IEEE Wireless Communications April 2014 57 1536-1284/14/$25.

00 2014 IEEE
Jiajia Liu is with Xidian
University. Part of this
work was finished when
he was a JSPS research
fellow at Tohoku Univer-
sity.
Yuichi Kawamoto, Nei
Kato, and Hiroki Nishiya-
ma are with Tohoku Uni-
versity.
Naoto Kadowaki is with
NICT.
ENHANCI NG SPECT RAL EF F I CI ENCY F OR LTE- ADVANCED
AND BEYOND CEL L UL AR NET WORKS
INTRODUCTION
To cope with the exponential growth of mobile
broadband data traffic and the unprecedented
consumer demand for faster data connectivity,
the Third Generation Partnership Project
(3GPP) has been developing an enhanced Long-
Term Evolution (LTE) radio interface called
LTE-Advanced (LTE-A, also known as Release
10 and beyond), aiming to significantly enhance
the current LTE and support much higher capac-
ity and coverage, higher throughput, and lower
latency, higher peak rates and better user experi-
ence, and so on.
In LTE-A networks, besides the overall cov-
erage provided by traditional macrocells, various
classes of low-power nodes (LPNs, e.g., pico
eNBs, femto eNBs, and relays) can be distribut-
ed throughout the macrocells as a more targeted
underlay to further enhance the areas spectral
efficiency, alleviate traffic hot zones, and thus
improve the end user experience. However, dif-
ferent from macro eNBs, which are usually
tower-mounted and equipped with high-speed
backhaul connections, the underlaid LPNs may
be subject to a backhaul bottleneck. Specifically,
as the femto eNBs are usually deployed in homes
and use backhaul connections such as digital
subscriber line (DSL) or cable modems, clearly,
a 10 MHz LTE femtocell is going to be limited
by the backhual, especially in the uplink [1].
Regarding pico eNBs to be deployed at building
corners or street poles, it may be financially pro-
hibitive to install and maintain a high-quality
backhaul connection at such locations.
Considering the limited backhaul connections
within small cells and the imbalanced traffic dis-
tribution among different tiers, one can see that
for a heterogeneous network (HetNet) consist-
ing of multi-tier small cells, it is highly possible
that some cells (areas) are severely congested
while the adjacent cells are very lightly loaded.
Consider a user equipment (UE; in 3GPP termi-
nology, a mobile station) that is providing some
online videos from the Internet (e.g., Youtube)
roams into a congested cell. How can the already
congested eNB provide continuous and desirable
Internet access for this newcomer? What about a
group of UEs? Therefore, it is of critical impor-
tance to achieve efficient load balancing among
multi-tier cells in LTE-A HetNets.
There has been a lot of research on load bal-
ancing in different types of networks, such as
wide-area networks (WANs) [2], wavelength-
division multiplexing (WDM)-based packet net-
works [3], WiFi networks [4, 5], backbone
networks [6], mobile ad hoc networks [7, 8],
satellite networks [9], and mesh networks [10].
JIAJIA LIU, YUICHI KAWAMOTO, HIROKI NISHIYAMA, NEI KATO, AND NAOTO KADOWAKI
ABSTRACT
In LTE-Advanced networks, besides the over-
all coverage provided by traditional macrocells,
various classes of low-power nodes (e.g., pico
eNBs, femto eNBs, and relays) can be distribut-
ed throughout the macrocells as a more targeted
underlay to further enhance the areas spectral
efficiency, alleviate traffic hot zones, and thus
improve the end-user experience. Considering
the limited backhaul connections within low-
power nodes and the imbalanced traffic distribu-
tion among different cells, it is highly possible
that some cells are severely congested while
adjacent cells are very lightly loaded. Therefore,
it is of critical importance to achieve efficient
load balancing among multi-tier cells in LTE-
Advanced networks. However, available tech-
niques such as smart cell and biasing, although
able to alleviate congestion or distribute traffic
to some extent, cannot respond or adapt flexibly
to the real-time traffic distributions among
multi-tier cells. Toward this end, we propose in
this article a device-to-device communication-
based load balancing algorithm, which utilizes
D2D communications as bridges to flexibly
offload traffic among different tier cells and
achieve efficient load balancing according to
their real-time traffic distributions. Besides iden-
tifying the research issues that deserve further
study, we also present numerical results to show
the performance gains that can be achieved by
the proposed algorithm.
DEVICE-TO-DEVICE COMMUNICATIONS ACHIEVE
EFFICIENT LOAD BALANCING IN
LTE-ADVANCED NETWORKS
LIU_LAYOUT_Layout 4/29/14 8:04 PM Page 57
IEEE Wireless Communications April 2014 58
For cellular networks, most prior offloading
techniques were based on borrowing channels
from adjacent lightly loaded cells, such as load
balancing with selective borrowing [11] and
channel borrowing without locking [12, 13].
Other offloading techniques include direct retry
[14, cell breathing [15], mobile-assisted call
admission [16], and overlaying ad hoc relays on
top of cellular networks [17, 18]. It is noticed
that the available schemes cannot be directly
applied to traffic offloading in LTE-A HetNets,
since the major target here is to balance traffic
load among multi-tier cells, which differ primari-
ly in terms of physical size, maximum transmit
power, and so on. Furthermore, compared with
previous works, device-to-device (D2D) commu-
nication-based offloading techniques have the
following unique features: first, D2D communi-
cations are fully controlled by the operator,
including the transmit power of end users and
D2D relays, transmit time slot, frequency
resources, and so on; second, D2D communica-
tions use the same frequency resources as cellu-
lar transmissions for better spectral efficiency;
thus, both inter- and intra-cell interference man-
agement are critical issues. In the integrated net-
work of cellular and ad hoc relays [17, 18], there
are only macrocells, and the overlaid ad hoc
relays are fixed and use the dedicated industrial,
scientific, and medical (ISM) band. Such scenar-
ios are much simpler than D2D-based offload-
ing, because there is no need to address the
challenging issues of relay mobility, inter- and
intra-cell interference, and so on.
Therefore, besides the common requirements
for traffic offloading over multiple mobile termi-
nals, D2D-based offloading techniques should sat-
isfy the following two basic requirements in
LTE-A networks: on one hand, in order to avoid
frequent change of offloading paths and provide
satisfactory quality of service (QoS) for end users,
it should be able to overcome various network
dynamics (e.g., movement of end users or D2D
relays, interference from surrounding terminals)
that may easily deteriorate newly established D2D
links; on the other hand, it should achieve effi-
cient interference management via proper physi-
cal resource block (PRB) assignment, transmit
power control, and so on, to effectively alleviate
the impact of D2D communications on adjacent
ongoing cellular transmissions.
Some techniques have been proposed for
alleviating congestion and balancing traffic in
LTE-A HetNets. One example is the idea of
equipping small cells with big memory blocks
and caching popular videos or other common
downloads, which can be updated periodically in
off-peak time periods [19]. As all mobile user
interactions will not have to traverse the back-
haul and Internet, it can alleviate the backhaul
congestion between small cells and the core net-
work to some extent, rather than the congestion
over the air. That is, the air interface between
UEs and eNBs can still be highly congested.
Another approach is biasing, which assigns small
eNBs a bias value and pushes load onto small
cells by replacing the usual max-SINR associa-
tion with a biased SINR [20]. Although the bias-
ing technique with preconfigured bias values is
simple and effective, it is difficult to respond or
adapt flexibly to the real-time traffic distribu-
tions among multi-tier cells. Furthermore, even
a simple update of the bias value in a small cell
may affect the cell associations of lots of already
connected UEs in the cell and other adjacent
cells, which is not favorable from the users per-
spective. Therefore, how to efficiently offload
traffic among different tier cells and achieve effi-
cient load balancing according to their real-time
traffic distributions remains a challenging prob-
lem.
Toward this end, in this article we show how
to achieve efficient and real-time load balancing
in LTE-A networks via D2D communications.
The rest of this article is organized as follows. In
the next section, an overview of LTE-A networks
and D2D communications is provided. This is
followed by an application of D2D communica-
tions for direct traffic offloading. In the follow-
ing section, we propose a D2D
communication-based algorithm for efficient
loading balancing in LTE-A networks, and dis-
cuss some research issues that deserve further
study. We present numerical results to show the
performance gains that can be achieved by our
D2D-based algorithm, and finally conclude the
article in the last section.
OVERVIEW OF LTE-ADVANCED NETWORKS
AND DEVICE-TO-DEVICE COMMUNICATIONS
LTE-ADVANCED NETWORKS
As an evolution of LTE (Releases 8 and 9), LTE-
Advanced is proposed to meet or exceed the
requirements of the International Telecommuni-
cation Union (ITU) for the fourth generation
(4G) cellular systems known as International
Mobile Telecommunications-Advanced (IMT-A).
LTE-A adopts orthogonal frequency-division mul-
tiple access (OFDMA) in the downlink (DL) and
single-carrier FDMA (SC-FDMA) in the uplink
(UL), along with spatial multiplexing using multi-
layer multiple-input multiple-output (MIMO)
[21]. In LTE-A, carrier aggregation is employed
to support flexible spectrum aggregation and
maximum deployment bandwidths of 100 MHz.
Via advanced MIMO techniques, LTE-A enables
peak data rates of 1 Gb/s in DL and 500 Mb/s in
UL, and performance gain of 1.41.6 from LTE
in both capacity and cell edge user throughput to
be achieved. Other features include coordinated
multipoint (CoMP) transmission in DL and
CoMP reception in UL, enhanced intercell inter-
ference coordination (eICIC), self-optimizing net-
works (SONs), multimedia broadcast/multicast
service (MBMS), and so on [22].
Among the techniques to achieve enhance-
ments of capacity, coverage, and spectral effi-
ciency, an important new development is the
deployment of the HetNet, which enables vari-
ous LPNs to be distributed across a macrocell
network as an underlay. The LPNs include pico
eNBs (i.e., BS for Hotzone cells with typical
transmit power of 33 dBm), home eNBs (i.e., BS
for femtocells with typical transmit power of 20
dBm), relays, and remote radio heads (RRHs,
also called distributed antenna systems, DASs).
Basically, pico eNBs, relays, and RRHs are
Among the techniques
to achieve
enhancements of
capacity, coverage, and
spectral efficiency,
an important new
development is the
deployment of
heterogeneous network,
which enables various
LPNs to be distributed
across a macrocell
network as an underlay.
LIU_LAYOUT_Layout 4/29/14 8:04 PM Page 58
IEEE Wireless Communications April 2014 59
placed indoors or outdoors and are open to all
mobile UEs, and home eNBs are placed indoors
and can be configured either as open access or
for only a closed subscriber group. Furthermore,
pico eNBs have typically planned deployment by
the operators, while home eNBs are usually ran-
domly deployed by customers [23].
Besides the above HetNets, another technol-
ogy component in LTE-A is D2D communica-
tions underlying a cellular infrastructure, to be
introduced in the next section.
DEVICE-TO-DEVICE COMMUNICATIONS
Generally speaking, a UE pair moving within
close proximity to each other in a LTE-A net-
work, can establish a D2D link with or without
the assistance of the serving eNB(s). Further-
more, the D2D communications can be operated
in the unlicensed spectrum band, such as the
industrial, scientific and medical (ISM) radio
bands, or in the same licensed band as cellular
UEs. When mobile UEs conduct D2D communi-
cations utilizing the unlicensed band, like the
WLANs band (i.e., 2.4 GHz and 5 GHz for IEEE
802.11/WiFi) and the IEEE 802.16 WiMAX band
(2.5 GHz, 3.5 GHz, 5.8 GHz, etc.), it is very simi-
lar to that in classic ad hoc mobile networks if
the cellular operators (i.e., the eNBs) are not
involved in the communication process. In order
to achieve efficient spatial reuse of the precious
wireless band resources, in this article we restrict
our interests to the case of operator assisted
D2D communications and provide for D2D UE
pairs access to the same licensed band as LTE-A
cellular UEs. That is, the LTE-A eNBs will assist
D2D pairs in the operations of peer discovery,
link establishment, PRB assignment, transmit
power control, intra-cell and inter-cell interfer-
ence management, and so on, and thus enable
the D2D communications as a controlled or con-
strained underlay of the existing LTE-A net-
works. Note that this is different from the case of
FlashLinQ [24], which works on dedicated
licensed band for D2D communications, causing
no interference to cellular connections.
Specifically, we focus on the application of
D2D communications in traffic offloading, and
illustrate its great potential to achieve efficient
load balancing in future LTE-A networks, as
detailed in ensuing sections. Note that in LTE-A
networks, multihop D2D communications
require very complicated procedures for resource
allocation and interference management at the
operator side, and also incur non-negligible sig-
nal overheads among D2D UEs and serving
eNBs. Therefore, we focus only on the case of
one-hop D2D communications in this article,
and show how to apply it to balance the network
traffic, increase system throughput, and enhance
spectral efficiency in LTE-A networks.
DIRECT TRAFFIC OFFLOADING VIA
D2D COMMUNICATIONS
Direct traffic offloading via D2D communica-
tions corresponds to the case in which two
mobile UEs share contents directly with each
other without routing the data via the serving
eNBs or core network. As shown in Fig. 1, the
D2D links can be established between mobile
UEs located within the same macrocell, picocell,
and femtocell, or between mobile UEs served by
different eNBs.
In such a scenario, the D2D session can be
initiated by either the mobile UE (i.e., the trans-
mitter) or the operator (i.e., the packet data net-
work, PDN, gateway). In the former case, the
transmitter explicitly requests to set up a D2D
session by selecting a specific Session Initiation
Protocol (SIP) uniform resource indicator (URI)
extended with a special D2D keyword, such as
.direct or .local, which notifies the System Archi-
tecture Evolution (SAE) network of the prefer-
ence for a local D2D connection. In the latter
case, the PDN gateway that actually routes IP
packets to the eNBs serving the destination UE
is able to detect potential D2D traffic after pro-
cessing the IP headers of the data packets. If the
data is destined to the same eNB or the eNBs
serving neighboring cells, the eNB(s) notify
mobile UEs of the potential D2D communica-
tion and request measurement of the D2D link
quality. One can find some suggested operation
procedures in [25].
The direct D2D offloading has been exten-
sively discussed in previous works [2527]. One
of its typical applications is mobile peer-to-peer
(P2P)-style content sharing, where each mobile
UE acts as a mobile P2P server, installs a big
memory block, stores lots of popular contents,
and registers its available contents to the opera-
tors. If the data requested by a mobile UE hap-
pens to be registered by a nearby UE, the
operator can then forward the UEs request to
the nearby UE and set up for them a D2D ses-
sion, thus offloading the traffic from the serving
eNB and core network.
The application of the above direct D2D traf-
fic offloading, although able to offload data from
the serving eNB, is subject to the limitation that
the receiver of the outgoing data (or the holder
of the requested data) is in close proximity to
the transmitter. Furthermore, it cannot detour
traffic from congested macro or pico eNBs to
adjacent lightly loaded (uncongested) eNBs. To
address these limitations, in the next section, we
propose a D2D-communication-based algorithm
to achieve efficient load balancing among differ-
ent tier cells in LTE-A networks.
Figure 1. Direct trafc ofoading via D2D communications, where the trans-
mitter and receiver in each D2D link directly share the contents without
routing the trafc via the serving eNBs or core network.
LIU_LAYOUT_Layout 4/29/14 8:04 PM Page 59
IEEE Wireless Communications April 2014 60
LOAD BALANCING AMONG MULTI-TIER
CELLS VIA D2D COMMUNICATIONS
A D2D COMMUNICATION-BASED ALGORITHM FOR
EFFICIENT LOAD BALANCING IN LTE-A NETWORKS
We present a load balancing algorithm that is
able to take advantage of D2D communications
to efficiently detour traffic from congested
macrocells, picocells, or femtocells to adjacent
uncongested cells. Without loss of generality, we
focus on a three-tier HetNet consisting of macro-
cells, picocells, and femtocells (with open
access), and take a congested macrocell as an
example to illustrate our D2D-communication-
based load balancing algorithm. For the case of
a congested picocell or femtocell in other multi-
tier HetNets, the algorithm can be followed sim-
ilarly. The algorithm has four steps in total, and
will proceed to the next step only after it fails in
the current step.
D2D Communication-Based Load Balancing Algorithm
Suppose that in the congested macrocell an asso-
ciated (or attached) mobile UE is requesting
access to Internet. As the serving macro eNB is
already fully loaded, there is no available PRB
for the requesting UE within the macrocell. The
macro eNB operates as follows:
Step 1 The macro eNB tries to offload the
requesting UE to an uncongested cell adjacent
to the UE via D2D relay. Specifically, the macro
eNB first obtains the location of the requesting
UE, and checks whether there are any uncon-
gested macro eNBs, pico eNBs, or femto eNBs
adjacent to the requesting UE. If so, the macro
eNB multicasts the location of the requesting
UE to all uncongested candidate eNBs via an
X2 interface; then the uncongested eNBs reply
with the information of their associated mobile
UEs that are in close proximity to the requesting
UE and willing to relay the traffic for it. The
macro eNB then instructs the requesting UE to
measure and report the D2D link quality
between itself and the potential relays. After
selecting the D2D relay, the macro eNB collabo-
rates with the eNB serving the relay to jointly
assign PRBs and schedule transmissions for the
D2D link between the requesting UE and the
relay, and the cellular link between the relay and
its serving eNB. Thus, the macro eNB manages
to detour the traffic of the requesting UE to a
neighboring lightly loaded cell via the D2D
relay. The details of D2D relay selection, PRB
assignment, and transmit power control are left
to be discussed in the next section.
Figure 2 shows an example of step 1. UEs 1,
3, and 5 request the congested macro eNB to
provide Internet access, and the macro eNB
manages to offload their traffic to adjacent light-
ly loaded picocells or macrocells via the D2D
relays UEs 2, 4, and 6, respectively.
On the other hand, if there are no other
eNBs around the requesting UEs, all neighbor-
ing eNBs are fully loaded, or the uncongested
neighboring eNBs fail to find an eligible D2D
relay, the macro eNB proceeds to step 2.
Step 2 The macro eNB tries to release some
occupied PRBs for the requesting UE by offload-
ing a currently served macro-tier UE and its
ongoing traffic to an adjacent uncongested cell
via combined D2D and cellular links. At the
operator side, the procedures of offloading a
connected UE are similar to that in step 1. How-
ever, besides the basic requirements for setting
up combined D2D and cellular links in detour-
ing traffic, the following condition should also be
satisfied such that the macro-tier UE can be
seamlessly offloaded: The newly established
D2D and cellular links should be able to provide
equivalent (or at least comparable) quality of
service (QoS) in terms of throughput and delay,
and thus result in non-observable changes in
quality of experience (QoE). According to this
requirement, one can see that only those con-
nected macro-tier UEs, each of which has good
link quality with the D2D relay and can also get
sufficient PRBs from the corresponding neigh-
boring eNB, can be considered as candidates for
traffic offloading. Otherwise, if the above
requirements cannot be satisfied and the macro
eNB fails to offload any being served macro-tier
UE (i.e., no PRBs can be released from the
macrocell for the requesting UE), the macro
eNB proceeds to step 3.
We use Fig. 3 to illustrate step 2 of the D2D-
communication-based load balancing algorithm.
As shown in Fig. 3, UE 1 requests Internet
access in the congested macrocell. Since the
macro eNB is already fully loaded and has no
free PRBs, it first releases the PRBs occupied by
UE 2 after offloading UE 2 to the uncongested
neighboring pico eNB 1 via the D2D relay UE 3,
and allocates the newly released PRBs to UE 1.
Step 3 The macro eNB tries to offload the
requesting UE to a congested eNB that is close
to the UE and able to release some PRBs by
offloading a currently served UE to its nearby
Figure 2. Illustration of step 1 in the D2D communication based load bal-
ancing algorithm. The congested macro eNB ofoads the data trafc of the
requesting UE 1, UE 3, and UE 5 to the adjacent uncongested picocells or
macrocells via the D2D relays UE 2, UE 4, and UE 6, respectively.
UE1
UE2
UE3
UE4
UE6
UE5
Macro
eNB
Pico
eNB
Femto eNB
(open access)
Uncongested:
Congested:
LIU_LAYOUT_Layout 4/29/14 8:04 PM Page 60
IEEE Wireless Communications April 2014 61
uncongested cell. For the requesting UE, the
QoS (e.g., throughput and delay) of its traffic
depends on the link quality between itself and
the D2D relay and also the number of PRBs
that can be released from the adjacent eNB. For
the eNB routing the traffic to (or from) the
requesting UE, the requirements and procedures
of offloading a served UE are similar to that in
step 2. However, the complexity of operation in
step 3 is much higher than that in step 2. In par-
ticular, one cellular link release and one D2D
link setup are required in step 2; step 3 requires
the release of one cellular link and setting up
two D2D links. Furthermore, step 2 involves col-
laboration between two eNBs and one D2D
relay, while step 3 involves three eNBs and two
D2D relays. As shown in Fig. 4, after the con-
gested pico eNB 1 offloads its service from UE 3
to the uncongested pico eNB 2, the congested
macro eNB offloads the requesting UE 1 to the
pico eNB 1, where UE 2 and 4 are the D2D
relays for offloading UE 1 and 3, respectively.
If the macro eNB fails to offload the request-
ing UE to an adjacent congested eNB due to
lack of good channel quality, D2D relay(s), suffi-
cient PRBs, and so on, it proceeds to the last
step: step 4.
Step 4 In the last step, the macro eNB tries to
allocate to the requesting UE PRBs newly
released by offloading a currently served macro-
tier UE to a nearby congested eNB, which is able
to offload a served UE to an adjacent uncongest-
ed cell. Similar to step 3, it requires collabora-
tions between three eNBs and two D2D relays in
UE offloading and PRB releasing. However,
besides setting up two D2D links, step 4 needs to
release two cellular links, which is one more than
in step 3. As shown in Fig. 5, the macro-tier UE
2 can be offloaded via D2D relay UE 3 to the
congested pico eNB 1 after UE 4 is offloaded to
the uncongested pico eNB 2 via UE 5.
From the above algorithm, one can see that
D2D communication has significant potential in
achieving efficient load balancing according to
the real-time traffic distributions among differ-
ent tier cells of LTE-A HetNets. For the case of
LTE-A networks without D2D communications,
it is possible to achieve a certain level of load
balancing by assigning small eNBs different bias
values and pushing load onto those less congest-
ed cells. However, since the preconfigured bias
values are averaged over the statistical analysis
of network traffic, such techniques cannot
respond or adapt flexibly to the real-time net-
work dynamics in terms of eNBs powered on/off,
mobile UEs coming/leaving, and so on. Further-
more, by utilizing D2D communications, it could
even directly route the traffic between a UE pair
without occupying the air interface between eNB
and UEs; this is never the case for LTE-A net-
works without D2D, where every single bit has
to be transmitted through the eNB.
It is noticed that in LTE-A networks, some
areas may have severe interference among nodes
depending on the node density there. Due to the
limited channel resources at the eNB and the
overly crowded nodes in the area, the wireless
link (uplink or downlink) between a mobile UE
and the eNB usually has very poor SINR (i.e.,
the air interface between UEs and the eNB is
severely congested). For such scenarios, the pro-
posed D2D-based traffic offloading algorithm
can also be utilized to alleviate the air interface
congestion, increase system throughput, and
improve user experience. Specifically, the serving
eNB first communicates with the surrounding
eNBs via the specific X2 interface to obtain the
list of neighboring eNBs that are relatively light-
ly loaded. Then the serving eNB either directly
offloads a requesting UE to an uncongested
Figure 3. Illustration of step 2 in the D2D-communication-based load bal-
ancing algorithm. The congested macro eNB rst ofoads a currently
served UE (i.e., UE 2) and its ongoing trafc to the adjacent uncongested
pico eNB 1 via D2D relay UE 3, then allocates the newly released PRBs
(occupied previously by UE 2) to the requesting UE 1. The dashed line
denotes a newly released link.
UE1
UE2
UE3
Pico eNB1
Figure 4. Illustration of step 3 in the D2D communication-based load bal-
ancing algorithm. After UE 3 is ofoaded from the congested pico eNB 1
to the uncongested pico eNB2 via D2D relay UE 4, the requesting UE 1 is
ofoaded from the congested macro eNB to the pico eNB 1 via UE 2. The
dashed line denotes a newly released link.
UE1
UE2 UE3
UE4
Pico eNB1
Pico eNB2
LIU_LAYOUT_Layout 4/29/14 8:04 PM Page 61
IEEE Wireless Communications April 2014 62
adjacent cell, as defined in step 1, or releases
some occupied PRBs by offloading a currently
served UE, as defined in step 2, depending on
the actual location of the UE. Note that after
applying steps 1 and 2, the eNB only needs to
allocate channel resources to a smaller number
of UEs remaining in the area (i.e., the UEs that
cannot be offloaded). Considering the much
smaller transmit power and communication
range adopted by D2D communications, the
radio interference within the area can effectively
be alleviated.
DISCUSSIONS AND RESEARCH ISSUES
In this section, we discuss the requirements and
difficulties in offloading traffic among multi-tier
cells via D2D communications, and identify the
challenging issues in terms of algorithm com-
plexity, D2D relay selection, PRB assignment,
transmission schedule, power control, interfer-
ence management, network dynamics, QoS satis-
faction, incentive stimulation, privacy and
security, analytical modeling, and other topics.
Prerequisite According to the SAE architecture
in LTE systems, an eNB can obtain from the
PDN gateway the distributions of adjacent eNBs
as well as the list of mobile UEs served by each
eNB. Therefore, the only prerequisite of our
algorithm is that each eNB is able to obtain the
current locations of its associated mobile UEs.
Clearly, this is not a technically challenging task
for the operator (e.g., the popular function of
locating ones iPhone). According to the triangu-
lation, generally at least three eNBs are required
to pinpoint a cell phone and its owner precisely.
The only problem is accuracy: in downtown or
urban areas where there are lots of eNBs, it is
easier for operators to accurately locate a UE; in
suburban or rural areas, it is less accurate, but
GPS can be a good alternative. On the other
hand, because rural areas are not as crowded or
capacity demanding, the desire to balance traffic
among eNBs is not as obvious.
Supposed Network Applications and Services Consid-
ering the mobility issues of mobile UEs, it is very
difficult for an eNB to provide stable bandwidth
for an end user through the selected D2D relay.
Therefore, the proposed D2D communication-
based traffic offloading algorithm may be uti-
lized for network applications or services that
are either delay-tolerant or not bandwidth-hun-
gry, such as browsing news, checking email, and
instant messaging. Some popular exemplary
applications include Twitter, Facebook, LINE,
Viber, and WhatsApp.
Complexity The details of network operations in
the proposed D2D-based traffic offloading are
summarized in Table 1. Clearly, from direct
offloading to step 4, a monotonically non-
decreasing varying trend can be observed for the
number of involved eNBs, the number of
involved UEs, the number of established D2D
links, and the number of released cellular links.
Due to the increasing complexity of operations,
particularly the number of established D2D links
and the number of released cellular links, we
consider only four steps in the D2D-based load
balancing algorithm, i.e., we allow a congested
eNB to try at most four different kinds of traffic
offloading to provide Internet access for a
requesting mobile UE. Generally speaking, one
can further proceed to more complicated opera-
tions of traffic offloading, which unavoidably
involves the collaborations between more eNBs
and more mobile UEs, the setup of more D2D
links, and the release of more cellular links.
Combined D2D Link and Cellular Link It is one of
the common routes in steps 14 for D2D-based
traffic offloading, which may largely affect the
algorithm performance. Specifically, the follow-
ing aspects deserve further study. First is a
method to select the best D2D relay. As two
UEs closer in space may not necessarily own a
better channel in practice, it is advisable to mea-
sure the actual D2D link quality before a relay
decision. In addition, the cellular link condition
between the relay and its eNB should also be
taken into account. Second is the assignment of
PRBs for D2D and cellular links. Basically, the
number of PRBs allocated for a D2D link should
be decided according to what is allocated for the
cellular link. Also, one should take into consid-
eration the difference between the UL and DL
for combined D2D and cellular links, and the
difference between the frequency bands
deployed in different tier eNBs to which the
operator may allocate different bands according
to their backhaul connections. Finally, research
needs to be done on scheduling transmissions
and managing interference. Besides the differ-
ence caused by time-division duplex (TDD) and
frequency-division duplex (FDD), another issue
is that the D2D link extends over two adjacent
cells, which means D2D transmission easily
affects the cellular transmissions in two cells.
Again, transmit power control can be an effec-
Figure 5. Illustration of step 4 in the D2D-communication-based load bal-
ancing algorithm. After pico eNB 1 ofoads the served UE 4 via UE 5 to
the uncongested pico eNB 2, the congested macro eNB allocates to the
requesting UE 1 the PRBs newly released by ofoading the currently
served macro-tier UE 2 to the pico eNB 1 via UE 3. The dashed line
denotes a newly released link.
UE1
UE2
UE3
UE4
UE5
Pico eNB1
Pico eNB2
LIU_LAYOUT_Layout 4/29/14 8:04 PM Page 62
IEEE Wireless Communications April 2014 63
tive technique to manage the intra-cell and
inter-cell interference.
Releasing a Cellular Link As a basic operation in
the proposed algorithm, it is of vital importance
that an eNB select the proper connected UE to
release and detour the ongoing traffic to neigh-
boring cells. Besides the basic requirements of
establishing, for the released UE, a combined
D2D and cellular link, the newly combined route
should provide for the UE at least comparable
QoE. This means it should be able to support
QoS in terms of throughput, delay, and jitter,
similar to that previously provided by the
released cellular link. Furthermore, depending
on the service contract customers sign with the
operator, some customers may simply not allow
being offloaded.
Network Dynamics and QoS Satisfaction The unpre-
dictable network dynamics may have a non-negli-
gible impact on the performance of the proposed
D2D-based traffic offloading algorithm. Note
that the communication range between a D2D
pair is very limited. A mobile UE may easily fail
to send/receive data through the newly estab-
lished offloading path, due to the movement of
itself or that of the D2D relay. Furthermore, the
channel quality of the D2D link may become
very poor due to the low battery power at a relay
node or interference from surrounding terminals,
especially in urban areas. Such network dynamics
should be carefully accounted in D2D based traf-
fic offloading, particularly in the selection of
D2D relays and released cellular links. How to
guarantee acceptable QoS for the end user and
avoid frequent change of offloading path remains
challenging and deserves further study. Also, it
would be meaningful to conduct extensive simu-
lations under the 3GPP recommended mobile
scenarios to further evaluate the performance of
the proposed algorithm.
Incentive, Privacy, and Security Incentive stimula-
tion refers to how to encourage a mobile UE to
participate in D2D communications as a D2D
relay. According to the actual contributed traffic
amount, the operator can appropriately reward a
mobile UE that has delivered data to other users
(in the case of direct offloading) or has forward-
ed data as a D2D relay (in the case of offloading
among multi-tier cells). On the other hand, in
D2D-based offloading, a mobile UE no longer
receives (sends) data directly from (to) the eNB,
but via another mobile UE instead, which may
cause a lot of difficult threats. It is very difficult
for operators to guarantee secure transmission
through multiple terminals, since the intermedi-
ate relays may attempt malicious attacks in lots
of ways. Actually, in order for the proposed
D2D-based traffic offloading algorithm to
address such issues, the eNBs and the end user
have to take some actions before starting to
transmit data. Suppose a congested macro eNB
has to offload a mobile user to an adjacent light-
ly loaded pico eNB. Using the route usermacro
eNBpico eNB, the user and the pico eNB can
quickly establish a session key, which can be
securely used for encrypting the data transmitted
through the offloading path. Note that we only
utilize a one-hop operator-assisted D2D commu-
nication for traffic offloading (i.e., we allow only
one D2D relay in each offloading path). There-
fore, the operator can effectively prevent the
intermediate D2D relay from accessing, tamper-
ing with, or falsifying the data, because the relay
has no idea of the session key and is selected by
the operator (i.e., the operator has the physical
information of the relay, e.g., IMEI). However,
for the case of offloading through more than
one D2D relays, a malicious D2D relay may
tamper with the packets, inject falsified data, or
simply drop some incoming packets. How to
identify and avoid such behaviors remains chal-
lenging and deserves further study.
Performance Modeling and Analysis The impor-
tance of developing models for performance
analysis can never be emphasized enough. Due
to the introduction of small cells, it appears to
be nontrivial to accurately simulate a multi-tier
LTE-A HetNet. The traditional hexagonal grid-
based model seems inapplicable, since the small
cells (e.g., femtocells) are usually irregularly
scattered or clustered within the existing macro-
cell area. On the other hand, recently, the Pois-
son point process (PPP)-based model has
received wide attention because it is analytically
tractable and appears to capture the main trend
of HetNet performance. It is reported that
besides giving similar shapes of SINR distribu-
tions, the grid model and the PPP model differ
mostly in absolute SINR, with the former being
optimistic and the latter being pessimistic [28].
NUMERICAL RESULTS
In this section, we present some numerical
results to illustrate the performance gains that
can be achieved by our D2D-based load balanc-
ing algorithm. We consider a simple scenario
that consists of a single macrocell underlaid by
two picocells and two femtocells (configured as
open access). The coverage area of a macrocell,
picocell, and femtocell is assumed to be circles
of radius 250 m, 100 m, and 50 m, respectively.
Since the cell size, especially in urban areas,
decreases to cover numerous number of users
with each macrocell, we set the cell size smaller
than the future environment of an urban area. A
marco eNB, pico eNB, and femto eNB are
assumed to have the same frequency resources
of 50 orthogonal channels. That is, each eNB of
a macrocell, picocell, and femtocell can provide
Internet access for 50 users at a time. In total,
Table 1. Summary of network operations in the proposed D2D-based traffic
offloading.
Direct Step 1 Step 2 Step 3 Step 4
Involved eNBs 1 or 2 2 2 3 3
Involved UEs 2 2 3 4 5
Established D2D links 1 1 1 2 2
Released cellular links 0 0 1 1 2
LIU_LAYOUT_Layout 4/29/14 8:04 PM Page 63
IEEE Wireless Communications April 2014 64
600 UEs are uniformly distributed in the cover-
age area of a macrocell, and each UE requests
for Internet access with a probability of 33.3 per-
cent. Moreover, the maximum distance allowed
for D2D communication between a UE pair is
set to 20 m.
With the above parameter settings, we obtain
the ratio of UEs that can access the Internet
(i.e., the UEs requesting Internet access) under
our D2D-based load balancing algorithm, and
compare it with the scenario where our algo-
rithm is not applied. The numerical results are
summarized in Fig. 6. From the figure, one can
clearly see that the ratio of UEs that can access
the Internet increases after applying more steps
of our algorithm. Specifically, applying steps 1, 2,
3, and 4 of the proposed algorithm enables
around 86 percent of the requesting UEs to
access the Internet simultaneously, while only 65
percent of UEs can be supported when not
applying the algorithm. This is because via D2D
communications, the requesting users can be
offloaded from the congested macro eNB to the
pico eNBs and femto eNBs, which are relatively
lightly loaded, thus enabling more users to be
served and achieving higher spectral efficiency.
Additionally, it is believed that we can get more
advantages from traffic offloading to the small
cells from the macrocell by our proposed
method. For example, in a situation where
numerous UEs are requesting to access the
Internet (e.g., during events like the Olympics),
the concentration of the traffic to the macrocell
causes the traffic congestion on the macrocell,
which brings on long delay, packet drop, and so
on. However, since our proposed method makes
it possible to balance the traffic load of each
cell, it can also avoid traffic congestion. Although
some problems remain to be considered, such as
load balancing between neighboring small cells,
inhomogeneous user distribution in time and
space, interference management, security issues,
and so on, it is understood that traffic offloading
is significant for the cellular network combined
with D2D communications technologies.
CONCLUSIONS
While we expect LTE-A networks to significantly
enhance the current LTE, and support much high-
er capacity and coverage, higher throughput and
lower latency, higher peak rates, better user expe-
rience, and more, it is necessary to consider effec-
tive technology to address the congestion caused
by imbalanced traffic distributions among multi-
tier cells. Different from available techniques, in
this article, we highlight a D2D-communication-
based technique for load balancing that is able to
efficiently offload traffic among multi-tier cells
according to their real-time traffic distributions. In
addition, we present numerical results to show the
great promise of applying the proposed algorithm.
Here we would like to emphasize also the chal-
lenges in applying D2D communications for traffic
offloading, such as PRB allocation and a transmis-
sion schedule for combined D2D and cellular
links, interference management, cellular link
releasing, mobility, incentive stimulation, and
security issues. Also, it would be meaningful to
further evaluate the performance of D2D-based
traffic offloading when applied together with other
techniques like cell biasing.
ACKNOWLEDGMENT
This research was partially funded by the project
R&D on Efficient Frequency Utilization for
Self-Forming Mobile Terminals Networks
(135002001), of SCOPE supported by the Min-
istry of Internal Affairs and Communications
(MIC), Japan.
REFERENCES
[1] J. G. Andrews, Seven Ways that HetNets Are a Cellular
Paradigm Shift, IEEE Commun. Mag., vol. 51, no. 3,
Mar. 2013, pp. 13644.
[2] J. He, An Architecture for Wide Area Network Load
Balancing, IEEE ICC, 2000.
[3] A. Narula-Tam and E. Modiano, Dynamic Load Balanc-
ing for WDM-Based Packet Networks, IEEE INFOCOM,
2000.
[4] P.-H. Hsiao et al., Load-Balancing Routing for Wireless
Access Networks, IEEE INFOCOM, 2001.
[5] H. Gong and J. Kim, Dynamic Load Balancing through
Association Control of Mobile Users in WiFi Networks,
IEEE Trans. Consumer Electronics, vol. 54, no. 2, 2008,
pp. 34248.
[6] L. Georgiadis et al., Lexicographically Optimal Balanced
Networks, IEEE/ACM Trans. Net., vol. 10, no. 6, Dec.
2002, pp. 81829.
[7] C.-F. Huang, H.-W. Lee, and Y.-C. Tseng, A Two-Tier
Heterogeneous Mobile Ad Hoc Network Architecture
and Its Load-Balance Routing, IEEE VTC-Fall, 2003.
[8] C. K. Toh, A.-N. Le, and Y.-Z. Cho, Load Balanced Rout-
ing Protocols for Ad Hoc Mobile Wireless Networks,
IEEE Commun. Mag., vol. 47, no. 8, 2009, pp. 7884.
[9] T. Taleb et al., Explicit Load Balancing Technique for
NGEO Satellite IP Networks with On-Board Processing
Capabilities, IEEE/ACM Trans. Net., vol. 17, no. 1,
2009, pp. 28193.
[10] S. Jung et al., Distributed Potential Field Based Rout-
ing and Autonomous Load Balancing for Wireless Mesh
Networks, IEEE Commun. Letters, vol. 13, no. 6, 2009,
pp. 42931.
[11] H. Jiang and S. Rappaport, CBWL: A New Channel
Assignment and Sharing Method for Cellular Communi-
cation Systems, IEEE Trans. Vehic. Tech., vol. 43, no.
2, May 1994, pp. 31322.
[12] S. K. Das, S. K. Sen, and R. Jayaram, A Dynamic Load
Balancing Strategy for Channel Assignment Using Selec-
tive Borrowing in Cellular Mobile Environment, Wire-
less Networks, vol. 3, no. 2, 1997, pp. 33347.
[13] , A Novel Load Balancing Scheme for the Tele-Traf-
fic Hot Spot Problem in Cellular Networks, Wireless
Networks, vol. 4, no. 2, 1998, pp. 32540.
Figure 6. The ratio of UEs that can access Internet to the UEs which request
for Internet access.
No D2D
offloading
Step 1 Steps 1-2 Steps 1-3 Steps 1-4
65
60
R
a
t
i
o

o
f

U
E
s

t
h
a
t

c
a
n

a
c
c
e
s
s

I
n
t
e
r
n
e
t

[
%
]
70
75
80
85
90
LIU_LAYOUT_Layout 4/29/14 8:04 PM Page 64
IEEE Wireless Communications April 2014 65
[14] B. Eklundh, Channel Utilization and Blocking Proba-
bi l i ty i n A Cel l ul ar Mobi l e Tel ephone System wi th
Directed Retry, IEEE Trans. Commun., vol. 34, no. 2,
Apr. 1986, pp. 32937.
[15] X. Wu, B. Mukherjee, and S. H. G. Chan, Maca-an
Efficient Channel Allocation Scheme in Cellular Net-
works, IEEE GLOBECOM, 2000.
[16] S. Das, H. Viswanathan, and G. Rittenhouse, Dynamic
Load Balancing Through Coordinated Scheduling in
Packet Data Systemsg, INFOCOM, 2003.
[17] H.Wu et al., Integrated Cellular and Ad-Hoc Relay
Systems: ICAR, IEEE JSAC, vol. 19, no. 10, Oct. 2001,
pp. 210515.
[18] E. Yanmaz and O. Tonguz, Dynamic Load Balancing and
Sharing Performance of Integrated Wireless Networks,
IEEE JSAC, vol. 22, no. 5, June 2004, pp. 86272.
[19] N. Golrezaei et al., Femtocaching: Wireless Video
Content Delivery through Distributed Caching Helpers,
IEEE INFOCOM, 2012.
[20] Q. Ye et al., User Association for Load Balancing in
Heterogeneous Cellular Networks, IEEE Commun. Let-
ters, vol. 12, no. 6, 2013, pp. 270316.
[21] Physical Layer Aspects for Evolved Universal Terrestrial
Radio Access (UTRA), 3GPP TR 25.814 V7.1.0, Sept. 2006.
[22] A. Ghosh et al., Lte-Advanced: Next-Generation Wire-
less Broadband Technology, IEEE Wireless Commun.,
vol. 17, no. 3, June 2010, pp. 1022.
[23] 3GPP, Evolved Universal Terrestrial Radio Access (E-
UTRA); Further Advancements for E-UTRA Physical Layer
Aspects, TR 36.814 V9.0.0, Mar. 2010.
[24] X. Wu et al., Flashlinq: A Synchronous Distributed
Scheduler for Peer-to-Peer Ad Hoc Networks, Allerton,
2010.
[25] K. Doppler et al., Device-to-Device Communication as
an Underlay to LTE-Advanced Networks, IEEE Com-
mun. Mag., vol. 47, no. 12, Dec. 2009, pp. 4249.
[26] G. Fodor et al., Design Aspects of Network Assisted
Device-to-Device Communications, IEEE Commun.
Mag., vol. 50, no. 3, Mar. 2012, pp. 17077.
[27] L. Lei, Z. Zhong, C. Lin, and X. S. Shen, Operator
Controlled Device-to-Device Communications in LTE-
Advanced Networks, IEEE Wireless Commun., vol. 19,
no. 3, June 2012, pp. 96104.
[28] J. G. Andrews, F. Baccelli, and R. K. Ganti, A Tractable
Approach to Coverage and Rate in Cellular Networks,
IEEE Trans. Commun., vol. 59, no. 11, Nov. 2011, pp.
312234.
BIOGRAPHIES
JIAJIA LIU [S11, M12] is currently a full professor at the
School of Computer Science and Technology at Xidian Uni-
versity. He was a recipient of the Chinese Government
Award for Outstanding Ph.D. Students Abroad in 2011, the
Tohoku University RIEC Student Award, and the Tohoku
University Professor Genkuro Fujino Award in 2012. He also
received the Yasujiro Niwa Outstanding Paper Award in
2012 and the Best Paper Award of IEEE WCNC 2012. He
was also the recipient of the prestigious Dean Award and
President Award of Tohoku University in 2013.
YUICHI KAWAMOTO [S] is pursuing his Ph.D. degree at the
Graduate School of Information Sciences (GSIS) at Tohoku
University, Sendai, Japan. He has received the Best Paper
Awards of some international conferences, including IEEE
GLOBECOM 13. Also, he was awarded the Satellite Com-
munications Research Award in 2011 from the Institute of
Electrical, Information, and Communications Engineers
(IEICE). He is a recipient of the Japan Society for the Pro-
motion of Science (JSPS) award in 2014.
HIROKI NISHIYAMA [M08, SM13] has been an associate pro-
fessor at GSIS, Tohoku University, since 2012. He has pub-
lished more than 100 peer-reviewed papers in many quality
publications, including prestigious IEEE journals and con-
ferences. He has received the Best Paper Awards from
many international conferences including many IEEE flag-
ship events, such as IEEE GLOBECOM in 2013 and IEEE
WCNC 12. He was also a recipient of the IEEE Communica-
tions Society Asia-Pacific Boards Outstanding Young
Researcher Award, the IEICE Communications Society Aca-
demic Encouragement Award 2011, and the 2009 FUNAI
Foundations Research Incentive Award for Information
Technology. His research interests cover a wide range of
areas including traffic engineering, congestion control,
satellite communications, ad hoc and sensor networks, and
network security. One of his outstanding achievements
includes Relay-by-Smartphone, which makes it possible to
share information among people using only the WiFi func-
tionality of smartphones. He is a member of IEICE.
NEI KATO [A03, M04, SM05, F13] is a professor with
GSIS, Tohoku University. He currently serves as Chair of
IEEE Ad Hoc & Sensor Networks Technical Committee, the
IEEE ComSoc Sendai Chapter, Associate Editor-in-Chief of
IEEE Transactions on Parallel and Distributed Systems and
IEEE Internet of Things Journal, an Area Editor of IEEE
Transactions on Vehicular Technology, and an Editor of
IEEE Wireless Communications and IEEE Network. He is
strategic adviser to the president of Tohoku University. He
is a fellow of IEICE.
NAOTO KADOWAKI [M] received his B.S. in communications
engineering, Masters degree in information engineering,
and Ph.D. from Tohoku University in 1982, 1984, and
2010, respectively. From April 1984 to March 1986, he was
with Mitsubishi Electric Corporation. He joined the Com-
munications Research Laboratory (CRL), later renamed the
National Institute of Information and Communications
Technol ogy (NI CT), i n 1986. He has been i nvol ved i n
research and development of high data rate satellite com-
munication systems, mobile and personal satellite commu-
ni cati ons, computer networks, and communi cati on
protocols in NICT. From July 1990 to June 1991, he was a
visiting researcher at AUSSAT, which was reformed as
OPTUS Communications, Sydney, Australia. After he was
with Advanced Telecommunications Research Institute
I nternati onal (ATR) as head of the Departments of
Autonomous Systems and Smart Networks from July 2004
to December 2006, he was managi ng di rector of the
Strategic Planning Department in NICT from January 2007
to June 2008, and director general of the Wireless Network
Research Institute of NICT from July 2008 to June 2013. He
is currently senior executive director and executive director
of the Strategic Planning Department, NICT. He is a mem-
ber of the AIAA and IEICE of Japan.
Although there are still
some problems
remaining to be
considered, it is
understood that the
traffic offloading is
significant for the
cellular network
combined the D2D
communication
technologies.
LIU_LAYOUT_Layout 4/29/14 8:04 PM Page 65