This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2018.2789902, IEEE
Transactions on Wireless Communications
1

Adaptive Mobility Load Balancing Algorithm for

LTE Small-Cell Networks
Md Mehedi Hasan, Sungoh Kwon, Member, IEEE, and Jee-Hyeon Na, Member, IEEE

Abstract—Small cells were introduced to support high data- and has become an integral part of future networks in order
rate services and for dense deployment. Owing to user equipment to support the upcoming data demand.
(UE) mobility and small-cell coverage, the load across a small- The deployment of small cells is growing fast, especially
cell network recurrently becomes unbalanced. Such unbalanced
loads result in performance degradation in throughput and in large shopping malls, stadiums, universities, multi-storey
handover success, and can even cause radio link failure. In residential apartments, and offices [7]. Based on the policies
this paper, we propose a mobility load balancing algorithm of service providers, the deployment of small cells can be
for small-cell networks by adapting network load status and planned or unplanned [8]. Due to their low cost, subscribers
considering load estimation. To that end, the proposed algorithm may have their own small cells and can deploy them anywhere,
adjusts handover parameters depending on the overloaded cells
and adjacent cells. Resource usage depends on signal qualities even to turn on and off at any time. Therefore, more or fewer
and traffic demands of connected UEs in Long Term Evolution small cells will be mostly randomly distributed throughout the
(LTE). Hence, we define a resource block-utilization ratio as a network.
measurement of cell load, and employ an adaptive threshold Even if the deployment of small cells is planned, due to
to determine overloaded cells, according to the network load their low service area, the small-cell network is vulnerable to
situation. Moreover, to avoid performance oscillation, the impact
of moving loads on the network is considered. Through system- the mobility of user equipment units (UEs). Since the small
level simulations, the performance of the proposed algorithm is cell has low transmission power, only a few UEs can be
evaluated in various environments. Simulation results show that served by each small cell, and mobility of UEs leads to an
the proposed algorithm provides a more balanced load across unbalanced load across the network. Moreover, the preferences
networks (i.e. smaller standard deviation across the cells) and of small cells during cell selection loads more traffic onto
higher network throughput than previous algorithms.
them, which also causes an overloaded traffic situation. Such
Index Terms—Small-Cell Network, Mobility Load-Balancing, unbalanced loads over the network results in performance
Self-Organizing Network, Measurement Reporting, Handover, degradation in capacity and handover success rates. When UEs
Cell Individual Offset, Load Estimation
try to move onto overloaded small cells, even if neighboring
cells remain underloaded, the deficit in resources results in
I. I NTRODUCTION handover failures or poor QoS. As a result, some cells cannot
satisfy the QoS requirements, while other neighboring cell
HE demand for mobile broadband services with higher resources remain unused. Hence, appropriate configuration and
T data rates and improved quality of service (QoS) in-
creases rapidly with the increases in smart devices and applica-
management of the network is required.
To overcome those unbalanced load issues, as well as
tions for information and communications technologies (ICT). to improve cellular network performance, the self-organized
The expected demand for wireless data in 2021 is 49 exabytes, network (SON) was proposed to configure and optimize the
which is seven times the demand in 2016 [1]. Consequently, network without human intervention [9]. Based on the SON
mobile networks need to prepare for massive traffic growth algorithm location, the SON is classified as centralized, dis-
over the next decade. tributed, or a hybrid. In a centralized SON, all functionality
To support the data demand, as well as to increase network is located in a separate subsystem. On the other hand, for
capacity, the small cell was introduced and will play an impor- a distributed SON, all the functionality is located in base
tant role in the future fifth-generation (5G) network [2] [3]. A stations. A hybrid SON is a combination of centralized and
small cell is a low-powered, low-cost radio-access node with distributed SONs, where some of the functionality is located in
a range that varies from ten to several hundred meters [4] [5]. base stations, and the rest of the functionality is located in the
Although the small cell was initially designed to extend service central SON subsystem. A SON comes with several features,
coverage by serving shadow areas of macro coverage, dense like mobility robustness optimization, mobility load balancing
deployment of small cells in a wireless network can signifi- (MLB), RACH optimization, interference management, and so
cantly increase the capacity and throughput of the network [6]. on [10]. MLB is responsible for load balancing in the network.
Hence, the small cell is now considered a prominent solution, The MLB algorithm for a SON optimizes the handover param-
eters and achieves load balancing without negatively affecting
M. M. Hasan and S. Kwon (Corresponding author) are with the School the user experience.
of Electrical Engineering, University of Ulsan, Ulsan 44610, Korea, e-mail: Previous work researched the mobility load-balancing prob-
[email protected], [email protected]
J. Na, is with the Radio Access Network S/W Research Team, ETRI, lem from various aspects. A joint MLB and inter-cell interfer-
Daejon, 34129, Korea, e-mail: [email protected] ence coordination (ICIC) algorithm [11] solves MLB and ICIC

1536-1276 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2018.2789902, IEEE
Transactions on Wireless Communications
2

problems together. Considering the link condition of target

cells, a rule-based cell individual offset (CIO) optimization
process was presented [12] to achieve a balanced network.
Both of the algorithms considered nth-tier non-adjacent neigh-
bors to overcome the load-balancing issue, with a fixed thresh-
old to determine overloaded cells. Since the authors considered
only the load status of the target cell (neglecting signal quality)
to find appropriate targets, inappropriate target cell selection
without considering the service quality from the handover of
UEs utilizes more resources or drops the service. Moreover,
due to the fixed threshold, the algorithms cannot perform Fig. 1. Access network architecture with a centralized SON.
load balancing adaptive to varying network environments. In
[13], a load balancing algorithm was proposed considering and a centralized SON (cSON), as shown in Fig. 1. The small
small cells and multi-traffic UEs in the network. However, cells operate in open access mode, and interconnect with each
the algorithm also adopted a fixed threshold to determine other via X2 interface [15]. When UEs are handed over from
overloaded cells. An algorithm proposed in [14] uses the one small cell to another, the small cells exchange handover-
network flow theory, which includes adjacent and non-adjacent related information through the X2 interface. The cSON is
overloaded neighborhood cells in the balancing process. The connected to all the small cells via S1 interface [16]. Based on
algorithm depends highly on underloaded neighboring cells, information gathered from small cells, the cSON periodically
and cannot ensure load balancing among neighborhood cells optimizes and updates handover parameters of the small cells
in a peak-load situation, when most of the cells are considered to balance the load across the network.
overloaded. However, these previous efforts are not able to In Long Term Evolution (LTE), the smallest unit of a
provide effective load balancing under various load situations, physical resource block (PRB) is defined as 180 kHz wide in
such as underloaded neighborhood situations and highly over- frequency and 0.5 ms long (1 slot) in time [17]. The frequency
loaded neighborhood situations. Hence, it is necessary to study is divided into 12 sub-carriers of 15 kHz, and one slot contains
a load-balancing algorithm that can adapt to various network a group of seven orthogonal frequency-division multiplexing
environments and avoid load ping-pongs. (OFDM) symbols (normal cyclic prefix)1 . A resource element
This paper proposes an algorithm for MLB across small is defined as one sub-carrier with one symbol period. Hence,
cells by conditionally adjusting the handover parameters. Un- a resource block is composed of 12 × 7 resource elements.
like the traditional fixed threshold approach, the algorithm In other words, a resource block is a group of seven OFDM
uses an adaptive threshold to find overloaded cells. By in- symbols; that is, 12×15 kHz wide in frequency. The scheduling
troducing maximum moveable load for the overloaded cells, time unit for allocating resources is 1 ms (2 slots), referred
the algorithm restricts the load released from the overloaded to as the transmission time interval (TTI), so we consider a
cells to the neighboring cells. In moving the load of the UEs, discrete system time of 1 ms resolution throughout this paper.
the algorithm estimates load status of the currently overloaded The total number of available physical resource blocks of a cell
cells and candidate target cells and decides UE handovers in is set according to the system bandwidth [18]. For example,
order to effectively distribute the load and avoid performance the total number of PRBs is 100 for a system bandwidth of
oscillation. In order to shift the candidate UEs, the CIOs of the 20 MHz.
serving and target cells are adjusted based on measurements
reported by UEs. Therefore, this approach is not limited only B. Measurement of Cell Load
to small-cell networks but is also applicable to macro cells
To optimize the performance of the network through load
and heterogeneous networks. The simulation results show that
balancing, it is important to find a proper measurement by
the proposed algorithm can ensure balanced loads.
which the load in a cell can be represented accurately. The
The remainder of the paper is organized as follows. Sec- load can be the radio load, the transport network load, or even
tion II describes the system model and assumptions we made. the processing load or the number of users served by a cell. To
Section III overviews load balancing-related handover param- determine whether a cell is overloaded or not, in this paper,
eters, while Section IV introduces a threshold adaptive to the resource block utilization ratio (RBUR) calculated from the
network load to decide overloaded cells, and a load estimate physical resource blocks (PRBs) of the cells is considered.
to consider the impact of the load to be moved on the serving The RBUR is defined as the ratio of PRBs used by a cell to
cell and its neighboring cells. The proposed algorithm is the total number of PRBs in the cell. For a given time duration,
explained in Section V, followed by system evaluation results T, the average RBUR of small cell i at time t is expressed as
in Section VI. Section VII concludes the paper.
t 1
RBiτ ,
X
RBU Ri = (1)
T · NP R B τ ∈(t−T, t )
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
A. System Model 1 The number of OFDM symbols depends on the cyclic prefix length, which
is configured by higher layer [17]. Under a normal cyclic prefix configuration,
In this paper, we consider a set of small cells N , denoted as the number of symbols is seven per slot whereas the number is six under an
N = {1, · · · , N } over which the load balancing is carried out, extended cyclic prefix configuration.

1536-1276 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2018.2789902, IEEE
Transactions on Wireless Communications
3

where RBiτ and NP R B are the number of allocated resource UEs are required to periodically report the measured signal
blocks and the total number of resource blocks in the cell at quality to the serving cell. This kind of measurement reporting
time τ, respectively. increases signaling overhead. To avoid high signaling over-
A higher average RBUR for a small cell indicates that the head problems, LTE provides a set of measurement report
cell has a higher load to serve and fewer available resources. mechanisms to be performed by UEs [19]. Each of those
If the RBUR of a cell reaches 1, the resources of the cell are predefined measurement report types is called an event. The
completely depleted, and UEs moving to that cell will either be measurement report includes the signal quality information
dropped or will experience low throughput. In this paper, we from the serving cell and the neighboring cell. In this paper,
did not consider any kind of call admission control (CAC) 2 . event-driven measurement reporting is considered in order to
Hence, when a new UE arrives in an overloaded cell, it will execute handovers and also to get edge UEs’ information.
be accommodated by that cell, but the per-UE throughput in
the cell will be impacted. Therefore, a shift is necessary to A. Events in LTE
reduce the load of the highly loaded cell by moving some of
LTE specifies eight events for measurement reporting: six
the UEs from it to a lightly loaded cell.
events (A1, A2, A3, A4, A5, and A6) for intra-LTE measure-
ments, and two events (B1 and B2) for inter-RAT measure-
C. Problem Formulation ments [20]. As we are considering intra-LTE load balancing,
The objective of this paper is to balance the load across we forgo inter-RAT measurement events.
the cells and improve network performance. To that end, our UEs report intra-LTE event-based measurements to the
problem is to reduce the load variance among the cells by serving cell when the criteria for a particular event has
moving edge UEs based on measurements. When the average been satisfied for a pre-specified time, called time to trig-
RBUR of a network of N small cells at time t during time ger (TTT) [21]. The triggering quantity and the reported
period T is measurement for the events can be either RSRP or RSRQ.
t 1 X t All the intra-LTE events (except A3 and A6) are triggered
RBU R N et = RBU Ri ,
N i ∈N when compared with predefined thresholds. However, events
A3 and A6 are triggered based on real-time signal quality
and the standard deviation of RBURs among the cells in the comparisons between cells. Among the threshold-based events,
network is a UE triggers event A1 when the signal quality of the serving
s
t 1 X t t cell becomes better than a pre-defined threshold. Event A2 is
σ(RBU R ) = (RBU Ri − RBU R N et ) 2 , (2)
N i ∈N triggered when the signal quality of the serving cell becomes
worse than a given threshold. When the signal quality of a
then the problem can be expressed as neighboring cell becomes better than a defined threshold, event
A4 is triggered. However, event A5 is triggered by a UE when
t the serving cell becomes worse than a defined threshold and
minimize σ(RBU R )
(3) a neighboring cell becomes better than another threshold.
subject to 0 ≤ RBiτ ≤ NP R B , ∀i ∈ N ,∀τ ∈ (t − T,t). On the other hand, event A3 is defined as a triggering event
In order to shift the load from an overloaded cell to an when a neighboring cell is offset better than the serving cell;
under-loaded cell and balance the network, we adaptively under carrier aggregation, event A6 is triggered by a UE when
adjust handover parameters, and hand UEs over by consid- the signal quality of a neighboring cell is offset better than
ering the impact of the shifting load on the network. In the the secondary cell. The primary cells are the cells selected
following sections, we explain the handover parameters and by the UEs during the initial connection establishment. The
the estimated shifting load, and we propose a load-balancing secondary cells can be added or removed by the primary cells
algorithm. For notational simplicity, we drop time t from the depending on the service demands of the UEs while carrier
following section, unless otherwise stated. aggregation is considered. In this paper, A3 and A4 event mea-
surements are used to trigger a handover and select candidate
III. M OBILITY C ONTROL PARAMETERS UEs for handover, respectively and RSRP is assumed reporting
signal quality for measurements.
To perform various functionalities, such as handover, mo-
bility management, capacity and coverage optimization, auto-
matic neighbor relations, and so on, each cell requires the B. Load Shifting Using A3 Event-Based Handover
measurements of signal qualities, reference signal received Event A3 is triggered, and the UEs report measurement
power (RSRP), or reference signal received quality (RSRQ) results to the serving cell when the signal of a neighboring cell
from the neighboring cells. To that end, a network lets UEs is offset better than that of the serving cell. Since the triggering
report signal quality to their serving cells either periodically logic of event A3 depends on relative signal quality of a
or as event-based reports. In periodic measurement reporting, neighboring cell, event A3 is the most suitable, even among the
intra-LTE events, for finding the best neighboring cell. Hence,
2 A CAC in a cell ensures the QoS of existing users by rejecting users
event A3 is widely used for triggering handovers in wireless
incoming to the cell due to instantaneous cell load in a dynamic environment.
The sacrifice of the handover users degrades the overall network performance network [22] [23]. The small cell decides to trigger a handover
even if the problem can be solved by load balancing. if the A3 triggering criteria remain satisfied for longer than the

1536-1276 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2018.2789902, IEEE
Transactions on Wireless Communications
4

TTT. The condition for an event A3 measurement report is as

follows:

Mn + O f n + Ocn − H yst > M p + O f p + Ocp + O f f , (4)

where Mn and M p denote the average RSRP values; O f n

and O f p are the frequency-specific offsets; Ocn and Ocp
are the cell individual offsets for the target and serving cells,
respectively; H yst is the hysteresis parameter; and O f f is the
A3 event offset between the primary and target neighboring
cells. In this paper, we forgo O f n and O f p because they are
used for inter-frequency handover.
By changing the values of Ocn, Ocp, and O f f , the han-
dover decision area can be shifted, and the handover can
be intentionally delayed or hastened according to the load
situation of the serving and target cells. If the value for Ocn in Fig. 2. The basic working procedure of the MLB algorithm: (a) optimization
of A3 event parameters to achieve load balancing, and (b) collecting candidate
a cell for a particular neighbor is increased, or the value of Ocp UEs and desired target cell information by setting the A4 threshold.
is decreased, then the cell range is reduced virtually, and UEs
of the cell can be handed over to the neighboring cell to reduce
the load. If the value for Ocn is decreased or the value of Ocp strength (or quality) for the serving cell and for the neighbor-
is increased, the cell will attract UEs from the neighboring cell ing cell. Hence, by setting a reasonable T hresh and selecting
to increase the cell load. For example, in Fig. 2 (a), Cell B only those UEs that report measurements, a small cell can
is overloaded with four UEs and its neighboring cells, A and obtain edge-UEs’ information and list candidate UEs to be
C, have a light load with two UEs. There are edge UEs (UEs moved for load balancing.
3 and 4) at the borders of cells A and B, but no UEs are at For example, in Fig. 2 (b), UE 3 is outside the A4 event
the borders of cells B and C. Hence, by decreasing Ocn for boundary of Cell B and reports measurements. When Cell B
Cell A, UE 3 hands over from Cell B to Cell A, and the load is overloaded, with the measurement report from UE 3, Cell
of the network is balanced. Although changing Ocp can delay B can hand UE 3 over to Cell A by reducing Ocn for Cell
or hasten the handover, it has an impact on the cell range of A. Hence, each small cell i makes a set, E i , of edge UEs
all the neighboring cells. Hence, for the sake of load-shifting based on A4 event reports. The centralized SON gathers edge-
with a particular neighboring cell, the parameter Ocn is more UEs’ information E from all the small cells, which is E =
suitable for adjustment. {E1 , · · · , E N }.
After a handover based on an A3 event, UEs move to
neighboring cells and load at the serving cell is lessened. IV. A DAPTIVE T HRESHOLD AND L OAD E STIMATION
However, the system does not have information on edge UEs To balance load over the network, we need a threshold
prior to handover for load balancing. To balance the load, to decide whether cells are overloaded or not, and we need
the system needs information about UEs that are on the transfer-load estimates to ensure system stability.
outskirts of the cell and that can be moved. Frequent periodic
measurement can provide accurate UE distribution but brings
redundant overhead, which deteriorates network performance. A. Adaptive Threshold
Hence, we utilize A4 events to gather edge-UEs’ information. Previous work introduced a fixed threshold to detect over-
loaded cells [11] [12] [22] [24]. However, the fixed threshold
is not applicable to all scenarios. For example, if the MLB has
C. Finding Edge UEs to Shift Using A4 Event a low threshold and all cells have more load than the threshold,
The SON achieves load balancing by transferring outskirts’ the previous MLB algorithms do not work. Although all the
UEs in overloaded cells to relatively low-loaded neighboring cells exceed the threshold, the variance of loads among the
cells. To that end, the SON requires information about edge cells may be large enough for load balancing. If the threshold
UEs in the overloaded cells. Since event A4 is triggered when is too high, the algorithm will hardly ever find an overloaded
the RSRP of a neighboring cell becomes better than a provided cell for load balancing. The load situation of a network also
threshold, it is the appropriate for collecting outskirts’ UEs changes over time and space. As a result, a fixed threshold
information from a cell. Using event A4, all the small cells to identify an overloaded cell is not reasonable. Therefore,
in the network collect the outskirts’ UEs’ information and to identify overloaded cells for different load situations, an
share it with the SON. The criterion for triggering event A4 adaptive threshold, T h Ada pt , is introduced, which is defined
is expressed as, as ( )
T h Ada pt = max RBU R N et , T h I nit , (6)
Mn + O f n + Ocn − H yst > T hresh, (5)
where T h I nit is a pre-defined acceptable load threshold, which
where T hresh is event A4’s threshold. UEs that satisfy the determines whether the network requires load balancing or not.
conditions in (5) will report measurements, including signal If at least one cell’s load status exceeds T h I nit , the algorithm

1536-1276 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2018.2789902, IEEE
Transactions on Wireless Communications
5

will initiate load balancing. The value of T h I nit should be where M s (i,e) and Mn ( j,e) are the measured RSRPs of serving
kept minimum in order to proceed to load balancing. small cell i and neighboring small cell j, respectively. Based
on edge-UEs and their load information gathered from small
B. UE Load Estimation cells, the SON can periodically compute the average load of
Required resource blocks (RBs) to serve edge UEs before edge UEs, which is expressed as
and after handover are different, since UEs experience dif- 1 X X
ferent signal qualities (RSRP or RSRQ) from different cells. ρedge = ρ (i,e) , (9)
kE k i ∈N e ∈E
In order to balance loads, the algorithm determines the load i

of the edge UEs to be moved from the overloaded cell and where kE k is the total number of UEs in set E.
the load to serve the edge UEs at the target neighboring cells
before performing the handover. V. P ROPOSED A LGORITHM
From the Shannon formula [25], channel capacity C is
This section introduces a mobility load-balancing algorithm
defined as C = B · log(1 + θ), where B is the transmission
for LTE small-cell networks, which adapts to the network load
bandwidth and θ is the signal-to-interference-plus-noise ra-
status and considers the impact of the UEs loads to be moved
tio (SINR). The load-balancing algorithm performs handovers
on the existing network load in order to avoid ping-pongs due
of UEs that is located in the edge areas of cells and that
to the moved load. The proposed algorithm is composed of
experience low SINRs of receiving signal quality. In the low
two parts: information gathering and load balancing.
SINR regime 3 , by Taylor series expansion [26], the channel
capacity can be approximated as
A. Information Gathering
C ' B · θ, (7)
In the information-gathering routine, the SON gathers in-
For a given SINR from small cell i and required capacity Cr eq , formation on the UEs that are available to move and that are
the required number of RBs for UE e is inversely proportional located at the borders of the small cells. To that end, the SON
to the SINR, which can be defined as adjusts T hresh in Equation (5) for all small cells (based on
Ps A3 event measurement reports of the neighboring cells) and
θ= ,
Pi + Pn gathers information on UEs in the edge areas of their serving
where Ps is the power of the signal, Pi is the average interfer- cells and on the neighboring cells of the UEs that are best for
ence caused as a result of collisions between resource blocks handover.
that are utilized by multiple cells simultaneously [27] [28] and Based on the A3-event measurement reports during a period,
Pn denotes the background noise. the algorithm establishes a set of UEs, U = {U1 , · · · , UN },
Signal power Ps is measured over 12 sub-carriers in the where Ui is the set of UEs that report measurements to
entire bandwidth over NP R B resource blocks, and is then cell i. To adjust a threshold for A4 of each cell, the algorithm
normalized to a single sub-carrier bandwidth [29] [30]. As averages the serving cell RSRPs reported by UEs at each cell,
RSRP is the average power of resource elements that carry which is expressed as follows, for small cell i:
cell-specific RS over the entire bandwidth, Ps can be modeled 1 X
using the RSRP and per-antenna sub-carrier activity factor x M si = M s (i,u) ,∀i ∈ N . (10)
kUi k u ∈U
as i

Ps = x · 12 · NP R B · RSRP. After averaging the RSRPs at which A3 events are triggered

For a full-load transmission when x = 1, SINR can be in a serving cell, the A4 threshold of small cell i is set to the
expressed as average of M s i of the neighboring cells, which is expressed
as
12 · NP R B
! 1 X
θ= RSRP. T hreshi = M si , (11)
Pi + Pn kBi k u ∈B
i

With identical interference and noise, SINR is linearly pro- where Bi is a set of neighboring cells of cell i, which
portional to RSRP at the edge area [31]. Since handovers are reported by UEs in Ui . Hence, the A4 thresholds are
take place at the edge of the cells, the UEs at the cell edge adaptively adjusted to the updates of A3-event parameters.
experience almost identical interference and noise. After setting adaptive thresholds for A4 events T hreshi ∀i ∈
Hence, for a given number of RBs (NP R B ), let us denote N , the SON then collects measurement reports under the A4
the average load to serve edge UEs e ∈ E i of small cell i as event. UEs reporting an A4 event are considered candidates
ρ (i,e) . Then, the estimated load, ρ̂ ( j,e) to serve the edge UEs for which the load can be shifted to a neighboring cell to
of small cell i at neighboring cell j can be expressed by mitigate the load of the serving cell. The SON establishes a
M s (i,e) database based on measurement reports triggered by A4 events
ρ̂ ( j,e) ≈ ρ (i,e) , (8)
Mn ( j,e) for each cell: UEs, RSRPs (M ss and Mns) of serving cells
and neighboring cells, and from the neighboring cells. Let
3 In wireless LTE networks, the SINR at the cell edge area remains low, and ( k E k)
E i = {ei(1) , ei(2) , · · · , ei i } denote the set of edge UEs that
the approximation of (7) is assumed. In the simulation section, we provide
the radio environment maps of our simulation environments in Fig. 3, which report A4 event measurements to cell i and which are sorted
show that the SINR at the edge area follow the assumption. in decreasing order of RSRP (Mn) of the neighboring cell.

1536-1276 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2018.2789902, IEEE
Transactions on Wireless Communications
6

( k E k)
A set Ti = {τi(1) , τi(2) , · · · , τi i } is denoted as the set of
neighboring cells reported under A4 event by the edge UEs ρo = RBU Ro − RBU Rl ow .
H
to serving cell i.
In other words, the moveable load is the excessive load
For example, edge UE ei(2) ∈ E i reports the RSRP of a
of the cell with respect to the lower bound. If there is no
neighboring cell, τi(2) , to serving cell i, and the RSRP of
load restriction for movement, an overloaded cell may shift
ei(n) is greater than or equal to the RSRP of ei(m) when an unconscionable amount of the load to become an under-
n < m. The routine for gathering information is summarized utilized underloaded cell, and a neighboring cell becomes
in Algorithm 1. overloaded. Moreover, the algorithm may enter an infinite loop
Algorithm 1 Information gathering due to ping-pongs.
1: Get A3 measurement reports
After computing the moveable load (H ρo ) of the overloaded
2: Compute the average M s i in (10) for i ∈ N
cell, o ∈ O, and initializing estimated loads of cells ( RBU FRi s)
3: Compute T hreshi in (11) for i ∈ N
to RBU Ri s for all i ∈ N , the algorithm then finds the possible
4: Set T hreshi as A4 threshold for Cell i
candidate UEs in E i of that particular overloaded small cell.
5: Get A4 measurement reports
To that end, for each of the edge UEs, eo(n) ∈ E o , the algorithm
6: Update candidate UE E i information of Cell i
computes the load contribution in the present serving cell,
o ∈ O, as ρ (o,e (n) ) and estimates the load D ρ (τ (n) ,e (n) ) in the
o o o
neighboring cell, τo , as in (8).
B. Load Balancing The algorithm then checks the conditions below:
In the load-balancing routine, the SON periodically per- ρo
H > ρ (o,e (n) ) (12)
o
forms load balancing by moving UEs from highly loaded
cells to lightly loaded neighboring cells based on the average RBU R net > FR (n) + D
RBU τ
ρ (τ (n) ,e (n) ) (13)
o o o

RBURs and estimated load to move. First, the centralized FRo − ρ (n)
RBU > FR (n) + D
RBU ρ (τ (n) ,e (n) ) (14)
(o,e ) o τ o o o
SON receives RBUR information from all the small cells and
calculates the average RBUR of edge UEs, ρedge , as (9). Next, Here, condition (12) restricts release of the load from over-
the SON sorts all the small cells in decreasing order of RBUR. loaded cell o so that cell o cannot become an underutilized
After sorting RBURs, the SON algorithm compares the low-loaded cell. If condition (12) is ignored, the overloaded
maximum RBUR, RBU Rma x , of the list with predefined cell will release the extra load to become underloaded, and the
threshold T h I nit in order to determine if the network needs neighbor will become overloaded, which leads the procedure
load balancing or not. If the maximum RBUR of a small cell into an infinite loop of load balancing. By comparing the
is greater than the threshold, the network is considered to be estimated after-handover load with the average network load,
overloaded, and the SON algorithm performs load balancing. conditions (13) and (14) ensure that the target neighboring
As mentioned in Section IV-A, according to the load over cell will not become overloaded and the load status of the
the network, loads to move to other cells are different. Hence, overloaded cell remains higher than the target neighboring
we set threshold T h Ada pt adaptive to network load status, as cell. If conditions (13) and (14) are ignored the target cell
in (6), in order to determine relatively overloaded cells in the will receive extra load from the overloaded neighbors and will
network. If the average load across the network, RBU R N et , become overloaded.
is greater than the prefixed threshold, T h I nit , the adaptive If the three conditions are satisfied, UEs eo(n) is accepted for
threshold is set to the network average load. Otherwise, the handover to the neighboring cell, and related cell individual
threshold is set to T h I nit . The algorithm establishes the set O offsets are updated as follows:
of overloaded cells such that RBU Ro ≥ T h Ada pt for o ∈ O CIO = M(o,e (n) ) − M(o,τ (n) ) + H yst o + ∆ (15)
o o
and O ⊂ N . Therefore, to balance the network load, the
Ocn (o,τ (n) ) = CIO (16)
algorithm sequentially takes the overloaded cells from O in o

the order of cell load, and lessens the load to underloaded Ocn (τ (n) ,o) = −CIO (17)
o
neighboring cells by moving UEs from the overloaded cells. where M(o,e (n) ) is the RSRP of overloaded cell o serving
During the load-balancing procedure, a moved load may o

make a underloaded cell overloaded, which induces ping- UE eo(n) , M(o,τ (n) ) is the RSRP of target cell τo(n) measured
o

pongs among multiple cells, which results in an infinite loop. by UE eo(n) , H yst o is a hysteresis parameter of cell o, and
To avoid such ping-pongs, we define a lower bound for cell ∆ is an increment step specified by LTE [19]. The algorithm
loads RBU Rl ow , which is defined as symmetrically updates the cell individual offsets of both the
1 serving and target cells at the same time in (16) and (17) in
RBU Rl ow = RBU R net − ρedge . order to maintain the level of ping-pongs.
2
Hence, after balancing the load, the overloaded cells do not Finally, the algorithm updates the load information of the
have an average load lower than the lower bound. serving and target cells as follows:
For an overloaded cell in O, the algorithm performs the RBU
FRo = FRo − ρ (n)
RBU (o,e ) o
following procedure. First, the algorithm computes the max-
imum load of overloaded cell o ∈ O to move, referred to as RBU
FR (n)
τ
= FR (n) + D
RBU τ
ρ (τ (n) ,e (n) )
o o o o
the moveable load, which is expressed as ρo
H = ρo − ρ (o,e (n) )
H
o

1536-1276 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2018.2789902, IEEE
Transactions on Wireless Communications
7

TABLE I work scenarios using ns-3 [32]. We considered a homogeneous

N OTATIONS USED IN THE PROPOSED MLB ALGORITHM network with small cells only. There are 10 small cells in
each of the network scenarios. In the first and second network
Notations Description
scenarios, we considered a uniform deployment of small cells,
R BU R i Average RBUR of small cell i
while considering irregularly deployed small cells in the third
R BU R N et Average RBUR of small cells in the network
scenario obtained from real network information 4 , as shown
R BU R ma x Average RBUR of the most overloaded small cell in
the network, i.e R BU R ma x = max R BU R i in Fig. 3.
i
We considered Scenario 1 to diagnose the impact of the
R
G BU R i Estimated RBUR of small cell i
proposed MLB algorithm on the network from different as-
T h I ni t Initial threshold for finding overloaded small cells
T h Ada p t Adaptive threshold for finding overloaded small
pects (Subsections VI-C, D, E, and F). However, to find the
cells , T h Ada p t = max (R BU R N et , T h I ni t ) effect of the mobility models and the velocity of users on MLB
ρ
Hi Maximum movable load from an overloaded small algorithms, we simulated Scenario 2 with different user mobil-
cell i ity models and various speeds (Subsection VI-G). Scenario 3
ρ (i, j ) Consumed load by UE j in serving cell i was considered to analyze the performance of the algorithm
ρ
D(k, j ) Estimated load of UE j at target cell k in an irregularly deployed small cell environment with pre-
defined streets and a parking area layout (Subsection VI-H).
Regarding an initial UE distribution in Scenario 1, 50%
and goes for the next candidate UE of the overloaded small of the mobile UEs were randomly distributed over the whole
cell. area, and the rest were fixed and uniformly distributed over
The above process continues for all the overloaded cells in the border areas of the small cells. The randomly distributed
the network according to the list. The notations used in the UEs followed the circular way (CW) mobility model [34],
proposed MLB algorithm are summarized in Table I, and the where UEs move in a circular path with a 10 m radius and a
process for load balancing is summarized in Algorithm 2. speed of 3.6 km/h. In Scenario 2, all the UEs were initially
Algorithm 2 Load balancing randomly distributed across the same network coverage in
Scenario 1. For the sake of mobility in Scenario 2, 70%
1: Get RBU R information from all small cells N
of the UEs were modeled with random way point (RWP)
2: Calculate ρedge of edge UEs
mobility, and the rest were modeled as random walk (RW)
3: Sort N according to RBUR and find RBU Rma x
UEs. However, UEs in Scenario 3 were randomly distributed
4: if RBU Rma x > T h I nit then
inside the predefined areas, which were modeled as streets and
5: RBU R N et ← mean(RBU Ri ) where i ∈ N
parking areas following the real data [33], and the mobility
6: T h Ada pt ← max(RBU R N et ,T h I nit )
of those UEs was modeled as RWP and RW inside that area.
7: while RBU Ri ≥ T h Ada pt do . i = {1, 2, · · · N }
Table II shows the mobility models we considered for different
8: O includes small cell i .O⊂N
network deployments.
9: end while
The bandwidth for each small cell was set to 20 MHz, which
10: for all o ∈ O do
corresponds to 100 PRBs, and transmission power was set
11: ρo = RBU Ro − RBU R net + 12 ρedge
H
to 24 dBm. The path loss was modeled using a non-line of
12: Get candidate E o . Algorithm I
FRo ← RBU Ro sight (NLOS) propagation loss [35]. We adopted a channel
13: RBU
( k E k) QoS-aware (CQA) scheduler to allocate the available resource
14: for all eo o do
blocks among the UEs, because it is more suitable for real-time
15: Compute ρ (o,e (n) )
o applications [36]. For an acceptable load threshold, we set the
16: Estimate D ρ (τ (n) ,e (n) )
o o initial value, T h I nit , to 0.75, which means that the algorithm
17: if (12), (13), and (14) are satisfied then
initiates load balancing if at least one cell exceeds 75% of
18: CIO ← M(o,e (n) ) − M(o,τ on ) + H yst o + ∆
o its RBUR. Once load balancing is triggered, the algorithm
19: Ocn (o,τ on ) ← CIO
calculates adaptive threshold T h Ada pt with (6), based on the
20: Ocn (τ on , o) ← −CIO
FRo ← RBU FRo − ρ (n) network load status. Each simulation was performed for a
21: RBU (o,e o ) duration equivalent to 10 minutes of real-time network oper-
22: FR (n) ← RBU
RBU FR (n) + D ρ (τ (n) ,e (n) ) ation, and the algorithm optimized the network every minute,
τo τo o o
23: ρo ← H
H ρo − ρ (o,e (n) ) based on measurements reported from the small cells during
o
24: end if the period. More specific parameters are given in Table III.
25: end for
26: end for
27: end if
B. Performance Evaluation Metrics
For performance evaluation, we considered two aspects:
load distribution across the network, and throughput. To mea-
VI. P ERFORMANCE E VALUATION sure load distribution, standard deviation was considered, as
A. Simulation Environments
4 The cell layout of Scenario 3 was modeled based on real 4G small cell
To evaluate the performance of the proposed algorithm, we deployment information on the campus of University of Ulsan, Korea. The
performed system-level simulations with three different net- cell layout and coverage information were collected from Sensorly [33].

1536-1276 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2018.2789902, IEEE
Transactions on Wireless Communications
8

TABLE II defined in (2). Even if the load is balanced across the network
M OBILITY M ODELS FOR T HE U SERS IN D IFFERENT S CENARIOS by MLB algorithms, throughput can be sacrificed from picking
up the wrong UEs. If an unsuitable UE is forced to move
Scenario Deployment Mobility of UEs to a neighboring cell by setting a high CIO value, the UE
1 Uniform 50% CW mobility users and 50% will experience very low SINR from the new serving cell.
static users
Consequently, the new serving cell may fail to maintain the
2 Uniform 70% RWP mobility users and 30%
RW mobility users data rate required by the UE. If a trade-off exists between
3 Irregular 70% RWP mobility users and 30% load balancing and throughput, the algorithm deviates from
RW mobility users the interests of network carriers. Hence, we also considered
the impact of load balancing on throughput.
To validate the performance of the proposed algorithm, we
TABLE III compared it with two previous algorithms: the joint MLB and
S IMULATION PARAMETERS
ICIC algorithm [11], and a traditional, fixed-threshold MLB
Parameters Values
approach [24]. We set the threshold to determine overloaded
Number of small cell 10
cells at 95%, and to perform handover at 85% for the joint
Tx power 24 dBm
MLB and ICIC algorithm [11]. There was no information
System bandwidth 20 MHz
about an optimal threshold for the fixed-threshold MLB [24].
Antenna mode Isotropic
Therefore, we conducted several simulations with different
Number of UEs 80 threshold values and found an optimal threshold value of 85%,
Path loss P L = 147.4 + 43.3l og10 (R) such that the algorithm achieved the best performance at that
Fading Standard deviation 4 dB, lognormal value. Hence, we set the threshold value at 85% for the fixed-
Resource scheduling CQA scheduler threshold MLB algorithm. For the sake of simplicity in the
C I O mi n and C I O ma x -6 dB, 6 dB figures, we denote the proposed MLB algorithm as Proposed
Hysteresis 2 dB MLB, the joint MLB and ICIC algorithm as ICIC + MLB, the
∆ 1 dB traditional fixed-threshold MLB algorithm as Fixed MLB, and
T h I ni t 0.75 (75% of total PRBs) simulations without an MLB algorithm as No MLB.

C. Impact of MLB on Load Distribution

To identify the impact of the MLB algorithms on load
distribution across the cells, we simulated a uniformly de-
ployed small-cell network with CW-mobility UEs and static
UEs (Scenario 1 in Table II). The scenario was simulated with
80 UEs, and data traffic for each of the UEs was set at a
guaranteed bit rate (GBR) of 1 Mbps. Fig. 4 expresses the load
distribution in terms of RBUR for the network scenario that
does not adopt the MLB algorithm, as well as for the scenario
with our proposed MLB algorithm. It is clear that, as time
goes on, the load among the small cells becomes more evenly
distributed when the proposed MLB algorithm is adopted.
Fig. 4a shows that after 10 minutes of network operation
(a) without considering an MLB algorithm in the network, cells 5
and 6 are extremely overloaded with an RBUR of 0.98.
However, cells 3 and 8, which are neighbors of cells 6 and 5,
respectively, remained underutilized at an RBUR of 0.72. The
difference between the maximum and minimum RBURs of
the network is 0.26. On the other hand, when the proposed
MLB algorithm is implemented, as can be seen from Fig. 4b,
the overloaded small cells released some of their load to
the underloaded neighboring small cells. As a consequence,
the load across the small cells in the network became more
balanced. The figure shows that the RBUR of cell 8 reached
the maximum 0.96 after 10 minutes, and cell 5 ended up at an
RBUR of 0.86, the minimum among the small cells. Hence,
the gap between the maximum and minimum RBUR of the
(b)
network was reduced from 0.26 to 0.10, which means the
Fig. 3. Radio environment map: (a) uniformly deployed small-cell network variance in the load among the cells in the network is low,
(b) irregularly deployed small-cell network.
and the network is well balanced.

1536-1276 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2018.2789902, IEEE
Transactions on Wireless Communications
9

1 1
Proposed MLB
ICIC + MLB
0.95 Fixed MLB
0.95 No MLB
0.9

RB Utilization
RBUR

0.85 0.9

0.8
0.85
0.75

0.7 0.8
13579 13579 13579 13579 13579 13579 13579 13579 13579 13579 1 2 3 4 5 6 7 8 9 10
1 min 2 min 3 min 4 min 5 min 6 min 7 min 8 min 9 min 10 min Time (min)

(a) (a)

1 8
Proposed MLB
ICIC + MLB
0.95 Fixed MLB
7.8
No MLB

Throughput (Mbps)
0.9
7.6
RBUR

0.85
7.4
0.8

7.2
0.75

0.7 7
13579 13579 13579 13579 13579 13579 13579 13579 13579 13579 1 2 3 4 5 6 7 8 9 10
1 min 2 min 3 min 4 min 5 min 6 min 7 min 8 min 9 min 10 min Time (min)

(b) (b)
Fig. 4. RBUR status of the network with 80 UEs and 1 Mbps connectivity Fig. 6. Impact of load balancing on resource utilization and throughput in
per UE: (a) without an MLB algorithm (b) with the proposed MLB algorithm. the network with 80 UEs at 1 Mbps data rate per UE: (a) Resource utilization
(b) Average cell throughput.

0.16 No MLB
Fixed MLB
0.14 ICIC + MLB hand, when an MLB algorithm is adopted, UEs are distributed
Proposed MLB
from the overloaded small cells to the low-loaded small cells,
Standard Deviation

0.12

0.1 and release resources in the overloaded cells. Therefore, the

0.08 small cells can fully allocate the required resources to the
0.06 UEs, which leads to increased overall resource utilization.
0.04
Moreover, while load balancing, the intentionally shifted UEs
0.02
may experience relatively low signal quality from the new
0
1 2 3 4 5 6
Time (min)
7 8 9 10 serving cells; therefore, the new serving cells may require
more resources to serve those shifted UEs, as estimated in
Fig. 5. Standard deviation of RBUR among cells in the network. (8). In Fig. 6a, resource utilization is higher with our proposed
algorithm.
The proposed algorithm increases network throughput as
Fig. 5 shows the performance of three different MLB
well. Fig. 6b shows the average cell throughput with different
approaches in terms of standard deviation of RBUR among the
MLB algorithms for the considered network scenario. Since
small cells in the network. The standard deviation of RBURs
the overloaded cells are unable to allocate the required re-
with the proposed adaptive MLB algorithm in the network
sources to the UEs, the required data rate for the UEs cannot
becomes smaller, compared with the other MLB algorithms.
be provided in those cells. Therefore, when the network has
no MLB algorithm, UE throughput in the overloaded cells
D. Impact of MLB on Resource Utilization and Throughput is limited by the resource deficit. However, MLB algorithms
While load balancing, the overall resource utilization of the evenly distribute the load across the cells and enable the cells
network may be affected. Fig. 6a shows the average resource to allocate the required resources to the UEs, which leads to
utilization with different MLB algorithms for the network increased network throughput. In the figure, the network has
scenario considered in Subsection VI-C. Resource utilization a higher average cell throughput when adopting our proposed
in the network increased while adopting the MLB algorithms. MLB algorithm.
In the network without an MLB algorithm, the overloaded
small cells are unable to allocate the required resources to the
UEs due to a resource deficit; meanwhile, the low-loaded small E. Impact of Offered Load Variation
cells remain underutilized, as shown in Fig. 4a. As a result, To study the effect of traffic load variation, we simulated the
resources in the network remain underutilized, but the UEs same scenario in the previous subsection with different traffic
are denied from the required resource allocation. On the other load per UE, but kept the total number of UEs unchanged.

1536-1276 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2018.2789902, IEEE
Transactions on Wireless Communications
10

The traffic loads per UE was set at 128 kbps, 256 kbps,
512 kbps, and 1 Mbps. Fig. 7 shows the performance of the
MLB algorithms under the four different traffic loads in the 0.1

network after 10 minutes of network operation. 0.08

Standard Deviation
In a low traffic load situation, when most of the small
0.06
cells remain underloaded, the previous algorithms show better
performance, as shown in Fig. 7a. However, when the traffic 0.04

load increases heavily and most of the small cells become No MLB
0.02 ICIC + MLB
overloaded, the joint MLB and ICIC algorithm and the fixed- Fixed MLB
Proposed MLB
threshold MLB algorithm fail to balance the load. On the 0
128kbps 256kbps 512kbps 1Mbps
Load per UE
other hand, by changing the threshold value according to the
network load conditions, the adaptive MLB algorithm balances (a)
the load among the cells.
8
Fig. 7b shows the average cell throughput with different
MLB algorithms in the network after 10 minutes of network 7

operation. Although in the low traffic situation the MLB 6

Throughput (Mbps)
algorithms balanced the load well, they have very little impact 5

on throughput. However, with the increase in traffic, average 4

cell throughput increases when the MLB algorithms are con- 3

Proposed MLB
sidered. Among the MLB algorithms, with the proposed MLB 2 ICIC+MLB
algorithm in the network, the increment is at maximum. 1
Fixed MLB
No MLB
We studied the change in the threshold value of the proposed 128kbps 256kbps 512kbps
Load per UE
1Mbps

algorithm with different loads in the network. Fig. 8 shows that

(b)
for a traffic load where each of the UEs has a 1 Mbps GBR,
the threshold is 0.89; however, when GBR is 128 kbps per Fig. 7. Performance of different MLB algorithms under different traffic loads
in the network after 10 minutes: (a) Standard deviation of RBUR among cells
UE, the threshold value changed to 0.81. The figure illustrates (b) Average cell throughput in the network.
that our proposed algorithm balances the load adaptive to the
variant load in time, and various levels of load, by changing
1
the threshold according to the network load status, whereas 1 Mbps
128 kbps
algorithms with a fixed threshold cannot. 0.95

0.9
Threshold

F. Variance in the Number of UEs 0.85

To determine the impact of per-UE traffic on MLB ap- 0.8

proaches, we simulated the evenly deployed small-cell network 0.75

with CW mobility and fixed UEs, as seen in Scenario 1 of
0.7
Table II. We simulated the network with different numbers 0 1 2 3 4 5
Time (min)
6 7 8 9 10

of UEs but kept the total traffic demand unchanged. In the

first simulation, we considered 80 UEs with connectivity of Fig. 8. The value of T h Ada p t for the network at two different traffic loads.
512 kbps, whereas 40 UEs with 1 Mbps connectivity was
considered in the second simulation. We kept the total data
demand in the network at 40 Mbps in both simulations. identical in both the simulations, the variation in the number of
Fig. 9 shows the performance comparison among the dif- UEs results in different average throughput. However, in both
ferent MLB approaches, which clarifies that per-UE traffic simulations, our proposed algorithm shows better performance
has a significant impact on load balancing. In Fig. 9a, the than the other MLB algorithms.
standard deviation for RBUR with 40 UEs at 1 Mbps per UE
is higher than the standard deviation with 80 UEs at 512 kbps
per UE. The performance of MLB algorithms with UEs that G. Impact of User Mobility on MLB
have low per-UE traffic demand is better than UEs with high In practice, mobile users may have different mobility, and
per-UE demand. UEs with a higher data rate required more velocity as well, which can affect the performance of the
RBs than UEs with a low data rate, and consequently, the MLB algorithm. Therefore, to scrutinize the effect of user
MLB algorithms have no chance to fine-tune the load among mobility we simulated Scenario 2, which has the same small
the cells. On the other hand, shifting UEs with low traffic cell deployment as Scenario 1 but different user mobility
demand transfers a small load from the overloaded cells to models. In Scenario 2, we adopted RWP and RW mobility
the under-loaded cells. As a result, the MLB algorithms can models instead of the CW and static models of Scenario 1.
distribute the load more evenly among the cells. For RWP, we considered two different UE velocities (low and
Fig. 9b shows the average cell throughput observed from high). At the low velocity, we randomly set the UE velocity to
both the simulations. Although we kept the total data demand a pedestrian speed of between 5 km/h and 10 km/h. However,

1536-1276 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2018.2789902, IEEE
Transactions on Wireless Communications
11

0.35 0.3
No MLB
ICIC + MLB
0.3 0.25
Fixed MLB
Proposed MLB

Standard Deviation
0.25
Standard Deviation

1Mbps−40UE
512kbps−80UE 0.2
0.2
0.15
0.15
No MLB
0.1 ICIC + MLB
0.1
Fixed MLB
0.05 Proposed MLB
0.05
Low velocity
High velocity
0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Time (min) Time

(a) (a)

8
3.8
7.8

Throughput (Mbps)
Throughput (Mbps)

Proposed MLB
3.7 Fixed MLB
ICIC + MLB 7.6
No MLB
3.6 1Mbps−40UE
512kbps−80UE 7.4 Proposed MLB
Fixed MLB
3.5 ICIC + MLB
7.2 No MLB
Low velocity
High velocity
3.4 7
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Time (min) Time (min)

(b) (b)
Fig. 9. Performance of different MLB algorithms in the network for different Fig. 10. Performance of different MLB algorithms while considering various
numbers of UEs keeping the total data demand for network throughput at UE velocity in the network: (a) Standard deviation of RBUR among the cells
40 Mbps: (a) Standard deviation of RBUR among the cells (b) Average cell (b) Average cell throughput in the network.
throughput in the network.

floor plans. Therefore, we scrutinized the robustness of the

at the high velocity, the speed of UEs was set between 10 km/h proposed MLB algorithm in an irregularly deployed small-
and 30 km/h, considered running and cycling speeds. cell network (Scenario 3 in Table II) with a realistic mobility
Fig. 10 shows the performance of different MLB algorithms pattern following real small cell deployment information [33].
in the network when different user velocities are considered. Furthermore, considering the variations in the number of
Since Scenario 1 has a high ratio of static UEs, and the speed users in reality, we also investigated the performance of the
of CW-mobility UEs is relatively low (3.6 km/h), the standard proposed MLB algorithm with different numbers of UEs in
deviation converges smoothly and remains steady in Fig 5. the considered network.
However, Fig. 10a shows fluctuations in the standard deviation 1) Impact of cell layout: To figure out the impact of cell
when considering different UE velocities in Scenario 2. Due to layouts and user mobility patterns we simulated Scenario 3
the higher user mobility in Scenario 2, more frequent change in with 80 UEs. We set the data rate at 1 Mbps GBR per UE, and
UE topology and handovers make an oscillated load situation the velocities of the UEs were set randomly between 5 km/h
among the cells. Hence, the MLB algorithms cannot achieve and 10 km/h (considered pedestrian speed). The performance
optimal load balancing, and the figure shows fluctuations of different MLB algorithms in the irregularly deployed small-
in standard deviation. However, despite user mobility, our cell network is shown in Fig. 11. Fig. 11a shows that our
proposed algorithm outperforms the other MLB approaches proposed MLB algorithm balanced the load across the network
in both mobile environments. well, even in the irregular deployment with various cell sizes,
Fig. 10b shows the average cell throughput in the network and outperforms the other MLB approaches, as seen in the
with the MLB algorithms. In the figure, the average cell previous scenarios. The more evenly load distribution by
throughput at the high velocities is lower than at the low the proposed algorithm leads to an increased average cell
velocities because the MLB algorithms are unable to achieve throughput in the network. Fig. 11b shows that the average
optimal load balancing due to load oscillation. However, with cell throughput in the network is also the highest when the
various MLB approaches, the average cell throughput in the proposed MLB approach is adopted in the network.
network increased. With our proposed MLB algorithm in both
2) Impact of the number of UEs: To study the effect of
velocity environments, the average cell throughput is higher
changes in the number of UEs in the network, we simu-
than the others.
lated Scenario 3 with various numbers of UEs. The standard
deviation and the average cell throughput were measured
H. Impact of Irregularity on MLB after 10 minutes of network operation, and the performance
The practical deployment of small cells may be irregular, of MLB algorithms are compared in Fig. 12. As discussed
and user mobility patterns may vary according to road and in Subsection VI-E, the proposed MLB algorithm updates

1536-1276 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2018.2789902, IEEE
Transactions on Wireless Communications
12

0.3 0.3
No MLB
ICIC+MLB
0.25 0.25 Fixed MLB
Proposed MLB
Standard Deviation

Standard Deviation
0.2 0.2

0.15 0.15

0.1 0.1
No MLB
0.05 ICIC+MLB 0.05
Fixed MLB
Proposed MLB
0 0
1 2 3 4 5 6 7 8 9 10 20 40 60 80 100
Time Number of users

(a) (a)

8 10

9
7.8
8
Throughput (Mbps)

Throughput (Mbps)
7
7.6
6

7.4 5

4
Proposed MLB Proposed MLB
7.2 ICIC+MLB 3 ICIC+MLB
Fixed MLB Fixed MLB
No MLB 2 No MLB
7
1 2 3 4 5 6 7 8 9 10 20 40 60 80 100
Time (min) Number of users

(b) (b)
Fig. 11. Performance of different MLB approaches in irregularly deployed Fig. 12. Performance of different MLB algorithms in irregularly deployed
small-cell network with 80 UEs at 1 Mbps per UE: (a) Standard deviation of small-cell network with various numbers of UEs: (a) Standard deviation of
RBUR among cells (b) Average cell throughput in the network. RBUR among cells (b) Average cell throughput in the network.

the triggering threshold adaptive to the network load, which ACKNOWLEDGMENT

enables the algorithm to cope with the different load situations,
This work was partly supported by Basic Science Re-
and distributes the load more evenly than the other MLB algo-
search Program through the National Research Foundation
rithms, as shown in Fig. 12a. Fig. 12b shows that the proposed
of Korea (NRF) funded by the Ministry of Science, ICT &
algorithm also outperforms the other MLB algorithms in terms
Future Planning (2015R1A1A1A05001069), and the Institute
of average cell throughput in all situations. Therefore, the
for Information & communications Technology Promotion
simulation results from different aspects demonstrate that in
(IITP) grant funded by the Korea government (MSIT) (No.
a realistic network scenario, the proposed MLB algorithm is
2014-0-00282, Development of 5G Mobile Communication
able to achieve more even load balancing.
Technologies for Hyperconnected smart services).
VII. C ONCLUSION
R EFERENCES
In this paper, we proposed an adaptive algorithm for mo-
bility load balancing in a small-cell network in order to [1] “Cisco visual networking index: Global mobile data traffic forecast
update, 2016-2021 white paper,” White Paper, Cisco, March 2017.
improve network performance, such as the service drop rate [2] J. Hoadley and P. Maveddat, “Enabling small cell deployment with
and throughput. Due to lack of resource, overloaded cells HetNet,” IEEE Wireless Communications, vol. 19, no. 2, pp. 4–5, 2012.
cannot allocate the amount of required resource to UEs even [3] J. G. Andrews, H. Claussen, M. Dohler, S. Rangan, and M. C. Reed,
“Femtocells: Past, present, and future,” IEEE Journal on Selected Areas
if underloaded neighboring cells are capable to serve. Since in Communications, vol. 30, no. 3, pp. 497–508, 2012.
such unbalanced loads bring throughput degradation, load [4] “Lte release 12 and beyond: Nokia white paper,” White Paper, Nokia
balancing can improve network performance. In the case of Network, 2014.
[5] “Small cells, what’s the big idea?” Small Cell Forum (SCF), Tech. Rep.
LTE, the resource usage depends on wireless signal quality and SCF030, Feb. 2012.
traffic demands of UEs. Hence, to develop a load balancing [6] Qualcomm: The 1000x data challenge. [Online]. Available:
algorithm, we defined a resource block-utilization ratio as a https://www.qualcomm.com/invention/1000x
[7] Y. Bian and D. Rao, “Small cells big opportunities,” Global Business
measurement of cell load, and employed an adaptive threshold Consulting. Huawei Technologies Co., Ltd, 2014.
to determine overloaded cells according to the network load [8] 3GPP, “Scenarios and requirements for small cell enhancements for E-
situation. To avoid ping-pongs of load among the cells, we UTRA and E-UTRAN,” 3rd Generation Partnership Project (3GPP), TS
36.932.
considered not only overloaded cells but also their neighboring [9] 3GPP, “Evolved universal terrestrial radio access network (E-UTRAN);
cells by estimating the moving load’s impact on the existing self-configuring and self-optimizing network (SON) use cases and
cell loads. Via simulations, we showed that our proposed solutions,” 3rd Generation Partnership Project (3GPP), TS 36.902.
[10] S. Feng and E. Seidel, “Self-organizing networks (SON) in 3GPP long
algorithm balanced more of the load across the network, and term evolution,” Newsletter, Nomor Research GmbH, Munich, Germany,
provided higher throughput than previous algorithms. Tech. Rep., May 2008.

1536-1276 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2018.2789902, IEEE
Transactions on Wireless Communications
13

[11] N. O. Tuncel and M. Koca, “Joint mobility load balancing and inter- with OMNET++,” in Proc. 12th International Conference Distributed
cell interference coordination for self-organizing OFDMA networks,” Computing and Artificial Intelligence. Springer, 2015, pp. 251–258.
in Proc. IEEE 81st Vehicular Technology Conference (VTC Spring), [35] J. B. Andersen, T. S. Rappaport, and S. Yoshida, “Propagation mea-
Glasgow, Scotland, May 2015, pp. 1–5. surements and models for wireless communications channels,” IEEE
[12] N. Zia and A. Mitschele-Thiel, “Self-organized neighborhood mobility Communications Magazine, vol. 33, no. 1, pp. 42–49, 1995.
load balancing for LTE networks,” in Proc. Wireless Days (WD), 2013 [36] J. M. Ruiz-Avilés, S. Luna-Ramírez, M. Toril, F. Ruiz, I. de la Bandera,
IFIP, Valencia, Spain, Nov. 2013, pp. 1–6. P. Muñoz, R. Barco, P. Lázaro, and V. Buenestado, “Design of a com-
[13] Z. Huang, J. Liu, Q. Shen, J. Wu, and X. Gan, “A threshold-based putationally efficient dynamic system-level simulator for enterprise LTE
multi-traffic load balance mechanism in lte-a networks,” in 2015 IEEE femtocell scenarios,” Journal of Electrical and Computer Engineering,
Wireless Communications and Networking Conference (WCNC), March vol. 2012 (2012).
2015, pp. 1273–1278.
[14] X. Zhang, Q. Li, X. Gu, W. Li, S. Jia, and L. Zhang, “The study of
load balancing based on the network flow theory in LTE-A systems,” in
Proc. IEEE International Symposium on Personal, Indoor and Mobile
Radio Communications (PIMRC Workshops), London, United Kingdom,
2013, pp. 91–95.
[15] 3GPP, “Evolved universal terrestrial radio access network (E-UTRAN);
X2 application protocol (X2AP),” 3rd Generation Partnership Project
(3GPP), TS 36.423.
[16] 3GPP, “Evolved universal terrestrial radio access network (E-UTRAN); Md Mehedi Hasan received the B.Sc. degree in
S1 application protocol (S1AP),” 3rd Generation Partnership Project computer science and engineering from the Univer-
(3GPP), TS 36.413. sity of Chittagong, Chittagong, Bangladesh, in 2009.
[17] 3GPP, “Evolved universal terrestrial radio access (E-UTRA); physical From 2012 to 2014 he worked as a software engi-
channels and modulation,” 3rd Generation Partnership Project (3GPP), neer in two different leading software development
TS 36.211. company in Bangladesh. Currently he is doing M.S.
[18] 3GPP, “Evolved universal terrestrial radio access (E-UTRA); FDD leading Ph.D. degrees in electrical engineering at
repeater radio transmission and reception,” 3rd Generation Partnership University of Ulsan, Korea. His current research in-
Project (3GPP), TS 36.106. terests include mobile network optimization, energy
[19] 3GPP, “Evolved universal terrestrial radio access (E-UTRA); radio efficient mobile communication, mobility manage-
resource control (RRC); protocol specification,” 3rd Generation Part- ment, traffic offloading and load balancing.
nership Project (3GPP), TS 36.331.
[20] S. Ahmadi, LTE-Advanced A Practical Systems Approach to Under-
standing the 3GPP LTE Releases 10 and 11 Radio Access Technologies.
Academic Press, 2013, ch. 5.4, pp. 257–261.
[21] Y. Lee, B. Shin, J. Lim, and D. Hong, “Effects of time-to-trigger
parameter on handover performance in son-based lte systems,” in Proc.
IEEE Asia-Pacific Conference on Communications (APCC),. IEEE,
2010, pp. 492–496.
[22] R. Kwan, R. Arnott, R. Paterson, R. Trivisonno, and M. Kubota, “On
Sungoh Kwon (S’05 / M’08) received the B.S. and
mobility load balancing for LTE systems,” in Proc. IEEE 72nd Vehicular
M.S. degrees in electrical engineering from KAIST,
Technology Conference (VTC Fall), Ottawa, Canada, 2010, pp. 1–5.
Daejeon, Korea, and the Ph.D. degree in electrical
[23] P. Legg, G. Hui, and J. Johansson, “A simulation study of lte intra-
and computer engineering from Purdue University,
frequency handover performance,” in Proc. IEEE 72nd Vehicular Tech-
West Lafayette, IN, in 1994, 1996, and 2007, respec-
nology Conference (VTC Fall), Ottawa, Canada, 2010, pp. 1–5.
tively. From 1996 to 2001, he was a research staff
[24] A. Lobinger, S. Stefanski, T. Jansen, and I. Balan, “Load balancing in
member with Shinsegi Telecomm Inc., Seoul, Korea.
downlink LTE self-optimizing networks,” in Proc. IEEE 71st Vehicular
From 2007 to 2010, he developed LTE schedulers as
Technology Conference (VTC Spring), Taipei, Taiwan, May 2010, pp.
a principal engineer in Samsung Electronics Com-
1–5.
pany, Ltd., Korea. He has joined to University of
[25] C. E. Shannon, “A mathematical theory of communication,” ACM Ulsan as an Assistant Professor with the School of
SIGMOBILE Mobile Computing and Communications Review, vol. 5, Electrical Engineering since 2010, and is now an Associate Professor with the
no. 1, pp. 3–55, 2001. same school. His research interests are in wireless communication networks.
[26] I. S. Gradshteyn and I. M. Ryzhik, Table of integrals, series, and
products. Academic Press, 2014, pp. 18–19.
[27] A. S. Hamza, S. S. Khalifa, H. S. Hamza, and K. Elsayed, “A survey on
inter-cell interference coordination techniques in OFDMA-based cellular
networks,” IEEE Communications Surveys & Tutorials, vol. 15, no. 4,
pp. 1642–1670, 2013.
[28] Y. Medjahdi, M. Terre, D. L. Ruyet, D. Roviras, J. Nossek, and
L. Baltar, “Inter-cell interference analysis for OFDM/FBMC systems,”
in Proc. IEEE Workshop on Signal Processing Advances in Wireless
Communications, Perugia, Italy, Jan. 2009, pp. 598–602. Jee-Hyeon Na is a director in radio access network
[29] M. L. Rocca. RSRP and RSRQ measurement in LTE. S/W section at ETRI, where she has been employed
[Online]. Available: http://laroccasolutions.com/training/78-rsrp-and- since 1989 after receiving BS degree in computer
rsrq-measurement-in-lte/ science from CheonNam National University. She
[30] M. L. Rocca. RSRQ to SINR relation. [Online]. Available: received her Ph.D. and MS in computer science
http://laroccasolutions.com/training/164-rsrq-to-sinr/ from ChungNam National University in 2008 and
[31] A. Farhana, S. Ramprasad, H. Roshanak, S. Kumbesan, and A. So- 2002 respectively. She is a member of IEEE and
laiman, “SINR, RSRP, RSSI and RSRQ measurements in long term IEICE communication part. Her research interests
evolution networks,” International Journal of Wireless & Mobile Net- are in the area of 4G, 5G small cells, enhanced
works (IJWMN), vol. 7, no. 4, 2015. Mobile Broadcast and Multicast Services and Self-
[32] Discrete-event network simulator for internet systems - ns3. [Online]. Organizing Networks.
Available: https://www.nsnam.org
[33] Sensorly: Unbiased wireless network information. [Online]. Available:
http://www.sensorly.com
[34] C. Ley-Bosch, R. Medina-Sosa, I. A. González, and D. S. Rodríguez,
“Implementing an IEEE 802. 15.7 physical layer simulation model

1536-1276 (c) 2017 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.