You are on page 1of 12

IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 4, NO.

1, JUNE 2007 39

A Probabilistic Approach for


Managing Mobile Ad-Hoc Networks
Remi Badonnel, Radu State, Olivier Festor

Abstract— A pure management approach where all the nodes requirements on the management plane: we propose a new
are managed at any time is too strict for mobile ad-hoc scheme where only some spatio-temporal connected subsets
networks. Instead of addressing the management of the whole of nodes are managed. Such assumptions are well suited for
network, we propose a probabilistic scheme where only a subset
of nodes is managed in order to provide a light-weight and an ad-hoc network, where only nodes that had a more or less
efficient management. These nodes are determined based on network presence are interesting enough to be managed.
their network behavior to favor subsets of well connected and Our main contributions can be summarized in a) definition
network participating nodes. With respect to such a selective of probabilistic management, b) concept formalization with a
management scheme, we derive probabilistic guarantees on the distributed algorithmic method for organizing the management
percentage of nodes to be managed. Our contribution is centered
on a distributed self-organizing management algorithm at the plane, c) integration of this method into a management archi-
application layer, its efficient deployment into a management tecture, d) extensive set of simulations to obtain quantitative
architecture and on a comprehensive simulation study. We will and qualitative results of the method.
show how to organize the management plane by extracting This article is consequently structured as follow: we intro-
spatio-temporal components and by selecting manager nodes duce the concept and the stakes of probabilistic management
with several election mechanisms based on degree centrality,
eigenvector centrality and K-means paradigm. in ad-hoc networks in Section II. We present the underly-
ing distributed algorithmic management method in Section
Index Terms— Network management, mobile ad-hoc networks, III by describing the spatio-temporal connectivity measure,
probabilistic analysis, management architecture.
the extraction of spatio-temporal connected components and
the election of manager nodes. We detail several manager
I. I NTRODUCTION election mechanisms based on degree centrality, eigenvector
centrality and K-means paradigm. We show in Section V how
A D-HOC networks are spontaneous networks deployed
without requiring a fixed network infrastructure such that
localized and temporal limited network connectivity can be
to integrate the probabilistic approach into the ANMP man-
agement framework. Section VI presents and discusses a set
assured in some dedicated target deployment environments. of simulations performed with a common network simulator.
Examples include military applications, emergency rescue Finally, we discuss related work in Section VII and conclude
teams, and other geographically challenging environments. the paper with pointers to future work in Section VIII.
Such networks are highly dynamic, since nodes might come
and go, based on user mobility, out-of-reach conditions and II. P ROBABILISTIC M ANAGEMENT
energy exhaustion. These factors interact strongly with the Monitoring and managing ad-hoc networks is challenged by
network service, requiring new paradigms for the management several constraints which are not encountered in the common
of ad-hoc networks. fixed networks:
Our paper proposes a new management approach for ad-
• Relevant management information: the challenge is to
hoc networks centered on a novel configuration scheme for
identify the essential pieces of relevant information but
the network management plane. This work is orthogonal to our
also to determine what management operations to be
previous work published in [1]. In that paper we addressed the
taken in this context. While we have a fair amount of
issue of what to manage in an ad-hoc network, leaving beside
understanding about what monitoring information should
the issue of the how-to manage it. We complete now our work
be collected, few have focused on what management
with a new approach for organizing the management plane. actions to be taken and on the demonstration of their
This approach can be applied to all management scenarios
effectiveness.
where the administrators do not want to monitor and configure
• Management domain related: since ad-hoc networks are
all the network nodes, but where they only need to manage the dynamically formed loosely coupled entities, the notion
network in a less fine-grained manner. It can cover multiple
of administrative domain, responsible for the overall
applications including mobility, connectivity, ressource usage,
management can be challenging for some application
quality of service, fault detection. The key idea is to relax the
domains. Even if we consider an hybrid ad-hoc network
Manuscript received October 20, 2006; revised February 22, 2007. The connected to the gateway of a company, cooperative
associate editor coordinating the review of this paper and approving it for and role-based schemes should be further investigated to
publication was J. Hellerstein. organize the management plane in an efficient manner.
The authors are with the MADYNES Research Team, LORIA-INRIA
Lorraine, France (email: {badonnel, state, festor}@loria.fr). • Management node reliability and willingness to cooperate
Digital Object Identifier 10.1109/TNSM.2007.030104. (illustrated by Fig. 1): an agent/manager node is limited
1932-4537/07$25.00 
c 2007 IEEE
40 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 4, NO. 1, JUNE 2007

by its inherent limits (connectivity, failures) such that define a probabilistic management scheme? The probabilistic
the provided management data may be biased. How to nature of the management plane derives from the guarantees
improve the reliability of management data which are that we can offer. Whereas in a fixed network, direct neighbor-
provided by other nodes, as well as the reliability of hood relationships among nodes are fixed and management is
data provided by the node itself? Additionally, malicious defined to comprise the totality of the network nodes, it is quite
nodes or non-cooperative nodes may provide false data natural to approach an ad-hoc network with a probabilistic
or no data at all. The key issue is to define distributed framework. We will not guarantee in this framework that all
monitoring and management paradigms, where we can nodes will be managed, but we can give probabilistic bounds
reducethe impact of bias or faulty information. on the percentage of nodes that are managed on average.
• Cost of monitoring: resources (battery power) are already Our probabilistic management is a selective scheme
scarce in ad-hoc networks and monitoring and man- which consists in only managing the most interesting
agement actions are additional consumers in an already nodes of the network based on a statistical analysis.
resource under-powered context. Defining lightweight The notion of interesting is define with respect to relative
schemes, where already available information is used and good network presence and topology relationship. We define
resource availability and consumption can be minimal, a spatio-temporal component to be a subset of the network
are the main issues. nodes, such that any pair of nodes have a high adjacency
probability. The term spatio-temporal is inspired by the spatial
dimension – nodes are required to be adjacent – and the
Malicious temporal one – nodes should have a high probability to be
v4
Behavior
adjacent. Since this probability is estimated over a given
Agent
time period, we identify pairs of nodes that tend to directly
interconnect. Such pairs can be aggregated in spatio-temporal
Non Cooperative Reliability?
components. This can be done by identifying all connected
Behavior v1 components in a graph built from the original network nodes,
Legitimity?
Manager where an edge between two nodes is constructed if and only
v2
if a significant probability of adjacency for these nodes was
Agent v3 Reliabitity? estimated.
We restrict the management plane to the first largest and/or
Agent
maybe the second largest connected component, assuming
that the nodes worth to be managed are located in these
components. Since these components are associated to the
largest node subsets having significant network presence and
v5 Fault behavior frequent adjacent relationships, we consider it is a viable
Agent approach for the specifics of ad-hoc networks. The notion of
spatio-temporal connectivity is related to the time averaged
connectivity of network nodes and does not represent an
Fig. 1. Management node reliability issue: an agent/manager node may instant snapshot connectivity view as proposed in [3]. We
provide biased management data because of its inherent limits such as relate the notion of adjacency to high probabilities of observed
connectivity and failures (node v5 ). It may also provide false data or no
data at all (malicious node v4 or non-cooperative node v2 ). The key issue is direct neighborhood, such that our definition is broader in
to define management paradigms where we can reducethe impact of bias or scope than the simple instant connectivity one.
faulty information. The management approach is centered around these con-
cepts. In the following sections, we will detail a distributed
The key assumption of the probabilistic management ap- management algorithmic method to estimate the adjacency
proach was motivated by the simple observation that ad- matrix and to construct the spatio-temporal connected com-
hoc network management should not comprise the totality of ponents. We will then show how this algorithmic method can
network elements (as it is typically the case in fixed networks). be supported by extending a management architecture.
If node behavior can be better described probabilistically
(a node might be on-line with a given probability), why III. D ISTRIBUTED A LGORITHMIC M ANAGEMENT M ETHOD
should we preserve a deterministic framework for the network The algorithmic method for probabilistic management is
management? based on the spatio-temporal connectivity measure. A net-
The first paper describing conceptual management ap- work node evaluates its spatio-temporal connectivity with
proaches for ad-hoc networks was published by Burgess in its neighborhood and communicates that information to the
[2], where a network model for ad-hoc networks is defined as other network nodes in order to construct the spatio-temporal
a probability matrix, where an element in row i and column connectivity matrix of the ad-hoc network. From this matrix,
j is equal to the probability of nodes i and j to be directly the node is capable to detect spatio-temporal connected com-
connected. Our work is partially motivated by that paper: if ponents of the network and to elect its network manager. We
we assume that such a matrix is known, why not restrict assume a minimal cooperation among nodes, where partial
the network management plane to the nodes that seem to control is allowed. If necessary, the cooperation could be
intercommunicate and have an effective network presence and stimulated by considering an incentive approach such as [4].
BADONNEL et al.: A PROBABILISTIC APPROACH FOR MANAGING MOBILE AD-HOC NETWORKS 41

A. Spatio-temporal connectivity measure were neighbors on the time interval [tl , tl+1 ] and the percent-
age of time pij is higher than the predefined threshold. The
An ad-hoc network is seen as a set of n mobile nodes
graph link is weighted by the percentage of time. Each network
V = {v1 , v2 , ..., vn } moving in a given surface during a time
node vi determines the connected component (noted CCvi ) of
period T . The time period T is split in k measurement interval
which he is part by running through the graph. A connected
[tl , tl+1 ] with tl = l × Tk for an integer l ∈ [0, k]. The choice
of optimal time period T and interval values [tl , tl+1 ] is not component corresponds to a subset of well-connected nodes
in V . The extraction of connected components in the spatio-
addressed in this paper but an excellent analysis on timescales
temporal connectivity graph is described by algorithm 1.
for information flow to monitor ad-hoc networks can be found
in [5]. Each node vi measures the percentage of time pij it was
neighbor with a network node vj . On a time interval [tl , tl+1 ], Algorithm 1: extraction of connected components
the measurements are locally stored in a list N l (vi ) composed Data: spatio-temporal connectivity matrix MST l
C
of tuples (vi , vj , pij ) and are subsequently exchanged and Result: connected component CCvi of node vi
merged among network nodes. The suffix l represents the time initialization
factor and means the measure was performed during the time a) initialize the set CCvi with node {vi } as the
interval [tl , tl+1 ]. single element; // i.e. CCvi = {vi }
The exchange of local measurements lead conceptually to repeat
l
a network spatio-temporal connectivity matrix MST C (or at b) add to the set CCvi all the nodes connected
least a partial view) obtained by concatenating the list of mea- to a node element of CCvi ;
surements performed by the network nodes. Rows/columns c) delete all the doubles of CCvi ;
l
stand for a network node. The i-th row of MST C represents until
the list of measurements N (vi ) of a node vi . If node vi was
l
no change in calculation of connected component CCvi
neighbor with node vj on [tl , tl+1 ], an entry MST l
C [i, j] exists
and contains the percentage of time pij that the pair (vi , vj )
was directly connected on that time interval. As the goal is The CCvi set is initialized with a single element: the node
to highlight network nodes presenting a high probability of vi itself. The key idea of the graph component extraction is
adjacency (and also to limit the management data), only the to add recursively to the set CCvi the neighbor nodes of
spatio-temporal connectivity values which are higher than a each node already part of CCvi . A node vi can extract the
threshold value λ are considered in the matrix. connected component CCvi , but can also extract the other
components CCvj (j = i) he is not part of, by initializing
a set CCvj with a node which is not already assigned and
Spatio-temporal v6 by following the same approach. Each element of a resulting
components Agent 0.45 v7 connected component has been a neighbor of another element
Agent
during at least a minimum percentage of time.
0.42

v1
Agent

C. Manager election in a connected component


0.50

0.47

v8
A connected component, such as CCvi , represents a set
Manager
0.6

v5 of network nodes with an high value of spatio-temporal


4

Manager connectivity. It represents a sub-domain to be managed by one


0.6

0.
2

40 manager, or possibly by several managers if the connected


v10 v2
0.44 v9
component size is significant. An election mechanism is
Agent Agent
Agent
required to determine the manager nodes in each connected
2
0.4

component. Assuming that each node in the network has an


v3
v4
unique identifier (MAC address), the simplest way is to define
an arbitrary election mechanism by electing the node with
Agent Percentage the minimal identifier. The Fig. 3 describes a spatio-temporal
of time
connected component where nodes are identified by a unique
name. We could consider that the node name is the election
Fig. 2. Extraction of spatio-temporal connected components: a spatio- criteria. In that case, node v1 will be elected as the manager
temporal component is a subset of network nodes, such that any pair of nodes
have a high adjacency probability. The term spatio-temporal is inspired by node even if this node is relatively isolated in the connected
the spatial dimension nodes are required to be adjacent and the temporal one component.
nodes should have a high probability to be adjacent. This mechanism is thus not efficient, as it does neither take
into account the structural properties of the connected com-
ponent nor the relative importance of nodes. We will define
more refined election mechanisms based on centrality and K-
B. Connected component extraction
means paradigm. These mechanisms will be applied to the data
l
The matrix MST C can be represented as a graph of n nodes set formed by the spatio-temporal connectivity measurements
as shown in Fig. 2. A graph edge exists between two nodes, limited to the nodes in the connected component. This data set
l
if the following double condition is respected: the two nodes is a submatrix noted SST C , where rows/columns represent the
42 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 4, NO. 1, JUNE 2007

v9 Node Degree Node Eigenvector

v3 2.16 v6 0.52

0.48
v6 1.90 v3 0.47
v7
0.62 v8 v7 1.81 v5 0.39
1
0.7
v4 1.27 v4 0.37
v2 v6 Eigenvector-based
Manager
v5 1.26 v7 0.36
7

0.
0.5
0.

62
54

v8 1.62 v2 0.17
v3 Degree-based
Manager
v5 v2 0.54 v8 0.15
0.6
42

0.64
0.

v4
v9 0.48 v1 0.14
v1
v1 0.42 v9 0.12
Spatio-temporal
Component

Fig. 3. Degree-based and eigenvector-based elections of manager nodes in a spatio-temporal connected component: the graph on the left is a spatio-temporal
connected component. The two tables synthesizes respectively the degree centrality and the eigenvector centrality measurements of the component nodes. In
each of these tables, nodes are classified by order of importance, so that the first entries of a given table correspond to the nodes elected as managers. The
degree-based election selects node v3 as manager, while the eigenvector-based election selects node v6 . The degree centrality is a local measure limited to
the one-hop neighborhood information, whereas the eigenvector centrality takes recursively into account the importance of the neighbor nodes.

m nodes part of a connected component. Let us consider the We applied this election mechanism to the scenario presented
scenario presented in Fig. 3. The spatio-temporal connected in Fig. 3. The degree measure for a node vi is computed as the
component is composed of 9 ad-hoc nodes and is associated sum of the elements of the ith line of matrix SST l
C . Results
the following 9 × 9 matrix SST
to ⎡ l
C: ⎤ are synthesized in the first table of Fig. 3. Node v3 has the
0 0 0.42 0 0 0 0 0 0 highest degree centrality with a value of 2.16. It will be elected
⎢ 0 0 0.54 0 0 0 0 0 0 ⎥ as the manager node. The time complexity of the degree-based
⎢0.42 0.54 0 0.63 0 0.57 0 0 0 ⎥
⎢ ⎥ election mechanism is of O(N ).
⎢ 0 0 0.63 0 0.64 0 0 0 0 ⎥
⎢ ⎥ 2) Eigenvector-based election mechanism: The degree cen-
⎢ 0 0 0 0.64 0 0.62 0 0 0 ⎥
⎢ 0 0 0.57 0 0.62 0 0.71 0 0 ⎥ trality is a local measure limited to the one-hop neighborhood
⎢ ⎥
⎢ 0 0 0 0 0 0.71 0 0.62 0.48⎥ information. An alternative measure is the eigenvector cen-
⎣ 0 0 0 0 0 0 0.62 0 0

trality defined by Bonacich in [7]. The key idea is to define
0 0 0 0 0 0 0.48 0 0
l the centrality measure in a recursive way: the more a node
The matrix SST C is a symmetric matrix because it represents is directly connected to central nodes, the more the node is
an undirected graph. The two following election mechanisms
central. For such a non-oriented connected graph, Bonacich
are inspired by research work on centrality. The concept of
demonstrates that the problem can be reduced to the typical
centrality allows to quantify the relative importance of a
eigenvector problem defined by equation 2.
node in a network or a graph structure. Several variants of
C ×
x = λ × x
l
this measure have been defined in social networking [6]. The SST (2)
objective is to identify the nodes having a central position The vectors that solve this equation are called eigenvectors
in the connected component, and to elect them as manager and are associated with corresponding eigenvalues. The princi-
nodes. pal eigenvector vpr is the solution vector which is paired with
1) Degree-based election mechanism: A well-known cen- the highest eigenvalue. The eigenvector centrality ceig of node
trality measure is the degree centrality. A node is central if it vi corresponds to the ith element of the principal eigenvector
is directly connected to a maximal number of neighbor nodes. l
of matrix SST C , as mentioned in equation 3.
The degree centrality of a node is the number of links linked to
this node. Each link being weighted by a percentage of time, ceig (vi ) = vpr (i) f or a given node vi (3)
the centrality degree measure is defined by equation 1. We We solved the eigenvector problem for the scenario defined
calculate, for each node vi in the component, the sum of the in Fig. 3 using a symbolic mathematics software. The highest
values (in percentage of time) of each link connected to the eigenvalue of SSTl
C is equal to 1.457 and is paired with the
node. Nodes having the highest centrality degree are elected principal eigenvector defined in equation 4:
as manager nodes.
vpr = [ 0.14 0.17 0.47 0.37 0.39 0.52 0.36 0.15 0.12 ]T (4)

j=m
cdeg (vi ) = l
SST C (i, j) f or a given node vi (1) The elements of the principal eigenvector are ranked by order
j=1 of importance, which leads to the second table of Fig. 3. Node
BADONNEL et al.: A PROBABILISTIC APPROACH FOR MANAGING MOBILE AD-HOC NETWORKS 43

v6 shows the highest value and is elected as manager node. Algorithm 2: election of manager nodes
The two centrality approach do not provide the same results: Data: spatio-temporal connectivity matrix SST l
C
the degree centrality is a local measure while the eigenvector Result: set of manager nodes for the connected
degree takes additionally into account the importance of the component
neighbor nodes themselves. A generalized version of the initialization
eigenvector centrality has been defined in [8] in the case a) select two clusters {c1, c2} arbitrarily;
of oriented connected graphs. The time complexity of the b) compute initial cluster centroids centroid(c1)
eigenvector-based election mechanism is of O(N 3 ). and centroid(c2) within those two clusters;
3) K-means election mechanism: As some ad-hoc nodes repeat
may provide biased or false management data, a key issue c) partition by assigning or reassigning all the points vj
is to provide a reliable election capable to out-weight or of the connected component to their closest cluster
discard faulty information. We therefore propose an alternative centroid;
election mechanism where the measurements performed by d) compute the new cluster centroid of each cluster;
each node in the connected component are compared in order until
to provide the best trade-off. The goal of the manager election no change in cluster centroid calculation
is basically to classify the connected component nodes in two finally
populations: agent nodes and manager nodes, favoring the the manager set is the cluster (c1 or c2) whose centroid
manager role to the nodes with the highest spatio-temporal has the highest norm;
connectivity.
We use the K-means classification method [9] to divide the
connected component nodes in two clusters noted {c1, c2}.
A node vj in the connected component (column of matrix minimize the objective function defined by equation 8 where
l
SST C ) is seen as a point in a m-dimensional vector space, clusters c1 and c2 represent the two node populations and
where a dimension corresponds to the view of a node vi (row where ||vj − centroid(c)|| is the distance between a point vj
j j
C ). The coordinates (x1 , x2 , .., xm ) of a point
l j
of matrix SST and the centroid of its cluster c.
vj correspond then to (SST l
C [1, j], S l
ST C [2, j], ...S ST C [m, j]).
l
j
The coordinate xj stands for the node self-measured up-time 1  
E= ||vj − centroid(c)||2 (8)
percentage. In this vector space, the norm of a point va noted 2
c∈{c1 ,c2 } vj ∈c
||va || is given by equation 5 and the distance measure between
two points va and vb noted ||va − vb || is given by equation 6.
After performing the algorithm, the two final clusters contain
network nodes with close spatio-temporal connectivity. The

k=m

 manager nodes are located in the cluster with the centroid
||va || = (SST
l
C [k, a])
2 (5) presenting the highest spatio-temporal connectivity norm.
k=1
An illustrative example is shown in Fig. 4 with a simple

k=m

 connected component composed of three nodes {v1 , v2 , v3 }
||va − vb || = (SST
l
C [k, a] − SST C [k, b])
l 2 (6) l
(presented on the left). The submatrix SST C is a 3 x 3 matrix
k=1
where rows/columns represents the three nodes in the con-
Intuitively, the K-means classification puts together ad-hoc nected component. The K-means classification is performed
nodes showing similar spatio-temporal connectivity values to in a three-dimensional vector space (presented on the right)
form two clusters. The centroid of a cluster c corresponds where an axis represents the view (spatio-temporal connec-
l
basically to the point whose coordinates are the mean of the tivity measure) of a given node (row of SST C ) and where
l
coordinates of all the points in the cluster. Assuming a cluster a point represents a node to be classified (column of SST C ).
c composed of k nodes {v1 , v2 , ..., vk }, the coordinates of The K-means algorithm divides the three nodes in two clusters
the centroid are thus given by equation 7 providing the i-th c1 and c2 represented by spheres in this three-dimensional
coordinate in the m-dimensional vector space. vector space. Cluster c1 is composed of nodes v2 and v3 ,
q=k q while cluster c2 is composed of a single node v1 . A numeri-
centroid(c) q=1 xi cal calculation shows the spatio-temporal connectivity norm
xi = (7)
k of centroid(c2) is higher than the norm of centroid(c1)
The K-means clustering for electing the manager nodes is (||centroid(c2)|| > ||centroid(c1)||). Manager nodes are lo-
described in algorithm 2 and can be divided in two main steps. cated in the cluster c2 whose centroid shows the highest norm.
In the first step of initialization, nodes of the connected com- The connected component will therefore be composed of a
ponent are arbitrarily assigned to a cluster c1 or c2 and the cen- single manager: node v1 . This election mechanism based on K-
troid of each cluster is calculated using previous equation 7. In means algorithm relies on the comparison of spatio-temporal
the second step, the algorithm reassigns iteratively each point connectivity measures performed by different nodes in a
vj to the cluster c ∈ {c1, c2} whose centroid is nearest and distributed manner. The comparison of these management data
recalculate the centroids. The algorithm stops when the cluster improve the election reliability by out-weighting biased/false
centroids are fixed (the coordinates do not change between two data. The time complexity of this election algorithm is of
iterations). The K-means algorithm corresponds actually to O(2N ).
44 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 4, NO. 1, JUNE 2007

v2

Cluster c2

X v1

v1

X v2

3
0.7

0.8
2
v2
X
0.12 v3

v1
Cluster c1
X v3

Spatio-temporal
Component

v3 Centroids

Fig. 4. K-means election of manager nodes in a spatio-temporal connected component: a simple scenario of a connected component composed of three
nodes {v1 , v2 , v3 } is presented on the left. The K-means classification method (presented on the right) divides the three nodes in two clusters c1 and c2
represented by spheres in this three-dimensional vector space. Ad-hoc nodes which will interact as managers are located in the cluster c2 whose centroid
shows the highest spatio-temporal connectivity norm. A single manager (node v1 ) is elected in this simple scenario.

IV. M ONITORING OF PATHOLOGICAL M OBILITY IN A {v1 , v2 , ..., vm }: management data reported by reliable agents
S PATIO - TEMPORAL C ONNECTED C OMPONENT will have higher weights. At the network deployment, the
weight of the agents is initialized with a similar weight
An ad-hoc node measures periodically the spatio-temporal
w1 = w2 = ... = wm , because the manager do not already
connectivity of the other network nodes. The manager nodes
know the agent nodes. When an agent vj is detected as patho-
can monitor the spatio-temporal connectivity variations over
logical, the manager decreases the weight wj of that agent. For
time of nodes located in a spatio-temporal connected compo-
instance, the weight can be divided by two, each time the agent
nent in order to detect pathological mobility behavior. While
shows an pathological mobility behavior. When the manager
normal mobility is difficult to define, in some specific target
will get management data from agents, it will more take into
deployment (for instance military applications), unpredicted
account in the management process the information provided
mobility patterns can seriously impact the network resilience
by agents with high reliable weights. This corrective operation
and service level and deteriorate the management plane or-
limit the propagation and the impact of biased information in
ganization. We mean by ”pathological mobility” a mobility
the management plane.
behavior that reveals underlying problems of ad-hoc nodes
such as: V. M ANAGEMENT F RAMEWORK
• Failures due to miss-configuration and errors at the We propose to integrate the probabilistic approach into the
physical layer may generate an atypical behavior, where ANMP (Ad-Hoc Network Management Protocol) management
nodes are intermittent although from a mobility point of framework [10] as presented in Fig. 5. ANMP defines a man-
view they did not change significantly, agement protocol for ad-hoc networks, but also describes the
• Routing failures which can be encountered when the underlying management architecture and information model.
routing process is affected by voluntary and malicious The protocol is fully compatible with SNMPv3 [11] and thus
activity (attacks against the routing plane), errors in its with most of current management architecture. The key idea
configuration or at the protocol stack level, of ANMP is to clusterize the management plane in order to
• Battery breakdown: spatio-temporal connectivity may be reduce the management message overhead. The management
lost for longer or shorter time periods because of battery plane is logically divided into a three-level hierarchy com-
lifetime. posed of a top-level manager, SNMP managers performed
Pathological mobility can be inferred by analyzing sta- by cluster-head nodes (CH) and SNMP agents performed
tistically the variations of the spatio-temporal connectivity by simple nodes (SN). ANMP introduces by default two
measurements pij over time. Ad-hoc nodes with such a clustering algorithms at the application layer to organize the
mobility can involuntary corrupt management data and report management plane:
biased or faulty information because of their pathology. A • Graph-based clustering algorithm (clusters of one hop
key issue is to define actions that compensate for the manage- neighbors). Each network node has a unique identifier
ment data provided by these problematic nodes. A possible and maintains the list of its one-hop neighbor nodes using
corrective operation consists in using a weight system to the network layer or the MAC layer. In a neighborhood,
estimate the reliability of agent nodes. The manager node the node with the minimum identifier is elected as the
vi assigns reliability weights {w1 , w2 , ..., wm } to agent nodes cluster-head, if it has not joined any other cluster.
BADONNEL et al.: A PROBABILISTIC APPROACH FOR MANAGING MOBILE AD-HOC NETWORKS 45

• Geographical-based clustering algorithm (clusters of up of mobile ad-hoc networks. As in a pure link state algorithm,
to three-hop neighbors). This algorithm uses the global each node performs two main operations: (1) it determines
positioning system in order to evaluate node positions. the list of direct-connected neighbor nodes by accomplishing
Network nodes are split into rectangular clusters based link sensing through periodic emission of hello messages and
on their spatial density. (2) it exchanges this link state information with the other
ANMP is defined at the application layer and draw a dis- nodes throughout the network by flooding topology control
tinction between clusters for management and clusters for messages. Due to the limited bandwidth of ad-hoc networks,
routing, as they aim at different goals (respectively organizing OLSR reduces the control overhead by selecting a subset
the management plane and maintaining routes). The authors of nodes called Multi-Point Relays (MPRs) which forward
of ANMP conclude the paper in [10] by pointing out it would broadcast messages during the flooding process. The selection
be interesting to explore other way of clustering or methods of MPRs is based on an heuristic which consists in selecting
to group the nodes in the network. for a given node the minimal subset of one-hop neighbors to
We therefore introduce an additional clustering algorithm at reach all the two-hop neighbors, in order to reduce the number
the application layer to implement the probabilistic approach of links used for forwarding topology information.
where clusters are defined by spatio-temporal connected com- During the link sensing operation, an OLSR node uses a
ponents. This extension is composed of three building blocks neighborhood base (see RFC 3626 section 4.3). and maintains
(1) to evaluate the spatio-temporal connectivity, (2) to extract the list of one-hop neighbor nodes by recording a set of tuples
the spatio-temporal connected components and (3) to elect the (N neighbor main addr, N status, N willingness) describing
spatio-temporal connected component manager(s). neighbors. The element N neighbor main addr stands for the
main address of a neighbor, N status specifies the link status
(symmetric or not) with the given neighbor and N willingness
is an integer specifies the willingness of the neighbor to carry
traffic on behalf of other nodes. We introduce an additional
entry in the neighbor tuple called N connectivity time perc to
maintain the percentage of time the neighbor was connected
to the node.
This new entry must then be integrated in the topology
control (TC) messages (defined in RFC 3626 section 9)
performing the advertisement of link states, so that the spatio-
temporal connectivity information is propagated in the ad-hoc
network. Including a sequence number, the format of a TC
message describes a set of Advertised Neighbor Main Address.
We define an additional field Advertised Neighbor Connec-
tivity Time Percentage to include the percentage of time the
neighbor was seen. Our spatio-temporal clustering algorithm
extension, defined over a piggybacked OLSR version, gives to
the ANMP architecture the capability to perform probabilistic
management.
The integration of the probabilistic approach into ANMP
implies to extend the management information base (MIB)
of this management framework and of the OLSR routing
protocol. A MIB corresponds basically to a database of
objects that can be monitored and managed by the network
management system [14]. We proposed an OLSR MIB module
Fig. 5. Integration of the probabilistic approach into the ANMP management in [13] which is partially described in Fig. 6. This partial
framework: ANMP is extensible to other clustering algorithms. In order to view details two groups: the first one called olsrGeneralGroup
organize the management plane based on the spatio-temporal approach, an
additional clustering algorithm is implemented over a piggybacked version of provides general information on the OLSR node, on its net-
the OLSR routing protocol. work interfaces compatible with OLSR (olsrInterfaceTable)
and on its willingness to interact as a MultiPoint Relay.
The management algorithms are executed locally by the The second group olsrLocalInformationGroup describes the
network nodes based on information obtained from the routing local information of the OLSR node including the one-hop
plane. The consistency of information and the robustness to neighborhood olsrNeighborSetTable and the two-hop neigh-
mobility of our management solution relies therefore on the borhood olsrTwoHopNeighborSetTable. The table olsrNeigh-
performances of the routing protocol. borSetTable enumerates the one-hop neighbors and their link
The evaluation of spatio-temporal connectivity is based on status. We added to this table an additional entry olsrNeigh-
the piggybacking of OLSR (Optimized Link State Routing borConnectivityTimePerc (shown in grey on the figure). This
Protocol) [12]. This routing protocol is an optimization of entry represents the percentage of time of connectivity in order
the classical link state algorithm adapted to the requirements to fit with the routing protocol piggybacking.
46 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 4, NO. 1, JUNE 2007

Fig. 6. Extension of the OLSR routing protocol MIB: The figure presents a
partial view of the OLSR MIB we proposed in [13]. The table olsrNeigh-
borSetTable enumerates the one-hop neighbors and their link status. We
added to this table an additional entry olsrNeighborConnectivityTimePerc
(shown in grey on the figure). This entry represents the percentage of time
of connectivity in order to fit with the routing protocol piggybacking.

Fig. 7. Extension of the ANMP management framework MIB: this figure


details only the anmpTopologyMaintenance group of that MIB, with the grey
The ANMP MIB is also extended to the probabilistic entries corresponding to our extension. We introduce an additional group
management. The anmpTopologyMaintenance group of that called anmpSpatioTemporalClustering for applying the probabilistic approach
to the management plane organization.
MIB is detailed in Fig. 7 with the grey entries corresponding
to the extension. This group contains objects related to the
topology information of the ad-hoc network. The anmpNeigh- the characteristics of a connected component including the
borTable defines the list of one-hop neighbors of a node list of its managers provided by the table anmpSTCCompo-
and is completed by the additional entry anmpNeighborCon- nentManagerTable and the list of its agents provided by the
nectivityTimePerc to ensure coherence with the OLSR MIB table anmpSTCComponentAgentTable. These MIB extensions
extension. The anmpProtocolTable maintains an entry for each allow to specify the management information implied by our
protocol which may be used for topology maintenance in probabilistic approach in a standardized way.
the management plane. Each of the two clustering algorithms
provided by default with ANMP is represented by a group: an-
mpGraphClustering and anmpGeographicalClustering. These VI. E XPERIMENTAL R ESULTS
groups are not detailed on the figure, but a more complete We describe in this section a set of simulations, in which
description can be found in [10]. We introduce an additional we apply the management algorithmic method to detect the
group called anmpSpatioTemporalClustering for applying the largest spatio-temporal connected components. The simula-
probabilistic approach to the management plane organization. tions are performed using the widely-accepted network sim-
This group maintains the information generated by the con- ulator ns-2 [15]. The data link layer is based on the wireless
nected component extraction algorithm (algorithm 1) and by IEEE 802.11 MAC model. The routing layer corresponds to
the manager election algorithm (algorithm 2) presented in the standardized OLSR routing protocol (NRL’s implemen-
Section III. In particular, the table anmpSTCComponentTable tation [16]). We consider an ad-hoc network composed of
contains the list of spatio-temporal connected components. n ∈ [5 − 30] mobile devices deployed in an urban area.
Each entry anmpSTCComponentEntry in the table inventories The nodes are moving in a squared surface with a 1000
BADONNEL et al.: A PROBABILISTIC APPROACH FOR MANAGING MOBILE AD-HOC NETWORKS 47

meters side according to the commonly-used random way- 110


point (RWP) model [17] (being aware of the model limits
100
[18]): each node moves at constant speed s ∈ [0.5−10] m/s to
a destination point selected uniformly in the squared surface 90
and then waits during a pause time p ∈ [1 − 50] s before
moving to a new destination. The simulation parameters which 80

Percentage of cases (simulations)


are not specified in the sum-up table I are set with the ns-2
70
default values. To avoid initialization discrepancy issues with
the RWP model [18], we used the steady-state mobility model 60
generator mobgen-ss where initial speeds and locations of
nodes are chosen from the stationary distribution to perform an 50
immediate convergence and provide more reliable simulations.
40
We determined experimentally the λ value (40%) based on an
initial set of simulations. 30
TABLE I
20
S IMULATION PARAMETERS
Parameter Value 10
Simulator ns-2
Simulation time 1800 s 0
0 10 20 30 40 50 60 70 80 90 100
Simulation area 1000 m x 1000 m
Minimal percentage of nodes in the largest spatio-temporal component
Number of ad-hoc nodes 5-30 nodes
Mobility model random waypoint
mobgen - steady state Fig. 8. Probability distribution for the ratio of the largest spatio-temporal
Speed 0.5 - 10 m/s connected component to the overall network: a point (x,y) on the graph
Pause time 1 - 50 s indicates the percentage y of simulations such that the largest spatio-temporal
component covers at least x percent of network nodes.
Physical layer FSP / 2-RGR
MAC layer IEEE 802.11
Routing layer NRL OLSR
λ value 40% different node sizes. For each node size, 150 simulations were
performed to avoid any bias. These results are summarized in
Fig. 9. We expected to have good results for low mobility (ie.:
In a first series of experiments we analyzed the probability small speeds and large pause times), where by good results
distribution of the ratio of the largest spatio-temporal con- we understand a high probability to have a large percentage of
nected component to the overall network. These results are the network nodes in the largest spatio-temporal component.
shown in figure 8 and represent an extensive set of simulations Such was the case indeed (note the case of speed = 0.5m/s
with different mobility parameters and node sizes. For each and pause time = 50s) where about 90% of the nodes are
individual setting, we performed 150 simulations to assure the located in the same component in about 90% of the cases. A
non-bias of the result. rather surprising result is however the case of small pause
On the x axis is plotted the percentage of the overall times (speed = 0.5m/s and pause time = 1s), where 90%
node size which is located in the largest spatio-temporal of the nodes are located in the same component in around
component, while the y axis stands for the percentage of 80% of the cases. If we analyze the cases with varying pause
cases (simulations). A point (x,y) on the graph stands for times and constant speed = 10m/s another interesting issue
the percentage of simulations (y) such that the largest spatio- is that the worst results are obtained with a high pause time.
temporal component was at least x. For instance we can This contradicts the initial hypothesis that a higher mobility
observe that the probability (y axis) of having at least 20% implies a smaller component size. It seems that in cases of
of nodes in the largest component is about 95%. We could higher node speeds, the spatio-temporal connectivity is better
have a best-effort type pragmatic management approach in if nodes do not rest/pause. Such a result could be summarized
the style: we manage around 20% of the nodes and in this in one phrase Move fast but do not rest.
case we will meet this requirement in 95% of the cases. We In the previous experiments we limited the analysis to
can assure the management of the half of the nodes within one the largest spatio-temporal component. A natural extension
single component with a probability close to 70%, which is is to consider the two largest spatio-temporal components.
a relative good result. Such a best-effort type of management Figure 10 summarizes an experiment suite analogous to the
could be implemented by one and only one manager station one illustrated in figure 8, but where the aggregated sizes of
following the spatio-temporal component. the most important components is considered. We can observe
A natural question is whether network mobility impacts the that in the extended case, about 50% of the nodes are located
management framework. Intuitively, higher mobility should in the two main components with a probability close to 90%.
make things worse (ie.: smaller components sizes for the That is, we can manage about 50% of the nodes with a
same probability), but a precise quantification of this effect very high probability. We see that with respect to the one
is required. A second series of experiments addressed this component scenario, we get about 20% more nodes at the
issue, where different mobility parameters were evaluated for same probability. If we consider in this case the requirement
48 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 4, NO. 1, JUNE 2007

110 110

100 100

90 90

80 80
Percentage of cases (simulations)

Percentage of cases (simulations)


70 70

60 60

50 50

40 (0.5, 50) 40
(0.5, 10)
(0.5, 1)
30 (1, 50) 30
(1, 10)
(1, 1)
20 (5, 50) 20
(5, 10)
(5, 1)
10 (10, 50) 10
(10, 10)
(10, 1)
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Minimal percentage of nodes in the largest spatio-temporal component Minimal percentage of nodes in the two largest spatio-temporal components

Fig. 9. Probability distribution for the ratio of the largest spatio-temporal Fig. 10. Probability distribution for the ratio of the two largest spatio-
connected component to the overall network according to the network temporal connected components to the overall network: the curve represents
mobility: each curve stands for different mobility parameters (speed, pause). the value of the aggregated ratios of the first and second spatio-temporal
connected components.

to manage at least 50% of the nodes, than we can achieve this


goal in about 90% of the cases. intelligence is spread over the network nodes according to
The last series of experiments presented in this paper are their capabilities. GUERILLA is organized in a two-tier ar-
illustrated in Fig. 11. Here we plotted the individual compo- chitecture. The higher-tier is composed of groups of peer-
nent sizes (expressed in percentage of the overall network node to-peer nomadic managers that maintain connectivity in the
size) where all simulations were done as previously described. management plane with the other managers. The lower-tier is
We conclude that the second largest component is relevant composed of agents and active probes (programmable scripts)
in about 80% of the cases, counting for around 15% of the which perform the localized management operations. This
network node size. This is relatively normal, since a small autonomic solution is conceptual and do not discuss the
sized first component implies a small sized second component. underlying clustering mechanism. However, it points out that
the management role depends on the node capability. The
probabilistic algorithm could take into account other factors
VII. R ELATED W ORK related to node state such as battery level and node process-
Lots of research efforts deal with connectivity issues in ing capability. In particular, a programmable middleware for
ad-hoc networks. The approaches in [19], [3], [20] define the dynamic deployment of services and protocols in ad-
how to optimize node connectivity in function of the network hoc networks is proposed by [23]. In this programmable
parameters such as power control parameters. However, they infrastructure, ad-hoc nodes can dynamically download and
focus on node connectivity at an instant time, while we are activate the required protocol and service software. Therefore,
interested in a probabilistic scheme based on the percentage of the platform can align the capabilities of heterogeneous ad-
time of connectivity, and while we consider the connectivity hoc nodes by using loadable plug-ins. The two management
is given as it is. An approach for organizing network nodes approaches ANMP and GUERILLA infers the use of n-hop
into clusters in which path availability can be bounded was clustering algorithms at the application layer. We consider that
proposed in [21], but this solution defined at the routing layer a pure and sound management approach, where all nodes are
aims at ensuring end-to-end network performance. managed at any time is too strict and we therefore propose
Among the pioneering approaches for managing ad-hoc a new relaxed clustering scheme where only some spatio-
networks, ANMP [10] presented in Section V proposes a temporal connected subsets of nodes are managed.
management protocol and a management architecture adapted Policy-based management approaches are also proposed in
to ad-hoc network specifics. Clustering algorithms (at the [24], [25], [26], [27]. For instance, Phanse in [25] elabo-
application layer) are used to organize the management plane rates a framework for policy-based management based on
and reduce the management messages overhead. We used three key mechanisms: service location, cluster management
ANMP as a support to integrate the probabilistic approach. and dynamic service redundancy. In particular, the hybrid
The GUERILLA management architecture proposed in [22] provisioning/outsourcing policy model is decentralized using
defines a self-management approach where the management a clustering algorithm to provide an efficient and robust
BADONNEL et al.: A PROBABILISTIC APPROACH FOR MANAGING MOBILE AD-HOC NETWORKS 49

110 intermittence due to mobility and other ad-hoc specifics,


1st largest component
2nd largest component while faulty behavior can generate pathological intermittence
100
behavior. The key issue addressed was how to differentiate
90 pathological intermittence from regular intermittence and thus
identify faulty nodes from regular non-faulty ones. While
80 fault detection in fixed wired networks is not hindered by
Percentage of cases (simulations)

the impossibility to observe a given node, ad-hoc networks


70
specifics do provide major challenges with respect to this
60
issue. We introduced in [32] an pathological intermittence
measure based on information theory allowing a local node to
50 detect pathological intermittent nodes by monitoring the rout-
ing plane. A distributed monitoring scheme, where network
40
nodes share their local measurements, is designed in order to
30
provide a reliable and robust detection of faulty nodes. Several
distributed methods of fault detection are defined including an
20 auto-configured detection method.

10
VIII. C ONCLUSIONS AND FUTURE WORK
0
0 10 20 30 40 50 60 70 80 90 100 Instead of defining a pure and sound management approach,
Minimal percentage of nodes in the largest spatio-temporal components we have presented in this paper a new probabilistic scheme for
configuring the management plane in an ad-hoc network. The
Fig. 11. Probability distributions for the ratio of the first and second largest underlying key idea is the notion of spatio-temporal connected
spatio-temporal connected components to the overall network: the second nodes. A spatio-temporal connected component is a subset of
largest component is relevant in about 80% of the cases, counting for around
15% of the network node size. the ad-hoc network, such that nodes within such a component
have a high probability of being directly connected. The
term spatial derives from this neighborhood notion, while
management. The approach in [24] considers an alternative the term temporal is related to the temporal behavior of this
policy management scheme where the policy distribution is neighborhood. In a store and forward oriented architecture,
based on the synchronization of policy databases when two such nodes are also capable to inter-communicate at a higher
network nodes meet. A probabilistic approach for policy-based hop-count. The management plane is limited to the largest
management can also be envisioned in ad-hoc networks. spatio-temporal connected components.
The DAMON architecture [28] defines a distributed moni- We have defined the underlying management self-
toring system based on agents for multi-hop networks: agents organizing algorithm for both extracting spatio-temporal con-
perform the network monitoring and send the measurements to nected components and electing manager nodes. We have
data repositories. DAMON supports multiple data repositories detailed several manager election mechanisms based on cen-
and includes an auto-discovery mechanism of data repositories trality degree, eigenvector centrality and K-means paradigm
by the agents. This generic architecture is not dedicated to and have shown how to apply them with realistic scenarios.
specific network parameters and could therefore be appro- These election mechanisms aim at selecting nodes having
priate for the storage of pathological mobility monitoring a structural importance in the spatio-temporal components,
data. A multi-tier architecture for efficient monitoring and and at limiting the propagation of biased and faulty infor-
management is proposed in [29] where nodes are grouped to mation in the management plane. The probabilistic approach
ensure that they can be reached via a member of its group. is integrated into the ANMP management framework over a
The approach is initially designed for sensor networks, but is piggybacked ad-hoc routing protocol. We have subsequently
applicable to ad-hoc networks, since they present the similar shown how to extend the ANMP management information
energy and bandwidth constraints. WANMon is a monitoring base.
tool described in [30] to monitor the resource usage in terms The management approach is called probabilistic since its
of network traffic, energy, memory and CPU, but its scope is behavior can only be guaranteed in a stochastic way. We
limited to the host-level monitoring. can ensure that a given percentage of the overall network
We introduced the concept of probabilistic management of will be managed with a fixed apriori known probability. We
ad-hoc networks in [31]. We detail further in this paper the have estimated these probabilities using extensive network
management algorithmic scheme, describing several manager simulations and we have performed a fine grained analysis
election mechanisms based on degree centrality, eigenvector of the mobility impact on our framework. Although network
centrality and K-means paradigm. The probabilistic scheme simulators present functional limits [33], the experimental
can also provide a basis for identifying the pathological be- methodology is generic and can be applied to other mobility
havior of nodes such as introduced in section IV. We proposed scenarios. We are working on alternative mobility models
in [32] a more refined approach for detecting faulty nodes including the Manhattan grid model and the reference point
in ad-hoc networks. Faulty behavior and intermittence are group mobility model. Our future work will consist in design-
closely related in ad-hoc networks: a node can have a regular ing an information theoric framework [32] for this probabilis-
50 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 4, NO. 1, JUNE 2007

tic scheme, and in extending the management algorithms to [23] S. Gouveris, S. Sivavakeesar, G. Pavlou, and A. Malatras, “Program-
take into account additional properties of nodes. mable Middleware for the Dynamic Deployment of Services and Proto-
cols in Ad-Hoc Networks,” in Proc. of the 9th IFIP/IEEE International
R EFERENCES Symposium on Integrated Network Management (IM’05), Nice, France,
Apr. 2005, pp. 3–16.
[1] R. Badonnel, R. State, and O. Festor, “Management of Mobile Ad- [24] A. Munaretto, S. Mauro, P. Fonseca, and N. Agoulmine, “Policy-based
Hoc Networks : Evaluating the Network Behavior,” in Proc. of the 9th management of ad-hoc enterprise networks,” in Proc. of HP Openview
IFIP/IEEE International Symposium on Integrated Network Manage- University Association 9th Annual Workshop, June 2002.
ment (IM’05), Nice, France, Apr. 2005, pp. 17–30. [25] K. Phanse, “Policy-Based Quality of Service Management in Wireless
[2] M. Burgess and G. Canright, “Scalability of Peer Configuration Manage- Ad-Hoc Networks,” Ph.D. dissertation, Faculty of the Virginia Polytech-
ment in Logically Ad-Hoc Networks,” IEEE eTransactions on Network nic Institute and State University, Aug. 2003.
and Service Management (eTNSM), vol. 1, no. 1, Apr. 2004. [26] R. Boutaba, Y. Iraqi, and B. Ishibashi, “Policy-Based Routing in Ad-
[3] P. Santi and D. M. Blough, “An Evaluation of Connectivity in Mobile Hoc Networks,” in Proc. of the First Ad-Hoc Networking Workshop
Wireless Ad-Hoc Networks,” in Proc. of the International Conference (Med-Hoc-Net’02), Sardegna, Italy, Sept. 2002.
on Dependable Systems and Networks (DSN’02). Bethesda, MD, USA: [27] R. Chadha, H. Cheng, Y.-H. Cheng, and J. Chiang, “Policy-based Mobile
IEEE Computer Society, June 2002. Ad-Hoc Network Management,” in Proc. of IEEE 5th International
[4] L. Buttyan and J. P. Hubaux, “Stimulating Cooperation in Self- Workshop on Policies for Distributed Systems and Networks (POL-
Organizing Mobile Ad Hoc Networks,” ACM/Kluwer Mobile Networks ICY’04), New York, USA, June 2004.
and Applications, vol. 8, no. 5, Oct. 2003. [28] K. Ramachandran, E. Belding-Royer, and K. Almeroth, “DAMON: A
[5] R. D’Souza, S. Ramanathan, and D. T. Land, “Measuring Performance Distributed Architecture for Monitoring Multi-hop Mobile Networks,”
of Ad-Hoc Networks using Timescales for Information Flow,” in Proc. in Proc. of IEEE International Conference on Sensor and Ad-Hoc
of IEEE International Conference on Computer Communications (IN- Communications and Networks (SECON’04), Santa Clara, CA, USA,
FOCOM’03), San Francisco, CA, USA, Apr. 2003. Oct. 2004.
[6] S. Borgatti and M. Everett, “A Graph-theoric Perspective on Centrality,” [29] M. Younis, P. Munshi, and E. Al-Shaer, “Architecture for Efficient
Social Networks, 2006. Monitoring and Management of Sensor Networks, Workshop on End-to-
[7] P. Bonacich, “Factoring and Weighing Approaches to Status Scores and End Monitoring Techniques and Services,” in Proc. of the 6th IFIP/IEEE
Clique Identification,” Journal of Mathematical Sociology, vol. 2, pp. International Conference on Management of Multimedia Networks and
113–120, 1972. Services (MMNS’03), Belfast, Northern Ireland, UK, Sept. 2003.
[8] P. Bonacich and P. Lloyd, “Eigenvector-like Measures of Centrality for [30] D. Ngo and J. Wu, “WANMON: a Resource Usage Monitoring Tool
Asymmetric Relations,” Social Networks, vol. 23, pp. 191–201, 2001. for Ad-Hoc Wireless Networks,” in Proc. of the 28th Annual IEEE
[9] J. B. MacQueen, “Some Methods for Classification and Analysis of Conference on Local Computer Networks (LCN’03), Bonn, Germany,
Multivariate Observations.” Berkeley, CA: Proc. of the 5-th Berkeley Oct. 2003, pp. 738–745.
Symposium on Mathematical Statistics and Probability, 1967, pp. 281– [31] R. Badonnel, R. State, and O. Festor, “Probabilistic Management of Ad-
297. Hoc Networks,” in Proc. of the 10th IEEE/IFIP Network Operations and
[10] W. Chen, N. Jain, and S. Singh, “ANMP: Ad-Hoc Network Management Management Symposium (NOMS’06), Vancouver, Canada, Apr. 2006.
Protocol,” IEEE Journal on Selected Areas in Communications (JSAC), [32] R. Badonnel, R. State, and O. Festor, “Using Information Theoric
vol. 17, no. 8, pp. 1506–1531, Aug. 1999. Measures for Detecting Faulty Behavior in Ad-Hoc Networks,” LORIA
[11] W. Stallings, SNMP, SNMPv2, SNMPv3, and RMON 1 and 2. Addison - INRIA Lorraine, Tech. Rep., May 2005.
Wesley Longman, 3rd edition, 1999. [33] J.-Y. L. Boudec and M. Vojnovic, “Perfect Simulation and Stationarity of
[12] T. Clausen and P. Jacquet, “Optimized Link State Routing Protocol a Class of Mobility Models,” in Proc. of IEEE International Conference
(OLSR),” Oct. 2003, IETF RFC 3626. on Computer Communications (INFOCOM’05), Miami, FL, USA, Mar.
[13] R. Badonnel, R. State, and O. Festor, “Management of Mobile Ad- 2005.
Hoc Networks: Information Model and Probe-based Architecture,” ACM
International Journal of Network Management (ACM IJNM), vol. 15,
no. 5, Sept. 2005. Remi Badonnel is a research assistant at INRIA, France working on network
[14] D. Perkins and E. McGinnis, Understanding SNMP MIBs. Upper and service management. He is preparing a PhD thesis in Computer Science
Saddle River, NJ, USA: Prentice Hall Inc., 1997. at Henri-Poincare University, Nancy, France. He received a Master Science
[15] “Ns-2 network simulator,” http://www.isi.edu/nsnam/ns/. Degree in Computer Engineering (2003) from ESIAL, Nancy, France. He
[16] “OLSR Extension for Ns-2,” Navy Research Laboratory OLSR Project, is working on designing, validating and implementing management models
http://pf.itd.nrl.navy.mil/projects/olsr/. and architectures for efficiently handling dynamic networks and services. His
[17] F. Bai, N. Sadagopan, and A. Helmy, “Important: a Framework to research interests are directed towards Ad-Hoc Networking, Network Man-
Systematically Analyze the Impact of Mobility on Performance of agement and Self-Configuration (at INRIA), but also Change Management,
Routing Protocols for Ad-Hoc Networks,” in Proc. of IEEE Interna- Service Delivery and Business Process (at IBM Research).
tional Conference on Computer Communications (INFOCOM’03), San
Francisco, CA, USA, Apr. 2003. Radu State is a permanent researcher at INRIA, France working on network
[18] J. Yoon, M. Liu, and B. Noble, “Random Waypoint Considered and security management. He has a Ph.D. degree (2001) from Henri-Poincare
Harmful,” in Proc. of IEEE International Conference on Computer University, Nancy, France and Master of Science in Engineering (1998) from
Communications (INFOCOM’03), San Francisco, CA, USA, Apr. 2003, the Johns Hopkins University, Baltimore, MD, USA. His research interests
pp. 1312–1321. are in the design of management models for highly dynamic environments
[19] D. M. Blough, M. Leoncini, G. Resta, and P. Santi, “The K-Neigh such as ad-hoc networks as well as in the security of the management plane.
Protocol for Symmetric Topology Control in Ad-Hoc Networks,” in He published more than 40 papers on network and service management issues
Proc. of the 4th ACM International Symposium on Mobile Ad-Hoc and serves in the TPC boards of several international conferences and journals,
Networking and Computing (MOBIHOC’03), Annapolis, MD, USA, including IFIP/IEEE IM, IFIP/IEEE DSOM, SAFIR, and A-ICT.
June 2003, pp. 141–152.
[20] O. Dousse and P. Thiran, “Connectivity vs Capacity in Dense Ad-Hoc
Networks,” in Proc. of IEEE International Conference on Computer Olivier Festor is a research director at INRIA. He has a Ph.D. degree (1994)
Communications (INFOCOM’04), Hong Kong, Mar. 2004. and an Habilitation degree (2001) from Henri-Poincare University, Nancy,
[21] A. B. McDonald and T. Znati, “A Mobility Based Framework for France. His research interests are in the design of algorithms and models
Adaptive Clustering in Wireless Ad-Hoc Networks,” IEEE Journal on for automated and scalable management for highly dynamic environments.
Selected Areas in Communications (JSAC), vol. 17, no. 8, pp. 1466– Member of the IRTF NMRG, he has published more than 70 papers in
1487, Aug. 1999. network and service management and serves in the technical program and
[22] C.-C. Shen, C. Jaikaeo, C. Srisathapornphat, and Z. Huang, “The organization committees as well as in the editorial boards of several interna-
GUERILLA Management Architecture for Ad-Hoc Networks,” in Proc. tional conferences and journals. He was the TPC Co-chair of IM2005. He is
of IEEE Military Communications Conference (MILCOM’02), Anaheim, currently leading the EMANICS European Network of Excellence dedicated
CA, USA, Oct. 2002. to Management Solutions for the Internet and Complex Services.

You might also like