
position.
1 if u = D then
2 exit /* success */ ;
3 next 1 ;
4 if u P then
5 dist |uD| ;
6 l u ;
7 else if l 6= 1 then
8 dist |lD| ;
9 else
10 dist +;
11 for v N(u) do
12 if |vD| < dist then
13 next v ;
14 dist |vD| ;
15 if next < 0 then
16 /* no route found in direct neighborhood */
17 broadcast(RREQ) ;
18 R R RREQ ;
19 W W RREQ ;
20 wait(RREQ TIMEOUT) ;
21 if ¬(r,r R |rD| < dist) r / W then
22 R R \ RREQ ;
23 /* no route found */
24 drop(data) ;
25 exit /* fail */ ;
26 else
27 for r R do
28 /* For each known route */
29 if |rD| < dist then
30 dist |rD| ;
31 next r ;
32 send(data, next) /* data transmitted */ ;
HGA does not guarantee delivery, indeed if no
node in the
k
-neighborhood of a node is aware of its
position, the node will not be able to route packets.
However, the route search mechanism allows to cir-
cumvent dead ends by searching new routes in the
k
-hops limit. Figure 3 shows an example of routing
from node
S
, which is not aware of its position, to
D
,
aware of its position. Figure 3(a) highlights a route
request with a maximum depth of
3
. Nodes receiving
the RREQ and knowing their position answer with a
RREP and
S
chooses the one which allow the max-
imum progression to destination. Once the packet
has been routed to the intermediate destination, a new
RREQ is initiated, the packet continues its progression
then it is routed using classical geographic routing
technique as intermediate destination knows a neigh-
bor allowing a progression through destination 3(b).
Finally, on Figure 3(c) a last RREQ is performed to
Algorithm 2:
Reception of a RREQ initiated by
S
with sequence number
seqnum
on node
u
for
destination D.
1 if (S,seqnum) R then
2 exit /* RREQ already sent */;
3 next
/
0 ;
4 dist |SD| ;
5 for v N(u) do
6 if |vD| < dist then
7 next v ;
8 dist |vD| ;
9 if next =
/
0 then
10 hop hop 1 ;
11 R R RREQ ;
12 if hop > 0 then
13 broadcast(RREQ) ;
14 else
15 send(RREP, u,next) ;
S
D
Reverse path
Forward path
Geographic routing
Node with position
Node without position
(a) First RREQ.
S
D
(b) Second RREQ and geograhic routing.
S
D
(c) Third RREQ.
Figure 3: Routing example in HGA.
allow the destination to be reached.
5 SIMULATION AND RESULTS
We choose to compare HGA to VCap, a routing algo-
rithm using a similar greedy routing technique but
SENSORNETS2014-InternationalConferenceonSensorNetworks
168
based on virtual coordinates. We also simulate a
greedy routing algorithm, slightly adapted to operate
with some nodes without position awareness. In this
variant, nodes without position knowledge do not send
HELLO messages, then they do not participate in rout-
ing but can emit data packets. When a node needs to
send a data packet (because it is the source or it needs
to transfer it), it chooses in its neighbors with position
the closest to the destination allowing a progress. If
the node sending the data packet is not aware of its
position, the notion a progress does not exist, then
the closest position aware neighbor to destination is
chosen.
We simulated HGA with route search depth of
1
,
2
,
3
and
5
hops. VCap has been simulated with
3
and
5 anchors.
Simulations have been performed using WS-
NET (Fraboulet et al., 2007) simulator. Each sim-
ulation lasts for
10
minutes. We generated random
connected topologies of
500
,
750
,
1000
,
1500
,
2000
,
2500
and
3000
nodes in a field of
500 m × 500 m
to
achieve average density of
10
,
15
,
20
,
30
,
40
,
50
et
60
nodes per communication range. Nodes have a com-
munication range of
40 m
. Each combination (algo-
rithm, average density) is launch
50
times on different
topologies. The same topology is used for a combina-
tion of average density and iteration number for each
different algorithm. For instance the first simulation
of HGA-1 uses the same topology as the simulation of
VCap-3, Greedy, etc.
To simulate a data traffic, every
10
seconds, each
node chooses randomly a destination node aware of
its position. The source node sends the data packet in
respect to the algorithm it executes. Nodes are desyn-
chronized in order to distribute emissions in time.
We measure delivery ratio, control message num-
ber and memory used. The delivery ratio is the number
of messages received divided by the number of mes-
sages sent. Control messages are HELLO, RREQ,
RREP and messages to initiate the coordinate sys-
tem. Finally, the memory used is the average num-
ber of neighbors saved by a node plus the number of
known routes for each node. Simulation parameters
are summed up in Table 1.
For the sake of clarity, in the following we refer to
HGA with a depth search of
1
hop by HGA-
1
(
k = 1
).
Moreover, HGA-
2
(resp. HGA-
3
and HGA-
5
) refer to
HGA algorithm with depth search of
2
(resp
3
and
5
)
hops. Finally VCap-
3
and VCap-
5
refer to the VCap
algorithm with
3
and
5
anchors. VCap results curves
have been duplicated on figures with
1%
,
5%
and
10%
located nodes in order to ease readings.
Table 1: Simulation parameters.
Parameter Value
Duration (m) 10
MAC layer idealmac
Interferences none
Data size (bytes) 10
Header size (bytes) 88
Field size 500 m× 500 m
Communication range 40 m
Number of nodes 500, 750, 1000, 1500, 2000, 2500, 3000
Density 10, 15, 20, 30, 40, 50, 60
Iterations 50
5.1 Delivery Ratio
Figure 4 shows delivery ratio for the four variants
of HGA, the two variants of VCap and the Greedy
algorithm. We can see on Figure 4(a) that whatever
the average density of the network, with
1%
of nodes
aware of their position, HGA-
5
has a delivery ratio
higher that the two variants of VCap and Greedy. We
also observe than from an average density of
40
, HGA-
5
reaches almost
100%
delivery ratio. As we expected,
HGA-
1
offers lower performances than HGA-
2
, which
offers lower performances than HGA-
3
while HGA-
5
gets the best performances. As well, VCap is better
with
5
anchors than with
3
as expected since it reduces
the This hierarchy is also observed with
5%
and
10%
of nodes aware of their position as seen on Figures
4(b) and 4(c).
We can observe that HGA-
1
performs much better
than Greedy. HGA-
1
has a delivery ratio
5
times higher
for a density of
10
and almost
15
times higher for a
density of
40
although only
1%
of nodes are aware of
their position. It shows that with only a search depth
of
1
hop (
k = 1
), we can achieve really interesting
performances. We will see how it impact the cost in
messages in the next section.
One notices that HGA algorithms take a better
advantage for higher densities of nodes. Indeed, for
1%
of nodes aware of their position (Fig. 4(a)), the
delivery ratio increases for densities from
10
to
20
. It
is similar for all variants of HGA and VCap. Besides,
from a density of
20
, the increase is stronger for HGA
than for VCap. It allows HGA-
2
to outperform VCap-
3
from a density of
25
and HGA-
3
is over VCap-
5
from a density of 35.
With
5%
of nodes aware of their position
(Fig. 4(b)), performances of HGA are improved. This
behavior is logical as a RREQ is more likely to find a
node aware of its position and then a node in the direc-
tion of the destination. It is interesting to highlight that
the delivery ratio increase is higher that with
5%
of
GeographicRoutingwithPartialPositionInformation
169
10 20 30 40
50 60
0
0.2
0.4
0.6
0.8
1
average density
delivery ratio
(a) 1% position.
10 20 30 40
50 60
0
0.2
0.4
0.6
0.8
1
average density
delivery ratio
(b) 5% position.
10 20 30 40
50 60
0
0.2
0.4
0.6
0.8
1
average density
delivery ratio
HGA 1
HGA 2
HGA 3
HGA 5
Vcap 3
Vcap 5
Greedy
(c) 10% position.
Figure 4: Delivery ratio.
nodes aware of their position. It allows all variants of
HGA to outperform both VCap variant from a density
of
25
. We can observe that this behavior is even more
true with a position knowledge of
10%
(Fig. 4(c)).
These results show that HGA benefits more of a higher
density of nodes than VCap. Indeed with VCap, the
number of nodes with the same coordinates do not nec-
essary decrease with the increase of node density while
RREQs of HGA have a higher probability to reach a
node in the direction of the destination as explained.
5.2 Control Messages Cost
Figure 5 shows the average number of control mes-
sages sent by nodes. For
1%
of nodes aware of their po-
sition (Fig. 5(a)), we observe for HGA that, the deeper
is the search, the more there is control messages. Like-
wise, the higher is density, the more messages are sent.
It can be explained by the fact that more nodes receive
the RREQs when the density is higher, each node re-
ceiving a RREQ potentially re-emits it, explaining the
exponential increase of the number of messages emit-
ted with the average density. It also explains why the
increase is stronger when the route search if deeper.
Figures show that the number of control messages sent
by Greedy is very low because only nodes aware of
their position emit control messages. For Greedy and
VCap, the number of control messages sent is constant
as it does not depend on density. For a low density, the
control messages number is close for all algorithms.
The gap increases with the density. For low densities,
it is interesting to prefer HGA-
5
which offers high de-
livery ratio, compared to other solutions, while adding
a reasonable amount of control messages. When the
average density increases, VCap is more suitable as
it offers a delivery ratio close to, if not better than
some variants of HGA while generating less control
messages.
When the number of nodes aware of their posi-
tion increases to
5%
, the number of control messages
strongly decreases for HGA, especially with deep
search. This behavior is explained because the proba-
bility of finding a route during a RREQ increases when
the number of nodes aware of their position increases.
This probability also increases when the density of
nodes increases (there is more nodes in the neighbor-
hood to satisfy the request). This explains the inversion
of curves for HGA for
5%
and
10%
of nodes aware of
their position (Fig. 5(b) and 5(c)). However, there is
no need to use deep variants for
10%
of nodes aware
of their position as the delivery ratio rapidly reaches
100% in that case.
5.3 Memory Cost
Figure 6 shows memory usage for the different algo-
rithms. We can see that memory needs are much higher
for HGA than VCap when
1%
of nodes are aware of
their position (Fig. 6(a)). With
5%
(Fig 6(b)), the gaps
for memory usage are lower and even really close to
VCap when average density increase. On Figure 6(c),
with
10%
of nodes with position knowledge, memory
needs are much lower for HGA and even lower than
VCap when density is over 40.
The memory size for Greedy is only the number of
neighbors aware of their position for each node. The
high amount of memory needed by HGA is explained
because in simulations, every route is kept until the
end of the simulation. It would be possible to lower
the memory needs by “forgetting” old routes with the
counterpart of having to send more control messages.
We can highlight that experimental conditions are de-
manding concerning memory needs. Indeed, sources
and destinations of data packet are chosen randomly
at a high frequency. In real applications, the number
of destinations of data packets are often limited if not
unique in case of data collection. In case of actuators,
one base station sends messages to different sources.
In that case the source is unique, thus the number of
routes to memorize for each node is reduced.
SENSORNETS2014-InternationalConferenceonSensorNetworks
170
10 20 30 40
50 60
0
2
4
6
·10
4
average density
control messages number per node
(a) 1% position aware nodes.
10 20 30 40
50 60
0
2,000
4,000
6,000
8,000
average density
control messages number per node
(b) 5% position aware nodes.
10 20 30 40
50 60
0
500
1,000
1,500
average density
control messages number per node
HGA 1
HGA 2
HGA 3
HGA 5
Vcap 3
Vcap 5
Greedy
(c) 10% position aware nodes.
Figure 5: Message control number.
10 20 30 40
50 60
0
500
1,000
1,500
average density
memory size
(a) 1% position.
10 20 30 40
50 60
0
100
200
300
average density
memory size
(b) 5% position.
10 20 30 40
50 60
0
20
40
60
80
100
120
average density
memory size
HGA 1
HGA 2
HGA 3
HGA 5
Vcap 3
Vcap 5
Greedy
(c) 10% position.
Figure 6: Average size in memory.
6 CONCLUSIONS
In this article we propose a novel solution for geo-
graphic routing, that allow to use nodes position when
available but also able to deal with unavailable po-
sitions. To the best of our knowledge, there is no
solution in the literature which relies on the same as-
sumptions. Indeed, we can find geographic routing
solutions for networks with all nodes aware of their
positions as well as solutions using virtual coordinates
where no node is aware of its position. We propose an
hybrid solution, taking advantages of real node posi-
tion when available but able to operate if not always
available. The geographic and the reactive methods
illustrated in this paper could be replaced by some
other methods from literature or new ones. Both part
should be chosen regarding application needs. The
underlying topology of the application will favor some
geographic routing techniques while the traffic pattern
will favor the classical routing part. For low data rates,
the classical routing part could be proactive or even
hybrid as proposed in ZRP (Haas et al., 2002).
Our solution,
HGA
offers much better perfor-
mances than Greedy geographic routing if part of the
nodes are not aware of their position. We also compare
HGA
with a routing solution based on virtual coordi-
nates, VCap, and show that for different scenarios, our
solution offer better performances with a limited or no
overhead.
For future works, we consider to validate our works
with realistic physical and MAC layers. The use of
real platform such as FIT or SmartSantander (Sanchez
et al., 2011) would allow us to validate these aspects
while testing realistic environments.
Studying node mobility is another interesting as-
pect, we could characterize the needed timeout before
routes expires with regards to nodes speed. Compari-
son with routing algorithm using virtual coordinates
would also bring some arguments for
HGA
since these
algorithms need to frequently rebuild coordinates sys-
tem, and then, increase number of control messages
exchanged in the network.
Finally it could be interesting to compare our solu-
tion with a variant in which each node keeps a
k
-hop
neighborhood table instead of using
k
-hop RREQ. The
study of control messages number would allow us to
find threshold to know which variant is better in some
cases.
REFERENCES
Al-Hader, M., Rodzi, A., Sharif, A., and Ahmad, N. (2009).
Smart City Components Architecture. In Proceedings
GeographicRoutingwithPartialPositionInformation
171
of the International Conference on Computational In-
telligence, Modelling and Simulation, CSSim, pages
93–97, Brno, Czech Republic. IEEE.
ˇ
Capkun, S., Hamdi, M., and Hubaux, J.-P. (2002). GPS-
free Positioning in Mobile Ad Hoc Networks. Cluster
Computing, 5(2):157–167.
Benbadis, F., Puig, J.-J., de Amorim, M. D., Chaudet, C.,
Friedman, T., and Simplot-Ryl, D. (2006). Jumps: En-
hancing hop-count positioning in sensor networks us-
ing multiple coordinates. International Journal on Ad
Hoc and Sensor Wireless Networks, abs/cs/0604105.
Benkic, K., Malajner, M., Planinsic, P., and Cucej, Z. (2008).
Using RSSI value for distance estimation in wireless
sensor networks based on ZigBee. In Proceedings
of the 15th International Conference on Systems, Sig-
nals and Image Processing, IWSSIP, pages 303–306,
Bratislava, Slovakia.
Caruso, A., Chessa, S., De, S., and Urpi, A. (2005). GPS
free coordinate assignment and routing in wireless sen-
sor networks. In Proceedings of the 24th Annual Joint
Conference of the IEEE Computer and Communica-
tions Societies, INFOCOM, pages 150–160, Miami,
FL, USA. IEEE.
Ch
´
avez, E., Mitton, N., and Tejeda, H. (2007). Routing in
Wireless Networks with Position Trees. In Kranakis,
E. and Opatrny, J., editors, Ad-Hoc, Mobile, and Wire-
less Networks, ADHOC-NOW, pages 32–45, Cancun,
Mexico. Springer Berlin Heidelberg.
Elhafsi, E. H., Mitton, N., and Simplot-Ryl, D. (2007). Cost
over Progress Based Energy Efficient Routing over
Virtual Coordinates in Wireless Sensor Networks. In
Proceedings of International Symposium on a World
of Wireless, Mobile and Multimedia Networks, WoW-
MoM, pages 1–6, Helsinki, Finland. IEEE.
Ermel, E., Fladenmuller, A., Pujolle, G., and Cotton, A.
(2005). On Selecting Nodes to Improve Estimated
Positions. In Proceedings of the 7th international Mo-
bile and Wireless Communication Networks, MWCN,
pages 449–460, Marrakech, Morocco. Springer US.
Fang, Q., Gao, J., Guibas, L., de Silva, V., and Zhang, L.
(2005). GLIDER: gradient landmark-based distributed
routing for sensor networks. In Proceedings of the
24th Annual Joint Conference of the IEEE Computer
and Communications Societies, INFOCOM, pages 339–
350, Miami, FL, USA. IEEE.
Fraboulet, A., Chelius, G., and Fleury, E. (2007). Worldsens:
Development and Prototyping Tools for Application
Specific Wireless Sensors Networks. In Proceedings
of the 6th International Symposium on Information
Processing in Sensor Networks, IPSN, pages 176–185,
Cambridge, MA, USA. ACM.
Haas, Z. J., Pearlman, M. R., and Samar, P. (2002). The
Zone Routing Protocol (ZRP) for Ad Hoc Networks.
IETF Internet Draft.
Liu, K. and Abu-Ghazaleh, N. (2008). Stateless and guaran-
teed geometric routing on virtual coordinate systems.
In Proceedings of the 5th International Conference
on Mobile Ad Hoc and Sensor Systems, MASS, pages
340–346, Atlanta, GA, USA. IEEE.
Mitton, N., Razafindralambo, T., Simplot-Ryl, D., and Stoj-
menovic, I. (2012). Towards a hybrid energy efficient
multi-tree-based optimized routing protocol for wire-
less networks. Sensors, 12(12):17295–17319.
Niculescu, D. and Nath, B. (2001). Ad hoc positioning sys-
tem (APS). In Proceedings of the Global Telecommuni-
cations Conference, GLOBECOM, pages 2926–2931,
San Antonio, TX, USA. IEEE.
Perkins, C. and Royer, E. (1999). Ad-hoc on-demand
distance vector routing. In Proceedings of the 2nd
Workshop on Mobile Computing Systems and Appli-
cations, WMCSA, pages 90–100, New Orleans, LA,
USA. IEEE.
Sanchez, L., Galache, J., Gutierrez, V., Hernandez, J., Bernat,
J., Gluhak, A., and Garcia, T. (2011). Smartsantander:
The meeting point between future internet research and
experimentation and the smart cities. In Future Net-
work Mobile Summit, FutureNetw, pages 1–8, Warsaw,
Poland.
Seada, K., Helmy, A., and Govindan, R. (2004). On the
effect of localization errors on geographic face routing
in sensor networks. In Proceedings of the 3rd interna-
tional symposium on Information processing in sensor
networks, IPSN, pages 71–80, New York, NY, USA.
ACM.
Stojmenovic, I. (2006). Localized network layer protocols
in wireless sensor networks based on optimizing cost
over progress ratio. IEEE Network, 20(1):21–27.
SENSORNETS2014-InternationalConferenceonSensorNetworks
172