Cooperative load balancing and dynamic

Cooperative Load Balancing and Dynamic
Channel Allocation for Cluster-Based
Mobile Ad Hoc Networks
Bora Karaoglu, Member, IEEE and Wendi Heinzelman, Senior Member, IEEE
Abstract—Mobile ad hoc networks (MANETs) are becoming increasingly common, and typical network loads considered for MANETs
are increasing as applications evolve. This, in turn, increases the importance of bandwidth efficiency while maintaining tight
requirements on energy consumption, delay and jitter. Coordinated channel access protocols have been shown to be well suited for
highly loaded MANETs under uniform load distributions. However, these protocols are in general not as well suited for non-uniform load
distributions as uncoordinated channel access protocols due to the lack of on-demand dynamic channel allocation mechanisms that
exist in infrastructure based coordinated protocols. In this paper, we present a lightweight dynamic channel allocation mechanism and
a cooperative load balancing strategy that are applicable to cluster based MANETs to address this problem. We present protocols that
utilize these mechanisms to improve performance in terms of throughput, energy consumption and inter-packet delay variation (IPDV).
Through extensive simulations we show that both dynamic channel allocation and cooperative load balancing improve the bandwidth
efficiency under non-uniform load distributions compared to protocols that do not use these mechanisms as well as compared to the
IEEE 802.15.4 protocol with GTS mechanism and the IEEE 802.11 uncoordinated protocol.
Index Terms—Mobile ad hoc networks, bandwidth efficiency, distributed dynamic channel allocation
Ç
1 INTRODUCTION
MOBILE ad hoc networks (MANETs) have been an
important class of networks, providing communica-
tion support in mission critical scenarios including battle-
field and tactical missions, search and rescue operations,
and disaster relief operations. Group communications has
been essential for many applications in MANETs. The
typical number of users of MANETs have continuously
increased, and the applications supported by these net-
works have become increasingly resource intensive. This, in
turn, has increased the importance of bandwidth efficiency
in MANETs. It is crucial for the medium access control
(MAC) protocol of a MANET not only to adapt to the
dynamic environment but also to efficiently manage band-
width utilization.
In general, MAC protocols for wireless networks can be
classified as coordinated and uncoordinated MAC protocols
based on the collaboration level [1]. In uncoordinated proto-
cols such as IEEE 802:11, nodes contend with each other to
share the common channel. For low network loads, these
protocols are bandwidth efficient due to the lack of over-
head. However, as the network load increases, their band-
width efficiency decreases. Also, due to idle listening, these
protocols are in general not energy efficient. On the other
hand, in coordinated MAC protocols the channel access is
regulated. Fixed or dynamically chosen channel controllers
determine how the channel is shared and accessed. IEEE
802.15.3 [2], IEEE 802.15.4 [3], and MH-TRACE [4] are exam-
ples of such coordinated protocols. Coordinated channel
access schemes provide support for quality of service (QoS),
reduce energy dissipation, and increase throughput for
dense networks. Extensively deployed cellular networks
also use a coordinated MAC protocol in which the channel
access is regulated through fixed base stations.
Some of the key challenges in effective MAC protocol
design are the maximization of spatial reuse and providing
support for non-uniform load distributions as well as sup-
porting multicasting at the link layer. Multicasting allows
sending a single packet to multiple recipients. In many
cases, supporting multicasting services at the link layer is
essential for the efficient use of the network resources, since
this approach eliminates the need for multiple transmis-
sions of an identical payload while sending it to different
destinations [5].
Spatial reuse is tightly linked to the bandwidth effi-
ciency. Due to the lossy nature of the propagation medium,
multiple devices can use the same channel resources in spa-
tially remote locations with minimal effect on each other.
Integrating spatial reuse into a MAC protocol drastically
increases bandwidth efficiency. On the other hand, due to
the dynamic behavior in MANETs, the traffic load may be
highly non-uniform over the network area. Thus, it is cru-
cial that the MAC protocol be able to efficiently handle
spatially non-uniform traffic loads. Uncoordinated proto-
cols intrinsically incorporate spatial reuse and adapt to the
changes in load distribution through the carrier sensing
mechanism. However, coordinated protocols require
B. Karaoglu is with Samraksh Company, Leesburg, VA 20175.
E-mail: bora.karaoglu@samraksh.com.
W. Heinzelman is with the Department of Electrical and Computer Engi-
neering, University of Rochester, Rochester, NY 14627.
E-mail: wheinzel@ece.rochester.edu.
Manuscript received 25 Mar. 2013; revised 29 May 2014; accepted 30 June
2014. Date of publication 14 July 2014; date of current version 30 Mar. 2015.
For information on obtaining reprints of this article, please send e-mail to:
reprints@ieee.org, and reference the Digital Object Identifier below.
Digital Object Identifier no. 10.1109/TMC.2014.2339215
IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 14, NO. 5, MAY 2015 951
1536-1233 ß 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

careful design at the MAC layer, allowing the channel con-
trollers to utilize spatial reuse and adopt to any changes in
the traffic distribution.
Similar to cellular systems, coordinated MANET MAC
protocols need specialized spatial reuse and channel bor-
rowing mechanisms that address the unique characteristics
of MANETs in order to provide as high bandwidth effi-
ciency as their uncoordinated counterparts. Due to node
mobility and the dynamic nature of the sources in a
MANET, the network load oftentimes is not uniformly dis-
tributed. In this paper we propose two algorithms to cope
with the non-uniform load distributions in MANETs:
a light weight distributed dynamic channel alloca-
tion (DCA) algorithm based on spectrum sensing,
and
a cooperative load balancing algorithm in which
nodes select their channel access providers based on
the availability of the resources.
We apply these two algorithms for managing non-uni-
form load distribution in MANETs into an energy efficient
real-time coordinated MAC protocol, named MH-TRACE
[4]. In MH-TRACE, the channel access is regulated by
dynamically selected cluster heads (CHs). MH-TRACE has
been shown to have higher throughput and to be more
energy efficient compared to CSMA type protocols.
Although MH-TRACE incorporates spatial reuse, it does
not provide any channel borrowing or load balancing
mechanisms and thus does not provide optimal support to
non-uniform loads. Hence, we apply the dynamic channel
allocation and cooperative load balancing algorithms to
MH-TRACE, creating the new protocols of DCA-TRACE,
CMH-TRACE and the combined CDCA-TRACE.
The contributions of this paper are: i) we propose a
light weight dynamic channel allocation scheme for clus-
ter-based mobile ad hoc networks; ii) we propose a coop-
erative load balancing algorithm; iii) we incorporate
these two algorithms into our earlier TRACE framework
leading to DCA-TRACE and CMH-TRACE; and iv) we
combine both algorithms to provide support for non-uni-
form load distributions and propose CDCA-TRACE. We
compare the performance of these algorithms for varying
network loads.
The rest of this paper is organized as follows. In Section 2,
we discuss related work. Section 3 presents the dynamic
channel allocation through spectrum sensing and coopera-
tive load balancing algorithms in detail. Section 4 discusses
the adaptation of these algorithms in the TRACE frame-
work. Starting with a brief introduction of the MH-TRACE
protocol in Section 4.1, in Section 4.2 and in Section 4.3, we
present the DCA-TRACE, CMH-TRACE and CDCA-
TRACE protocols. In Section 5, the performance of CDCA-
TRACE, DCA-TRACE, CMH-TRACE, MH-TRACE and
IEEE 802.11 are compared for various network topologies.
Finally, we conclude the paper in Section 6.
2 RELATED WORK
The responsibility of the MAC layer is to coordinate the
nodes’ access to the shared radio channel, minimizing con-
flicts. In a multi-hop network, obtaining a high bandwidth
efficiency is only possible through exploiting channel reuse
opportunities. Indeed, efficient utilization of the common
radio channel has been the center of attention since the early
development stages of wireless communication [6].
Cidon and Sidi [7] present a distributed dynamic channel
allocation algorithm with no optimality guarantees for a
network with a fixed a-priori control channel assignment.
Alternatively, there are various game-theoretic approaches
to the channel allocation problem in ad hoc wireless net-
works [8], [9]. Gao and Wang [8] model the channel alloca-
tion problem in multi-hop ad hoc wireless networks as a
static cooperative game, in which some players collaborate
to achieve a high data rate. However, these approaches are
not scalable, as the complexity of the optimal dynamic chan-
nel allocation problem has been shown to be NP-hard [10],
[11], [12], [13].
In multi-hop wireless networks, CSMA [14] techniques
enable the same radio resources to be used in distinct loca-
tions, leading to increased bandwidth efficiencies at the cost
of possible collisions due to the hidden terminal problem
[15]. Different channel reservation techniques are used to
tackle the hidden terminal problem. Karn [16] use an RTS/
CTS packet exchange mechanism before the transmission of
the data packet. 802.11 distributed coordination function
(DCF) uses a similar mechanism. Although this handshake
reduces the hidden node problem, it is inefficient under
heavy network loads due to the exposed terminal problem.
Several modifications to the RTS/CTS mechanisms have
been proposed to increase the bandwidth efficiency [17], [18]
including use of multiple channels such as [19], [20], [21].
However, these approaches attempt to solve the problem
of channel assignment when there is a single intended desti-
nation of each transmission, and they do not cover group
communication. In many cases, using link layer multicast-
ing/broadcasting increases the efficient use of network
resources [5]. Indeed, many MANET applications such as
military field communications [22] and inter vehicle com-
munication systems [23] make use of broadcast services. In
this paper, we particularly focus on link layer broadcasting
and consider MANET scenarios where the destination of
the generated packet is not a specific node in the local neigh-
borhood but all the nodes in the immediate neighborhood of
the transmitter. The IEEE 802.11 [24] standard defines and
allows link layer broadcasting services for both infrastruc-
ture and ad hoc modes. In ad hoc broadcast communication
mode, the IEEE 802.11 MAC DCF specification disables the
RTS/CTS mechanism as well as acknowledgments (ACKs).
There is no MAC-level recovery or re-transmission for
broadcast frames. The broadcast performance of IEEE
802.11 has been studied through simulations [25], [26] as
well as analytically [27].
In coordinated MAC protocols, channel assignment is
performed by channel coordinators. Spatially separated
coordinators can simultaneously use the same channels
with the channel reuse concept. The cellular concept [28]
that regulates channel access through fixed infrastructure
called base stations also forms the basis of the widely
deployed GSM systems [29].
The types of strategies for on-demand dynamic channel
allocation used in cellular systems can be divided into two
categories: centralized and distributed schemes. In central-
ized dynamic channel allocation schemes [30], the available
952 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 14, NO. 5, MAY 2015

channels are kept in a pool and distributed to various cells
by a central coordinator. Although quite effective in maxi-
mizing channel usage, these systems have a high overhead
and cannot be applied to MANETs due to the lack of high
bandwidth and low latency links between the cluster heads
for coordination.
Distributed dynamic channel allocation for cellular net-
works has also been studied extensively [31], [32], [33]. In
distributed dynamic channel allocation, each cell is assigned
a number of channels. These channels can be exchanged
among adjacent cells through message exchange mecha-
nisms between the channel regulators (cell towers) in an on
demand basis. This approach, too, is not directly applicable
to MANETs. Unlike in the cellular case, in MANETs, the
message exchanges between the channel regulators also
consume network resources. Due to node mobility and the
dynamic behavior of the network, the large overhead associ-
ated with the frequent message exchanges may overwhelm
the network and decrease the bandwidth efficiency.
Load balancing has also been studied within the context
of heterogeneous networks. In the case of excess demand,
part of the network load can be offloaded to other networks
using heterogenous gateway nodes. Song et al. [34] present
a policy framework for such resource management in a
loosely coupled cellular/WLAN integrated network.
Although dynamic channel allocation and channel handoff
are studied extensively within the context of cellular net-
works, they have not been studied much in the context of
MANETs, where the bandwidth efficiency and load balanc-
ing are mostly studied at the network layer [13], [35]. Wu
et al. [35] extend the AODV protocol to include a distributed
system to infer the network status and to optimize routes
considering bandwidth efficiency and stability. A central-
ized load aware joint channel assignment and routing algo-
rithm is proposed in [13].
At the MAC layer, Tseng et al. [36] propose a location
aware dynamic channel allocation scheme for MANETs.
However, their protocol mandates that location information
be provided to each node. Namboothiri and Sivalingam [37]
study the capacity of the IEEE 802.15.4 protocol for linear
and grid topologies and calculate the optimal channel
assignment yielding the maximum possible channel reuse.
However, the results are not generalizable to the complex
and dynamic topologies of typical MANETs. Primary Colli-
sion Avoidance type channel allocation algorithms [38], [39],
[40], [41] assign channels to the nodes one by one, mitigat-
ing the conflict relationships in a connection graph at each
iteration. Finally, Chowdhury et al. [42] propose a dynamic
channel allocation scheme for IEEE 802.15.4 systems using
a single hop overlay weight-based clustering structure.
Although the proposed system reduces the message
exchanges over previously built Primary Collision Avoidance
algorithms, the proposed system is entirely message driven
and requires the construction of clusters. Also this system is
susceptible to topology changes during the channel alloca-
tion phase. To the best of our knowledge, our work is the
first attempt to solve the dynamic channel allocation prob-
lem solely based on carrier sense measurements (i.e., spec-
trum sensing), greatly reducing the overhead.
We first introduced the preliminary concept of dynamic
channel allocation for TRACE systems in [43]. In this paper,
we extend the concept and analyze the non-uniform load
distribution problem from both the perspective of member
nodes and the clusterheads. We also introduce a collabora-
tive load balancing algorithm for TRACE. By combining the
dynamic channel allocation and collaborative load balanc-
ing algorithms, we propose the CDCA-TRACE protocol
that has the highest bandwidth efficiency among the
TRACE family of protocols.
We investigate the performance of the dynamic channel
allocation and collaborative load balancing algorithms, by
comparing them to MH-TRACE[4], which implements the
basic multi-hop MAC protocol of the TRACE system, as
well as the beacon enabled IEEE 802.15.4 protocol in GTS
mode of operation and the well known IEEE 802.11 [24] pro-
tocol. Thanks to the popularity of IEEE 802.11, the literature
consists of many references comparing the performance of
IEEE 802.11 with many other existing protocols.
3 BANDWIDTH EFFICIENCY TECHNIQUES FOR
COORDINATED MAC PROTOCOLS
In this section we describe the lightweight dynamic channel
allocation mechanisms based on channel sensing and the
cooperative load balancing algorithms. We begin with a dis-
cussion of our assumptions:
Single transceiver. The nodes in the network are
equipped with a transceiver that can operate in one
of two modes: transmission or reception. Nodes can-
not simultaneously transmit and receive.
Channel sensing. The receiver node is able to detect
the presence of a carrier signal and measure its
power even for messages that cannot be decoded
into a valid packet.
Collisions. In the case of simultaneous transmissions
in the system, neither of the packets can be received
unless one of the transmissions captures the receiver.
The receiver can be captured if the power level of
one of the transmissions is significantly larger than
the power level of all other simultaneous transmis-
sions. Such a capturing mechanism is the driving fac-
tor of the advantages gained through channel reuse.
Channel coordinators. The channel resources are man-
aged and distributed by channel coordinators. These
coordinators can be ordinary nodes that are selected
to perform the duty, or they can be specialized
nodes. The channel is provided to the nodes in the
network for their transmission needs by these chan-
nel coordinators. The system is also assumed to be a
closed system where all the nodes comply with the
channel access rules.
Networks operating under these assumptions and incor-
porating a channel reuse scheme can achieve relatively
higher bandwidth efficiency under uniform network loads.
However, the system needs additional mechanisms to tackle
the problem of non-uniform distribution of the network load.
3.1 Dynamic Channel Allocation Algorithm
The first mechanism that we propose is a dynamic chan-
nel allocation algorithm similar to the ones that exist in
cellular systems. Under non-uniform loads, it is crucial
KARAOGLU AND HEINZELMAN: COOPERATIVE LOAD BALANCING AND DYNAMIC CHANNEL ALLOCATION FOR CLUSTER-BASED MOBILE AD... 953

for the MAC protocol to be flexible enough to let addi-
tional bandwidth be allocated to the controllers in the
heavily loaded region(s).
Dynamic channel allocation systems in cellular systems
[31], [32], [33] depend on higher bandwidth back-link con-
nections available to cell towers. The cell towers are coordi-
nated using these back-link connections in order to provide
dynamic channel allocation and spatial reuse simulta-
neously. On the other hand, in MANETs, the channel coor-
dinators can only communicate by sharing common
channel resources, reducing the resources available for data
transmission. In addition to this, the interference relation-
ships between channel coordinators are highly dynamic.
Hence, implementing a tight coordination would be too
costly for a MANET system. Instead, we adopt a dynamic
channel borrowing scheme that utilizes spectrum sensing.
In this algorithm, the channel controllers continuously
monitor the power level in all the available channels in the
network and assess the availability of the channels by com-
paring the measured power levels with a threshold. If the
load on the channel controller increases beyond capacity,
provided that the measured power level is low enough, the
channel coordinator starts using an additional channel with
the lowest power level measurement. Once the channel
coordinator starts using the channel, its transmission
increases the power level measurement of that channel for
nearby controllers, which in turn prevents them from
accessing the same channel. Similarly, as the local network
load decreases, controllers that do not need some channels
stop the transmissions in that channel, making it available
for other controllers.
In this dynamic channel allocation algorithm, channel
coordinators react to the increasing local network load by
increasing their share of bandwidth. Although being
effective in providing support for non-uniform network
loads, the reactive response taken by the channel coordina-
tors increases the interference in the entire system.
3.2 Cooperative Load Balancing
The DCA algorithm approaches the problem of non-uni-
form load distribution from the perspective of the channel
coordinators. The same problem can also be approached
from the perspective of the other nodes in the network.
Using cooperative nodes smooths out mild non-uniformi-
ties in the load distribution without the need for the adjust-
ments at the channel coordinator side.
The load on the channel coordinators originate from the
demands of the ordinary nodes. Many nodes in a network
have access to more than one channel coordinator. The
underlying idea of the cooperative load balancing algorithm
is that the active nodes can continuously monitor the load of
the channel coordinators and switch from heavily loaded
coordinators to the ones with available resources. These
nodes can detect the depletion of the channels at the coordi-
nator and shift their load to the other coordinators with
more available resources. The resources vacated by the
nodes that switch can be used for other nodes that do not
have access to any other channel coordinators. This
increases the total number of nodes that access the channel
and hence increases the service rate and the throughput.
4 APPLYING DISTRIBUTED CHANNEL ALLOCATION
AND COOPERATIVE LOAD BALANCING TO
TRACE
4.1 Protocol Overview: MH-TRACE
This section briefly describes the MH-TRACE protocol. The
complete protocol description is available in [4]. Also vari-
ous protocol parameters are optimized in [44].
In MH-TRACE, certain nodes assume the roles of chan-
nel coordinators, here called cluster-heads. All CHs send
out periodic Beacon packets to announce their presence to
the nodes in their neighborhood. When a node does not
receive a Beacon packet from any CH for a predefined
amount of time, it assumes the role of a CH. This scheme
ensures the existence of at least one CH around every node
in the network.
In MH-TRACE, time is divided into superframes of equal
length, as shown in Fig. 1, where the superframe is repeated
in time and further divided into frames. Each clusterhead
operates using one of the frames in the superframe structure
and provides channel access for the nodes in its communi-
cation range.
Each frame in the superframe is further divided into sub-
frames. The control sub-frame is used for signaling between
nodes and the CH, and the data sub-frame is used to trans-
mit the data payload. In the Beacon slot, CHs announce
their existence and the number of available data slots in the
current frame. The CA slot is used for interference estima-
tion for CHs operating in the same frame (co-frame CHs).
During the CA slot, CHs transmit a message with a given
probability and listen to the medium to calculate interfer-
ence caused by other CHs operating in the same frame. By
monitoring the interference levels in the medium during
Fig. 1. A snapshot of MH-TRACE clustering and medium access. CHs
are represented by diamonds. CH-frame matching, together with the
contents of each frame, is depicted.

the Beacon and CA slots of each frame, CHs switch to the
least noisy frame from their perspective.
Contention slots are utilized by the nodes to send their
channel access requests to the CH. A node that wants to
access the channel randomly selects a contention slot and
transmits a contention message in that slot. After listening
to the medium during the contention slots, the CH becomes
aware of the nodes that have channel requests and forms
the transmission schedule by assigning available data slots
to the nodes. After that, the CH sends a Header message
that includes the transmission schedule.
There are an equal number of IS slots and data slots in
the remainder of the frame. During the IS slots, nodes send
short packets summarizing the information that they are
going to be sending in the corresponding data slot. By lis-
tening to the relatively shorter IS packets, receiver nodes
become aware of the data that are going to be sent and may
choose to sleep during the corresponding data slots. These
slots contribute to the energy savings mechanism by letting
nodes sleep during the relatively longer data slots whose
corresponding IS packets cannot be decoded. IS packets can
also carry routing information. However, for the purposes
of this paper, we assume that all the nodes that can success-
fully receive the IS packet listen to the corresponding data
slot, since we are testing the performance of the MAC layer
only. Routing considerations addressed in [45] are out of
the scope of this paper.
Another use for the IS packets is to notify the CH about the
utilization of the slot by the assigned node. CHs automati-
cally reserve a data slots for nodes that had a reservation in
the previous superframe and actively used it. CHs drop the
reservation in the case of either missing IS packets or an IS
packet with an end-of-stream instruction. In the beginning of
its frame, each CH calculates the available data slots and
includes this information in its Beacon packet. We utilize this
information in both the dynamic channel allocation and the
cooperative load balancing algorithms.
4.2 Dynamic Channel Allocation for TRACE
In MH-TRACE, each CH operates in one of the frames in the
superframe. Since the number of data slots is ﬁxed, the CH
can only provide channel access to a limited number of
nodes. Due to the dynamic structure of MANETs, one CH
may be overloaded while others may not be using their data
slots. In that case, although there are unused data slots in
the superframe, the overloaded CH would provide channel
access only to a limited number of nodes, which is equal to
the number of data slots per frame, and the CH would deny
the channel access requests of the others. Thus, the system
needs a dynamic channel allocation scheme to provide
access to a larger number of nodes.
DCA-TRACE lets CHs operate in more than one frame
per superframe, if they are overloaded. Instead of choosing
and operating in the least noisy frame as in MH-TRACE, in
DCA-TRACE, based on the load level, CHs decide on the
number of frames they require and opportunistically choose
that many frames from the least noisy frames.
DCA-TRACE includes two additional mechanisms on
top of MH-TRACE: i) a mechanism to keep track of the
interference level from the other CHs in each frame; and ii)
a mechanism to sense the interference level from the
transmitting nodes in each data slot in each frame. These
mechanisms make use of existing messages and do not add
complexity other than slightly increasing memory require-
ments to store the interference levels.
The MH-TRACE structure provides CHs the ability to
measure the interference from other CHs in their own frame
and in other frames through listening to the medium in the
CA slot of their own frame and the Beacon slots of other
frames. In MH-TRACE, CHs use this mechanism to choose
the minimum interference frame for themselves. DCA-
TRACE makes use of the same structure. However, in order
to accommodate temporary changes in the interference lev-
els that may occur due to CH resignation or unexpected
packet drops, an exponential moving average update mech-
anism is used to determine the current interference levels in
each frame. At the end of each frame, the interference level
of the Beacon and CA slots are updated with the measured
values in that frame using
Ik;t ¼
Mk;t if Ik;tÀ1 Mk;t;
1 À að ÞIk;tÀ1 þ aMk;t o:w:;

(1)
where Ik;t and Ik;tÀ1 are the interference levels of the kth slot
in the current and the previous superframe, respectively.
Mk;t is the measured interference level of the kth slot in the
current superframe, and a is a smoothing factor, which is
set to 0:2 in our simulations. The interference level of the
frame is taken as the maximum interference level among
the interference levels of the Beacon and CA slots.
In DCA-TRACE, CHs mark a frame as unavailable if there
is another cluster that uses the frame and resides closer than
a certain threshold, Trintf , measured through the high inter-
ference value of that frame. Even under high local demand,
CHs refrain from accessing these frames that have high inter-
ference measurements, in order to protect the stability of the
clustering structure and the existing data transmissions. At
the end of each superframe, CHs determine the number of
frames that they need to access, m, based on the reservations
in the previous frame. Depending on the interference level of
each frame, they choose the least noisy m frames that have
an interference value also below a common threshold, Thintf .
If the number of available frames is less than m, the CHs
operate only in the available frames. Thintf prevents exces-
sive interference in between co-frame clusters that can poten-
tially destabilize the clustering structure.
Although the TRACE algorithm quickly re-elects CHs,
causing interference among clusters and hence triggering
cluster-head resignation may lead to service outage until
new clusters are formed. In order to prevent this, the thresh-
old should be set to ensure that the most distant node can
receive packets from the CH given the other clusters operat-
ing on the same frame. On the other hand, such an over-con-
servative Thintf setting diminishes the beneﬁts of dynamic
channel allocation. We tested DCA-TRACE with no thresh-
old, and with two Thintf values set equal to the power levels
of packets at 350 m away and 750 m away from a transmit-
ter that transmits at the given power level for our operation.
We observed that the intermediate selection of Thintf leads
to the a higher number of data packet receptions. Thus, in
our simulations, Thintf is set to a level that corresponds to
the power of a packet 350 m away from the transmitter at

the given transmission power level and propagation model.
However, it is important to note that the system perfor-
mance can further be optimized for a given performance
metric such as maximum transmission rate or minimal
energy consumption and a given scenario.
Another mechanism that DCA-TRACE adds on top of
MH-TRACE is the dynamic assignment of data slots. In
MH-TRACE, data slots are assigned in a sequential order.
On the other hand, since DCA-TRACE introduces channel
borrowing, the CH has to refrain from reallocating a data
slot that has been borrowed by another CH and instead
must allocate another data slot that has a lower interference
value. In order to do this, CHs keep track of the interference
levels of each IS slot of each frame in the superframe. In
order to accommodate temporary changes, the exponential
moving average smoothing mechanism of (1) is also used
for IS frames. Knowing the interference values of all IS slots,
the CH opportunistically assigns the available data slots to
the nodes that request channel access beginning with the
slot that has the lowest interference value. This mechanism
helps to reduce any possible collisions between the trans-
missions sharing the same data slot.
Channel sensing and assignment in DCA-TRACE is simi-
lar to cognitive radio systems. However, since we do not
distinguish between the primary CH of the frame and the
CH that borrows a channel, we treat them equally in having
access to the available data slots in any frame.
4.3 Collaborative Load Balancing for TRACE
In the previous section, we described DCA-TRACE, which
tackles non-uniform load distribution by allowing the CHs to
access more than one frame in the superframe. The same prob-
lem can also be tackled from the member nodes’ perspective.
In our previous work[1], we determined that the majority
of the nodes in a TRACE network are in the vicinity of more
than one CH (they are in the vicinity of two, three or four
CHs with probabilities of 52, 19 and 1 percent). The nodes
that are in the vicinity of more than one CH can ask for
channel access from any of these CHs. Using a cooperative
approach and a clever CH selection algorithm on the nodes,
the load can be migrated from heavily loaded CHs to the
CHs with more available resources.
In the TRACE protocols, nodes contend for channel access
from one of the CHs that have available data slots around
themselves. After successful contention, they do not monitor
the available data slots of the CHs around them. Due to the
dynamic nature of the network load, a cluster with lots of
available data slots may become heavily loaded during a
data stream. In order to tackle this issue, nodes should con-
sider the load of the CH not only when they are first contend-
ing for channel access but also after securing a reserved data
slot during the entire duration of their data stream.
In order to further elaborate this, consider Fig. 2. Nodes
A-G are source nodes and need to contend for data slots
from one of the CHs. Each CH has six available data slots.
In MH-TRACE, if their contentions go through in alphabeti-
cal order, node G would mark CH1 as full and would ask
for channel access from CH2. However, if node G secures a
data slot from CH1 before any of the nodes A-F, one of the
source nodes would not be able to access to the channel.
In DCA-TRACE, once CH1 allocates all of its available
slots, it triggers the algorithm to select an additional frame.
However, accessing one additional frame might not always
be possible, if the interference levels on all the other frames
are too high. Moreover, accessing additional frames
increases the interference in the Beacon and Header slots of
these frames and may trigger CH resignations and reselec-
tions in the rest of the network that temporarily disturbs
ongoing data streams on the resigned CHs. Finally, access-
ing additional frames increases interference on the IS and
data slots of the new frame and decreases the potential
extent these packets can reach.
In order to overcome these difficulties, we propose
CMH-TRACE and CDCA-TRACE, which add cooperative
CH monitoring and reselection on top of MH-TRACE and
DCA-TRACE, respectively. In CMH-TRACE and CDCA-
TRACE, nodes continuously monitor the available data slots
at the CHs around themselves announced by the Beacon
messages. When all the available data slots for a CH are
allocated, with a probability p, the active nodes attempt to
trigger the cooperative load balancing algorithm. When the
cooperative load balancing is triggered, the node that is cur-
rently using a data slot from the heavily loaded CH con-
tends for data slots from other nearby CHs while keeping
and using its reserved data slot until it secures a new data
slot from another CH.
The additional contention overhead introduced to neigh-
boring CHs by the cooperative load balancing is limited. It
is important to note that only the active nodes that have
access to another CH with free resources can trigger cooper-
ative load balancing algorithm. Probabilistically triggering
the algorithm further reduces this load. Considering the fact
that TRACE already has a low contention overhead [1]
thanks to its automatic channel reservation algorithm for
active nodes, the slight increase in the contention overhead
does not have a significant effect on protocol performance.
Cooperative load balancing does not alter the clustering
structure, and it is desirable over selecting an additional
frame at the CH. However, cooperative balancing does not
completely solve the problem. The source nodes may not be
in the vicinity of another CH, and hence their load cannot
be transferred to another CH. In that case, triggering the
DCA algorithm is required. Thus, in CDCA-TRACE, we
include the additional frame selection algorithm of DCA-
TRACE with some delay. A fully loaded CH resets a
counter, NDCA ¼ 0, and starts incrementing it at the begin-
ning of each superframe while it remains fully loaded. The
CH attempts to (subject to the interference levels in the
frames) access an additional frame when NDCA ¼ TDCA.
Fig. 2. Demonstration of a scenario for the collaborative load balancing
algorithm.

This provides time for the active member nodes to trigger
the cooperative load balancing algorithm and transfer their
load to nearby CHs.
The parameters for p and TDCA determine the response
time of the algorithms. A small p value leads to a slower
response time for the cooperative load balancing algorithm,
while a large p value potentially increases the contention
overhead on the neighboring CHs. On the other hand, a
small TDCA value leads to CHs triggering dynamic channel
allocation before cooperative load balancing has a chance to
free up resources in the CH’s current frame, while a large
TDCA value decreases the system response time. In our sim-
ulations, we used a threshold value of TDCA ¼ 3 and
p ¼ 0:5. These parameters can be further optimized for a
given scenario and a desired optimal performance metric.
5 PERFORMANCE EVALUATION
The size of the region over which the nodes are located, the
number of nodes in the network, and their data generation
patterns are all important in optimizing the design parame-
ters [1]. However, due to the dynamic nature of MANETs
this information might not be available a priori, and some
of these parameters may change over the course of the net-
work lifetime. Thus, it is necessary for the protocols to
dynamically adjust to changing conditions.
In uncoordinated MAC protocols such as IEEE 802.11
[46], the common channel resource is shared among the
nodes in the network based on carrier sensing. This simple
behavior is well suited for handling any non-uniformities in
the load distribution. However, these protocols do not scale
well as the load in the network increases due to the increas-
ing number of collisions. On the other hand, coordinated
MAC protocols such as the TRACE protocols and IEEE
802.15.4 (GTS mode) minimize or eliminate collisions by
allocating dedicated channel resources to transmitters.
Unlike MH-TRACE, the channel allocation for DCA-TRACE
and CDCA-TRACE can be adjusted on the fly, making them
more flexible protocols compared to their predecessor. By
adjusting the channel access scheme, they are more capable
of adapting to: i) shrinking network dimensions, and ii)
non-uniformities in load distribution.
In this section we compare the performances of the
CDCA-TRACE, DCA-TRACE, CMH-TRACE protocols as
well as their predecessor, MH-TRACE and two IEEE stan-
dard protocols, namely IEEE 802.15.4 and IEEE 802.11. IEEE
802.15.4 module with GTS assignment extension for ns2
simulator is based on the Guerreiro et al. [47] available
online [48]. The system model and the assumptions are
described in Section 5.1.
Due to the movement of the nodes in the network, the
diameter of the network may shrink over the course of net-
work operation. At one extreme, when the largest distance
between any two nodes in the network is below the commu-
nication radius, nodes form a single hop connected net-
work. The bandwidth efficiency of MH-TRACE sharply
reduces for such an operation, as MH-TRACE cannot adjust
the number of frames in each superframe dynamically, and
each CH can only utilize a single frame per superframe.
However, the dynamic channel allocation mechanism of
DCA-TRACE enables adaptation of the protocol to this
environment by letting the single CH access all the frames
and all the data slots. We investigate this scenario in
Section 5.2. Cooperative load balancing is not effective in
this simple scenario since there is only a single CH. Hence,
CMH-TRACE and CDCA-TRACE perform similar to their
predecessors, namely MH-TRACE and DCA-TRACE,
respectively. Thus, we omit the CMH-TRACE and CDCA-
TRACE results for this scenario.
Due to the dynamic environment, the network load
might not be distributed uniformly among the clusters. In
Section 5.3, we study a scenario in which the network load
is localized in a limited portion of a multi-hop network. We
investigate the effects of cooperative load balancing and
dynamic channel allocation and compare CMH-TRACE
and DCA-TRACE with MH-TRACE. We also analyze the
combined improvements of both algorithms through
CDCA-TRACE. Finally, we compare the performance of all
of these protocols with another coordinated protocol, IEEE
802:15:4, and a typical uncoordinated protocol, IEEE 802:11.
We study random load distributions in Section 5.4. The
performances of CDCA-TRACE, DCA-TRACE, CMH-
TRACE, MH-TRACE, IEEE 802:15:4 and IEEE 802:11 are
compared in a scenario with randomly selected source
nodes in a multi-hop network with randomly distributed
mobile nodes.
5.1 System Model
For comparison purposes, we conduct ns-2 simulations of
all of the protocols. The system model is discussed in this
section.
We addressed various routing layer considerations of
TRACE systems in our previous work [45]. In this paper,
we focus on the performance of the MAC layer only. Hence,
we utilize simple network and transport layer protocols
that provide local broadcasting. A connection-less transport
layer model is assumed in which the transport layer directly
connects the upper and lower layers. All data packets are
assumed to be destined to the local neighborhood (i.e., local
broadcasting). All received data packets are passed to the
application layer and are not relayed further.
Matching the network layer algorithm, link layer broad-
casting is assumed. All the nodes in the vicinity of the trans-
mitter receive the packet as long as the power levels permit
successful decoding. Ad hoc DCF mode for link layer
broadcasting traffic is used for IEEE 802.11. Note that in this
mode, the RTS/CTS and ACK mechanisms are disabled.
Similarly, no ACK mechanism is used in the TRACE proto-
cols either, and there are no packet retransmissions. For
IEEE 802.15.4, beacon enabled mode of operation is used
with guaranteed time slot (GTS) mechanism. The ACK
mechanism is disabled for the data packets but is active for
the control messages.
The TRACE protocols require time synchronization at
the MAC layer. In our simulations, nodes are assumed to be
perfectly synchronized. TRACE does not implement a node
synchronization algorithm. In real life implementations,
synchronization should be provided either using special-
ized systems such as GPS or external synchronization algo-
rithms implemented alongside TRACE. It is possible to
obtain high synchronization accuracy on the order of

nanoseconds by using GPS systems [49]. The synchroniza-
tion algorithms that are based on packet exchanges are less
accurate and introduce synchronization errors to the sys-
tem, especially for larger networks. These synchronization
errors may reduce the performance of implementations that
use inaccurate synchronization algorithms.
The default propagation model (two-ray ground
model) that is available in ns-2 [50] is used. For all simu-
lations, we used a constant transmit power that results in
a maximum receiving range of 250 m under zero interfer-
ence. In the case of interference, all packets received dur-
ing the interference period are dropped unless one of the
packets captures the receiver with a power value at least
10 times larger than the power of any interfering packets.
The source application generates real-time traffic in con-
stant bit-rate (CBR) mode that generates packets every
25 ms. Due to real-time communication constraints, packets
become obsolete and are discarded at the source if they are
not sent within 25 ms. The channel rate is set to 2 Mbps for
TRACE and 802.11 while the default channel rate of
250 Kbps is used for 802.15.4 in order to ensure consistency
with various internal timer values such as association time-
outs and ACK timeouts. In order to account for the data rate
difference, a source coding rate of 4 Kbps is used for
802.15.4 while 32 Kbps is used for the other protocols.
Starting at ts¼2 s (80th packet generation interval),
every five packet generation interval one source node starts
generating packets, thereby increasing the number of active
sources and the load in the network.
For TRACE, the superframe duration is matched to the
source packet generation interval of 25 ms. Each superframe
consists of six frames with six data slots each. For 802.15.4,
superframe order SO is set to 1, leading to a superframe
duration of 30.72 ms.
For node mobility, the random way-point mobility model
[51], [52] is used, where the node speeds are chosen from a
uniform random distribution between 0.0 and 5.0 m/s with
zero pause time. The energy model discussed in [1] is used.
Multi-hop extensions of the IEEE 802.15.4 protocol use
full functioning devices (FFDs) that transmit their own
beacons and respond to association requests. However,
managing a large number of FFDS in an IEEE 802.15.4 net-
work is problematic due to the overhead associated with
the increased number of control messages. Our initial sim-
ulations showed that under targeted node densities, over-
head overwhelms the system resources, reducing the
system performance severely. Efficient cluster tree creation
and maintenance for multi-hop 802.15.4 networks is an
open problem and is out of the scope of this paper. In
order to isolate the problems, for the multi-hop scenarios
of Sections 5.3 and 5.4, we pre-deploy 25 stationary coordi-
nators in a uniform grid formation to cover the entire net-
work. The dimensions in the grid are selected to have each
coordinator separated by less than the communication
radius from all the adjacent coordinators. We allow an ini-
tialization period for these nodes by turning them on
20 seconds before the other nodes.
5.2 Single Hop Network
In this section, the performance of DCA-TRACE, MH-
TRACE and IEEE 802.11 are compared for a single hop
connected network in which 100 nodes, including 40
sources, are stationary and distributed over a 100 m x
100 m region with a uniform grid formation. Considering
a receiving range of 250 m, the nodes form a single hop
network. Figs. 3a and 3b the average number of transmit-
ted packets, TX, and average number of received packets,
RX, in each packet generation interval, averaged over 80
iterations throughout the simulation duration of 20 sec-
onds. Total rate of data transmissions and receptions are
also depicted considering the packet size of 100 bytes. For
802.15.4, both the rate and packet sizes are reduced eight
fold. Hence, normalized rate figures are presented for the
IEEE 802:15:4 protocol.
We omitted CDCA-TRACE and CMH-TRACE results for
this scenario. Due to the small size of the network, the
Fig. 3. Rate and number of data transmissions and receptions in each packet generation interval for (a,b) a single hop network (c,d) multi-hop net-
work with localized load distribution, (e,f) multi-hop network with random load distribution.1
1. The rate values for IEEE 802:15:4 are normalized to match 8 fold
difference in the channel rate.

TRACE framework operates with a single cluster. Thus, col-
laborative load balancing is ineffective for this scenario, and
CDCA-TRACE operates similar to DCA-TRACE with the
exception of reaction time. In CDCA-TRACE, CHs wait
three superframes before accessing to an additional frame.
This duration is smaller than the rate of increase in network
load and hence does not alter the results. Conﬁrming this,
we observed same performance in both protocols and omit-
ted the CDCA-TRACE results. For similar reasons, the
CMH-TRACE results are also omitted.
Due to the CH resignation mechanism in MH-TRACE,
only a single CH can operate in such a scenario. Since each
CH only accesses one of the frames in each superframe and
hence has access to only six data slots, TX saturates at a
value of 6. The access time for all the remaining frames is
not used and is thus wasted. On the other hand, DCA-
TRACE adapts to this situation by letting the single CH
access all six frames and all the data slots. Hence, DCA-
TRACE saturates at a value of 36 providing channel access
to six times more nodes. Since channel access is fully coordi-
nated under a single CH, both MH-TRACE and DCA-
TRACE eliminate all the collisions and a similar gain can be
observed also in RX.
Similar to MH-TRACE, IEEE 802.15.4 also has a limit on
the maximum number of GTS slot allocations and allows a
maximum of seven GTS slots out of 16 in its 30.72 ms long
superframe. Since the superframe of 802.15.4 is larger than
the packet generation interval (25 ms), TX varies between 7
and 4. GTS allocation eliminates collisions, and each trans-
mission can be received by all 99 remaining nodes. Thus, for
each interval RX is exactly 99 times TX for the same interval.
As the number of source nodes increases, TX increases in
IEEE 802.11. However, due to the lack of coordination, addi-
tional source nodes in IEEE 802.11 increase the collisions in
the network. Hence, the number of receptions does not
increase in the same proportion as the number of transmis-
sions. Although the collision probability of 802.11 DCF is large
as shown in [53], due to receiver capturing, many successful
receptions are possible even under simultaneous transmis-
sions. Still, at the maximum load, IEEE 802.11 yields around
30 percent fewer receptions compared to DCA-TRACE.
5.3 Localized Load Distribution
In this section, the performances of CDCA-TRACE, DCA-
TRACE, CMH-TRACE, MH-TRACE, IEEE 802:11, and
IEEE 802:15:4 are compared for a network in which 40
source nodes are stationary and distributed over a 100 m
x 100 m square centered in the middle of the 1,000 m x
1,000 m region with a uniform grid formation. The
remaining 200 nodes are mobile and deployed randomly.
In IEEE 802:15:4, we do have an additional 25 controllers
in a grid formation covering the entire network. Figs. 3c
and 3d present the average number of transmitted pack-
ets per packet generation interval, TX, and the average
number of received packets per packet generation inter-
val, RX, as well as total rate of data transmissions and
receptions averaged over 80 iterations throughout the
simulation duration of 20 seconds.
In order to investigate the effect of dynamic channel allo-
cation, we compare DCA-TRACE and MH-TRACE. In the
beginning of the simulation, the number of active sources in
the network is low and there are unused data slots in the
frames of almost all the CHs. Hence, TX increases at the
same pace in all four protocols as the number of sources
increases. As the number of sources increases, in MH-
TRACE, CHs allocate available data slots to the source
nodes. After all available data slots are assigned, further
channel access requests are denied and hence TX converges
to around 15. This number is greater than the number of
data slots in one frame as multiple CHs can provide access
to the source nodes depending on random selection of the
CHs. On the other hand, TX in DCA-TRACE converges
around a value of 26. The dynamic channel allocation mech-
anism of DCA-TRACE adapts the channel allocation based
on the load and enables the protocol to provide channel
access to 73 percent more nodes compared to MH-TRACE
at the highest load level of 40 source nodes.
Compared to MH-TRACE, DCA-TRACE also leads to a
gain of similar magnitude in the number of receptions, as
CHs choose the frames they access and the data slots they
allocate based on the interference levels in the medium. The
average number of data receptions in MH-TRACE and
DCA-TRACE are around 1,300 and 2,175, respectively.
Thus, DCA-TRACE leads to a gain of 67 percent in the num-
ber of receptions compared to MH-TRACE.
Next, we focus on the cooperative load balancing algorithm
by comparing MH-TRACE and CMH-TRACE. Both protocols
converge as the load in the network increases. However, at
the highest load, TX and RX in CMH-TRACE converge to
values 10 percent higher than TX and RX in MH-TRACE.
The improvements of cooperative load balancing and
dynamic channel allocation are combined in CDCA-
TRACE. Under high load, CDCA-TRACE improves TX by 3
and 80 percent compared to DCA-TRACE and MH-TRACE,
respectively. Similarly, RX is improved in CDCA-TRACE
by 3 and 77 percent compared to DCA-TRACE and MH-
TRACE, respectively.
IEEE 802:15:4 has the lowest performance compared to
the other protocols. Although there are multiple coordina-
tors serving the region over which the source nodes are
located, source nodes can only request channel access from
the coordinator with which they are associated. Moreover,
the control messages in other parts of the network interfere
with the GTS allocation procedure and further reduce the
performance. TX for IEEE 802:15:4 varies between 0 and 7.
Compared to IEEE 802:15:4, CDCA-TRACE improves TX
and RX by four fold and nine fold, respectively.
Next, we compare the performance of CDCA-TRACE
and IEEE 802.11. Unlike the TRACE protocols, the overhead
for signalling between member nodes and the CHs, namely
Beacon, CA, Contention slots, and Header, does not exist in
IEEE 802.11. Moreover, IEEE 802.11 does not divide the
channel spatially, and hence it is not effected by the larger
region over which the passive nodes are distributed. The
entire bandwidth is shared only among the active nodes in
the smaller localized region through the channel sense
mechanism. On the other hand, TRACE dynamically selects
and maintains CHs in the entire network, including the pas-
sive part. Hence, at the maximum load, IEEE 802.11 can pro-
vide channel access to about 33 nodes, which is 22 percent
higher than the average number of nodes for which CDCA-
TRACE provides channel access. However, some of the

transmissions cannot be received at the receiver side due to
collisions. The lack of coordination in IEEE 802.11 leads to a
larger number of collisions compared to CDCA-TRACE.
Thus, at the load level of 40 source nodes, RX in CDCA-
TRACE is 29 percent higher compared to 802.11, which has
an average RX value of 1,750.
As mentioned previously, another key performance mea-
sure for the MAC protocols serving MANETs is energy con-
sumption. A desirable MANET MAC protocol should be
not only bandwidth efficient but also energy efficient.
CDCA-TRACE consumes 56 percent less energy compared
to IEEE 802.11.
Inconsistency in packet delays is not desirable from the
perspective of real time communication, since the con-
struction of the stream at the receivers would be problem-
atic. A high delay variation would require a large anti-
jittering buffer, which would increase overall latency of a
real-time application. In order to measure jitter, we con-
sider inter packet delay variation (IPDV) for consecutive
packets, as described in [54]. Thanks to the coordinated
channel access, CDCA-TRACE provides smoother opera-
tion and leads to four orders of magnitude smaller aver-
age absolute IPDV compared to IEEE 802.11, as can be
observed from the data in Fig. 5. The reduction in the
IPDV makes CDCA-TRACE more suitable for real time
applications compared to IEEE 802.11.
5.4 Random Load Distribution
In this section, the performance of CDCA-TRACE, DCA-
TRACE, CMH-TRACE, MH-TRACE, IEEE 802:11, and IEEE
802:15:4 are compared for a network of 400 nodes including
200 source nodes. There are an additional 25 controllers for
IEEE 802:15:4 simulations. All the nodes are mobile with
randomly distributed initial locations over a 1,000 m x 1,000
m region. Figs. 3e and 3f present the average number of
transmitted packets per packet generation interval, TX, and
the average number of received packets per packet genera-
tion interval, RX, as well as total rate of data transmissions
and receptions averaged over 80 iterations throughout the
simulation duration of 60 seconds.
Similar to the previous scenario, as the network load
increases, with a decreasing pace, all protocols provide
channel access to more nodes, resulting in an increase in TX
up to a saturation point. Beyond this point, TX saturates as
the number of sources increases.
Thanks to the dynamic channel allocation mechanism,
DCA-TRACE is not affected from the non-uniformities in
the load distribution as much as MH-TRACE. Hence, the
rate of increase of TX is higher for DCA-TRACE compared
to MH-TRACE.
Dynamic channel allocation also helps dynamically
adjust the spatial reuse ratio on the fly based on the
channel interference measurements. Under low loads, it
allows the protocol to operate with reduced interference
by reducing the level of spatial reuse used by MH-
TRACE. However, under high loads spatial reuse is
increased up to the point limited by the frame availabil-
ity interference threshold, Thintf:
2
Thanks to the dynamic channel allocation, DCA-TRACE
can provide channel access to a larger number of nodes
compared to MH-TRACE, as seen by the higher saturation
point in Fig. 3e. At the highest simulated load level of 200
source nodes, DCA-TRACE provides channel access to an
average of 139 nodes while MH-TRACE can only provide
channel access to an average of 77:7 nodes. Hence, DCA-
TRACE provides channel access to 79 percent more nodes
under high load in a multi-hop scenario compared to MH-
TRACE. However, due to the increased interference caused
by the higher spatial reuse, the number of collisions also
increases. Thus under high load, the improvement in RX is
lower than the gain in TX. At the highest load, RX in DCA-
TRACE is 19 percent higher than that in MH-TRACE. Due
to the same phenomena, under low levels of traffic load, the
number of receptions in DCA-TRACE is slightly lower than
the number of receptions in MH-TRACE, although the num-
ber of transmissions are approximately equal. However, the
maximum difference is less than 1 percent and thus is
insignificant.
Next, we focus on the cooperative load balancing by
comparing CMH-TRACE and MH-TRACE. TX increases
faster for CMH-TRACE compared to MH-TRACE since
non-uniformities in source distribution caused by the ran-
dom source selection are smoothed out in CMH-TRACE.
CMH-TRACE improves TX by as much as 4 percent. How-
ever, for very large network loads, all the clusters in the net-
work are fully occupied. Nodes cannot use cooperative load
balancing as none of the clusters in their neighborhood
have available resources. Thus, both protocols converge to
the same value under very high network loads.
Looking at the combined performance of dynamic chan-
nel allocation and cooperative load balancing, we compare
CDCA-TRACE and DCA-TRACE. In terms of TX, the effect
of the addition of cooperative load balancing is only mar-
ginal. Both methods are effective in tackling the problem of
non-uniform load distribution for medium load levels, how-
ever, cooperative load balancing is not effective when the
network load is very high. Nonetheless, cooperative load
balancing does not alter spatial reuse and hence does not
increase the interference and the collisions. Thus, in Fig. 3f,
an improvement of 2 percent can be observed in RX for
CDCA-TRACE compared to the RX for DCA-TRACE.
Similiar to the results in the previous section, IEEE
802:15:4 has the lowest performance compared to the other
protocols. Compared to the scenario in the previous section,
this scenario has a larger number of nodes that further
increases the collisions due to an increased number of con-
trol messages. Moreover, due to the mobility of the source
nodes, association and GTS allocation should be repeated
whenever the source nodes lose contact with the coordina-
tor. However, the collisions on the association and GTS allo-
cation control messages further reduce the performance of
IEEE 802:15:4 in this scenerio. Thus, compared to IEEE
802:15:4, CDCA-TRACE improves TX and RX by 20 fold
and six fold, respectively.
Furthermore, we also compare the performances of
CDCA-TRACE and IEEE 802:11. Despite the clustering con-
straints and the signaling overhead of TRACE, CDCA-
TRACE outperforms IEEE 802.11. In Fig. 3e, under heavy
load, TX in CDCA-TRACE is 14 percent higher than that of
2. Note that this threshold setting can be arbitrated for a tradeoff
between fewer collisions and higher TX and vice versa.

IEEE 802.11, which provides channel access to only 122
nodes at a load of 200 source nodes. In addition to this,
IEEE 802.11 suffers from packet collisions due to the lack of
coordination. Packet collisions increase with increasing net-
work load, and an increased number of transmissions does
not necessarily correspond to an increased number of recep-
tions. As observed in Fig. 3f, for IEEE 802.11, RX starts to
decrease as TX increases above 50 (400th packet generation
interval) and reduces to 1; 700 at the maximum number
of sources.
Furthermore, being a coordinated protocol, CDCA-
TRACE keeps the advantages of low energy consumption
and very low jitter. The average energy consumption per
node per second for all four protocols are presented in
Fig. 4. DCA-TRACE consumes only 54 percent of the energy
consumed by IEEE 802.11, even though the number of
receptions is significantly larger. The 16 percent increase in
the average energy consumption in DCA-TRACE compared
to MH-TRACE is the result of the increased number of
transmissions and receptions.
Fig. 5 presents the average absolute IPDV for all four pro-
tocols averaged over all transmitter and receiver pairs and
over the simulation set. DCA-TRACE leads to a three orders
of magnitude smaller average absolute IPDV compared to
802.11, thanks to the channel reservation scheme in TRACE.
Compared to MH-TRACE, DCA-TRACE has a larger aver-
age absolute IPDV due to the CHs actively monitoring IS
slots for minimum interference and changing slot reserva-
tions accordingly with changing conditions. However, the
trade off between minimum interference point of operation
and minimum packet delay variation can be resolved
according to the requirements of the application by modify-
ing the slot reservation mechanism at the CHs. For instance,
CHs could preserve the transmission schedule unless the
interference of the slot is above a threshold in order to
decrease the variation in the packet delays.
To sum up, under heavy and randomly distributed
loads, CDCA-TRACE not only increases the number of
source nodes that can get channel access compared to an
uncoordinated protocol, IEEE 802.11, but it also reduces the
number of collisions, average energy consumption, and
average absolute IPDV drastically, leading to a higher num-
ber of receptions and significant energy savings, thanks to
the coordination mechanisms.
6 DISCUSSIONS AND CONCLUSIONS
In this paper, we studied the problem of non-uniform load
distribution in mobile ad hoc networks. We proposed a light
weight dynamic channel allocation algorithm and a
cooperative load balancing algorithm. The dynamic channel
allocation works through carrier sensing and does not
increase the overhead. It has been shown to be very effective
in increasing the service levels as well as the throughput in
the system with minimal effect on energy consumption and
packet delay variation. The cooperative load balancing algo-
rithm has less impact on the performance compared to the
dynamic channel allocation algorithm. We showed that
these two algorithms can be used simultaneously, maximiz-
ing the improvements in the system. The combined system
has been shown to perform at least as well as the systems
with each algorithm alone and performs better for many
scenarios. Both of the algorithms as well as the combined
system also have a fast response time, which is on the order
of a superframe duration of 25 ms, allowing the system to
adjust under changing system load.
We proposed a novel MAC protocol, CDCA-TRACE,
that combines dynamic channel allocation and cooperative
load balancing algorithms into the TRACE framework.
CDCA-TRACE, which controls channel utilization through
the dynamically selected distributed channel coordinators,
is compared to beacon enabled IEEE 802.15.4 in GTS mode
of operation and IEEE 802.11, which controls channel utili-
zation in a fully distributed manner. The carrier sensing
mechanism enables CDCA-TRACE to select the channel
coordinators more effectively compared to IEEE 802.15.4.
CDCA-TRACE provides channel access to 20x more nodes
and improves the number of receptions 6x compared to
IEEE 802.15.4. Having channel coordinators, CDCA-
TRACE is shown to require about 60 percent less energy
consumption and three orders of magnitude lower packet
delay variation. Moreover, thanks to the dynamic channel
allocation algorithm and the cooperative load balancing
algorithm, in a network with randomly distributed sour-
ces, CDCA-TRACE provides channel access to 14 percent
more nodes and improves the number of receptions 3:5x
compared to IEEE 802.11.
The results presented in this paper are based on simula-
tion studies of the proposed protocols. Hence this study,
like all simulation studies, relies on simplifications of real
life phenomena. There are many additional challenges in
real life implementation of all protocols. In addition to syn-
chronization, imperfect physical layers, and interference
from devices out of the system, limited memory and com-
puting capabilities of the hardware are among the chal-
lenges to be tackled. In order to test the feasibility of the
CDCA-TRACE protocol in real life scenarios that have these
additional challenges, we recently implemented CDCA-
TRACE on Microsoft SORA [55] software defined radio sys-
tems communicating on the 2.4 GHz ISM band. In our
Fig. 4. Average energy consumption per node per seconds of the
protocols.
Fig. 5. Average absolute inter packet delay variation of the protocols.

implementation, we successfully used a synchronization
algorithm that does not rely on a GPS system and instead
uses existing Beacon, Contention and IS packets. We tested
the CDCA-TRACE system with 4 platforms positioned in
multi-hop, multi-cluster formation. Our preliminary results
show that the algorithm achieves a synchronization accu-
racy of 60 ms in this environment. We observed that the
CDCA-TRACE protocol maintains its stability with this
accuracy level and successfully runs an application provid-
ing real time voice communication.
In this paper, we focused on bandwidth utilization and
leave full adaptation of the system for delay sensitive
communications as future work. For instance, ”channel
handover” is not implemented. Systems incorporating
channel handover provide uninterrupted channel access
for source nodes that travel away from one channel coor-
dinator towards another one by transferring their load.
The cooperative load balancing algorithm can be extended
to provide such channel handover capability. In this pro-
visioned system, moving active nodes are required to
change their channel coordinator not only based on the
load on the channel coordinators but also based on the
RSSI measurements of Beacon packets from each channel
coordinator. Similar to cooperative load balancing, the
nodes would not drop the reserved channel resources
before they secure new channels in order to have uninter-
rupted channel access during the transition. In addition to
this, CHs may reserve a certain percentage of resources
for transfers and not use them for new calls, should it be
desirable to prioritize preventing dropped calls at the
expense of an increased number of blocked calls.
In this paper we have not investigated the effects of
upper layers such as the routing layer, and instead focused
on the MAC layer capability and local broadcasting service.
Packet routing has a significant impact on the load distribu-
tion. Local link layer broadcasting service is directly used
by some routing algorithms such as network flooding.
Moreover, it can be used alongside with network coding
and simultaneous transmission techniques for cooperative
diversity. In general, joint optimization of the MAC and
routing layers may enable even more efficient solutions.
Investigation of the effects of routing is left as future work.
ACKNOWLEDGMENTS
This work was supported in part by Harris Corporation, RF
Communications Division and in part by CEIS, an Empire
State Development designated Center for Advanced
Technology.
REFERENCES
[1] B. Karaoglu, T. Numanoglu, and W. Heinzelman, “Analytical per-
formance of soft clustering protocols,” Ad Hoc Netw., vol. 9, no. 4,
pp. 635–651, Jun. 2011.
[2] Part 15.3: Wireless medium access control (MAC) and physical layer
(PHY) specifications for high rate wireless personal area networks
(WPAN), IEEE Draft Standard, Draft P802.15.3/D16, Feb. 2003.
[3] Wireless Medium Access Control (MAC) and Phys. Layer (PHY) Speci-
fications for Low-Rate Wireless Personal Area Networks (WPANs),
IEEE Standard 802.15.4-2006, Sep. 2006.
[4] B. Tavli, and W. B. Heinzelman, “MH-TRACE: Multi hop time res-
ervation using adaptive control for energy efficiency,” IEEE J. Sel.
Areas Commun., vol. 22, no. 5, pp. 942–953, Jun. 2004.
[5] M. Kappes, “An experimental performance analysis of {MAC}
multicast in 802.11b networks for VOIP traffic,” Comput. Commun.,
vol. 29, no. 8, pp. 938–948, 2006.
[6] B. Leiner, D. Nielson, and F. Tobagi, “Issues in packet radio net-
work design,” Proc. IEEE, vol. 75, no. 1, pp. 6–20, Jan. 1987.
[7] I. Cidon and M. Sidi, “Distributed assignment algorithms for
multi-hop packet-radio networks,” in Proc. 7th Annu. Joint Conf.
IEEE Comput. Commun. Soc., Mar. 1988, pp. 1110–1118.
[8] L. Gao, and X. Wang, “A game approach for multi-channel alloca-
tion in multi-hop wireless networks,” in Proc. 9th ACM Int. Symp.
Mobile Ad Hoc Netw. Comput., 2008, pp. 303–312.
[9] M. Felegyhazi, M. Cagalj, S. Bidokhti, and J.-P. Hubaux, “Non-
cooperative multi-radio channel allocation in wireless
networks,” in Proc. IEEE 26th Conf. Comput. Commun., May.
2007, pp. 1442–1450.
[10] R. Ramaswami and K. Parhi, “Distributed scheduling of broad-
casts in a radio network,” in Proc. 8th Annu. Joint Conf. IEEE Com-
put. Commun. Soc.. Technol.: Emerging Converging, Apr. 1989, vol. 2,
pp. 497–504.
[11] A. Ephremides, and T. Truong, “Scheduling broadcasts in multi-
hop radio networks,” IEEE Trans. Commun., vol. 38, no. 4, Apr.
1990, pp. 456–460.
[12] A. Sen and M. Huson, “A new model for scheduling packet radio
networks,” in Proc. 15th Annu. Joint Conf. IEEE Comput. Soc. Netw.
Next Generation, Mar. 1996, vol. 3, pp. 1116–1124.
[13] A. Raniwala and T. cker Chiueh, “Architecture and algorithms for
an IEEE 802.11-based multi-channel wireless mesh network,” in
Proc. IEEE 24th Annu. Joint Conf. IEEE Comput. Commun. Soc., Mar.
2005, vol. 3, pp. 2223–2234.
[14] L. Kleinrock, and F. Tobagi, “Packet switching in radio channels:
Part I–carrier sense multiple-access modes and their throughput-
delay characteristics,” IEEE Trans. Commun., vol. 23, no. 12,
pp. 1400–1416, Dec. 1975.
[15] F. Tobagi, and L. Kleinrock, “Packet switching in radio channels:
Part II–the hidden terminal problem in carrier sense multiple-
access and the busy-tone solution,” IEEE Trans. Commun., vol. 23,
no. 12, pp. 1417–1433, Dec. 1975.
[16] P. Karn, “Maca-a new channel access method for packet radio,” in
Proc. ARRL/CRRL Amateur Radio 9th Comput. Netw. Conf., 1990,
vol. 140, pp. 134–140.
[17] S. Toumpis, and A. Goldsmith, “New media access protocols
for wireless ad hoc networks based on cross-layer principles,”
IEEE Trans. Wireless Commun., vol. 5, no. 8, pp. 2228–2241, Aug.
2006.
[18] D. Shukla, L. Chandran-Wadia, and S. Iyer, “Mitigating the
exposed node problem in IEEE 802.11 ad hoc networks,” in Proc.
12th Int. Conf. Comput. Commun. Netw., Oct. 2003, pp. 157–162.
[19] N. Jain, S. Das, and A. Nasipuri, “A multichannel CSMA mac pro-
tocol with receiver-based channel selection for multihop wireless
networks,” in Proc. 10th Int. Conf. Comput. Commun. Netw., 2001,
pp. 432–439.
[20] A. Nasipuri and S. Das, “Multichannel CSMA with signal power-
based channel selection for multihop wireless networks,” in Proc.
52nd Veh. Technol. Conf., 2000, vol. 1, pp. 211–218.
[21] J. So and N. H. Vaidya, “Multi-channel MAC for ad hoc networks:
Handling multi-channel hidden terminals using a single trans-
ceiver,” in Proc. 5th ACM Int. Symp. Mobile Ad Hoc Netw. Comput.,
2004, pp. 222–233.
[22] K. Tang and M. Gerla, “Mac layer broadcast support in 802.11
wireless networks,” in Proc. 21st Century Mil. Commun. Conf.,
2000, vol. 1, pp. 544–548.
[23] J. Zhu, and S. Roy, “MAC for dedicated short range communica-
tions in intelligent transport system,” IEEE Commun. Mag., vol. 41,
no. 12, pp. 60–67, Dec. 2003.
[24] Wireless LAN Medium Access Control (MAC) and Phys. Layer (PHY)
Specifications, IEEE Standard 802.11-2007, Jun. 2007.
[25] M. Torrent-Moreno, D. Jiang, and H. Hartenstein, “Broadcast
reception rates and effects of priority access in 802.11-based vehic-
ular ad-hoc networks,” in Proc. 1st ACM Int. Workshop Veh. Ad Hoc
Netw., 2004, pp. 10–18.
[26] J. Xie, A. Das, S. Nandi, and A. Gupta, “Improving the reliability
of ieee 802.11 broadcast scheme for multicasting in mobile ad hoc
networks,” in Proc. IEEE Wireless Commun. Netw. Conf., 2005,
vol. 1, pp. 126–131.
[27] X. Ma, and X. Chen, “Performance analysis of IEEE 802.11 broad-
cast scheme in ad hoc wireless LANs,” IEEE Trans. Veh. Technol.,
vol. 57, no. 6, pp. 3757–3768, Nov. 2008.

[28] V. H. MacDonald, “Advanced mobile phone service: The cellular
concept,” Bell Syst. Tech. J., vol. 58, pp. 15–41, 1979.
[29] E. Standard, “Global system for mobile communications (GSM)
standard for mobile phones,” 1990.
[30] I. Katzela, and M. Naghshineh, “Channel assignment schemes for
cellular mobile telecommunication systems: A comprehensive
survey,” IEEE Pers. Commun., vol. 3, no. 3, pp. 10–31, Jun. 1996.
[31] C.-L. I, and P.-H. Chao, “Distributed dynamic channel allocation
algorithms with adjacent channel constraints,” in Proc. 5th IEEE
Int. Symp. Personal, Indoor Mobile Radio Commun., Wireless Netw.-
Catching Mobile Future, Sep. 1994, vol. 1, pp. 169–177.
[32] C.-L. I and P.-H. Chao, “Local packing-distributed dynamic chan-
nel allocation at cellular base station,” in Proc. IEEE Global Telecom-
mun. Conf., Commun. Theory Mini-Conf. Tech. Program Conf. Rec.,
1993, pp. 293–301.
[33] J. Jiang, T.-H. Lai, and N. Soundarajan, “On distributed dynamic
channel allocation in mobile cellular networks,” IEEE Trans. Paral-
lel Distrib. Syst., vol. 13, no. 10, pp. 1024–1037, Oct. 2002.
[34] W. Song, W. Zhuang, and Y. Cheng, “Load balancing for cellular/
WLAN integrated networks,” IEEE Netw., vol. 21, no. 1, pp. 27–33,
Jan./Feb. 2007.
[35] C. Wu, K. Kumekawa, and T. Kato, “A MANET protocol consider-
ing link stability and bandwidth efﬁciency,” in Proc. Int. Conf.
Ultra Modern Telecommun. Workshops, Oct. 2009, pp. 1–8.
[36] Y.-C. Tseng, C.-M. Chao, S.-L. Wu, and J.-P. Sheu. (2002).
Dynamic channel allocation with location awareness for
multi-hop mobile ad hoc networks. Comput. Commun., vol. 25,
no. 7, pp. 676–688, 2002. [Online]. Available: http://www.
sciencedirect.com/science/article/B6TYP-4442SB0-5/2/
e7d943053b3685f73c2e4ebae3b06583
[37] P. Namboothiri and K. Sivalingam, “Capacity analysis of multi-
hop wireless sensor networks using multiple transmission chan-
nels: A case study using ieee 802.15.4 based networks,” in Proc.
IEEE 35th Conf. Local Comput. Netw., Oct. 2010, pp. 168–171.
[38] A. Bertossi, and M. Bonuccelli, “Code assignment for hidden ter-
minal interference avoidance in multihop packet radio networks,”
IEEE/ACM Trans. Netw., vol. 3, no. 4, pp. 441–449, Aug. 1995.
[39] R. Battiti, A. A. Bertossi, and M. A. Bonuccelli, “Assigning codes
in wireless networks: Bounds and scaling properties.” Wireless
Netw., vol. 5, pp. 195–209, 1999.
[40] L. Hu, “Distributed code assignments for CDMA packet radio
networks,” IEEE/ACM Trans. Netw., vol. 1, no. 6, pp. 668–677, Dec.
1993.
[41] C. Schurgers, G. Kulkarni, and M. Srivastava, “Distributed on-
demand address assignment in wireless sensor networks,” IEEE
Trans. Parallel Distrib. Syst., vol. 13, no. 10, pp. 1056–1065, Oct.
2002.
[42] K. Chowdhury, P. Chanda, D. Agrawal, and Q.-A. Zeng, “DCA-a
distributed channel allocation scheme for wireless sensor
networks,” in Proc. IEEE 16th Symp. Personal, Indoor Mobile Radio
Commun., Sep. 2005, vol. 2, pp. 1297–1301.
[43] B. Karaoglu and W. Heinzelman, “A dynamic channel allocation
scheme using spectrum sensing for mobile ad hoc networks,” in
Proc. IEEE Global Telecommun. Conf., 2012.
[44] B. Karaoglu, T. Numanoglu, and W. Heinzelman, “Adaptation of
TDMA parameters based on network conditions,” in Proc. IEEE
Wireless Commun. Netw., Apr. 2009, pp. 1–9.
[45] B. Karaoglu, and W. Heinzelman, “Multicasting vs. broadcasting:
What are the trade-offs?” in Proc. IEEE Global Telecommun. Conf.,
Dec. 2010, pp. 1–5.
[46] Wireless LAN medium access control (MAC) and physical layer (PHY)
speciﬁcations, IEEE Standard, vol. 802, no. 1, p. 999, 1999.
[47] J. R. Andr Guerreiro, J. L. R. Souza, “Improving NS-2 network
simulator for IEEE 802.15.4 standard operation,” in Proc. 5th Simp-
sio de Informtica, Sep. 2013, pp. 432–443.
[48] NS-2 GTS extension. [Online]. Available: http://www.karyon-
project.eu/wp-content/uploads/2013/11/ns-2_v2_35_802_15_4_
gts_extension.zip, 2013.
[49] P. Vyskocil and J. Sebesta, “Relative timing characteristics of GPS
timing modules for time synchronization application,” in Proc.
Int. Workshop Satellite Space Commun., 2009, pp. 230–234.
[50] Network simulator (NS-2). [Online]. Available: http://www.isi.
edu/nsnam/ns, 2000.
[51] T. Camp, J. Boleng, and V. Davies, “A survey of mobility models
for ad hoc network research,” Wireless Commun. Mobile Comput.,
vol. 2, no. 5, pp. 483–502, 2002.
[52] C. Bettstetter, and C. Wagner, “The spatial node distribution of
the random waypoint mobility model,” in Proc. Mobile Ad-Hoc
Netzwerke, 2002, pp. 41–58.
[53] H. L. Vu and T. Sakurai, “Collision probability in saturated IEEE
802.11 networks,” in Proc. Australian Telecommun. Netw. Appl.
Conf., 2006, pp. 21–25.
[54] A. Morton and B. Claise. (2009, Mar.). Packet delay variation
applicability statement. RFC 5481 (Informational), Internet Eng.
Task Force [Online]. Available: http://www.ietf.org/rfc/rfc5481.
txt
[55] K. Tan, H. Liu, J. Zhang, Y. Zhang, J. Fang, and G. M. Voelker,
“Sora: High-performance software radio using general-purpose
multi-core processors,” Commun. ACM, vol. 54, no. 1, pp. 99–107,
Jan. 2011.
Bora Karaoglu received the BS degrees in elec-
trical and electronics engineering (major) and
industrial engineering (double major) from Middle
East Technical University, in 2006 and 2007,
respectively, and the MS and PhD degrees in
electrical and computer engineering from the Uni-
versity of Rochester, in 2008 and in 2014, respec-
tively. He is currently the wireless networking
researcher at The Samraksh Company, Virginia,.
His current research interests include wireless
communications and networking and mobile
computing. He is a member of the IEEE.
Wendi B. Heinzelman received the BS degree in
electrical engineering from Cornell University in
1995 and the MS and PhD degrees in electrical
engineering and computer science from MIT, in
1997 and 2000, respectively. She is currently an
associate professor in the Department of Electri-
cal and Computer Engineering, University of
Rochester, and has a secondary appointment as
an associate professor in the Department of
Computer Science. She is also the Dean of Grad-
uate Studies for Arts, Sciences and Engineering,
University of Rochester. Her current research interests lie in the areas of
wireless communications and networking, mobile computing, and multi-
media communication. She is a senior member of the IEEE.
For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.

Cooperative load balancing and dynamic

More Related Content

What's hot (17)

Similar to Cooperative load balancing and dynamic (20)

Recently uploaded (20)

Cooperative load balancing and dynamic