An Analysis of Bluetooth Scatternet Topologies
Rohit Kapoor, M. Y. Medy Sanadidi, Mario Gerla
3803H, Boelter Hall, UCLA
Abstract: Bluetooth technology is intended primarily as a
replacement of cables between electronic devices, as in
Personal Area Networks (PANs), or for connecting the
components of a computer system. In addition, larger
topologies of so-called “scatternets” are targeting wider
geographical area applications in factories, warehouses,
shopping malls and various sensor network applications.
Though some earlier work has looked at scatternet formation
and scheduling issues, less attention has been given to
optimizing scatternet topologies. Sizing a scatternet in terms
of minimizing the number of piconets has been addressed.
We consider in this paper topological design of scatternets,
taking into consideration application traffic requirements. We
study appropriate topologies, and size the network in terms of
piconets. We develop a scatternet queuing model and use it
to compare the delay-throughput characteristics of various
topologies. We find that the best topology is application
dependent. The analytical model is also used to determine the
optimal point to operate a scatternet, i.e., the traffic load that
saturates the network. We validate all our analytical results
by simulation experiments, and find that all analytic results
sufficiently match the simulation results.
1. Introduction
Bluetooth [1] is a low-cost, low-power radio
technology intended as a replacement of cables
between electronic devices. Bluetooth devices can form
small networks called “piconets” and information is
exchanged seamlessly among devices in the piconet.
An example of a piconet application is for connecting a
mobile phone, laptop, palmtop, headset, and other
electronic devices that a person carries around in his
every day life. Such a network is sometimes called a
Personal Area Network (PAN). A piconet may, from
time to time, also include devices that are not carried
along with the user, e.g., an access point for Internet
access or sensors located in a room.
Bluetooth piconets may be inter-connected to form
larger networks called scatternets. This requires some
units, called gateways, to time–division their presence
among the piconets they belong to. The formation of
such “scatternets” enhances the networking capabilities
of Bluetooth and makes it suitable for a wide range of
applications. Also, Bluetooth devices are expected to
be widely diffused in almost every electronic device in
the future. Such a large diffusion is likely to bring
down the cost of Bluetooth devices and enable
Bluetooth-enabled devices to be used in a wide array of
applications in various scatternet architectures.
In general, there will be two distinct modes in
which Bluetooth-based scatternets will be used [2].
One will be in dynamic environments, such as a
conference, where a scatternet will be formed on-the-
fly and will allow Bluetooth-enabled users to share
information such as visiting cards, multimedia files etc.
The second mode of use of scatternets will be in static
environments. In this mode, scatternets will be
configured similar to the way we configure wired
networks. An example of a static mode is a network
connecting household appliances. Another example is a
shopping mall, in which a scatternet covering the
whole mall may direct the client towards the products
he is interested in, or advise him about “special offers.”
Static deployments also incorporate many sensor
network scenarios [13] [14]. For example, a scatternet
of sensor-enabled Bluetooth devices may be deployed
in a warehouse to provide inventory tracking, i.e.,
locating items, reporting on need to restock, etc [3]. A
scatternet can also be used in “monitoring systems”, as
in patient monitoring in a medical services context. Yet
another example of static mode is security applications
using a network of camera-enabled Bluetooth nodes.
Such environments will typically involve nodes
operating on small batteries. The low power and
medium bandwidth of Bluetooth (1 Mbps) make it a
good fit in such environments. Two important
characteristics of such environments are: (a) unlike
typical sensor networks, the required node density is
not very high and (b) medium bandwidth requirements,
higher than typical sensor networks, but lower than
wired networks. It is primarily these kind of static
environments that this paper focuses on.
Scatternet architectures have received attention in
the literature recently. The scheduling of gateways
among piconets has been addressed in [4] [5].
Scatternet formation has been addressed in [6] [7] [12],
but the environments dealt with are typically dynamic
where nodes enter and leave the network throughout its
lifetime. However, gateway scheduling and scatternet
structures studies reported so far do not present any
quantitative analytic results regarding network
performance behavior.
In [7], the authors suggest that it is desirable to
minimize the number of piconets, since a large number
of piconets will incur more collisions. On the other
hand, each added piconet would increase system
capacity. Another issue is the optimal number of
gateways to be used in connecting piconets. In this
paper, we present an analytic approach for designing a
scatternet. We evaluate analytically the impact of the
number of piconets and the number of gateways on the
performance of a scatternet under specific application
traffic patterns. We develop an analytic model using an
open network of queues methodology. The “service
nodes” in the model are the gateways. The external
traffic to the gateways, is assumed to be Poisson; an
assumption that we validate using realistic simulations
of relevant applications. The analytic model is able to
correctly estimate delays in a scatternet given an end-
to-end traffic load. We further show how the model is
used to identify desirable characteristics of scatternets,
and to compare different scatternet structures for a
given traffic pattern. We also determine the maximum
traffic load that may be supported by a scatternet while
providing acceptable delays.
In Section 2, we discuss relevant Bluetooth
technology and present some fundamentals of
scatternet scheduling. In Section 3, we verify that the
arrival process at gateways is Poisson. In Section 4, we
develop the queuing model and in Section 5, we
present performance results for various scatternet
topologies considering two application traffic patterns.
Section 6 presents our concluding remarks and
suggestions for potential future work.
2. Bluetooth
The Bluetooth system [1] operates in the
worldwide-unlicensed 2.4 GHz Industrial–Scientific–
Medical (ISM) frequency band. To make the link
robust to interference, it uses a Frequency Hopping
(FH) technique with 79 carriers. Two or more
Bluetooth units sharing the same channel form a
piconet. Each piconet consists of a master unit and up
to seven active slave units. The master unit polls each
slave unit according to a polling algorithm, and a slave
is only allowed to transmit after the master has polled
it. Bluetooth packets may be one, three or five slots
long.
Two or more piconets can be interconnected to
form a scatternet. This requires a unit, called a
gateway, to belong to more than one piconet. Such a
unit can simultaneously be a slave member of multiple
piconets, but a master in only one, and can transmit and
receive data in only one piconet at a time. Gateway
participation in multiple piconets has to be on a time-
division multiplexing basis. The manner in which the
time of the gateway is divided between its piconets
depends upon the scheduling algorithm adopted by the
masters and the gateways.
The concept of Rendezvous Points (RPs) and
Rendezvous Windows (RWs) has been introduced in
[4]. A rendezvous point (RP) is a time slot at which a
master and a gateway have decided to “meet”, i.e., at
this slot, the master has agreed to address a packet to
the gateway and the gateway has agreed to listen to the
master. The RW determines the number of slots for
which a master and a gateway exchange data, starting
from their RP. By defining RPs and RWs
appropriately, it is possible to define a scheduling
scheme in which the gateway gets a certain fraction of
bandwidth in each of its piconets. We further specify
the scheduling assumed in our models in Section 5.
2.1 Bluetooth Simulation Model
The Bluetooth model used for the simulation
experiments in this paper is built on NS-2 [8]. The
model implements the standard features of Bluetooth
like Frequency Hopping, Multi-Slot Packets, and Fast
ARQ (Automatic Retransmission Query). In addition,
it also has support for defining scatternets, and gateway
scheduling algorithms.
3. The Poisson Assumption
In this section, we model some possible
applications of Bluetooth scatternets and go on to show
the existence of the Poisson property in the traffic
characteristics of scatternets for these applications. The
first category of applications is that of “monitoring
systems”. Monitoring applications include a variety of
sensor network scenarios in security, medical services,
factory control, warehouse management, etc. In such
applications, nodes typically send data to a central
node. The second category of applications includes, for
instance, users meeting for an ad hoc conference. In
such applications, any user will, over a period of time,
send a more-or-less equal amount of data to any other
user. Using simulation experiments, we verify that, in
both these kind of applications, the arrival process to
gateways exhibits the Poisson property.
A gateway is an aggregation point of data from
different piconets, each of which is in turn aggregating
traffic from a number of slaves. Thus, we believe that
the resulting arrival process to gateways is likely to be
Poisson, where dependencies that may have existed in
traffic from a single slave are weakened or eliminated.
The “monitoring system” applications are modeled
in the following manner: At random instants of time
(e.g. when an unusual state is detected), nodes send out
packets regarding the state of the machine they are
monitoring to a central node in the network. The
packets sent are assumed to be of random length since
they are typically generated by different types of
sensors. Packet sizes and transmission instants are
generated from uniform distributions.
The second category of applications is modeled in a
similar manner to the monitoring application except
that nodes send out packets not to a central node, but
randomly to any node in the network. Though these
assumptions are a little simplistic, they help in keeping
the analytical model (that we develop later) simple.
Also, as shown later, simulation results match closely
with the results obtained from the model, thus gateway nodes only, and not including other node types
validating it. in the model.
We simulated the above applications over several 1
topologies varying the number of masters, slaves and 0.9 Gatew ay 1
gateways. We assume here and in the rest of the paper 0.8 Gatew ay 2
Gatew ay 3
that gateways may only be slaves, and are not used as 0.7

Fraction of Pkts
masters simultaneously. This is a simplifying 0.6
assumption, and we intend to extend the model in the 0.5
future to study the alternative of deploying gateway 0.4
and master functionality in the same node. The 0.3

simulated configurations were generated as follows: 0.2

0.1
• The number of piconets was varied from 5 to 25, 0
with a step size of 4. 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Pkt inter-arrival tim e (sec)
• The number of slaves in each piconet was
randomly chosen from a uniform distribution, and Fig 1(a): Inter-arrival times of packets at gateway
the “mean” number of slaves was varied from 2 to queues
4.
1
• The mean number of gateways per piconet was 0.9 Gatew ay 3
varied from 2 to 4 and each gateway belongs to 2 0.8 Negative Exponential
piconets. 0.7

All data is sent as 1-slot Bluetooth packets, and a Fraction of Pkts

0.6

gateway can belong to a maximum of two piconets. 0.5

The Poisson process [9] simplifies much of the
analysis required in queuing systems. We tested the
packet arrival process at gateways and all experiments
showed some Poisson behavior as discussed below.
In Fig. 1(a), we show the distribution of the inter-
arrival times of packets at three gateways in a
configuration with 13 piconets, mean number of slaves
= 4, and mean number of gateways = 2, for the
monitoring application. Fig 1(b) shows a negative
exponential distribution that matches the distribution of
one of the gateways of Fig 1(a).
We see that the inter-arrival time of packets at
gateways follows a negative exponential distribution,
which means that the arrival process is Poisson. We
also performed Chi-Square tests for the arrival process
and these showed a good matching (ranging from 50%
to 80%) between the simulation and theoretical results.
We observed the same behaviour for gateways in all
experiments for both the applications. We do not show
the results for the second category of applications;
these also exhibit the Poisson property for arrival
process at gateways.
Another observation regarding gateways in our
applications is that they carry much more traffic than
the non-gateway slaves. Also, the traffic carried by
gateways is of the same order as that carried by
masters, but the lower service rate of gateways makes
them bottlenecks (service rate of a gateway will
roughly be less than that of the master by a fraction
proportional to the number of gateways in a piconet).
We use these observations to model the entire
scatternet as a network of queues representing the
0.4
0.3
0.2
0.1
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Pkt inter-arrival tim e (sec)
Fig 1(b): Negative Exponential distribution matching
the distribution at a gateway
We remark that applications similar in nature to
these two categories of applications will exhibit
Poisson arrivals to the gateways. Further, the gateways
will remain the bottleneck nodes in the network. In
other environments, for example, where a large
proportion of traffic is internal to a piconet, either of
the above two properties may not apply.
4. Queuing Analysis
In this section, we present an analytical model for
studying delay-throughput properties in Bluetooth
networks. We first present a queuing model for a
gateway and then perform the analysis by treating a
Bluetooth network as a network of queues, where each
queue represents a gateway.
We verified in the previous section that in certain
applications that we modeled, gateways are the
bottleneck nodes. Thus, we focus on the gateways and
model them as service centers in a network of queues,
with “external traffic” fed into the network by masters
and non-gateway slaves. As an example, consider the
Bluetooth network and its equivalent model shown in
Fig. 2. Note that the piconet in Fig 2 (b) only serves as
a source of traffic, while the gateway becomes a
service center.
Master Traffic Source
Non-gateway slave (Piconet)
Gateway slave Gateway
Piconet slave
(a) (b)
Fig 2: A Bluetooth network (a) and the corresponding
network of service centers model (b).
4.1 Queueing Model of a Gateway
To develop the queueing model for a gateway node,
we make use of the following three observations
regarding the behaviour of a gateway:
(i) The service rate of a gateway depends upon the
polling it gets in each piconet it belongs to. Thus, if a
gateway gets a fraction x of the polling in a piconet, its
service rate for the piconet is (BW*x), where BW is the
Bluetooth bandwidth. Also, if a gateway belonging to
two piconets gets the same fraction of polling x in each
piconet, it may be seen as a link between the two
piconets with a bandwidth equal to (BW*x). In other
words, it will behave as a service center with a service
rate of (BW*x). (We assume here and in the rest of the
paper that a gateway can belong to a maximum of 2
piconets; our analysis does not address the behavior of
the gateway when it belongs to more than 2 piconets).
(ii) The gateway can be modeled as a shared
processor.
The gateway shares its time between each of its
piconets (as determined by the scheduling policy). The
gateway may be viewed as a processor servicing each
of its piconets in a Round-Robin manner. Ideally, in a
processor-sharing system, the quantum of service given
to each customer tends to zero. The gateway may be
seen as a practical processor sharing system, where the
quantum of service given to each customer is finite and
customers are given the “full capacity” of the processor
on a “part-time” basis. This assumption is validated by
the close matching of the analytical and simulation
results in Section 5.
(iii) The gateway behaves as a service center with
multiple servers.
When the gateway “visits” a piconet, let the time it
spends in each piconet be T Bluetooth slots. A
maximum of T Bluetooth packets may be
accommodated in these T slots (we assume for the sake
of convenience that Bluetooth packets are 1 slot in
length; the analysis would be very similar if the packets
were of 3 or 5 slots).
Thus, the number of packets that the gateway
transfers each time it visits a piconet depends upon the
value of T and the length of the gateway queue. If the
gateway queue has only one packet, this would fill up
one out of the T slots and the remaining T-1 slots
would be wasted. If there were 2 packets in the
gateway queue, there would be T-2 wasted slots and so
on. This system, thus, behaves like a multiple server
system, where each server can serve a single customer
(a 1-slot Bluetooth packet) in T time slots, and the
number of servers is T. The service rate ’ for a single
server in the service center is thus given by:
’ = (BW*x)/T (1)
Combining the above three observations, we model
the gateway as a “processor-sharing multiple-server
service center”, where each server has a service rate as
given in Eq. 1. The service rate for the gateway, , is
thus given by:
= (BW*x*k)/T 0<=k<=T
= (BW*x) T<k (2)
where k is the number of packets in the queue
of the gateway.
4.2. Analytical Model of a Bluetooth scatternet
We now develop an analytical model for a
Bluetooth scatternet, treating the set of gateways as a
network of queues. The product-form solution for an
open Jacksonian network of queues is given by:
p(k1, k2, ... , kn) = p1(k1) p2(k2) ... pn(kn) (3)
where ki is the number of customers at node i,
pi(ki) is the marginal distribution of finding ki
customers at node i , and n is the number of nodes.
Moreover, of the four types of service centers for
which the product form applies [10], the gateway is a
node of Type II: it is a shared processor and its service
process has a rational Laplace transform. The extension
of Type II service centers to include multiple servers
(as is the case with the gateway) is given in [11]. The
equilibrium probabilities at a node are given by:
n
p(n) = p(0) (max (T, j) ) ( / )n(4)
j=1
n!
where is the arrival rate at the node, is the service
rate of a single server of the node and T is the number
of servers.
 We now apply Eq. 4 to a scatternet and derive a
relation between average delay and throughput. Let M
be the number of links (i.e., gateways) in the network.
Let be the total traffic to the network and i be the
traffic rate at link i. Using Eq. 4, we get:
p(k) = (p(0) * ( / )k) 0< = k <= T
k!
(5)
p(k) = TT p(0) ( / T)k T<k
T!
Using Eq. 5 (and after some algebra), we get:
T ‘j’, the number of piconets each gateway belongs to. In
p(0) = [ ( / T)T ( / ( T – ) ) TT + ( / )k ] -1 (6)
k=0
T! k!
The average number of packets on link i, Ni is
given by:
Ni = (k * p(k))
T-1
Ni = p(0) [( / ) + ( / ) ( / )k] +
k=1
k!
p(0) TT [T ( / T)T ( /( T - )) + ( / T)T-1 ( /( T- ))2] (7)
T!
From [10], denoting the average delay by Tavg, we
know that:
M M
Tavg = ( i/ * Ti) = (Ni/ ) (8)
i=0 i=0
Equation (8) gives the relation between the
throughput and the average delay in a scatternet. Note
that even though Ni and hence, Tavg does not have a
closed form, it is easy to calculate since the value of T
(and hence, the number of elements in the summation)
is known.
5. Results
In this section, we apply the analytical model in two
typical applications, each having a different traffic
pattern. Our aim in considering both these traffic
patterns is to evaluate how different scatternet
topologies (different number of slaves and gateways
per piconet) affect the performance of the application,
determine the delay-throughput characteristics for a
particular topology, identify topologies which produce
better delay-throughput characteristics, etc.
5.1 Uniform Traffic Pattern
We first consider an application of Bluetooth
scatternets where any node in the network may have an
equal tendency to communicate with any other node.
Such an application may be typical of an ad-hoc
network of users meeting for a conference. In such an
environment, any user will, over a period of time, send
a more-or-less equal amount of data to any other user
and the traffic in the network is thus, uniform in nature.
It behooves us, given a uniform traffic pattern, to
consider topologies that are symmetric. Such
topologies could be created by making all piconets
have the same number of slaves and gateways and
connecting the gateways appropriately.
We analyze various such symmetric topologies
keeping the total number of nodes fixed in all
topologies. We consider a fixed number of nodes, N,
which may be connected using piconets with different
values of ‘i’, the number of gateways per piconet and
other words, one can assume that there are a certain
number of nodes that need to be connected into a
network to serve an application. The analysis provides
an approach to determine how these nodes may be
connected so as to give best delay-throughput
characteristics.
In some topologies, the number of nodes may not
be exactly equal to the value of N, since a certain value
of i and j may not allow all values of N. In such cases,
we assume that the network has the closest number of
nodes, smaller than N, that the values of i and j allow.
To accommodate this in the analysis, we adjust the
arrival rate and the service rate of a gateway
accordingly, as is explained later in this section.
A scheduling policy, will, in general, give a certain
fraction of bandwidth to each slave (non-gateway or
gateway) in a piconet. Due to the uniform traffic
pattern, all non-gateway slaves send the same amount
of data. Thus, our scheduling policy gives each non-
gateway slave the same amount of bandwidth. Also,
due to the symmetric topologies, each gateway is also
given the same amount of bandwidth. Since a gateway
is expected to carry more traffic than non-gateway
slaves, it requires more bandwidth than non-gateway
slaves. We assume that the gateway is given a polling
priority of ’p’ with respect to the non-gateway slaves,
which means that it gets ‘p’ times the bandwidth
allocated to non-gateway slaves. Thus, if there are ‘x’
non-gateway and ‘y’ gateway slaves in a piconet, the
fraction of polling given to a gateway is (1/(x/p) + y).
Thus, the bandwidth (service rate) given to a gateway
in a piconet, is given by (from Eq 2):
= (BW * x) = (BW/(x/p) + y) (9)
where BW is the Bluetooth bandwidth. We obtain
analytic results for various values of p, and hence
different values of . If the actual number of nodes is
less than the value of N, then is multiplied by a factor
m/N, where m is the actual number of nodes. Since the
number of nodes has to be made equal to N (for the
topology to have the same number of nodes as other
 i=5, j=2 i=4, j=2
Master
Non-gateway slave
Gateway
i=4, j=3
Fig 3: Different topologies formed by varying values of i and j when N = 28
topologies), the number of slaves would be multiplied
by N/m. This would change by a factor of m/N.
As an example, consider the topology with i = 5
and j = 2. Other topologies may be analyzed in a
similar manner. For (i = 5, j = 2, N = 28) there are 4
gateway links. Suppose each node sends data at a rate
of x packets/second to each of the other nodes. We
need to calculate the delay at one such gateway link
since the delay at all gateway links will be the same, by
symmetry. The two quantities that need to be
calculated are, , the total traffic in the network and ,
the traffic on each gateway link.
Calculation of : Each non-gateway slave sends data to
i - 1 other slaves and one master in its piconet. Note
that the data sent by a slave to any other non-gateway
slave (or master) in the same piconet does not
contribute to traffic in our analytical network (since it
does not travel on a gateway link). The data sent by a
slave to a gateway in the same piconet contributes to
the traffic since it goes on a gateway link. Thus, the
external traffic due to each non-gateway slave is to N–i
other nodes outside its piconet. Similarly, each master
sends data to i slaves in its piconet and N-i other nodes
outside its piconet. The data sent by a gateway to all
other nodes (in the same piconet and outside; there are
N-1 such nodes) contributes to traffic on the gateway
link.
Thus, = (N-G) (N-i) + G (N-1), where N is the
number of nodes and G is the number of gateway links.
In the case of the i = 5 and j=2 topology, this value is:
= (28-4) (28-5) x + 4*28x = 664x.
Calculation of : We assume that each data packet
travels along the shortest path. If there are two or more
shortest paths, traffic is equally divided between them.
It is easy to calculate the total traffic on a gateway link
as the traffic generated by all non-gateway slaves,
i=2, j=2
i=2, j=3 i=1, j=3
masters and gateway slaves, which involves the
gateway in the route. In the case of the i = 5, j=2
topology, this value may be calculated as:
= (Traffic from masters and non-gateway slaves
of Piconet i going through the gateway) + (Traffic
from gateway i going through the gateway)
= (6 * 11 + 6 * 11 + 6 * 4 + 6 * 4) x + (27 + 7.5 + 7.5 +
1) x = 180x + 43x = 223x.
If the actual number of nodes is less than the value
of N (as will be the case if the values of i and j do not
allow the value of N), then is multiplied by a factor
of N/m, where m is the actual number of nodes. Since
the number of nodes has to be made equal to N (for the
topology to have the same number of nodes as other
topologies), the number of nodes would be multiplied
by N/m. We, thus, need to multiply by N/m.
and may be calculated in a similar manner for each
of the topologies.
5.1.1 Validation with simulation experiments
Fig 4 shows the analytical and simulation results for
the delay-throughput characteristics of two topologies.
The traffic load is defined as the value of x, i.e., the
rate at which data is sent by each node to every other
node. We see that there is a very close matching
between the analytical and experimental results. We
observed such a close matching for all the topologies,
which validates the analytical model.
5.1.2 Comparison of Topologies
Figs 5 (a), (b) and (c) show the delay-throughput
characteristics for different values of i when j = 2, 3
and 4 respectively, with p = 1 (p is priority of
gateway). It can be seen from the figure that for the
type of applications considered, networks with a
smaller number of slaves (and hence larger number of
piconets) show smaller average delays. Moreover, the
breakdown point, i.e., the knee of the curve, for
networks with smaller number of slaves is larger than
that for networks with larger number of slaves. These
networks are thus, also more stable. From an intuitive
network may accept before delay deteriorates
precipitously, i.e. the optimal point to operate a
network [11] (which is the knee of the curve, or the
point where the tangent to the curve meets it).
9 i=5, j=2
8 i=4, j=2

Average Pkt Delay (sec)

i=2, j=2
point of view, a smaller number of slaves means a 7 i=1,
i=0,
j=2
j=2
larger bandwidth for the gateway and hence, a larger 6
5
service rate and lower delays on the gateway link.
4
3
9 i=5, j=2 2
i=1, j=3
8 1
Experimental i=5, j=2
Average Pkt Delay (sec)

7 Experimental i=1, j=3 0

0 2 4 6 8 10
6
Traffic Load
5
4 (a)
3 10
i=4, j=3
2 9 i=2, j=3

Average Pkt Delay (sec)

8 i=1, j=3
1 i=0, j=3
7
0 6
0 2 4 6 8 10 5
Traffic Load 4
3
Fig 4: Analytic and Simulation results for two 2
topologies 1
0
0 2 4 6 8 10 12
Moreover, as the number of gateways is increased Traffic Load
from 2 to 3, the delays get smaller and the breakdown
point becomes larger. On the one hand, the gateway (b)
bandwidth decreases when j is increased from 2 to 3, 10
i=2, j=4
but on the other hand, there is reduced traffic on each 9
i=1, j=4
Average Pkt Delay (sec)

8
gateway link due to the following two reasons: 7 i=0, j=4
• In topologies with j=3, the average number of hops 6
between two nodes is reduced due to a larger 5
number of gateways (more connectivity). This 4
3
means that the total traffic on the network is less as 2
compared to topologies with j=2. 1
• Topologies with j=3 have a larger number of 0
gateways than those with j=2. This means that 0 2 4 6 8 10
Traffic Load
there are more gateways to carry the traffic, which
means lower traffic on each gateway. (c)
Fig 5 (a), (b) and (c): Delay-throughput characteristics
These two reasons together more than make up for for different values of i when j = 2, 3 and 4
the decreased gateway bandwidth and hence, the
system performance is better. 5.1.3 Effect of varying the gateway priority
For topologies with j=4, on the other hand, the
effect of the above two reasons is not enough to We now show the effect of giving different
counter the reduced bandwidth when compared with priorities to gateways, i.e., changing the value of p in
topologies with j=3. These topologies thus, show lower Eq. 9. Fig 6 (a) and (b) show the characteristics when
performance than those with j=3. In fact, we observed p=2 and p = 5 respectively.
that topologies with j > 4 show a worse behavior than The results show that a higher priority for the
those with j = 4. Thus, we see that there is an optimum gateways tends to decrease the delays. As earlier, a
number of gateways for the application. Another smaller number of slaves still gives lower delays and
important point is that the above analysis also enables higher breakdown points. In fact, the priority can only
the network designer to determine the maximum load a be increased to a certain point; if the priority is
increased beyond a value, the slave may become a 5.1.4 Variation in service rate of gateway due to
bottleneck and the gateway may cease to be a interference
bottleneck. This will violate the basic premise of the
model, which assumes that the gateways are the In the gateway model that we have considered so
bottlenecks in the system. far, we have not taken into account the interference
caused by different piconets. In fact, the service rate for
10 i=5, j=2
the gateway link will be slightly reduced due to this
9 i=4,
i=2,
j=2
j=2
interference. Moreover, since different topologies will
8 have different number of piconets, the reduction in the
Average Pkt Delay (sec)

i=1, j=2
7 i=0, j=2
value of service rate due to interference will not be the
6
5
same for different topologies. We now incorporate the
4 effects of this interference in the calculation of the
3 service rate.
2 We assume a very simple collision model. For a
1 piconet, the probability that a packet collides in
0
0 2 4 6 8 10
frequency when there are n other piconets present
Traffic Load within range = (n-1)/79. Moreover, each lost (collided)
packet leads to a loss of two packets since the return
(a)
packet carries the acknowledgement. Thus, the new
10 service rate, coll for a gateway is given by:
9
coll = [1 – 2(n-1)/79] (10)
Average Pkt Delay (sec)

i=5, j=2
8 i=4, j=2
7 i=2, j=2
6 i=1, j=2 Now, the number of piconets within range of a
i=0, j=2
5 piconet depends upon the number of piconets in the
4 topology and the density of nodes. If the density of
3
2
nodes is d, then on the average, 100d nodes will be
1 present in the range of a piconet (since range of
0 Bluetooth is 10 m). When the number of nodes is 28
0 2 4 6 8 10 (as in our example case), a density of 0.28 means that
Traffic Load
all nodes (and hence, all piconets) are within range of
(b) each other.
Fig 6 (a) and (b): Characteristics for p =2 and p=5 In Figs 7 (a) and (b), we plot the delay-throughput
characteristics for various topologies with coll for d =
0.28. We see the increase in delays and the smaller
Breakdown Breakdown Breakdown ‘knee-point’ of the curves compared to the cases when
point = 6sec point = 8sec point = 10sec interference was not taken into account. Our model is
i = 5, j = 2 4.5 14.5 - now able to determine delays with interference
i = 2, j = 2 1.8 6 - between piconets also taken into account. In fact,
i = 1, j = 2 0.9 2.9 - different densities of nodes may also be easily taken
i = 4, j = 3 1.5 3.1 9 into account now.
i = 2, j = 3 0.75 1.55 4.5 We validated the results of this first application by
i = 1, j = 3 0.37 0.77 2.3 simulations and found a close matching. We also
Table 1: Gateway priorities needed to achieve certain repeated the experiments for various value of N, from
breakdown points for different topologies N=25 to N=60 and found similar behavior, i.e.,
topologies with a smaller number of slaves and j=3
Another observation is that the breakdown point show smaller delays. These results, are thus, quite
increases with an increase in priority of the gateway. general, for this application.
Table 1 shows the priority required for various
topologies to achieve a certain breakdown point when 5.2 Centralized Traffic Pattern
N =28. The entries with a ‘-’ imply that this breakdown
point cannot be achieved for the topology. Thus, We now consider a second application of Bluetooth
topologies with j=2 cannot achieve a breakdown point scatternets where each node in the network wishes to
of 10sec, whereas topologies with j=3 can. Also, it can communicate with one central server. Traffic rates
be seen that a smaller value of i requires a lower from all nodes are assumed equal. Such a traffic pattern
priority to achieve a certain breakdown point. may be typical of a factory or a warehouse where these
nodes are used in tasks like inventory tracking. Again,
we would like to see how different scatternet server. Gi and Gi are the traffic rate and service rate
topologies affect the delay-throughput characteristics in respectively of gateway Gi.
such an application.
G3 Type III piconet
10 i=5, j=2
i=4, j=2
9 i=2, j=2 G2 Type II piconet
8 i=1, j=2
Average Pkt Delay (sec)

i=0, j=2
7
6 G1
5
4
3
2
1
0 Stage 1
0 2 4 6 8
Traffic Load Master
(a) Gateway
Stage 2
Stage 3 Piconet
9
i=4, j=3
8 i=2, j=3 Central piconet
Average Pkt Delay (sec)

7 i=1, j=3
i=0, j=3 Fig 8: Central server topology with i=0 and j=2
6
5
4 G3 needs to carry traffic (to the central server) for
3 one master node and itself. Thus, G3 needs twice the
2 bandwidth as a non-gateway slave node. If a node
1 sends traffic at the rate of x, then BWG3 = 2x.
0 Now, G2 needs to carry traffic for two gateways of
0 2 4 6 8 10 the G3 category, one master node and itself. Thus,
Traffic Load
BWG2 = (2*BWG3) + 2 = 6x. Similarly, BWG1 = 14x.
(b) There are two G1s at the central piconet and each
Fig 7 (a) and (b): Characteristics taking interference carries traffic equal to 14x. Thus, each gets 14/28 of
into account for d=0.28 the central piconet’s bandwidth (the central master’s
polling). Thus G1 = 14x; and G1 = (14/28) * BW.
Since each node sends an almost equal amount of Now, at piconet of type II, G1 gets 14/28 of the
data to the server, it needs to be given the same polling. Thus, bandwidth given to each gateway, G2 =
bandwidth to the central server. (1 – 14/28)/2 = 7/28. Each gateway G2 carries 6x of
We first consider a topology with (i=0, j=2, and N the traffic => G2 = 6x; G2 = (7/28) * BW.
= 29), as shown in Fig 8. We assume, as earlier, that a At piconet of type III, G2 gets 7/28 of the polling.
gateway belongs only to two piconets. G1, G2 and G3 Thus, bandwidth given to each gateway G3 = (1-
represent gateway categories that may have different 7/28)/2 = 10.5/28. Each gateway G3 carries 2x of the
traffic and service rates. As shown in the figure, the traffic. => G3 = 2x; G3 = (10.5/28) * BW.
topology consists of various stages. Note that the Also, total traffic rate in the network is = 28x
topology is symmetric at each level (i.e., at each level, (each node sends x to the master of the central piconet).
piconets have the same number of slaves and Figure 9 shows a topology with i=0 and j=3. Recall
gateways). If a certain topology (number of slaves and that we compare topologies keeping the number of
gateways per piconet) shows best delay-throughput nodes in the scatternet fixed. As earlier, in some
characteristics, then applying the same topology at topologies, the number nodes may not be exactly equal
each stage gives a topology that is better than any to N.
other. In fact, if any of the stages in the scatternet had a
different topology, then it could be improved by
replacing its topology with the best one. We, thus, have
the same topology at each level.
We now apply the analysis of Section 3 to the i=0,
j=2 topology. BW refers to the Bluetooth bandwidth,
while BWGi is the bandwidth required by a gateway so
that each node gets the same bandwidth to the central Fig 9: Central server topology with i=0 and j=3
 Fig 10 shows the delay vs. traffic load curves for experiments, which show a very close matching with
different values of i (while keeping j fixed), with N= the analytical results.
29. It can be seen that for the type of applications
considered, networks with a larger number of slaves 3
per piconet (and hence a smaller number of piconets)
2.5 i=0, j=2
show smaller average delays. In this application,

Average Pkt Delay (sec)

i=0, j=3
gateways that serve a small number of slaves (as those 2
i=0, j=4
in networks with smaller number of slaves) have a high
1.5
capacity but a small amount of data to transfer. This
means that a large part of the gateway capacity is 1
wasted and there is no gain from the large capacity of
0.5
the gateway. On the other hand, a small number of
slaves leads to a larger number of average hops 0
traveled by a packet. This results in higher delays. 0 50 100 150 200
Traffic Load

3 (a)
2.5
2.5
Average Pkt Delay (sec)

i=0, j=2

Average Pkt Delay (sec)

2 2 i=3, j=2
i=2, j=2
i=4, j=2
i=3, j=3
1.5 1.5

1
0.5
0
0 50 100
Traffic Load
Fig 10: Characteristics for topologies with different
values of i and j=2
The breakdown points, or the knees of the curves,
for networks with different numbers of slaves are
almost the same. The breakdown point is determined
by the ratio of ( / ) at the G1 gateways (since these are
the bottleneck gateways). If there are n G1 gateways,
then each G1 gateway gets 1/n of the piconet
bandwidth and has to serve 1/n of the traffic. Thus,
( / ) at the G1 gateways is almost the same for all
topologies.
Figs 11 (a) and (b) show the effect of varying j on
the delay-throughput profile. Networks with a smaller
number of gateways show smaller delays and higher
breakdown points. A smaller number of gateways leads
to a higher capacity per gateway. The effect, though, is
very marginal, since giving gateways a large capacity
may not be very beneficial, as explained above. Thus,
increasing the number of gateways does not have too
much of an effect, though, a smaller number of
gateways shows a better behavior.
We conducted the same tests for different number
of nodes, varying from N=25 to N=60. The results
were quite similar to the results for N=29. We can thus,
conclude, that for applications with centralized traffic,
the results shown for N=29 are valid, in general. We
also validated the above results with simulation
1
0.5
150 200 0
0 50 100 150 200
Traffic Load
(b)
Fig 11 (a) and (b): Characteristics for different values
of j
6. Conclusions and Future Work
In this paper, we addressed several issues relating
to topology of scatternets. We presented an analytic
approach to determining the number of non-gateway
and gateway slaves per piconet to ensure acceptable
delay characteristics. We developed a queuing model
for a scatternet and validated the model via simulation
experiments. We studied two typical applications of
Bluetooth scatternets with different traffic patterns. In
each application, we used analytic results to compare
the delay-throughput characteristics of various
topologies and determine the optimal point to operate
the network.
We intend to generalize, in the near future, our
analytic model to study topologies where a gateway is
shared among an arbitrary number of piconets. We also
intend to include the effects of various scheduling
schemes. For example, when the traffic is non-uniform,
a dynamic scheduling scheme could be used to modify
the bandwidth fraction allocated to the gateways based
on traffic patterns. More sophisticated models of
“processor sharing” and the effect of interference on a
gateway are also planned.

References

1) J. Haartsen, "BLUETOOTH - the universal radio

interface for ad hoc wireless connectivity", Ericsson
Review, n.3, 1998, pp. 110-117.
2) M. Gerla, Y. Lee, R. Kapoor, T. Kwon, A. Zanella
“UMTS-TDD: A Solution for Internetworking Bluetooth
Piconets in Indoor Environments”, ISCC 2002.
3) D. Estrin, R. Govindan, J. Heidemann, S. Kumar, “Next
century challenges: scalable coordination in sensor
networks”, Mobicom 1999.
4) P. Johansson, M. Kazantzidis, R. Kapoor, M. Gerla,
“Bluetooth: An Enabler for Personal Area
Networking”, IEEE Network, September 2001, Vol.15
No.5.
5) A. Rácz, G. Miklós, F. Kubinszky, A. Valkó,
“Bluetooth: A pseudo random coordinated scheduling
algorithm for Bluetooth scatternets”, Mobihoc 2001.
6) T. Salonidis, P. Bhagwat, L. Tassiulas, R. LaMaire,
“Distributed topology construction of Bluetooth
personal area networks”, Infocom 2001.
7) C. Law, A. Mehta, K. Siu, “Bluetooth: Performance of
a new Bluetooth scatternet formation protocol”,
Mobihoc 2001.
8) “Network Simulator (NS-2)“, www-
mash.cs.berkeley.edu/ns/
9) L. Kleinrock, “Queueing Systems, Volume I: Theory”.
10) L. Kleinrock, “Queueing Systems, Volume II: Computer
Applications”.
11) F. Baskett, K.M. Chandy, R.R. Muntz, F.G. Palacios,
“Open, Close and Mixed Networks of Queues with
Different Classes of Customers”.
12) G. Zaruba, S. Basagni, I. Chlamtac, “Bluetrees-
scatternet formation to enable Bluetooth-based ad hoc
networks”, ICC 2001.
13) “CrossBow Technology”, www.xbow.com
14) “Oceana Sensor”, http://www.oceanasensor.com/
Bluetooth/sensing-systems.htm