Characterizing the Effects of Rapid LTE

Deployment: A Data-Driven Analysis

Kareem Abdullah? , Noha Korany? , Ayman Khalafallah? , Ahmed Saeed† , Ayman Gaber‡
? Alexandria University, † Georgia Institute of Technology, ‡ Vodafone Egypt

Abstract—The deployment of LTE, while old news to developed can be very helpful for community-operated cellular networks
countries, is ongoing in developing countries with some starting that service underdeveloped regions (e.g., [11]).
LTE deployments only a year ago. LTE is essential for developing Standard 4G cellular network deployment implements the
countries where broadband access is mostly available through
wireless connection. The process of LTE deployment is costly following steps [12, 13, 14]. First, network dimensioning
and time consuming. Network operators attempt to minimize takes into account user demand and density. It also accounts
both overheads. With the current challenges facing network for the desired Quality of Service (QoS) provided by the
operators to acquire new sites, 4G deployment always depends on operator. Then, network planning determines optimal radio
reusing the already existing 2G/3G macro layer. In this paper, sites locations and parameters. To achieve rapid deployment,
we study the approach of rapid 4G deployment by analyzing
data from a major network operator in a developing country. In it is typical for mobile operators to reuse the already existing
particular, we look at network measurements from a single large 3G radio sites and infrastructure for 4G deployment. This
cluster right after LTE deployment. We analyze the impact of approach capitalizes on two factors: 1) having a single Radio
the deployment approach on user throughput and find that LTE Access Network (RAN) that can already support both 3G and
throughput could be improved further for the majority of LTE LTE networks, and 2) the overlap in supported spectrum for
users in more than 50% of the studied cells by improving either
interference or coverage. We find that this is mainly because the both networks. However, 3G and 4G technologies are signif-
LTE system is more sensitive to interference when compared to icantly different. This requires several extensive optimization
3G. Finally, we show a data-driven approach to detect the affected operations to be conducted before and after 4G activation.
cells and mitigate the issue through physical optimization, this The optimization operations are determined based on site
approach balances employed LTE transition best practices with measurements, collected terrain information, and capacity es-
cost efficiency and rapid deployment.
timation. The outcome is a set of new radio antenna param-
eters for eNodeBs (4G radio sites) to ensure that each site
I. I NTRODUCTION
achieves its coverage without interference. This operation is
The deployment of 4G/LTE networks marked a historic tran- time consuming and is known to be costly [15] which conflicts
sition in how people consumed digital content on their phone with any operator objectives, motivating further reduction in
[1]. This is especially critical for developing countries where deployment steps. In this paper, we assess the side effects
reliable and affordable broadband access might be limited to of reducing the deployment overhead by reducing the cost of
cellular networks [2]. This transition is somewhat old news post-activation optimization.
in developed countries (e.g., US and Japan had their initial Post-activation optimization is needed due to the difference
deployments in 2010 [3]). However, in developing countries between physical layer protocols used in 3G and 4G networks.
this transition is fairly recent and ongoing (e.g., El Salvador In particular, 3G networks tolerate overlapping coverage be-
[4] and Egypt [5] had their initial deployments in 2017). The tween multiple cells, while such conditions cause problems
deployment of 4G/LTE was marked by fierce competition for 4G networks. In 3G networks, Soft Handover (SHO)
between network operators trying to reach the market first. handles interference gracefully by allowing a single user to
This competition can determine the benefits gained by the receive data from multiple cells at the same time [16]. SHO is
operator from the network upgrade [6]. That requires rapid enabled by Code Division Multiple Access (CDMA). On the
deployment of 4G network, cost efficiency and attempting to other hand, graceful handling of interference is not feasible in
exploit the full benefit of 4G spectrum efficiency. 4G/LTE networks because it relies on Orthogonal Frequency-
With 5G deployments already underway, 4G deployment Division Multiplexing (OFDM) which does not have the same
techniques remain relevant for several reasons. First, 5G macro-diversity as CDMA. Overlapping coverage areas in
deployment includes expansion of use of 4G bands, relying 4G networks lead to several problems including ping-pong
on similar modulation and spectrum access techniques used handovers, handover failures, waste of the network resources,
in 4G deployments [7]. Second, cellular operators all over the and degradation in signal quality [17].
world are currently just starting to perform “3G shutdown” Performing rigorous post-activation optimization to handle
[8, 9]. This operation is performed, hand in hand with 5G the difference between 3G and 4G is one of the costliest
deployment, in order to keep up with customer demand by steps in the deployment. That is because it involves reviewing
refarming spectrum used by 3G networks to be used by 4G net- the coverage of tens of thousands of cells based on some
works instead [10]. Finally, rapid 4G deployment techniques propagation and interference models then verifying the plans
using rigorous drive tests. This cost can be reduced when
3G
taking into account the effects of overlapping coverage areas
on 3G networks. In particular, overlapping coverage areas F1 F1
waste resources in 3G networks. Moreover, SHO has limi- UE 1
tations in terms of maximum number of overlapping cells it
can handle. Hence, a well-planned 3G network should have 3G NodeB 3G NodeB

minimal overlap between coverage areas. Site A Site B

Soft Handover
In this paper, we study the rapid deployment of 4G, testing Region
the idea of reusing the existing 3G sites as is1 , minimizing
the extensive post-activation optimization steps efforts. This (a) In 3G networks, impact of overlapping coverage is mitigated
through SHO.
is enabled by relying on a well-planned 3G network with
minimal overlap in coverage areas. Our goal is to characterize 4G
the side effects of the rapid deployment. We refer to these
side effects as bad radio conditions. Then, we determine F1 F1
their extent in terms of both their effect on user performance UE2
as well as their prevalence in the studied sites. Finally, we
present a simple approach to handle affected site. Our aim 4G eNodeB 4G eNodeB

is to reach the sweet spot between cost reduction, rapid Site C Site D
Serving Site Inter-cell Neighbor Site
deployment and optimum user experience. We conducted our Interference Region

study based on analysis of network traces and measurements (b) In 4G networks, overlapping coverage areas cause interference,
from a major network operator. We collected data from a single hurting user observed performance.
cluster of cells covering two large cities, spanning two months
in 2017 which represents the initial state of the network Figure 1: Comparison between the effect of overlapping cov-
right after the rapid LTE deployment. The collected traces erage areas in 3G and 4G networks.
and measurements capture a snapshot of the operator’s 4G
deployment just before the implementation of post-activation with poor radio conditions. Furthermore, we validate that
optimization. This enables us to assess the value and side the main cause of the problem is overlapping coverage areas
effects reducing the cost of post-activation optimization. between cells by looking at sectors of the network with zero
Our work is motivated by the appeal of reusing sites and overlap (§V).
equipment with minimal to no hardware modifications to move • What is a cost-effective resolution of this problem? We
from 3G to 4G. This low cost operation should benefit cellular employ physical optimization, which is discussed in details
operators deploying 4G for the first time as well as performing in (§VI), guided by our developed approach of detecting the
3G shutdown. It can also be useful for community operated affected cells, to improve their performance. In particular,
cellular networks (e.g., [11]). Furthermore, our work motivates we employ antenna parameters optimization techniques to
analytical models and software tools that can determine the reduce or eliminate the effects of interference, leading to
parameters of the 4G deployment based on the 3G network user throughput improvement by up to 114%. We believe
the operator starts with, as well as Inter-Cell Interference this approach to be the best practice in rapid 4G deployment
Coordination (ICIC) [18] techniques that takes into account as it balances performance and cost efficiency.
the deployment approach.
II. BACKGROUND
In brief, our work in this paper aims at answering the
following questions about the rapid 4G deployment approach: User devices, such as smartphones, tablets or modem cards
• What is a good indicator of bad radio conditions? We
connect to a radio cell over a certain radio frequency or a
study correlation between different cell Key Performance carrier. Each cell covers a geographic area with a directional
Indicators (KPIs) and user throughput. We find that the antenna and it is common to find 3 such cells covering a full
most contributing factor affecting user throughput is the circle, approximately 120 degrees each, but there can be more
number of users within a cell using high-order modulation or fewer cells with different coverage areas. Multiple cells
and coding scheme (§IV). covering the same direction and area can be called a sector.
• How prevalent are poor radio conditions, caused by cov-
For coverage and capacity, there are typically multiple cells
erage overlap between neighboring sites, in the operator’s per base station, anywhere from 3 to 12, sometimes even more.
network? Our goal is to help network operators to gain an There can be hundreds of thousands of cells in the network. In
insight of the extent of side effects of transitioning from 3G LTE, the site which spans all LTE sectors is called eNodeB,
to 4G keeping post-activation optimization to a minimal. which is the evolution of 3G NodeBs. A Cluster is a group
We use the bad radio condition indicators to identify cells of radio sites covering a certain geographical area defined by
the network operator.
1 We omit the details of site architecture as it is part of our operators Typical LTE deployment relies on network dimensioning in
confidential information. which user demand, density, along with the operator’s desired
QoS provisioning are used to plan base stations placement and 88
capacity. This step goes hand in hand with a coverage planning

Accuracy (%)
84
phase where the parameters of the antenna of each base station 80
is determined such that all areas of interest are covered and 76
inter-cell interference is reduced. This approach is standard 72
and discussed in several textbooks (e.g., [12, 13, 14]). Cover- 0 2 4 6 8 10 12 14 16 18
age planning is one of the costliest steps in the deployment. Number of Features
That is because it involves reviewing the coverage of tens of Figure 2: Throughput estimation accuracy improvement using
thousands of cells based on some propagation and interference sequential forward selection, showing that the first six features
models then verifying the plans using rigorous drive tests. This are enough to provide best estimation accuracy.
operation is both costly and time consuming which conflicts
with the goals of the operators that are eager to be the first to
deploy LTE in a market that is hungry for more bandwidth. III. DATA S ET
LTE Rapid deployment leverages two advantages the operator
has: 1) a single RAN that operates both 3G and LTE networks We use a set of KPIs collected by a top-tier mobile
allowing the same equipment to be used for either of the two operator with more than 30 million subscribers serving an
technologies, and 2) re-farming of 2G and 3G spectrum for entire country. All of our measurements are collected from the
LTE operation, which is a technique by which spectrum of operator network point of view. The data we use is collected
older technologies with diminishing demand is used by newer from a cluster within the operator’s network that spans two
technologies [10]. Spectrum re-farming allowed the utilization major cities. We collect data spanning two months in 2017 that
of the same antenna configuration used in 3G to be used in represent the state of the network immediately post rapid LTE
LTE as well. Following this approach, the upgrade from 3G to deployment on that cluster. All data reported are anonymized
LTE became a matter of upgrading infrastructure equipment and collected in aggregate.
as most of the other components, including antennas and radio The collected KPIs are measured at the eNodeBs, capturing
modules where used as is for LTE. The main challenge of this both network KPIs and per-cell aggregate end-user KPIs. Our
approach is handling the difference between how 3G and LTE measurements are limited to KPIs collected and aggregated by
networks handle overlapping coverage areas. eNodeB equipment. We follow this approach to maintain the
ability to collect data at scale by avoiding tracking KPIs of
In 3G networks, users falling in the area of coverage overlap individual users. This allows us to provide an approach that
between multiple cells don’t exhibit severe problems. This can be realistically used by the operator to detect cells affected
is due to the macro-diversity technique that’s employed by by the deployment methodology.
the CDMA and W-CDMA standards. This makes use of a KPI measurements are collected at 15-minute interval span-
feature called soft handover (SHO) where a cell phone is ning two consecutive months. We believe that this period and
simultaneously connected to two or more cells extending monitoring granularity to be representative of normal network
its capacity, shown in Figure 1a [16]. While this approach operations capturing hourly load variability and trends. We
mitigates the effects of interference from the point of view of filter the data to focus on trends at peak hours because that
the user, it is still problematic from the point of view of the is what the network provisions for. As indicated earlier, user
operator. SHO comes at the expense of available resources, information is collected in aggregates per cell and does not
as one user will utilize those resources for the two, or more, contain any personal or identifiable information about owners
base stations, while it’s preferred to connect from one cell of devices or exact base stations info. The data is in a columnar
only. Hence, overlapping areas should be minimized as much format, with 4.2 million records, each record represents the
as possible, though, it won’t affect the customer experience in KPIs of a certain cell over the 15-minute period.
3G as we’ve illustrated. Our goal is to characterize the impact of rapid 4G deploy-
ment. For that to happen, capitalizing on the studied cluster’s
On the other hand, overlapping coverage areas in LTE dataset, we devised a supervised machine learning model to
is not preferable from both the user perspective and the classify the cells into two groups: good and poor from the
network operator perspective. From the user perspective, the user throughput perspective. The throughput threshold used to
access scheme used in LTE (i.e., OFDM) does not handle label these cells was picked by the mobile operator reflect-
interference gracefully, as shown in Figure 1b. This can lead to ing the minimum throughput required to achieve a targeted
several problems including lower user throughput, ping-pong customer experience and we omit the exact threshold value
handovers, and handover failures. From a network perspective, the operator uses. Then, multiple feature selection algorithms
overlapping coverage areas in LTE waste resources similar to were applied to choose the most impacting features (KPIs)
the case of 3G in addition to making performance at users on user throughput out of the available KPIs to the operator.
worse. Hence, when LTE is deployed on top of an already The best performing algorithms were the Sequential search
operational 3G network, the effect of overlapping coverage methods [19, 20] especially, the ”floating” methods [21]. Fig-
areas is amplified and becomes a main issue in the network. ure 2 shows the sequential floating forward selection algorithm
performance for a decision tree classification algorithm [22]. • [CI] RRC-Connected Users: This KPI reflects the average
It’s apparent that a selected group of six features achieved the number of users that establish a connection with the RRC
best estimation accuracy, we list them below and we will study (Radio Resource Control) layer. This includes all users
each of them closely against the user throughput in (§IV). within a cell whether they have any data to send or not.
We categorize the KPIs we found into three categories: Generally, as the number of users increases, this leads to a
1) Quality of Experience Indicators (QoEI), 2) Quality of decrease in the average user throughput within the cell due
Coverage Indicators (QoCI), and 3) Capacity Indicators (CI). to the finite radio resources.
Each of the available KPIs had different aggregation levels • [CI] Number of Active Users: Active users are users with
(e.g., average and max values per cell), and the following are data remaining in the transmit buffer of the cell at a certain
the selected KPIs, all of which are averaged per cell: transmission time interval (TTI). This KPI is the average
number aggregated over the collection interval. Note that
• [QoEI] User Downlink (DL) Throughput: This KPI is this number does not reflect all users utilizing the DL as it
the main User Quality of Experience indicator used by only counts users with buffered data.
the operator in the radio domain [23]. The higher the user
throughput, the more it is indicated that the customers served
by that cell in the downlink direction are having a good Our goal is to assess the network behavior when the rapid
experience, and vice versa. One of our goals is to find how 4G deployment mentioned earlier is used. In particular, we are
other KPIs affect this one. During network planning and interested in assessing user quality of experience of users in
maintenance, this KPI is used as the main drive for upgrades the 4G deployment, represented by average user throughput
and provisioning of resources to different cells. per cell. We are interested in understanding the relationship
• [QoCI] High-Order MCS Penetration Rate (HOMPR): between network parameters and user experience. To that end,
This KPI captures the ratio of end users in a certain we compare the effect of Capacity Indicators (CIs) and Quality
cell employing high-order Modulation and Coding Scheme of Coverage Indicators (QoCIs) on the Quality of Experience
(MCS). In particular, this KPI indicates the percentage of Indicator (QoEI), which is the average user throughput in
users using 64-QAM modulation, the highest modulation our case. Figure 3 shows the relationship between Average
available at the current network implementation. Users are User Throughput and the different KPIs. We use hexagonal
able to utilize high-order MCS when radio conditions are binning to visualize the distribution of collected 15-minute
good. The higher the employed MCS, the higher the data data samples from all cells. Hexagonal binning allows us to
rate the user can use. Cases where individual users utilize compare the distribution of data across two KPIs.
lower-order MCS are typical (e.g., user at the border of the
cell or behind a signal reflector). Also, inter-cell interference We start with QoCIs. Figure 3a shows that the higher the
has a major impact on the granted MCS, the higher the High-Order MCS Penetration Rate the higher the through-
interference is, the lower the MCS order will be. Thus, when put users might experience. The highest concentration of
a large ratio of users of a certain cell rely on lower-order samples is along the area corresponding to this relationship.
MCS, it indicates that there is a problem at the cell level. Samples with high rate of High-Order MCS Penetration and
• [QoCI] Channel Quality Indicator (CQI): This KPI is low throughput can be justified by low available capacity.
reported from users to the cell to indicate channel quality Invertedly, the rare occasions of having bad radio conditions
seen by the user. This report helps the cell scheduler deter- and still relatively high throughput occur when high bandwidth
mine the MCS to be used by that user. This metric is part and low user count occur. We observe a similar trend the
of the LTE standard unlike signal-to-noise-and-interference- CQI metric (Figure 3c). However, CQI exhibits less correlation
ratio (SINR). The way CQI is calculated takes into account with throughput than the HOMPR. Figure 3b shows the Block
a vendor-based definition of SINR along with the decoding Error Rate (%) which naturally has negative correlation with
capabilities of the user’s device. When the CQI is high, this throughput. Visually, it is clear that among all the QoCIs, the
indicates good experienced radio conditions within the cell. HOMPR has the highest correlation with the user throughput.
• [QoCI] Block Error Rate (BLER): This is the fraction
of blocks delivered on the channel that fail CRC (Cyclic Figure 3d shows a clear relationship with PRB utilization.
Redundancy Check). A high BLER percentage indicates a The higher the utilization, the lower the attained throughput as
problem within the cell. contention for resources increases. The samples indicating the
• [CI] PRB Utilization: A Physical Resource Block (PRB) is high utilization and still relatively high user throughput occur
the smallest unit of resources that can be allocated to a user when spectrum resources are utilized by a few users. CIs are
which is specified in terms of radio sub-carriers and time. less correlated with total cell population compared to QoCIs.
The PRB Utilization KPI captures the ratio of the utilized Figures 3e and 3f both relate the number of users to user
physical resource blocks with respect to all the available throughput. In case of the RRC-connected users, there is no
resource blocks. This KPI reflects how overloaded a specific clear correlation as connected users do not necessarily utilize
cell is which can be used to attribute lower user throughput and resources. On the other hand, the negative correlation is
to lack of resources at the cell when it is high. clear between active users and user throughput.
  105 80 105 80 105
$YHUDJH8VHU7KURXJKSXW0ESV

Average User Throughput (Mbps)

Average User Throughput (Mbps)

 70 70
 104 60 104 60 104

6DPSOH&RXQW

Sample Count

Sample Count
 50 50
103 103 103
 40 40
 102 30 102 30 102
 20 20
101 101 101
 10 10
 0 0
     0 10 20 30 40 6 8 10 12 14
+LJK2UGHU0&63HQHWUDWLRQ Block Error Rate (%) Channel Quality Indicator (CQI)

(a) HOMPR (b) BLER (c) CQI

 105  105  105
$YHUDJH8VHU7KURXJKSXW0ESV

$YHUDJH8VHU7KURXJKSXW0ESV

$YHUDJH8VHU7KURXJKSXW0ESV
  
 104  104  104
6DPSOH&RXQW

6DPSOH&RXQW

6DPSOH&RXQW
  
103 103 103
  
 102  102  102
  
 101  101  101
  
              
35%8WLOL]DWLRQ 1XPEHURI55&&RQQHFWHG8VHUV 1XPEHURI$FWLYH8VHUV

(d) PRB Utilization (e) RRC-Connected Users (f) Number of Active Users
Figure 3: Hexbin scatter comparing distribution of values of different KPIs and Average User Down Link Throughput over
collected data.
Correlation with Average PRB RRC Number of
HOMPR CQI BLER
User Throughput Utilization Connected Users Active Users
All cells 0.454163 0.401373 -0.316519 -0.311759 -0.120421 -0.18782
Cells in good
0.265883 0.236271 -0.312613 -0.370343 -0.162819 -0.229172
radio conditions

Table I: Correlation between different KPIs and Average User Down Link Throughput.

IV. R ELATIONSHIP B ETWEEN BAD R ADIO C ONDITIONS 80 105 80 105

Average User Throughput (Mbps)

Average User Throughput (Mbps)

70 70
AND L OW U SER T HROUGHPUT
60 104 60 104
The first row of Table I summarizes our finding showing the
Sample Count

Sample Count
50 50
103 103
Pearson correlation coefficient between all KPIs. As clear from 40 40
30 102 30 102
Figure 3a, HOMPR has the highest correlation between the
20 20
other KPIs. Our hypothesis is that this correlation is an artifact 10 101 10 101
of the LTE deployment approach which leads to coverage 0
20 40 60 80
0
0 5 10 15 20
overlapping between cells. To understand our hypothesis, PRB Utilization (%) Number of Active Users
consider an ideally operating network. In the ideal case, user (a) PRB Utilization (b) Number of Active Users
throughput should be a factor only of available capacity in
a cell (i.e., low throughput should only occur when demand Figure 4: Hexbin scatter comparing distribution of values
exceeds capacity). However, in cases of bad radio conditions, different KPIs and Average User Down Link Throughput over
the coverage quality becomes an factor in determining user sample with Higher-Order MCS Penetration Rate is higher
throughput. This relationship wastes capacity by having nodes than 60%.
operating at lower throughput than available at the cell. It also
leads to poor user expience. with the second row of Table I show that under good radio
To validate our hypothesis, we started by looking at cells conditions capacity becomes the dominant factor controlling
in good radio conditions. Our goal is to show that when average user throughput, which is the expected behavior. This
the coverage is good, capacity is the only contributor to shows that poor radio conditions lead to a change in factors
user throughput. In our experience, good radio conditions are affecting user throughput. This motivated further analysis as
associated with HOMPR higher than 40%. Figure 4 along shown in the next section.
 Order MCS Penetration Rate (HOMPR) over all LTE cells
Throughput (Mbps)
Average User 12
10 of the studied cluster. We use HOMPR as a proxy for cells
8 suffering from bad radio conditions as it is the KPI with the
6
4 Band 1 highest correlation with average user throughput. In particular,
2 Band 2
0 a cell with very low High-Order MCS Penetration Rate reflects
-118 -116 -114 -112 -110 -108 -106 -104 -102 -100 that the majority of users can only decode robust lower-
LTE RSRP
order modulation schemes. This is unlikely to happen due to
Figure 5: Comparing average user throughput for users poten- individual users having bad radio conditions (e.g., large scale
tially experience bad radio condition (Band 1 users) and users fading), rather it is more likely that a large population of cell
not experiencing bad radio conditions (Band 2 users). users are facing a common source of interference.
1.00 The figure shows that 50% of the cells have a HOMPR
lower than 40%, which is the level we found to make average
0.75 user throughput more dependent on capacity than radio con-
CDF

ditions. To better understand the impact of the problem, we

0.50 look at the top 10% busiest cells and find that they exhibit a
0.25 Top 10% busy cells very similar distribution. Hence, we conclude that bad radio
Whole cells conditions are quite common in cases of rapid deployment and
0.00
0 20 40 60 80 can lead to significant deterioration in user throughput.
High Order MCS Penetration(%) VI. R ESOLVING P OOR R ADIO C ONDITIONS

Figure 6: CDF of ratio of users operating at high modulation Typically, mobile operators target a certain customer ex-
schemes for all studied cells and the top 10% busiest cells perience each year, trying to maintain an edge over other
showing similar distribution in both cases. competitors. And after estimating year-over-year expected
traffic increase, taking inputs from commercial teams and
meeting some certain strategies, they dimension the needed
V. P REVALENCE OF P OOR R ADIO C ONDITIONS capacity across all the mobile network’s domains, including
Our next step is to look at the extent of this issue in terms RAN, Transport and Core Networks. For RAN domain, the list
of its effect on throughput as well as the number of affected of cells likely to be congested (i.e., hot spots) are estimated in
cells when rapid deployment of 4G is employed. Our goal is advance, to be handled with a proper expansion according to
to determine whether bad radio conditions are likely to occur their configurations in term of number of carriers and sectors
when rapid deployment techniques are employed, and if so, at the site [24]. Huge investments are dedicated, in terms of
quantify its effect. As 3G cells are planned with minimal spectrum, license and hardware expansions to work proactively
overlap, one might expect that an LTE network deployed on those cells to make sure they would accommodate the
on top 3G coverage planning should have little problems. foreseen capacity needs.
However, we find that this is not always true. Apart from that, dealing with congested cells is a normal
We start with assessing the extent of effect of bad radio day to day activity for the RF optimizers at the mobile operator
conditions on average user throughput. In particular, we study [25]. Several approaches are implemented for the hot spot
a set of cells where users can access 4G through two different resolution. Usually, the solution involves bandwidth expansion
frequency bands. We refer to them as Band 1 and Band 2. after assuring the highest possible spectrum utilization, adding
A load balancing algorithm ensure that, for each cell, the more cells/sectors or even planning new sites. These expansion
number of users operating at each band is roughly the same. procedures are very costly but this can be afforded by the
The main difference between the two frequency bands is the operator to assure a better experience for the end user, and
cells had neighbouring cells operating in Band 1, potentially consequently, a higher revenue. However, these approaches to
having overlap in their coverage area in that band. On the other handling hot spots are not necessarily the right approaches in
hand, all studied cells had no neighbouring cells operating in when LTE is deployed following the methodology discussed
the Band 2, ensuring having no bad radio conditions due to in Section V.
interference. We compare average user throughput for both From practical observations, it was found that hot spots,
scenarios where Reference Signal Received Power (RSRP) is following a soft LTE deployments, are mostly formed due
low (i.e., areas with poor coverage). Figure 5 shows the result to overlapping coverage areas between different cells. Hence,
of the experiment. We find that average throughput is up to 3x relief of this congestion can be done by re-planning coverage
higher for users operating in Band 2 where there are no bad areas of affected cells. This approach requires: 1) detecting
radio conditions due to overlapping coverage areas. Hence, we cells suffering from this problem, and 2) providing some
conclude that cell overlap has significant negative impact on guidelines for handling the interference problem.
user throughput. Affected Cells Detection: We rely on High-Order MCS
Next, we look at the prevalence of the problem in the Penetration Rate to detect cells affected by our LTE deploy-
studied cluster of cells. Figure 6 shows the CDF of High- ment approach. In particular, cells with 40% or low High-
 Penetration Rate (%)
Throughput (Mbps)
24 70

High-Order MCS
Average User
20 Avg. User Throughput 60
16 MCS Pen. Rate 50
Down-tilt 12 40
Angle 8 30
0 ̊̊
Azimuth 4
Angle 0 5 10 15 20 25 30 35 40
Time (hours)
Antenna
Height (a) Case 1

Penetration Rate (%)

Throughput (Mbps)
28 70

High-Order MCS
Average User
Main Lobe
24 60
20 50
16
Distance 12 40
8 30
0 5 10 15 20 25
Figure 7: Illustration of cell physical parameters. Time (hours)
(b) Case 2
Order MCS Penetration Rate are considered affected. This is
consistent with our finding in Section IV where we found that Figure 8: Two case studies showing the effect of physical
for cells with High-Order MCS Penetration Rate larger than optimization that improved High-Order MCS Penetration Rate
40%, the main factors affecting user throughput are capacity on improving average user throughput by up to 114%.
factors not interference factors.
Physical Optimization: To address the problem of overlap- networks. We focus on recent measurements from major oper-
ping coverage areas, the parameters determining the coverage ators. For example, in [26] sector KPIs used to infer user QoE.
area of each cell are optimized using standard planning tools This allows the network operator to detect underperforming 3G
that simulate the coverage of affected cells. These tools sectors. The work in [24] uses sector KPIs to forecast which
identify optimal parameters for each antenna, in terms of cells will underperform under specific loads. Our approach
the elevation, azimuth and height (Figure 7), to make sure falls under this category to augment earlier work. In particular,
the dominance of the serving cells and isolation from the it provides a specific metric to detect underperforming cells
neighboring cells as much as possible. And those predictions due to 4G transition. It can be used in conjunction with other
can be calibrated with performing a field drive test to capture approaches to provide better cell planning and provisioning.
the coverage measurements and the results are sent back 2. Measurements taken at the network side, capturing user
for the planners to analyze and confirm their coverage plan. traces: Several systems rely on user traces to detect [27],
Finally, this parameters are applied to the affected site. predict [28] and react to [29] user throughput in 4G networks.
We demonstrate the effectiveness of this approach by show- 3. Measurements taken at the user side: This approach relies
ing the following two cases where the methodology was on measurements that are either collected by the operator
applied. Both cases are identified by having HOMPR below through wardriving (e.g., [30]), or crowdsourcing (e.g., [31]).
30%. Physical optimization was applied at hour 17 and 11 This direction of work is orthogonal to ours that can alert
for case 1 and 2, respectively. Figure 8 shows that, for both operators to coverage gaps that cannot be detected from the
cases, once physical optimization was applied both HOMPR operator side.
and average user throughput improve significantly. Average Interference cancellation and mitigation: Such techniques
user throughput is improved by 97% and 114% on average. have been deployed with varying degrees of success for more
This is a significant impact in throughput that is achieved than 20 years. They are generally categorized into three major
without provisioning any extra resources. Hence, we note that techniques as shown in [32]: interference cancellation, inter-
this approach allowed our operator to save costs in two ways. ference averaging and interference avoidance techniques. Al-
First, it allowed for a fast deployment of LTE where not all though in theory many proposed interference mitigation tech-
cells needed re-planning. Second, it allows the operator to niques have shown promising results, in practice the gains do
have an approach to resolve hot spots other than the costly not seem to materialize and better approaches need to be eval-
resource provisioning (e.g., creating more cells or licensing uated in realistic scenarios suitable for implementation[33].
more spectrum) to check before going to that step. For example, interference cancellation promises significant
gains but is likely limited to the LTE UL due to processing
VII. R ELATED W ORK
complexity, and will require real-time exchange of information
Data-driven Cellular Network Optimization: We cate- between base stations every few milliseconds to maximize the
gorize those approaches based on where measurements are gain for an LTE system. Beam-forming technologies such as
collected into the following categories: organized beam hopping show significant theoretical promise;
1. Measurements taken at the network side, capturing cell however, as practical technologies for a deployed LTE system
KPIs: Earlier work in this direction has focused on 3G they are still unproven [18].
 Also, LTE offers the capability to provide a flexible dy-
[12] A. ElNashar, M. A. El-Saidny, and M. Sherif, Design, de-
namic inter-base station approach to interference coordination ployment and performance of 4G-LTE networks: A practical
through the use of inter-base station signaling capabilities approach. John Wiley & Sons, 2014.
using the X2 interface defined between the base stations, [13] J. T. Penttinen, The LTE-Advanced Deployment Handbook: The
including the use of UL reactive overload indicators (OIs) Planning Guidelines for the Fourth Generation Networks. John
and proactive high interference indicators (HIIs) that provide Wiley & Sons, 2016.
[14] T. V. Krishnamurthy and R. Shetty, 4G: Deployment Strategies
bit maps of interference conditions on a per RB basis. DL and Operational Implications: Managing Critical Decisions in
inter-cell interference coordination (ICIC) is supported through Deployment of 4G/LTE Networks and their Effects on Network
the use of DL relative narrowband transmit power (RNTP) Operations and Business. Apress, 2014.
bit maps providing a coarse power indication on a per RB [15] Y. Chen, L. Duan, and Q. Zhang, “Financial analysis of 4G
basis [18]. Although this capability was already employed, network deployment,” in Proc. of IEEE INFOCOM, 2015.
[16] A. J. Viterbi, A. M. Viterbi, K. S. Gilhousen, and E. Zehavi,
further physical optimization was still required achieving extra “Soft handoff extends CDMA cell coverage and increases
enhancement in user throughput as we showed earlier. reverse link capacity,” IEEE Journal on selected areas in
communications, 1994.
VIII. C ONCLUSION [17] G. P. Pollini, “Trends in handover design,” IEEE Communica-
We use a large-scale LTE data set from a major telecom op- tions magazine, 1996.
erator to assess the value of a rapid LTE deployment approach [18] G. Boudreau, J. Panicker, N. Guo, R. Chang, N. Wang, and
S. Vrzic, “Interference coordination and cancellation for 4g
that relies on transitioning from 3G to 4G, minimizing the networks,” IEEE Communications Magazine, 2009.
overhead of extensive post-activation optimization steps. We [19] A. L. Blum and P. Langley, “Selection of relevant features and
also show a data-driven approach to detect underperforming examples in machine learning,” Artificial intelligence, 1997.
cells and present the physical optimization approach to handle [20] S. Das, “Filters, wrappers and a boosting-based hybrid for
the detected cells. Overall, the approach balances LTE transi- feature selection,” in Proc. of ICML, 2001.
[21] P. Pudil, J. Novovičová, and J. Kittler, “Floating search methods
tion best practices with cost efficiency to allow the operator in feature selection,” Pattern recognition letters, 1994.
to provide a high performing network rapidly. We believe this [22] S. R. Safavian and D. Landgrebe, “A survey of decision tree
approach will be beneficial for operators deploying 4G for the classifier methodology,” IEEE transactions on systems, man,
first time and performing 3G shutdown as well as community- and cybernetics, 1991.
operated networks servicing rural underdeveloped areas. [23] P. Casas, B. Gardlo, R. Schatz, and M. Mellia, “An Educated
Guess on QoE in Operational Networks Through Large-Scale
R EFERENCES Measurements,” in Proc. of Internet-QoE Workshop, 2016.
[24] J. Serrà, I. Leontiadis, A. Karatzoglou, and K. Papagiannaki,
[1] I. E. Sørensen, “Content in context: The impact of mobile media
“Hot or not? Forecasting cellular network hot spots using sector
on the British TV industry,” Convergence, 2018.
performance indicators,” in Proc. of IEEE ICDE, 2017.
[2] B. Sanou, “ICT Facts and Figures 2017,” Geneva: The Interna-
[25] L. Song and J. Shen, Evolved cellular network planning and
tional Telecommunication Union (ITU), 2017.
optimization for UMTS and LTE. CRC press, 2010.
[3] J. Place, S. Keeping, S. Akayama, and A. Schlautmann, “The
[26] I. Leontiadis, J. Serrà, A. Finamore, G. Dimopoulos, and
Business Benefits of 4G LTE,” Arthur D. Little UK, 2012.
K. Papagiannaki, “The good, the bad, and the KPIs: how to
[4] I. PLC, “El Salvador embraces LTE, gets its first 4G networks,”
combine performance metrics to better capture underperforming
2017, https://ovum.informa.com/resources/product-content/
sectors in mobile networks,” in Proc. of IEEE ICDE, 2017, pp.
te0009-001664.
297–308.
[5] D. N. Egypt, “4G service to launch on Thursday,” 2017, https:
[27] T. Mangla, E. Zegura, M. Ammar, E. Halepovic, K.-W. Hwang,
//dailynewsegypt.com/2017/09/24/4g-service-launch-thursday/.
R. Jana, and M. Platania, “VideoNOC: Assessing Video QoE
[6] L. Duan, J. Huang, and J. Walrand, “Economic analysis of 4G
for Network Operators Using Passive Measurements,” in Proc.
network upgrade,” in Proc. of IEEE INFOCOM, 2013.
of ACM MMSys, 2018.
[7] N. Bhushan, J. Li, D. Malladi, R. Gilmore, D. Brenner,
[28] J. Huang, F. Qian, Y. Guo, Y. Zhou, Q. Xu, Z. M. Mao,
A. Damnjanovic, R. T. Sukhavasi, C. Patel, and S. Geirhofer,
S. Sen, and O. Spatscheck, “An in-depth study of LTE: effect
“Network densification: the dominant theme for wireless evo-
of network protocol and application behavior on performance,”
lution into 5G,” IEEE Communications Magazine, 2014.
ACM SIGCOMM Computer Communication Review, 2013.
[8] A. Weissberger, “AT&T to shut down 3G network
[29] X. K. Zou, J. Erman, V. Gopalakrishnan, E. Halepovic, R. Jana,
in 2022; Verizon at end of 2019,” IEEE ComSoc
X. Jin, J. Rexford, and R. K. Sinha, “Can accurate predictions
Tech Blog, 2019, http://techblog.comsoc.org/2019/02/22/
improve video streaming in cellular networks?” in Proc. of ACM
att-to-shut-down-3g-network-in-2022-verizon-at-end-of-2019/.
HotMobile, 2015.
[9] A. Choros, “Australian 3G Network Shutdown:
[30] A. Balasubramanian, R. Mahajan, and A. Venkataramani, “Aug-
Everything you need to know,” Whistle Out, 2019,
menting Mobile 3G Using WiFi,” in Proc. of ACM MobiSys,
https://www.whistleout.com.au/MobilePhones/Guides/
2010.
Australian-3G-network-shutdown-what-you-need-to-know.
[31] “OpenSignal,” https://opensignal.com/, 2018.
[10] S. Han, Y.-C. Liang, and B.-H. Soong, “Spectrum refarming: A
[32] M. Rahman, H. Yanikomeroglu, and W. Wong, “Interference
new paradigm of spectrum sharing for cellular networks,” IEEE
avoidance with dynamic inter-cell coordination for downlink
transactions on communications, 2015.
LTE system,” in Proc. of IEEE WCNC, 2009.
[11] S. Hasan, M. C. Barela, M. Johnson, E. Brewer, and K. Heimerl,
[33] D. J. Ramirez, “Interference mitigation techniques for 4G
“Scaling community cellular networks with communitycellular-
networks,” Ph.D. dissertation, Supélec, 2014.
manager,” in Proc. of USENIX NSDI, 2019.