You are on page 1of 8

Performance Model of Flooding in OLSR

+
Andres Medina
Dept. of Electrical and Computer Eng.
University of Delaware
medina@ece.udel.edu
Stephan Bohacek
Dept. of Electrical and Computer Eng.
University of Delaware
bohacek@ece.udel.edu
ABSTRACT
OLSR is one of the most well developed routing protocols
for MANETs. OLSR as a link state routing protocol relies
on its ooding mechanism to disseminate topology informa-
tion to all nodes in the network. However, the performance
of the ooding algorithm is poorly understood. In this pa-
per models of key performance metrics of this algorithm are
presented. Specically, it describes a model for the control
overhead generated by the ooding algorithm and a model
of eciency in terms of network coverage. The models are
validated against simulations. The results presented in this
paper have signicant implications for the standardization
process.
Categories and Subject Descriptors
C.2.2 [Computer-Communication Networks]: Network Pro-
tocolsRouting protocols
General Terms
algorithms, design, performance
Keywords
MANETs, routing, ooding, OLSR, model, performance
+
The research reported in this document/presentation was
performed in connection with contract DAAD19-01-C-0062
with the U.S. Army Research Laboratory. The views and
conclusions contained in this document/presentation are
those of the authors and should not be interpreted as pre-
senting the ocial policies or position, either expressed or
implied, of the U.S. Army Research Laboratory of the U.S.
Government unless so designated by other authorized docu-
ments.
Citation of manufacturers or trade names does not consti-
tute an ocial endorsement or approval of the use thereof.
The U.S. Government is authorized to reproduce and dis-
tribute reprints for Government purposes notwithstanding
any copyright notation hereon.
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for prot or commercial advantage and that copies
bear this notice and the full citation on the rst page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specic
permission and/or a fee.
PE-WASUN10, October 1718, 2010, Bodrum, Turkey.
Copyright 2010 ACM 978-1-4503-0276-0/10/10 ...$5.00.
1. INTRODUCTION
The Optimized Link State Routing protocol (OLSR) [1]
for mobile ad hoc networks has been widely implemented
and studied. It has been chosen by the IETF MANET
Charter as the proactive routing protocol that will continue
the standardization process. There are implementations of
OLSR for almost all platforms including Windows, MAC,
Linux, Windows Mobile, Android, etc. A variation of OLSR
has also been chosen by the 802.11s IEEE TG [2] as the
proactive protocol of choice. However, its performance is
not completely understood.
OLSR stands for optimized link state routing protocol.
The optimization refers to three methods present in the
protocol to decrease the control overhead generated by the
topology dissemination mechanism. All three methods rely
on a subset of neighbors selected by each node known as the
Multipoint Relay (MPR) set of a node, which are chosen fol-
lowing the algorithm in [1]. Although dierent algorithms
may be used, the MPR subset must ensure that all nodes
reachable in two hops are still reachable through MPRs.
The rst optimization method shortens the length of topol-
ogy control packets. Contrary to other link state protocols
like OSPF [3], where nodes advertise all of their neighbors
to all nodes in the network, nodes in OLSR advertise only
the subset of neighbors that have selected them as MPRs.
The eect of this optimization is that all nodes will maintain
only partial topology information. However, if the MPR set
is up-to-date, then this partial topology information is su-
cient to determine the shortest paths. The impact of partial
topology information in realistic scenario where the MPR
sets are not up-to-date is unknown. However, this impact is
not the focus of this paper.
The second optimization supported by OLSR reduces con-
trol overhead by limiting the number of nodes that generate
topology control packets. Only nodes that are MPRs of an-
other node generate such packets. An exception to this rule
are nodes that were recently selected as MPRs, but are no
longer MPRs. These nodes continue to send empty topol-
ogy control packets for the validity time of the last topology
information packet they sent as MPRs. Thus, in mobile sce-
narios and lossy scenarios where the MPR set changes fre-
quently, a signicant fraction of nodes either are or recently
were MPRs, and the impact of the optimization method is
limited. In this paper, a model for the fraction of nodes that
generate a Topology Control (TC) packet is presented and
validated. The model accounts for dierent node densities,
node mobility, neighbor discovery methods, and dierent
physical radios.
The third method reduces control overhead by ooding
topology control packets only over MPR links. In [1] it is
stated that only if a node receives a packet for the rst
time from an MPR selector it will forward it. However,
some popular implementations of OLSR like OLSRD [4] dif-
fer from the standard. In [4], nodes forward the packet if it
has received the message from an MPR selector, regardless
of whether it had previously received the message from a
non-MPR selector. In this paper, we present a model of the
fraction of nodes that retransmit a packet for both ooding
methods. The paper also describes a model of the fraction
of nodes covered by a ood. This metric shows how ecient
is the ooding method to reach all nodes in the network.
It is shown in the paper that there is a tradeo between
the amount of overhead generated and the level of coverage
achieved by a ooding method.
Several eorts have been made to evaluate the perfor-
mance of OLSR. In [5], a model for control overhead is
presented. However only orders of magnitude are given to
estimate the overhead and no validation of the model is in-
cluded. In [6] authors propose a simple model to estimate
the amount of control overhead used by OLSR and compare
it to that of reactive protocols such as DSR. In [7] provides
a very simple model of control overhead. In [8] MPR ood-
ing is studied. They provide results of simple simulations
of the fraction of nodes that retransmit a packet and the
fraction of nodes covered by a ood. In [9] authors propose
a model of control overhead in OLSR as a function of node
mobility, whoever this impact is restricted to the optional
feature of OLSR to react to topology changes and also dis-
regards neighbor discovery dynamics. All of these models
neglect the dynamics of the neighbor discovery mechanism
and thus disregards the impact of incorrect local topology
information has on ooding, which requires up-to-date lo-
cal topology information. One complicating factor is that,
as mentioned, several used implementations of OLSR (e.g.,
OLSRD and QualNet) include signicant bugs. Thus, it is
sometimes unclear which version of OLSR is being modeled.
Because of the accuracy of the models presented here, cross-
validation detected the bugs, which are described in [10].
The rest of this paper is organized as follows. In section
2, the neighbor discovery model is briey described with
references to more detailed descriptions. Section 3 describes
the base overhead model. The components of this model
are developed in subsequent sections. Section 4 describes the
model of the fraction of nodes that generate topology control
packet, and Section 5 the fraction of nodes that retransmit
topology control packets. Section 5 also describes a model
for the fraction of nodes covered in the propagation of a TC
message. Section 6 compares the model to other models.
Finally, Section 7 concludes the paper.
Because of page limitations, some parts of the model are
only briey described. However, complete details, including
Matlab functions and precomputed data are available online
[11].
2. NEIGHBOR DISCOVERY MODEL
Proactive routing protocols rely on a Neighbor Discovery
(ND) mechanism to estimate for which neighbors bidirec-
tional communication is possible. ND utilizes past observa-
tions from passive and active probing of neighbors. Thus,
when nodes move fast as compared to the reaction speed of
ND, the estimation of link states can become stale. This
degradation of the estimate of link state impacts the qual-
ity of information in link state announcements as well as
ooding algorithms that rely on local topology information.
OLSR employs an ecient algorithm for ooding topology
control (TC) messages. OLSRs algorithm utilizes link state
in two ways. First, a node will only accept TC messages that
have been transmitted over a link that ND has marked as a
symmetric link. Second, the symmetric links of a node and
the symmetric links of the neighbors, are used to select a
set of multipoint relays. As explained in more detail below,
nodes will only forward TC messages that were transmitted
by nodes that are MPR-selectors
1
. Hence, if NDs estimate
of the set of symmetric neighbors is incorrect, the ooding of
TC messages can be signicantly impacted. Modeling this
impact is a key contribution of this paper.
There are various methods to detect neighbor nodes. RFC
3626 [1] species two neighbor discovery method that are
described next. The model described here is detailed in [12]
and can be used to model a wide range of neighbor discovery
methods besides those dened in RFC 3626.
The rst method described in [12] is a generalization of
the method presented in RFC 3626 is referred to as Event
Counting (EC) ND. In this method, a link is declared to be
up if l Hello messages are received over the link. Once the
link is up, the link remains up until 1 consecutive Hellos are
lost. Nodes advertise which links are up in Hello messages.
By utilizing this information, a node can determine whether
a link is bidirectional. A bidirectional link is declared to be
symmetric. The algorithm described in RFC 3626 is equiv-
alent to a specic case of this method where l = 1 and
1 = 3.
The second method described in RFC 3626 uses an ex-
ponential moving average of a measure of link quality |q to
establish whether the link is up. We refer to this algorithm
as Exponential Moving Average (EMA) ND. In this algo-
rithm, a link is declared to be up when |q rst exceeds a
threshold 1Qup. Once the link is up, the link status will
change to down only when |q drops below 1Q
down
, where,
of course, 1Q
down
1Qup. If a link quality measure is not
directly available from the MAC layer, then RFC 3626 pro-
poses a method based on the reception of Hello messages
via
|q =
_
|q n + (1 n) , if received Hello
|q n, if Hello is missed
,
where n is the weight of the exponential moving average. In
RFC 3626, the values suggested are 1Qup = 0.8, 1Q
down
=
0.3, n = 0.5.
The dynamics of both ND algorithms can be represented
by Markov chains that depend on the link loss probability.
For example, in EC, the state of a link between node A
and nodes B is given by the tuple (&A, dA, cA, &B, dB, cB),
where &A is the number of consecutive Hellos received by
node A. Once &A = l, it remains xed until dA = 1, where
dA counts the number of Hellos missed. cA = 1 if node
A believes that the link is symmetric. &B, dB, and cB are
dened similarly, but for node 1. Note that node As and
node Bs estimate of whether the link is symmetric are not
synchronized. Hence, it is possible that cA ,= cB.
The probability of moving from state (I, 0, 0, ), 0, 0) to
state (I + 1, 0, 0, ), 0, 0) is the probability of a successful
transmission of a Hello over the link. By ordering all the
1
A nodes MPRs are a set of nodes such that if the node
and its MPRs broadcast a message, then all nodes within
the 2-hop neighborhood will receive the message. If a node
has been selected to be an MPR, then its MPR selectors is
the nodes that selected it to be an MPR. See RFC 3626 for
details [1].
C
B A
Figure 1: Trajectories of nodes 1 and C relative to . Node
1 was outside communication range (packet delivery prob.
below 1%) when last change in trajectory took place. How-
ever, this is not the case for node C.
possible values of (&A, dA, cA, &A, dA, cA), we can dene a
vector A where Ai is the probability of being in the ith state.
We can then determine the transmission probability matrix
(j), which depends on the transmission probability j. Let-
ting Ai (t) be the probability of being in state i at time t,
we have A (t +TH) = A (t) (j (t)), where TH is the Hello
period and j (t) is the transmission probability at time t.
Thus, given a trajectory of transmission probabilities, j (t),
one can compute the probability of the ND declaring the link
to be symmetric. [12] provides signicantly more details on
computing link state probabilities.
Note that the probability of the link state estimate de-
pends on the trajectory of the transmission probabilities.
While this trajectory depends on the full trajectory of the
nodes, in random waypoint mobility, it can be estimated
from the nodes locations and velocities. That is, given two
nodes position and velocity at time t, the distance between
the nodes can be determined for time t t0 to t, where
neither of the nodes changed direction or speed during the
interval (t to, t). Assuming that the nodes were well out
of communication range at time t t0, the probability of
the link state can be accurately computed. However, if the
nodes are within communication range at time t to, then
the estimate of the link state is less accurate (See Figure 1).
In the case of random waypoint mobility over a large region,
nodes change direction infrequent. However, over small re-
gions, the nodes frequently change directions, resulting in
some errors. However, for moderate sized regions, the es-
timate of model quality is good. See [12] for details. Note
that the performance of ad hoc network is of little interest
when the region is very small, e.g., when most nodes are
within one hop.
3. OVERHEAD MODEL
OLSR allows the option of generating TC messages when
the topology has changed. In this case, the rate that TC
messages are generated depends on the node speed. How-
ever, here we do not consider this option and assume that
TC messages are generated as a rate of `TC. TC messages
announce topology information, and thus, the size of these
messages depends on the number of detected neighbors.
With this notation, the overhead in packets generated by
each node is given by
O1
pkts
= `TC )GenTC )RET `, (1)
where ` is the number of nodes in the network, )GenTC is
the fraction of nodes that generate a TC packet, and )RET
is the fraction of nodes that forwards a packet. Models of
)GenTC and )RET are key contributions of this paper and
are given in Sections 4 and 5, respectively.
4. FRACTION OF NODES THAT GENER-
ATE TC MESSAGES
In OLSR, nodes that are MPRs of at least one neighbor
generate topology control packets. When a node receives
a Hello packet of its last MPR selector indicating it is no
longer an MPR, it will continue to send empty TC packets
to invalidate previously distributed topology information. It
continuous to generate empty TC messages for the validity
time of its last disseminated topology control packet, where
the validity time is a user dened parameter. Thus, the
fraction of nodes that generate topology control packets is
given by
)GenTC = )MPR +)EmptyTC,
where )MPR is the fraction of nodes in the network that are
MPRs and )EmptyTC is the fraction of nodes that generate
empty TC packets. In this section, models for these two
quantities are presented. First, an idealized model is de-
scribed. Then, a sub-model to account for imperfect neigh-
bor discovery mechanism and channel losses is added.
4.1 Idealized Model
The two quantities )MPR and )EmptyTC are modeled us-
ing an on/o process, where nodes goes from the state "se-
lected as MPR by at least one node (A11)" to the state
"not selected as MPR (`oA11)". Denote by 1NoMPR (t[c, , 1)
the CDF of the time a node remains in state `oA11 when
the average node speed is c, average number of neighbors is
, and the length of one side of the simulated region is 1
(i.e., the simulated region is 11). Let tNoMPR (TE, c, , 1)
and tMPR (TE, c, , 1) be the expected duration that a node
is in the states `oA11 and A11, respectively, where TE
is the duration that a node continues to generate TC mes-
sages when it is no longer an MPR. The expected time that
a node sends empty TC packets is given by
tEmptyTC (TE, c, , 1) = (1 1NoMPR (TE[c, , 1)) TE+
1NoMPR (TE[c, , 1)
_
T
E
0
1
1NoMPR (t[c, , 1)
1NoMPR (TE[c, , 1)
dt.
To see this, note that the rst term on the right accounts
for the duration of TE when a node is in state `oA11 for
longer than TE. The second term on the right accounts for
the duration in state `oA11 when the node is in the state
`oA11 for less than TE, where 1
F
NoMPR
(tjs;;L)
F
NoMPR
(T
E
js;;L)
is
the probability of being in state `oA11 for no more than
t, conditioned on the duration being no more than TE.
Thus,
)EmptyTC (TE;c, , 1)
=
tEmptyTC (TE, c, , 1)
tNoMPR (TE, c, , 1) +tMPR (c, , 1)
and
)MPR (TE, c, , 1)
=
tMPR (c, , 1)
tNoMPR (TE, c, , 1) +tMPR (c, , 1)
.
Note that by scaling time, we can eliminate the depen-
dence on speed c. Specically, if c = c c
ref
, then
)EmptyTC (TE;c, , 1) = )EmptyTC (TEc, c
ref
, , 1)
)MPR (TE;c, , 1) = )MPR (TEc, c
ref
, , 1) .
0.5 1 1.5 2
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
time [secs]
C
D
F
T
i
m
e
N
o
-
M
P
R =6, 4 x 4
=6, 7 x 7
=6, 10 x 10
=18, 4 x 4
=18, 7 x 7
=18, 10 x 10
(a)
3 4 5 6 7 8 9 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Sideof simulationsquare[tx. ranges]
E
[
T
i
m
e

A
s

M
P
R
]
=6
=9
=12
=15
=18
(b)
0 0.05 0.1 0.15 0.2
0.5
0.6
0.7
0.8
0.9
1
speed [tx.ranges/hello interval]
F
r
a
c
t
i o
n
N
o
d
e
s

G
e
n
e
r
a
t e
T
C
=6, 4 x 4
=6, 7 x 7
=6, 10 x 10
=18, 4 x 4
=18, 7 x 7
=18, 10 x 10
(c)
Figure 2: (a) CDF of time a node is not an MPR of some other node. (b) Expected time as MPR. (c) OLSR Optimization of reducing
the fraction of nodes that generate TC is degraded with speed.
Thus, )EmptyTC and )MPR are functions of three variables,
TE, , and 1.
To obtain )EmptyTC and )MPR, idealized simulations over
these three parameters was performed. By idealized simu-
lations, we mean where the radio is modeled as on/o with
transmission range v = 1. Nodes move at a speed c
ref
= 1
tr.vonjcc and follow the random waypoint mobility [13].
Moreover, for these simulations, a perfect neighbor discov-
ery mechanism is assumed. That is, once two nodes are
in range, nodes instantly detect each other as symmetric
neighbors and update their MPR sets accordingly. Figure
2(a) shows 1NoMPR for dierent values of , and 1. Fig-
ure 2(b) shows tMPR as a function of density. As can be
observed in both Figures 2(a) and 2(b), both 1NoMPR and
tMPR converge for large 1. Hence, 1NoMPR and tMPR do
not need to be determined for all 1, but only for small 1,
e.g., for 1 10.
Figure 2(c) shows )GenTC, the fraction of time that a node
generates a TC message, when TE is three TC intervals. It
can be seen that the OLSR optimization of reducing the
number of nodes that generate at TC packet is ineective
as speed increases. For example, as seen in Fig. 2(c), at 0.2
tr.vonjcIc||o intcvo| almost all nodes generate topology
packets. This indicates the importance of accounting for the
overhead generation by empty TC messages.
Data and a Matlab function for estimating )GenTC for
4 20 are available for download [11].
4.2 Realistic Model
The idealized scenario discussed above is unrealistic in
two ways; it neglects transmission errors, and it neglects
the impact of ND. When transmission errors are considered,
even when nodes are not moving, ND might determine that
links break. Thus, to estimate tEmptyTC in a more realistic
setting, we use a tEmptyTC from an idealized scenario. The
key question is, which idealized scenario should be used to
model the realistic scenario? We use the idealized scenario
where
1. the average number of symmetric neighbors is the same
as the realistic scenario,
2. 1 is such that the number of nodes in the network is
the same as the realistic scenario, and
3. the average link lifetime is the same as the realistic
scenario.
The model of the ND in [12] provides an estimate of the
average number of symmetric neighbors in the realistic sce-
nario, while the average number of symmetric neighbors in
the idealized scenario is simply the number of nodes within
communication range.
Note that when ND is such that only strong links are
considered symmetric, the number of symmetric neighbors
is reduced. Essentially, such a ND increases the size of the
network, in that a message will need to cross more sym-
metric links to reach a destination. Hence, 1, the length
of one side of the simulated region is impacted by ND and
the radio. In order to solve for 1, we need to determine
`
idealized
(1, ), the number of nodes as a function of 1
and when the on/o radio model is used. This function
is easily determined from simulation. Given this function,
we solve for 1 in ` = `
idealized
(1, sym), where sym is
the average number of symmetric neighbors found in step 1
and ` is the number of nodes in the network.
Matching the average link lifetime is done as follows. In
the realistic case, the link lifetime depends on the node
speed, on the ND method, and on the radio. Again, the
model of ND presented in [12] provides an estimate of the
link lifetime. In the idealized case, the link lifetime can be
determined from the simulations described in the previous
subsection. Note that these simulations need to be only per-
formed for one speed, denoted by c
ref
. The link formation
rate at some other speed, c, can be found by scaling, i.e.,
1T (c) =
s
ref
LT
ref
s
, where 1T
ref
is the link lifetime found
from simulations, and 1T (c) is the link lifetime at speed c.
In summary, using the model in [12] for a particular node
density, node speed, radio, and ND method, one can com-
pute 1TND be the average link lifetime and sym be the av-
erage number of symmetric neighbors. Then, the fraction of
nodes that generate TC messages is )GenTC (c, , 1), where
c = c
ref
LT
ref
LT
ND
, = sym, and 1 solves ` = `
idealized
(1, sym).
4.3 Model Validation
To validate the model of the fraction of nodes that gener-
ate TC messages, extensive QualNet [14] simulations were
performed. The values derived from QualNet simulations
throughout this paper were found by averaging over 120 sim-
ulation trials where the simulation time is 180 seconds, but
data is only saved from the last 60 seconds (the rst 120 sec-
onds allow the protocols to stabilize). Here, the nodes were
constrained to be within a 1125i1125i region. 802.11gs
54A/jc bit-rate was used. With this bit-rate, the packet
loss probability for 801tcc packets probability is 1% when
the transmitter and receiver are 196i apart, and is 99%
when they are spaced 237i apart. Thus, 1125i is approx-
imately four "transmission ranges." For some experiments,
background trac as included. To this end, nodes broad-
casted 801tcc packets at Poisson distributed times such
the average rate that a node generates background trac is
01jc, 12501jc, 25001jc or 50001jc. Fig. 3(a) show ex-
0 5 10 15 20 25
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Speed[m/s]
P
r o
b
.
G
e
n
e
r a
t e
T
C

p
k
t
N=57, datarate: 0
N=73, datarate: 0
N=73, datarate: 1250
N=73, datarate: 2500
N=91, datarate: 0
(a)
0.05 0.1 0.15
0
0.2
0.4
0.6
0.8
1
Relativ e Er r or
C
D
F
E
r
r
o
r

P
r
o
b
. G
e
n
e
r
a
t e
T
C
p
k
t
Model
Model ( No P. Speed)
(b)
Figure 3: (a)Validation of Model of the fraction of nodes that
generate TC packets. (solid) Model (dashed) Qualnet. (b) CDF
of model error when full model is used (solid) and real speed is
match to that of the idialized case (dashed).
amples of the quality of t provided by this model. Figure
3(b) shows the CDF of the relative error over all scenarios
considered. Each scenario is characterized by the number
of nodes, `, the node speed, c, the ND method and its pa-
rameters, and the intensity of the background trac. Here
we considered ` = 52, 66, 73, 84, or 91 and c = 0, 5, ...,
25 (meters/sec), and interference 01jc, 12501jc, 25001jc
and 50001jc. The two ND methods from RFC 3626 were
considered, namely the EC ND and the EMA ND. In the
case of EC, (l, 1) = (1, 3) and (l, 1) = (4, 3). For the
EMA method, 1Qup = 0.8, 1Q
down
= 0.3, n = 0.5.
For reference, Fig. 3(b) shows the CDF of the relative
error when step 3 is replaced as follows
3
0
. the speed c in the idealized case is the same as the
speed in the realistic case.
Note that step 3
0
is a reasonable approach. However, this
approach results in signicant error. On the other hand,
matching link lifetimes results in a much smaller error.
5. FRACTIONOF NODES THATRETRANS-
MIT A PACKET AND COVERAGE
In this section we present models of the fraction of nodes
that retransmit a packet, vct, and the fraction of nodes that
receive the ooded message, or coverage, cocv. Two model-
ing approaches are followed. First, a detailed rst-principles
simulation-based model is developed in the next subsection.
This approach relies on an analytic model of neighbor dis-
covery but makes use of simulations of node locations. These
simulations can be quickly performed in Matlab and allows
vct and cocv to be estimated for a wide range of topology
and protocol parameters. Based on these simulations, black-
box models are developed in Section 5.2. These black-box
models give analytic formulas for vct and cocv.
5.1 First Principles Simulation-based Model
A large number of methods have been proposed for ood-
ing messages in MANETs. RFC 3626 details one method
[1], while the popular implementation of OLSR available at
[4] utilizes a slightly dierent approach. According to RFC
3626, when a node receives a TC message from a symmet-
ric neighbor, two conditions are checked: 1. that the TC
message has never been received from a symmetric neighbor
before, 2. that the symmetric neighbor is also an MPR se-
lector. If both of these conditions are met, then the node
forwards the message. On the other hand, according to the
OLSRD implementation, a TC message is forwarded if the
message has not been forwarded before and the second con-
dition is met. Thus, according to RFC 3626, if the message
is rst received from a symmetric neighbor that is not an
MPR selector, then the message is never forwarded. On the
other hand, in the OLSRD method, a node may have re-
ceived the message rst from a symmetric neighbor, and yet
still forward the message when it is receives from a MPR se-
lector. Intuitively, the OLSRDs implementation generates
more overhead, but reaches more nodes than the method
described in RFC 3626. Since both ooding methods are
widely deployed, we develop models for both of them and
validate this intuition.
It turns out that simulations of ooding can be performed
very eciently. For example, simulating ooding over a 100
node network takes about 50 msec on a standard desktop
2
.
Hence, Monte Carlo evaluation of ooding is reasonable.
On the other hand, determining which nodes are symmetric
neighbors and hence which nodes are MPRs is more com-
plicated to simulate. However, the model of ND presented
in [12] can be employed to determine the probability that
nodes are symmetric neighbors. Thus, the following steps
can be used to compute the fraction of nodes that relay
the message, vct, and the fraction of nodes that receive the
message, cocv.
1. For each node i, select its location, (ri, i), and its
velocity, (i, ni), according to the random waypoint
stable node distribution [13].
2. Use the ND model to determine the probability that
nodes i and ) are symmetric neighbors. Specically,
let cii;j denote the event that node i has marked
node ) as a symmetric neighbor and let ci
i;j
be
the event that node i does not declare node ) as a
symmetric neighbor. The ND model estimates the
of probabilities 1 (cii;j, cij;i), 1
_
ci
i;j
, cij;i
_
,
1
_
cii;j, ci
j;i
_
, 1
_
ci
i;j
, ci
j;i
_
.
3. Select a realization of symmetric neighbors, i.e., based
on the probabilities computed in Step 2, select which
nodes have been declared symmetric neighbors.
4. Given the set of symmetric neighbors, compute the
MPRs.
5. Estimate which nodes believe they are MPR selectors
6. Flood a message
Several comments are in order. Regarding Step 5, a node
will inform a neighbor that it has been selected as an MPR
via Hello messages. However, these Hello messages are sub-
ject to transmission error. Consequently, a node might not
be aware that it has been selected as a MPR. Of course,
the next Hello might also inform the neighbor that it is a
MPR. Eventually, either the node becomes aware that it
has been selected as an MPR, or the neighbor stops being
a symmetric neighbor. Unfortunately, a complete model of
this behavior is complicated. Thus, we seek a simple model
that captures the essential behavior as follows. We make
the approximation that there are two cases.
Prefect MPR Knowledge: In this case, we assume that
nodes that have been selected as MPRs are aware that they
are MPRs.
2
Simulation times in this paper are obtained from programs
and scrpits running on an Intel Core i7 CPU running 64bit
OS and 12GB RAM.
0 5 10 15 20 25
0.65
0.7
0.75
0.8
0.85
0.9
0.95
Speed[m/s]
OLSRD
C
o
v
e
r a
g
e
o
f
F
l
o
o
d
i n
g
N=57, datarate: 0
N=73, datarate: 0
N=73, datarate: 1250
N=73, datarate: 2500
N=91, datarate: 0
(a)
0 5 10 15 20 25
0.4
0.5
0.6
0.7
0.8
0.9
Speed[m/s]
RFC
C
o
v
e
r
a
g
e
o
f

F
l
o
o
d
i n
g
N=57, datarate: 0
N=73, datarate: 0
N=73, datarate: 1250
N=73, datarate: 2500
N=91, datarate: 0
(b)
0.05 0.1 0.15
0.2
0.4
0.6
0.8
1
Relative Error
C
D
F

E
r
r
o
r
C
o
v
e
r
a
g
e
o
f
F
l
o
o
d
i
n
g
Model
Model (Perfect MPR Knowledge)
(c)
Figure 4: Fraction of nodes reached by a TC message ood as predicted by the model (soild curves) and QualNet simulation (dashed)
for OLSRDs ooding method (a), and ooding method specied by RFC 3626 (b). (c) The CDF of the relative error between the
fraction of nodes that a TC message ood reaches as predicted by the model and QualNet simulations.
Imperfect MPR Knowledge: In this case, we assume that a
neighbor is aware that it is a nodes MPR only if it receives
a Hello message from the node. Thus, after a node has
selected a subset of its neighbors to be MPRs, we assume
that the probability that each of these MPRs is aware that it
is an MPR is equal to the probability of the MPR receiving
a Hello message from the node, which depends on the radio,
the distance to the MPR, and the interference.
The Perfect MPR Knowledge case is valid when nodes
remain are MPRs for an extended period of time. In this
case, a MPR has several chances to receive Hello messages
and learn that it has been selected as an MPR. The Imper-
fect MPR Knowledge case is valid when a nodes MPR fre-
quently change. While there are scenarios that are between
these cases, but as an approximation, we only consider these
two cases. We use the Perfect MPR Knowledge model when
nodes are stationary and use the Imperfect MPR Knowl-
edge model when nodes are mobile. The utility of Step 5 is
explored in Section 5.1.1.
When simulating ooding a message (i.e., Step 6), either
the method described by RFC 3626 or the method imple-
mented in OLSRD is used. In both cases, packet trans-
missions are subject to transmission errors, which depend
on the radio, distance between nodes, and interference. As
mentioned, in both ooding methods, a node will only con-
sider messages that are received from symmetric neighbors.
It is important to note that the nodes do not have synchro-
nized estimates of which links are symmetric (i.e., each end-
node of a link might have dierent estimates as to whether
the link is symmetric). We have found signicant errors
result when it is assumed that nodes estimates are synchro-
nized. Similarly, signicant errors result when it is assumed
that the end-nodes estimates are independent, i.e., signi-
cant errors result if one assumes that 1 (cii;j, cij;i) =
1 (cii;j) 1 (cij;i).
5.1.1 Model Quality
Figure 4 (a) and (b) shows the average fraction of nodes
that receive a TC message ood. As can be observed, the
model accurately predicts the values and the trends in cover-
age. Figure 4 (c) shows the CDF of the relative error of the
coverage predicted by the model and QualNet simulations.
Here the CDF is over all scenarios discussed in Section 4.3
where the coverage exceeds 70%. When the coverage is less
than 70%, then ooding protocol is performing poorly and
likely impacts the utility of the network. In such cases, we
are not interested in the degree of poor performance, but
rather whether the performance is poor or not.
As can be observed in Figure 4 (c), the model and QualNet
results yield similar results over a wide range of scenarios.
The median error is less than 1% and the 90th percentile
error is less than 5%.
Figure 4 (c) also shows the CDF of the relative error when
Step 5 from Section 5.1 is not used and instead we assume
Perfect MPR Knowledge, that is, we assume that nodes are
always aware when they have been selected as MPRs. Note
that this assumption causes signicant increase in the error.
Figure 5 is similar to Figure 4, except that Figure 5 consid-
ers the fraction of nodes that forward a TC message, which
is )RET in (1). As is the case of the coverage, the model ac-
curately predicts the values and the trends observed in Qual-
Net results. Similarly, Figure 5 (b) indicates that uniformly
assuming Perfect MPR Knowledge results in signicantly
larger error, especially in the tail of the error distribution.
Figure 5(c) show some samples of the time to compute the
ooding performance with the simulation-based model de-
scribed above and with QualNet simulations. Recall that the
simulation-based model, does require some Monte Carlo sim-
ulations. Specically, for each scenario, 120 sample topolo-
gies were used. For each topology, 10 TC messages were
ooded from each node. Similarly, the QualNet simulations
use 120 trials where each trial is run long enough for the
protocols to stabilize and then for enough time for each
node to generate 10 TC message oods. 5(c) shows that
the simulation-based model computes the performance of
the ooding between one and three orders of magnitude
faster than QualNet. For example, the estimated time to
compute the ooding performance over a 500 node network
with QualNet takes 123 days on a single processor, while the
simulation-based model requires only about 4800 sec. Con-
sequently, accurate performance estimates of ooding over
large networks via packet simulation is currently computa-
tionally intractable. However, the simulation-based mod-
eling approach described here allows accurate performance
estimates even for large networks, which is not possible with
packet simulation.
5.2 Black-Box Model
While the model described in Section 5.1 provides accu-
rate predictions of performance, the computations are rea-
sonably involved. Thus, to enable fast computation, a black-
box approach is utilized. A two stage approach is utilized
for developing black-box models. First, graph theoretic met-
rics are selected. Second, curve tting is used to model the
relationship between the ooding performance metric and
the graph theoretic metric. A well selected graph theoretic
metric improves the model quality of t. No methodology
0 5 10 15 20 25
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
0.55
0.6
Speed[m/s]
F
r
a
c
.
f o
r
w
a
r
d
T
C
p
k
t
OLSRD
RFC
N=57, datarate: 0
N=73, datarate: 0
N=73, datarate: 1250
N=73, datarate: 2500
N=91, datarate: 0
(a)
0 .0 5 0 .1 0 .1 5 0 .2
0 .2
0 .4
0 .6
0 .8
1
R e la tiv e Er r o r
C
D
F
E
r
r
o
r
F
r
a
c
. N
o
d
e
s
F
w
d
. F
l o
o
d
P
k
t .
Mo d e l
Mo d e l ( Pe r fe c t MPR Kn ow le d g e )
(b)
0 100 200 300 400 500
10
0
10
2
10
4
10
6
10
8
Number of Nodes
T
i m
e
t o
C
o
m
p
u
t e
E
s
t i m
a
t e
s
( s
e
c
)
Model (0m/s,0Bytes/s)
Model (25m/s,5KBytes/s)
Qualnet (0m/s,0Bytes/s)
Qualnet (25m/s,5KBytes/s)
(c)
Figure 5: (a) Frraction of nodes that forward a TC message ood as predicted by the model (soild curves) and QualNet simulation
(dashed) for OLSRDs ooding method (the upper part of (a)), and the ooding method specied by RFC 3626 (lower part of (a)). (b)
CDF of relative error between the fraction of nodes that forward a TC message ood as predicted by the model and QualNet simulations.
(c) Computation time for estimating performance of ooding via simulation-based model and full packet-level simulations with QualNet.
is presented for selecting the graph theoretic metric; we use
a trial and error approach. Similarly, the form the curves
used to t the relationship between the graph theoretic met-
ric and the ooding performance metric was selected by trial
and error, while the curve parameters were selected to min-
imize the mean square error.
A range of graph theoretic metrics are relevant to ooding.
We consider the following metrics. is the average number
of neighbors that receive a broadcasted message, hence
= 1
_
_

v6=w, w2N
jtrans (

n)
_
_
,
where A is the set of nodes in the network, jtrans (

n) is the
probability of successfully transmitting from to n and the
expectation is over arbitrarily select nodes in an arbitrarily
topology.
is the average number of nodes reached by a one-hop
ood. We dene 2 to be the number of nodes reached
in a two hop ood. and 2 are dened similarly, but
over the MPR-subgraph. Specically, given a node &, let
G
0
(&) = (\
0
(&) , 1
0
(&)) be the subgraph where \
0
(&)
if is a MPR of & or if there exists a node n where n is a
MPR of & and is an MPR of n. Similarly,

n 1
0
(&) if
n is a MPR of . is the number of nodes reached by a
one-hop ood that is originated at &, i.e.,
:= 1
_
_

v6=w,v;w2V
0
(v)
jtrans (

n) Pr (n is MPR of )
_
_
.
And 2 is the average number of nodes reached by a two-hop
ood over G
0
(&) originated at &, where the average is over
all nodes & and topologies. Finally, we dene similarly to
and , but only symmetric links are considered, i.e.,
= 1
_
_

v6=w,v;w2N
jtrans (

n) 1 (ciw;v)
_
_
Through trial and error, the following models of the cover-
age (cocv) and fraction of nodes that retransmit (vct) were
found as a function of the metric vector A = (, 2, , 2, ),
cocvRFC (A) = 1 exp
_
6.50296
_
2

_
1:81899
_
vctRFC (A) =
_
1.02654

0.0023
_
cocvRFC (A)
cocvOLSRD (A) = 1 exp
_
0.14208
2:8161
_
vctOLSRD (A) = (0.006228 (2 ) + 0.522713)
cocvOLSRD (A) .
Here the parameters were selected by minimizing the square
error between the black-box model and the simulation-based
model developed in Section 5.1.
Figure 6 shows the quality of t this black-box model
achieves. Specically, Figure 6 shows the CDF of the relative
error between the black-box model and the results derived
from QualNet simulations. For reference, these gures in-
clude the CDF of the relative error between the simulation-
based model developed in Section 5.1 and the QualNet re-
sults. As can be observed, the quality of t of the two mod-
els is similar, with the simulation-based model resulting in
slightly better quality of t.
6. COMPARISON TO OTHER MODELS
While there has been extensive simulation-based perfor-
mance evaluation of OLSR, to the best of our knowledge,
there has been no research on modeling the coverage and
only limited research on modeling the fraction of nodes that
forward TC messages. In [9] a model of OLSR was devel-
oped. This model is designed to model overhead when TC
messages are generated in response to link failures. While
our model is focused on periodic TC message generation,
the models have similar goals. Specically, in [9], the frac-
tion of nodes that forward a TC message is modeled as No.
MPRs/No. Neighbors. In [9], the number of neighbors was
dened to be the number of nodes within "communication
range." Note that we have found that the dynamics of neigh-
bor discovery play a signicant role in a nodes estimate of
the number of neighbors, and hence, the concept of com-
munication range is insucient to determine the number of
neighbors. Nonetheless, we can dene the No. Neighbors to
be the number of nodes that are close enough that the prob-
ability of successfully receiving a 801tcc packet is above
0.5. We have found that a good estimate of the number
of MPRs is No. MPRs = 1.47 + 2.77 (No. Neighbors)
1=3
,
which agrees with the order estimate given in [15]. With
0 0.1 0.2
0
0.5
1
|err|
c
d
f

c
o
v
e
r
a
g
e

R
F
C
Model
Simple Model
(a)
0 0.1 0.2
0
0.5
1
|err|
c
d
f

c
o
v
e
r
a
g
e

O
L
S
R
D
Model
Simple Model
(b)
0 0.1 0.2
0
0.5
1
|err|
c
d
f

r
e
t

R
F
C
Model
Simple Model
(c)
0 0.1 0.2
0
0.5
1
|err|
c
d
f
r
e
t
O
L
S
R
D
Model
Simple Model
(d)
Figure 6: CDF relative error between black-box model prediction and QualNet simulations of coverage achieved by the ooding algorithm
described in RFC 3626 (a), the coverage achieved by OLSRDs implementation (b), fraction of nodes that forward a TC message under
the algorithm described in RFC 3626 (c), and fraction of nodes that forward a TC message under the algorithm implemented in OLSRD
(d). For reference, the CDF of the relative error between the model described in Section 5.1 and QualNet simulations is included.
0.2 0.4 0.6 0.8 1 1.2 1.4
0
0.2
0.4
0.6
0.8
1
Relative Error
C
D
F

E
r
r
o
r
F
r
a
c
.
N
o
d
e
s

F
w
d
.
P
k
t
.
RFC
Model
Model
0-Speed Model
0.1 0.2 0.3 0.4
0.2
0.4
0.6
0.8
1
Relative Error
C
D
F

E
r
r
o
r
F
r
a
c
.
N
o
d
e
s

F
w
d
.

P
k
t
.
OLSRD
Model
Model
0-Speed Model
Figure 7: The CDF of the relative error between the fraction of
nodes that forward a TC message ood as predicted by dierent
models and QualNet simulations. The model presented in this
paper is refered to as Model, the model presented in the paper by
Wu et al. is Model. The 0-Speed Model is based on a modeling
approach found in several papers.
these denitions, we can apply the model developed in [9],
which we refer to as the , Model, following the notation in
[9].
In several other papers, the fraction of nodes that forward
TC messages is assumed to be independent of speed. How-
ever, no model is given for the fraction of nodes that forward
a TC message at zero speed. Instead, it is suggested that
packet simulation can be used. Recall that the dynamics of
ND cause the ooding to be impacted by node speed, as can
be observed in Figures 4 and 5. Nonetheless, we can apply
such a modeling approach by using the fraction of nodes
that forward TC messages at zero speed as an approxima-
tion of the fraction of nodes that forward TC messages at
any speed. We call this model the 0-Speed Model.
Since we have only recently discovered the bug in the OL-
SRD implementation (which was also a bug in QualNet), it
is sometimes unclear which version of OLSR is being mod-
eled in other papers. Thus, Figure 7(a) shows the CDF of
the relative error between these models and a RFC compli-
ant version of OLSR, while Figure 7(b) shows the CDF of
the relative error between these models and the OLSRD ver-
sion. As can be observed, the model presented in this paper
provides a much more accurate model. The 0-Speed Model
is the second best model. However, this model requires run-
ning full packet level simulations, which is computationally
dicult for large networks (see Section 5.1.1).
7. CONCLUSIONS
In this paper we present a detailed model of the over-
head generated by the topology information dissemination
mechanism of OLSR. We study how each of the three opti-
mizations present in OLSR vary with network parameters,
such as speed, density, channel usage and neighbor discov-
ery mechanism. The model shows that OLSR optimizations
are dramatically diminished in mobile and lossy scenarios.
Also, dierent implementations of OLSR generate signi-
cantly dierent overhead rates. The implications of these
models is currently under investigation.
Disclaimer
The views and conclusions contained in this document are
those of the authors and should not be interpreted as rep-
resenting the ocial policies, either expressed or implied, of
the Army Research Laboratory or the U. S. Government.
8. REFERENCES
[1] T. Clausen and P. Jacquet, RFC 3626: Optimized link
state routing protocol (OLSR), October 2003.
[2] (2010, May) IEEE p802.11 TGs. IEEE. [Online]. Available:
http://www.ieee802.org/ 11/Reports/tgs_update.htm
[3] J. Moy, RFC 2328: OSPF version 2, April 1998.
[4] A. Tnnesen, T. Lopatic, H. Gredler, B. Petrovitsch,
A. Kaplan, S.-O. Tcke, and Others. (2010, May) An adhoc
wireless mesh routing daemon OLSRD. [Online]. Available:
http://www.olsr.org/
[5] Z. Ye and A. A. Abouzeid, A unied model for joint
throughput-overhead analysis of mobile ad hoc networks,
in MSWiM 08. ACM, 2008.
[6] L. Viennot, P. Jacquet, and T. H. Clausen, Analyzing
control trac overhead versus mobility and data trac
activity in mobile ad-hoc network protocols, Wirel. Netw.,
vol. 10, no. 4, July 2004.
[7] Y. Huang, S. Bhatti, and S. A. Sorensen, Analysing the
impact of topology update strategies on the performance of
a proactive manet routing protocol, in ICDCSW 07, 2007.
[8] A. Qayyum, L. Viennot, and A. Laouiti, Multipoint
relaying: An ecient technique for ooding in mobile
wireless networks, INRIA, Tech. Rep. RR-3898, February
2000. [Online]. Available: ftp://ftp.inria.fr/
INRIA/publication/publi- pdf/RR/RR-3898.pdf
[9] X. Wu, H. Sadjadpour, and J. Gracia, Modeling of
topology evolutions and implication on proactive routing
overhead in manets, Comp. Comm., vol. 31, no. 4, March
2008.
[10] A. Medina and S. Bohacek, Technical report 02: Bugs
found in popular implementations of olsr. University of
Delaware, Tech. Rep., June 2010. [Online]. Available:
http://cbr.eecis.udel.edu/ publications
[11] S. Bohacek, A. Medina, and C. Aponte. (2010, June)
Component based routing. [Online]. Available:
http://cbr.eecis.udel.edu/
[12] A. Medina and S. Bohacek, Technical report 01: A
performance model of neighbor discovery in proactive
routing protocols. University of Delaware, Tech. Rep.,
June 2010. [Online]. Available: http://cbr.eecis.udel.edu/
publications
[13] W. Navidi and T. Camp, Stationary distributions for the
random waypoint mobility model, IEEE Trans. on Mobile
Computing, vol. 3, no. 1, 2004.
[14] (2006, October) Qualnet simulator 4.5. [Online]. Available:
http://www.scalable-networks.com/ products/qualnet/
[15] P. Jacquet, A. Laouiti, P. Minet, and L. Viennot,
Performance analysis of olsr multipoint relay ooding in
two ad hoc wireless network models, INRIA, Tech. Rep.,
September 2001. [Online]. Available: ftp://ftp.inria.fr/
INRIA/publication/publi-pdf/RR/RR-4260.pdf

You might also like