Dynamic Local Clustering for Hierarchical Ad Hoc Networks

Satu Elisa Schaeffer Stefano Marinoni

Laboratory for Theoretical Computer Science Laboratory for Theoretical Computer Science
Helsinki University of Technology Helsinki University of Technology
P.O. Box 5400, FI-02015 T KK, Finland P.O. Box 5400, FI-02015 T KK, Finland
[email protected] [email protected]
Mikko Särelä Pekka Nikander
Laboratory for Theoretical Computer Science Nomadiclab
Helsinki University of Technology Ericsson Research
P.O. Box 5400, FI-02015 T KK, Finland Jorvas, Finland
[email protected]

Abstract only switch from one cluster to another in order to mini-

mize intra-cluster routing and maintenance costs. Avoiding
Hierarchical, cluster-based routing greatly reduces rout- cluster changes helps stabilize routing by cluster hops in
ing table sizes compared to host-based routing, while reduc- comparison to routing based on individual links.
ing path efficiency by at most a constant factor [9]. More
importantly, the amount of routing related signalling traffic
is reduced [7, 11, 19]. On the other hand, address changes It is common for many clustering algorithm proposals
caused by nodes changing their cluster produces address that nodes are at most two hops away from the members of
management traffic. In this paper, we present a new lo- their corresponding clusters [2, 4, 5, 6, 12]. Methods differ
cal clustering method that produces dense and stable clus- for example in the usage of cluster heads and possible clus-
ters, thereby minimizing address changes and allowing bet- ter overlaps. Ohta et al. [16] propose a clustering algorithm
ter and more stable network conditions for ad hoc routing. similar to the one presented in this paper, where the clus-
ters are chosen from neighboring ones, bounding the size
of each cluster. Our contribution is in choosing the clusters
based on a particular method for local graph clustering that
helps achieve dense clusters [18]. Our clustering protocol
1 Introduction does not impose explicit constraints on the cluster diameter
and hence the intra-cluster hop counts are not limited. The
When clustering is introduced to an ad hoc routing sys- goal is to produce such a clustering where topology changes
tem, locally computable clustering is a necessity in order to are concentrated inside clusters and changes in inter-cluster
avoid generation of excess control traffic. In the ideal case, connectivity are avoided.
each arriving node is able to determine the appropriate clus-
ter simply by consulting its immediate neighbors, who will
not need to communicate further to determine the best clus- We aim at clusters with high local density and only few
ter. Proposals for and analysis of cluster-based routing in links to the rest of the network desirable as they simplify
dynamic networks include [10, 19]. the routing task. Link state algorithms, such as O LSR [3],
Within a clustered network, routing can be divided into require dense and relatively small networks in order to be
two subproblems: finding a route of clusters to the destina- efficient [17] and perform well for intra-cluster routing with
tion node and finding a route within each cluster either to dense and stable clusters. Inter-cluster routing, on the other
the next cluster or to the destination node within the cluster. hand, may well use on-demand routing protocols that con-
If two previously disconnected clusters become connected struct routes based on cluster hops and gain the advantage
or vice versa, the inter-cluster routing is affected. Desir- of more stable routes, as the clustering hides many route-
ably inter-cluster connectivity changes are rare and nodes breaking topology changes that occur within single clusters.

1
2 Cluster fitness

In this paper, we model ad hoc networks as dynamic

graphs, consisting of nodes and edges (bidirectional links).
The focus is on the clustering protocol. We use a graph-
theoretical fitness measure [18] to locally select the cluster
of an arriving node. We adopt the following notation to de-
fine the fitness measure used: in a graph G = (V, E), a
cluster candidate is a set of nodes C ⊆ V , and the set of
edges of the subgraph induced by C is Ec = {(m, n) ∈
E | m, n ∈ C}. The size of the cluster is the number of Figure 1. Stationary nodes with fixed range
nodes included in the cluster, denoted by |C|. The (local) have been added one by one, with existing
density δ` (C) of a cluster C is |Ec | / |C| nodes updating their clusters (indicated with


2 for clusters with
more than one node and zero otherwise. The density of the colors) after the newcomer selects a cluster.
entire graph δ (G) is simply |E| / |V2 | . Clusters for which Cluster heads have a black border. On the

δ` (C)  δ (G) can be considered good. The relative den- right, a more anomalous network structure.
sity δr (C) [14] is defined in terms of the internal degree
degint (C) = |Ec | and external degree
singleton cluster and becomes its cluster head. This allows
degext (C) = |{(m, n) ∈ E | m ∈ C, n ∈ V \ C}| (1)
the clustering to initialize in a distributed fashion.
of a cluster candidate C as the fraction of the internal degree If replies do arrive within a beacon frame, the node
of the total number of edges incident on the cluster. It is chooses among the neighboring clusters by optimizing the
commonly acknowledged that a good graph cluster should change in cluster fitness, choosing the cluster for which its
have many edges connecting the included nodes to each join would cause the highest increase (or smallest decrease).
other, and as few as possible connecting the cluster to the The node declares its selection by broadcasting a C LUSTER
rest of the graph, and hence, high relative density [8, 14]. J OIN message containing the cluster identifier of the cho-
We want each node to be connected to each member of their sen cluster. Upon the creation of a singleton cluster, the
cluster by at least one path within the cluster, preferably di- node sends a C LUSTER J OIN message containing the clus-
rectly linking to many cluster members, and linking to only ter identifier it chose.
few nodes outside its cluster. The first criterion is fulfilled The C LUSTER R EQUEST and C LUSTER R EPLY mes-
if only connected subgraphs are considered as cluster can- sages are then used periodically to maintain up-to-date
didates. We choose to optimize the product of the relative neighborhood information and to make decisions of leav-
and local densities to achieve clusters that fulfill the other ing and joining clusters. Generally, a node only performs a
two criteria: cluster switch (through a join operation) when it is quality-
2
increasing: a node i executing the cluster-selection protocol
2 degint (C) switches from its current cluster Ci to another cluster C if
f (C) = . (2)
|C| (|C| − 1)(degint (C) + degext (C)) the sum of the cluster fitnesses Ci and C grows as i switches
from Ci to C.
With respect to this measure, a good cluster is both dense
In addition, we impose upper and lower bounds on the
and “introvert”, and the combination avoids counterintuitive
cluster sizes so that nodes primarily choose clusters that are
clusterings produced by optimizing either one of the two
within the bounds. If there are no such neighboring clusters,
density measures alone.
a node prefers clusters below the lower bound, and in their
absence, then will create a new cluster. No node may join a
3 Clustering protocol cluster whose size is at or above the upper bound.
A node stays in the same cluster until it either announces
The clustering algorithm initiates, e.g. after the node first a join to another cluster or the cluster splits. The periodic
wakes up, by probing the neighborhood with a C LUSTER cluster request and subsequent cluster join messages are
R EQUEST message to which all neighboring nodes respond the basic cluster maintenance mechanisms. We utilize the
with a C LUSTER R EPLY. The response message consists cluster-head status of a node in coping with cluster splits
of the node identifier, cluster identifier and three integers: and having the cluster heads periodically broadcast keep-
the number of nodes in the cluster |C|, the internal degree alive messages that are flooded only within the respective
degint (C) of the cluster, and the external degree degext (C) clusters. The lack of a keep-alive message indicates to a
of the cluster. If no responses arrive, the node creates a node that it has disconnected from its cluster head and it
 Diff. in cluster order (%) Diff. in internal degree (%) Diff. in external degree (%)

1000 1000 1000

100 100 100

Average
10 10 10

1 1 1

1000 1000 1000

100 100 100

Variation
10 10 10

1 1 1

Simulation time Simulation time Simulation time

Figure 2. Differences in |C| (left), degint (C) (middle), and degext (C) (right) over a set of N = 10 runs
of D = 250 seconds. Each line relates to a single run. The upper plots show the average differ-
P30 abs(Ep (i)−Vp [C(i)])
ence Avgp = 100 ∗ 1/30 i=1 Vp [C(i)] over time, for p ∈ {size,in,out}, and the lower plots the
corresponding deviation.

must reinitiate the cluster selection protocol. the clustering, as well as the frequency with which updated
information is propagated in the network.
4 Experiments We traced a set of ns-2 runs and computed at each time
step the true values of the above measures and compared
those to the “belief” of each node, calculating the distance
We sketched a small-scale simulator to visualize clus- in percentage of the real value. Formally, in every instant of
terings [20] (examples shown in Figure 1). We also built time, every node i belongs to a precise cluster C(i). This
an ns-2 implementation [13] of the algorithm for larger cluster has its order Vsize [C(i)], its internal degree Vin [C(i)]
scale experiments. Our experiments with simulation tools and its external degree Vout [C(i)]; they are the actual values.
are promising: the clusters achieve a proper sense of lo- Likewise, at every instant of time, each node i, holds its
cality in space and their structure corresponds well to the estimated values respectively as Esize (i), Ein (i), Eout (i).
intuitive global clusterings of the network. Figure 2 shows that the estimate for cluster size does
In the ns-2 simulations, we used networks of 30 nodes not diverge over time and the internal degree often “re-
in a one square-kilometer area. The minimum cluster order stores” the correct value, but the estimates for the external
was set to five and the maximum to eight nodes; the simula- degree remain far from the true value. However, we seem
tor was very slow for larger networks. Each node probed its to achieve a practical clustering even with the problems in
neighborhood, with a range of 250 meters, on five-second determining the external degree. With additional control
intervals and the cluster heads broadcasted a status message overhead, the accuracy could be improved.
for intra-cluster flooding on five-second intervals.
One reason for the problematic estimation of the external
degree is that in cluster splits, there is a risk that the origi-
4.1 Effects of outdated information nal cluster will not notice the departure of some nodes. In
situations where splits are frequent and the departing nodes
Observing the behavior of the clustering method on the will often become completely detached from the old cluster,
simulators, it also seems feasible to approximate the fitness not even remaining in the neighborhood, the cluster heads
function using estimates of |C|, degint (C) and degext (C). should send out time-stamped beacon messages contain-
Such “lazy updates” would allow for a more relaxed control ing the cluster member list that are propagated by broad-
traffic within the cluster, as not all nodes need to be imme- cast within the respective clusters, and the member nodes
diately aware of newcomers, departing nodes, or changes respond (through a broadcast tree formed by the order in
in edges. The effects of outdated information can be de- which the nodes received the beacon message from each
duced from the fitness function (Equation 2); the magnitude other) by stating which of those members are currently their
of the difference between the actual value, and the assump- neighbors and how many other neighbors they have.
tion made at a single node depends on the rate of change in Such a mechanism allows for the entire cluster to main-
 Diff. of local and global density Average cluster order Average cluster fitness Diff. of local and global density Average cluster order Average cluster fitness
1 8 1 1 8 1

0.5 6 0.5 6

Average

Average
0 4 0.5 0 4 0.5

-0.5 2 -0.5 2

-1 0 0 -1 0 0

1 6 0.6 1 6 0.6
0.8 0.8
Variation

Variation
4 0.4 4 0.4
0.6 0.6
0.4 0.4
2 0.2 2 0.2
0.2 0.2
0 0 0 0 0 0
Simulation time Simulation time Simulation time Simulation time Simulation time Simulation time

Group-mobility model Random way-point model

Figure 3. The difference of the average local density and the global density, cluster order, and cluster
fitness averaged over the set of clusters for each time step in {0, 2, 4, . . . , 600} (average over the 10
runs drawn thick).

tain a more up-to-date view on the cluster topology. The 4.2 Cluster quality
cluster head should not send out a new beacon before it
receives the replies to the previous ones; the waiting time We studied the quality of the clusterings produced by a
should be reset upon the arrival of a reply and the com- series of ns-2 simulations, studying cluster density, fit-
putation of the current values should only be done after a ness, and stability as the main indicators. We ran N =
timeout occurs with no further reply arrivals. If however 10 simulations with 30 nodes. The mobility models uti-
the cluster head receives a reply after the timeout, it should lized were reference-point group mobility (GM) model with
increase the waiting time for the next beacon round. A nodes moving in small groups, random direction (RD)
mechanism for reducing the time if all replies arrive quickly model, random walk model (RW), and random way-point
could also be included. Note that by adding a hop counter (RWP) model [1, 15].
to the beacon messages, incremented by each forwarding In all our scenarios the nodes move with speed uniformly
node, nodes can include the value of the counter upon their distributed in [0, 15] m/s after an initial period of [0, 5]
first reception of the message to their replies and thus in- seconds. In GM, each individual node moves as in RWP,
form the cluster head of their “effective” distance from the but within a restricted area of 200m2 surrounding the group
cluster head; this information could also be used to adjust imaginary reference point, while reference points also move
the waiting time at the cluster head. as in RWP, but within the whole simulation area. For RW,
nodes change direction on one-second intervals.
We report averages, and variations of some measured in-
As described above, cluster formation is based on an ex- dicators; formally, we denote the average of a set of k values
change of simple messages that contain the cluster identi- k Pk
{y1 , y2 , . . . , yk } as Avg [yi ]1 = k1 1 yi , and the variation
fier and three integers: the size of the existing cluster, the
as:
internal degree of the cluster, and the external degree of the v
uP  2
u k k
cluster. If a link state routing protocol is used within the i=1 Avg [y i ]1 − y i
k
t
cluster, the nodes can use link state information to produce % [yi ]1 = (3)
k
the current values, and do not need to exchange any extra
messages for intra-cluster information. Using these figures, where, k denotes the cluster count at a certain time step,
together with information about the new or deleted edges, and yi = m(Ci ) is an instantaneous measure (concern-
each node under consideration is able to estimate the clus- ing the cluster Ci ) for a certain metric m. With regard
ter quality for each cluster candidate. to our clustering algorithm, we considered particularly sig-
nificant to monitor the overall conditions (average over all
clusters) in terms of density, order and fitness. Hence, the
For moderately sized clusters (at most 256 nodes) and metrics measured over a period of D = 600 seconds were
64-bit cluster identifiers, all of the required information the following, with a measurement taken for each time step
can be fit into 16 bytes. This could be included in exist- t ∈ {0, 2, 4, . . . , 600}: the local density of the clusters ver-
ing link-layer frames, IP layer address resolution, neighbor- sus the density of the graph (δ` (Ci ) − δ (G)), the cluster
discovery messages, routing messages, or in Wireless L AN order|Ci |, and the cluster fitness f (Ci ). For the density dif-
beacon frames. ference, the range is [−1 , 1] and a positive value indicates
 Table 1. Measures of graph (Equation 4) and cluster stability (Equation 5) for the mobility models
(MM), averaged over N = 10 experiments of duration D = 600 seconds.

MM B̃/D B̃int /D Ẽ/D Ẽint /D T̃ /D T̃int /D Q̃ F̃ S̃ T̃ · S̃

GM 1.53 0.25 1.55 0.21 3.08 0.46 0.03 0.01 0.04 77
RD 1.04 0.43 1.06 0.20 2.10 0.63 0.10 0.08 0.18 227
RW 1.13 0.47 1.15 0.42 2.28 0.90 0.04 0.02 0.07 89
RWP 1.50 0.59 1.51 0.26 3.01 0.86 0.09 0.07 0.16 289

that dense subgraphs have been selected as clusters; if the The results in Table 1 show that group mobility model
value is close to one, almost all links present in the graph and random walk show have the most stable clustering
are internal to some cluster. For the cluster order the range structure of the four, although the reasons differ. Random
is [0,8], its value over time it is a first indicator of cluster walk creates mainly local movements, which means that the
stability as stable clusters must have few fluctuations. The overall topology of the graph will tend to stay the same with
fitness varies in [0, 1] with values close to one indicating small variations. It has as low rate of topology changes
optimal clusters. as random direction, but causes much less changes in the
The results are shown in Figure 3; results for RD, and clustering structure. This is due to the local movements
RW mobility models were similar and omitted. All mobil- of nodes in random walk vs. global movements of nodes
ity models produced clusters with much higher local den- in random direction. Group mobility model creates global
sity than the density of the entire graph. Unexpectedly, movements, but with certain groups of nodes staying close
the group mobility model produced large clusters with very to each other. This causes a high rate of changes in the
high density, whereas group mobility scenario had consis- topology, but low rate of changes in the clustering. The
tently much better fitness than in other mobility models. random-direction model also produces global movements.
Both random way-point and random direction acted simi- We experienced very few clusters splits and changes in
larly, producing small, but dense clusters. general. Overall, the rate of changes in clustering is small.
Group mobility and random walk cause changes in cluster-
4.3 Cluster stability ing in 4% and 6% of the cases where topology changes and
random direction and random way-point models in 16% and
We also studied the stability of the graph and cluster 18% respectively.
topologies (results are shown in Table 1), recording the total
amount of link breakages Bi and new link establishments Ei 5 Conclusions
and the average topology change rate T /D by considering
the graph variations occurred per second in experiment i,
We introduced a new local measure for clustering qual-
i ∈ {1, 2, . . . , N }:
ity, and outlined a simple protocol for local cluster manage-
B̃ = Avg [Bi ]N N ˜
1 , Ẽ = Avg [Ei ]1 , T̃ = B̃ + E. (4) ment. The simulations show that the clustering algorithm is
capable of creating a clustering structure, which hides most
We additionally recorded the number of topology changes
of the topology changes within the network and thus mak-
that were internal to clusters, denoting these by Bint , Eint ,
ing inter-cluster routing task easier. The algorithm is capa-
and Tint , respectively 1 . Cluster stability was measured by
ble of capturing the structure that may exist in the move-
the number of cluster changes, distinguishing between two
ments of nodes. This is especially marked by the group
categories: Qi is the number of quality-increasing cluster
mobility model having the highest rate of topology change,
switches and Fi is the number of switches due to a cluster
while having least changes in clustering both per topology
split;
change and per unit of time. The algorithm also managed to
N N
form clusters with high local density, allowing us to parti-
 
Qi Fi
Q̃ = Avg , F̃ = Avg (5) tion the network into smaller subnetworks which are easily
B i + Ei 1 B i + Ei 1
managed by proactive routing algorithms such as O LSR [3]
We denote S̃ = Q̃+ F̃ ; note that as T̃ is the average number that are designed especially for dense networks with small
of topology changes, S̃ · T̃ is the average amount of cluster diameter.
changes in a single simulation run. As future work we plan to study how the proposed clus-
1 Inter-cluster topology changes could be deduced as a difference be- tering algorithm could be used to further optimize routing
tween T and Tint and address management. On top of a base-layer cluster-
ing, we could form a hierarchy of clusterings with a very [14] M. Mihail, C. Gkantsidis, A. Saberi, and E. Zegura. On the
similar cluster-formation protocol, relying on routing the semantics of Internet topologies. Technical Report G IT -CC-
higher-level cluster requests to the cluster heads. Such a 02-07, College of Computing, Georgia Institute of Technol-
layering would however introduce additional duties to the ogy, Atlanta, GA, U SA, 2002.
[15] J. Nuevo. Mobility generator program for NS-2,
cluster heads, but is an interesting area for further work.
2002. http://externe.inrs-emt.uquebec.ca/
users/nuevo/NSmobgenerator.htm.
Acknowledgments [16] T. Ohta, S. Inoue, and Y. Kakuda. An adaptive multihop
clustering scheme for highly mobile ad hoc networks. In
The first author was supported by the Academy of Finland (un- The Sixth International Symposium on Autonomous Decen-
der grants 202205 and 206235), the Nokia Foundation, and the tralized Systems (ISADS’03), pages 293–300, Apr 2003.
[17] C. A. Santivanez, R. Ramanathan, and I. Stavrakakis. Mak-
Rotary Foundation.
ing link-state routing scale for ad hoc networks. In Proceed-
ings of the Second ACM international symposium on Mobile
References ad hoc networking & computing, pages 22–32, Long Beach,
CA, U SA, 2001. ACM Press.
[1] T. Camp, J. Boleng, and V. A. Davies. A survey of mobility [18] S. E. Schaeffer. Stochastic local clustering for massive
models for ad hoc network research. Wireless Communica- graphs. In T. B. Ho, D. Cheung, and H. Liu, editors, Pro-
tion and Mobile Computing, 2(5):483–502, Sept. 2002. ceedings of the Ninth Pacific-Asia Conference on Knowledge
[2] C.-C. Chiang, H.-K. Wu, W. Liu, and M. Gerla. Routing Discovery and Data Mining (PAKDD-05), volume 3518 of
in clustered multihop, mobile wireless networks with fading Lecture Notes in Computer Science, pages 354–360, Berlin/
channel. In Proceedings Of IEEE SICON’97, pages 197– Heidelberg, Germany, 2005. Springer-Verlag GmbH.
211, Apr. 1997. [19] J. Sucec and I. Marsic. Clustering overhead for hierarchi-
[3] T. H. Clausen and P. Jacquet. Optimized link state routing cal routing in mobile ad hoc networks. In Proceedings
protocol (OLSR). Technical Report R FC 3626, Internet En- of the Twenty-first Annual Joint Conference of the I EEE
gineering Task Force, Reston, VA, U SA, 2003. Computer and Communications Societies, volume 3, pages
[4] A. Ephremides, J. E. Wieselthier, and D. Baker. A design 1698–1706, Los Alamitos, CA, U SA, 2002. I EEE Computer
concept for reliable mobile radio networks with frequency Society Press.
hopping signaling. Proceedings of the IEEE ?, 75:56–73, [20] S. E. Virtanen and P. Nikander. Local clustering for hierar-
Jan. 1987. chical ad hoc networks. In Proceedings of WiOpt’04: Mod-
[5] M. Gerla and J. T.-C. Tsai. Multicluster, mobile multimedia eling and Optimization in Mobile, Ad Hoc and Wireless Net-
radio network. ACM-Baltzer Journal of Wireless Networks, works, pages 404–405, Los Alamitos, CA, U SA, 2004. I EEE
1:255–265, 1995. Computer Society.
[6] T.-C. Hou and T.-J. Tsai. An access-based clustering proto-
col for multihop wireless ad hoc networks. IEEE Journal on
Selected Areas in Communications, 19(7):1201–1210, July
2001.
[7] F. Kamoun and L. Kleinrock. Stochastic performance eval-
uation of hierarchical routing for large networks. Computer
Networks, 3:337–353, Nov. 1979.
[8] J. M. Kleinberg and S. Lawrence. The structure of the web.
Science, 294(5548):1849–1850, Nov. 2001.
[9] L. Kleinrock and F. Kamoun. Hierarchical routing for large
networks: Performance evaluation and optimization. Com-
puter Networks, 1(3):155–174, 1977.
[10] P. Krishna, N. H. Vaidya, M. Chatterjee, and D. K. Prad-
han. A cluster-based approach for routing in dynamic net-
works. ACM S IGCOMM Computer Communication Review,
27(2):49–64, Apr. 1997.
[11] G. S. Lauer. Hierarchical routing design for SURAN. In
Proceedings of the I EEE International Conference on Com-
munications (I CC), pages 93–102, Los Alamitos, CA, U SA,
1986. I EEE Computer Society Press.
[12] C. Lin and M. Gerla. Adaptive clustering for mobile wireless
networks. IEEE Jour. Selected Areas in Communications,
15(7):1265–1275, Sep 1997.
[13] S. McCanne, S. Floyd, K. Fall, and K. Varadhan. The net-
work simulator ns-2. The V INT project, http://www.
isi.edu/nsnam/ns/.