A Data Analysis Methodology for Obtaining

Network Slices Towards 5G Cellular Networks

Feyzullah Kalyoncu, Engin Zeydan and Ibrahim Onuralp Yigit
Türk Telekom Labs, Istanbul, Turkey 34889
E-mail: {feyzullah.kalyoncu, engin.zeydan, ibrahimonuralp.yigit}@turktelekom.com.tr

Abstract—Mobile Network Operators (MNOs) are investigat- based on their policies, key triggers, drivers and use cases.
ing new business opportunities and planning to launch new Most of the time, MNOs have different target architectures and
revenue generating services in the context of 5G deployments. In deployment plans in future. For example, some MNOs may
order to achieve this, MNOs are willing to open up new services
and applications to end users as well as vertical industries such want to leverage their fixed line incumbency and infrastructure
as automotive, health, entertainment in order to obtain new rev- while others may rely on their previous Long Term Evolution
enues. Thanks to recent advances in virtualization technologies, (LTE) investments or already deployed small cell or Internet
network slicing is one promising approach that can be utilized of Things (IoT) network infrastructures. On the other hand, in
by MNOs for generating new services and applications that are order to increase revenue, MNOs also want to have early time-
tailored to demands of end users and vertical industries. Although
there have been many works on network slicing applied to mobile to-market advantage for embracing incremental revenue with
network infrastructure, data analytic approach using real world new business models, creating new value opportunities and
network dataset that investigates how many network slices are utilizing the cost per bit advantage of 5G systems. Therefore,
needed for each MNOs is still missing. In this paper, we propose it is important to know a priory how to drive business as well
a methodology to extract the number of network slices specific to as network strategy towards implementation of 5G.
each MNOs via applying clustering on Key Parameter Indicators
(KPIs) of major telecommunication operators. Our large scale End-to-end (e2e) network quality becomes important when
data analysis results indicate that the number of network slices MNOs open up new services and applications for their sub-
may differ depending on network quality of MNOs. For some scribers in 5G networks. Network quality of a MNO depends
MNOs up to five different network slices can be obtained which on radio, transport and core network performances. In order to
correspond to launching different services and applications on accommodate some of the networking requirements of certain
each selected network slice.
Index Terms—clustering, mobile operator, KPI, 5G, network applications, e.g. Ultra reliability and low latency communi-
slicing. cation (URLLC) mission critical services, each MNO should
have the required network upgrades and capacity for enabling
I. I NTRODUCTION connectivity of service equipments (UEs). Considering the fact
that there exists wide range of applications/services that can
Telecommunication providers are already generating vast cover different use cases, it is necessary to know how to extend
amounts of data. In order to gain insights into user experience the network slicing concept and address how many slices
or the network quality, this data needs to be collected in data need be initiated. However, there exists a gap in identifying
lakes and analyzed for proper operation. At the same time, next the required number of network slices to be launched for
generation applications & services are expected to be flexible, increasing e2e network quality of MNOs. In the literature,
automated and agile thanks to new infrastructure with network there have been numerous works on network slicing in the
slicing concept utilizing network virtualization technologies context of mobile networks [2]–[4]. Similarly, utilizing data
with Software Defined Networking (SDN) and Network Func- analytics together with network key performance indicator
tion Virtualization (NFV) in 5G. Fast service instantiating (KPI) performances have also been studied in previous works
infrastructure will enable Mobile Network Operators (MNOs) such as [5], [6]. However, combining real network data and
to offer customized services and deploy new services such analytical techniques with benefits of network slicing concept
as Massive Machine Type Communications (mMTC) in order for mobile networks is still an open research area.
to boost revenues. Therefore, network slicing and gaining
insights via data analytics is slowly shaping the future of A. Our Contributions
telecommunication. Although there have been different works that concentrate
5G era is starting to revise the business models as well on concept of network slicing for MNOs and emphasizing
as network architecture of future communication of MNOs. the importance of network slicing for MNOs in general, no
5G is expected to impact every sector such as automotive, prior efforts have concentrated on comprehensive data analysis
e-health, media & entertainment, energy and factories of over KPI of major MNOs using long duration and nationwide
future [1]. At the same time, for real deployments, MNOs real data measurement in order to determine an appropriate
are looking for the most optimized path towards 5G. It is a number of network slices. Different than previous works, in
well know fact that each MNO’s strategy will be different this work we perform data analysis using Machine Learning

978-1-5386-6355-4/18/$31.00 ©2018 IEEE

 Fig. 1: Network slicing for multiple MNOs having different services and applications.

(ML) algorithms over the KPIs collected from UEs that yield order to meet the specific demand of a service, MNOs need the
the end-to-end network performance of three MNOs collected control and flexibility of assigning network resources to each
over 18 months period. Our analysis is utilized to gain insights service or application. This can ensure MNOs to deliver high
into obtaining appropriate number of network slices that can quality-of-experience (QoE) with differentiated services for
accommodate various application and services for different their UEs. By exploiting the recent advances in virtualization
MNOs . Our contributions in this paper can be summarized technologies such as NFV and SDN, resources can be isolated
as follows: (i) We utilize end-to-end network performance into programmable sets of slices in order to enable certain
of three MNOs in Turkey using location area code (LAC) network functions and services.
based geographical and temporal properties of real data set Fig. 1 shows an architectural diagram of the network slicing
of 4G cellular networks. (ii) We utilize a clustering method concept applied for multiple MNOs. Each MNO is assumed
for obtaining insights into potential number of network slices to have different number of slices where each slice may serve
that can be launched for each MNOs. (iii) We find out that to different applications and services of MNO. For example,
depending on the KPI performances of each MNOs, the connected vehicles service which can be launched by a MNO
number of network slices of each MNO can significantly differ can utilize one or multiple slices underneath. Formally, we
where some MNO KPI performances allow higher number of assume that in our system there are M MNOs and each
network slices, while others may have low number of network MNO has Km total number of network slices. We denote
slices due to each MNOs subscribers’ data utilization and the MNO set M = {1, 2, . . . , M } and network slice set
network infrastructure characteristics. Km = {1, 2, . . . , Km }. Each MNO is assumed to have Sm
LACs with LAC set Sm = {1, 2, . . . , Sm }. Note that inside
II. S YSTEM M ODEL AND A RCHITECTURE each LAC-s ∈ Sm , there are many Base Stations (BSs) of
In this section, we investigate the network slicing archi- MNO-m ∈ M. Our focus in this paper is on investigating the
tecture, system concepts and some of the prominent 5G use performance KPIs of MNOs operating nationwide in different
cases. regions or cities to extract the potential number of network
A. Network Slicing: Concept and Architecture slice that each MNO can provide. Focusing on LAC-based
KPI measurements, we identify the impact of KPI on number
In present mobile infrastructures, MNOs are configuring of network slices for each MNOs’ performance.
Access Point Names (APNs) in order to allocate network
resources for each service that they provide. However, this B. 5G Application Use Cases
is both inefficient and time demanding operation where each 5G offers different capabilities to distinct use cases. The
device’s configuration should be performed across many net- most important three main categories of 5G are based on
work elements. This solution is also not providing fine-grained differentiation on latency, throughput and reliability. 5G In-
network resource control. This problem puts limitations on frastructure Public Private Partnership (5G-PPP) has identified
proper network operation which greatly reduces the net income three main use cases for 5G network slices [7]: (i) URLLC,
of MNOs as network utilization increases. A feasible approach (ii) enhanced Mobile Broadband (eMBB), (iii) mMTC.
in terms of increasing network flexibility is to make it easy (i) URLLC applications: URLLC is one generic mode
for MNOs to build network slices across whole network. In of Machine Type Communications (MTC) where it can
 be used for mission-critical applications and is driven after silhouette method based K-means clustering for a given
by requirements of fast and reliable interactions. Some MNO-m. As input to Algorithm 1, we provide nationwide
examples are reliable remote robotic actions, real time latency and DL speed values of each measurements performed
remote surgery or coordination among vehicles, factory by MNOs’ UEs. The output is Km together with LAC-s
automation, smart grids, reliable emergency systems, self- ∈ Sk and its corresponding cluster index for all MNOs. In
driving. Some of the requirements are: less than 10 Mbps Algorithm 1, Pm denotes the MNO-m’th matrix with rows
average throughput, 99.999% reliability, 1-10 ms latency, representing LAC-s ∈ Sm and columns representing average
subscriber densities up to 1,000 devices per square km latency and DL speeds. P represents the matrix formed by
and mobility support (up to 200 kmph). concatenating Pm , ∀m ∈ M in rows.
(ii) eMBB applications: This slice type is expected to pro-
vide high data rate and high traffic densities (capacity) Algorithm 1 Utilized K-means clustering with silhouette
and is driven by video and Virtual Reality (VR) re- method
quirements. Typical applications are in high resolution 1: procedure K- MEANS (Latency, DL Speed)
video streaming, real time work in cloud environment, 2: for each MNO-m ∈ M do
Augmented Reality (AR)/VR immersive games, 3D holo- 3: for each LAC-s ∈ Sm do
graphic presence. Some of the requirements are: less than 4: Group latency, DL speeds based on LAC-s
1-10 Gbps peak throughput and 50-100 Mbps average 5: Calculate mean of latency and DL speed
throughput, 50-100 ms latency, subscriber densities up to 6: Construct matrix Pm .
150,000 devices per square km and mobility support (up 7: end for
to 120 kmph). 8: end for
(iii) mMTC applications: mMTC investigates challenges 9: Construct matrix P and normalize.
for ubiquitous coverage, massive connection support for 10: for each MNO-m ∈ M do
delay-tolerant traffic and is driven by ability to scale and 11: Run K-means with silhouette on Pm of P
ensure efficient signalling traffic control. Some of the 12: Return Km with highest score
requirements are: less than 1 Mbps average throughput, 13: Return all LAC-s ∈ Sm and its cluster index;
50-100 ms latency, subscriber densities up to 1 million 14: end for
devices per square km and mobility on demand and 10 15: end procedure
years of battery lifetime.
III. C LUSTERING A LGORITHM FOR N ETWORK S LICING IV. P ERFORMANCE E VALUATIONS
In this section, we perform clustering on existing 4G cellular For evaluation results, anonymous data is collected offline
network data in order to extract insights into potential number on daily basis and transferred into the database for a period of
of slice that each MNO can offer to their subscribers on the 18 months ranging from January 2016 to June 2017 of MNOs
path towards 5G. operating nationwide in Turkey. The statistical distributions
of the observed test data are given in Table I for all MNOs
A. Silhouette Method based K-means Clustering which yield a relatively large sample size for fair evaluations.
In order to obtain the number of clusters from the data, Moreover, Table II shows the Coefficient of Variation (CoV)
we utilize silhouette method of [8]. The method of silhouette (= stdv/mean) value comparisons of some KPIs, i.e. DL,
works as follows: First, for each point l representing latency Upload (UL) and latency nationwide. The total number of
and Download (DL) speed in x and y axis of a MNO LACs is 689, 770 and 556 for MNO-1, MNO-2 and MNO-
respectively, we find the average distance between l and all 3 respectively. In our analysis, we utilize latency and DL
other points inside the cluster, which is a measure of cohesion, speed KPIs from our data set for performing our data analysis.
denoted by A. Second, we find the average distance between As a matter of fact, combining latency and DL speed can
point l and all other points in the nearest cluster, which is release important new set of applications and use cases that can
a measure of separation, denoted by B. Then, the silhouette provide new and unseen opportunities for MNOs. These can
coefficient ρl for l is calculated as, include low latency and low throughput applications such as
system control, low latency and high throughput applications
A−B
ρl = (1) such as AR/VR, high latency and low throughput applications
max(A, B) such as IoT and sensor devices and high latency and high
Then, the cluster coefficient of each point ρl is calculated throughput applications such as video.
and averaged in order to obtain “overall” cluster coefficient ρ. In Fig. 2, we plot the silhouette analysis for K-means
clustering on MNO-1 data where K1 = 5 is selected as
B. Proposed Methodology to Obtain Network Slices cluster number yielding largest silhouette coefficient score of
The details of the utilized K-means clustering with silhou- ρ = 0.4861. We can select K1 = 5 as the number of network
ette method customized for the utilized dataset is given in slices that MNO-1 can offer based on the existing data analysis
Algorithm 1. We assume that we obtain Km number of clusters of LTE networks. We run K-means algorithm using average
 TABLE I: Nationwide 4G Test Speed Data Statistics.
# of measurements 462,638 # of cities 81
# of districts 940 DL (average, Mbps) 42.766
# of neighborhoods 13,163 UL (average, Mbps) 16.891
# of MNOs 3 Latency (average, msec) 31.606
Obser. Duration 18 months Total # of test UEs 117,059

TABLE II: Comparisons of CoV values of observed KPIs for

4G nationwide.
CoV MNO-1 MNO-2 MNO-3
DL speed (Mbps) 0.8610 1.2078 0.7609
UL speed (Mbps) 0.7578 0.6566 0.8363
Latency (ms) 0.9802 1.0376 1.0938

DL speed data and latency values over LACs of MNO-1 as

input. Fig. 3 shows the distribution of DL speed and latency
distributions on x and y-axis respectively. Fig. 3 also marks
the corresponding cluster IDs of all LACs. As can be observed
from this figure, MNO-1 has wide range of latency and DL
speed values which corresponds to different number of clusters
(i.e. network slices) in the network. One key observation is that
average latency and DL speed performance vary significantly
nationwide based on the KPIs on each LAC and MNO-1 can
offer variety of services to its subscribers. As a starting point
for deploying next generation 5G cellular networks, based on Fig. 2: Determining the number of clusters using silhouette
the existing data analysis, it’s obvious that MNO-1 can start analysis for K-means clustering on MNO-1 data.
to offer five different slices with corresponding latency and
DL speed values serving various applications and services
across nation. For example, according to Fig. 3, MNO-1’s
existing infrastructure is suitable to provide five different
network slices with 1-) medium throughput-low latency, 2-)
low throughput-low latency, 3- high throughput-low latency, 4-
) low throughput- high latency and 5) low throughput-medium
latency demanding applications and services.
In Fig. 4, we demonstrate the MNO-2’s average silhouette
score distribution over different number of clusters. Based
on these values, for MNO-2, K2 = 2 is selected to be the
highest scoring cluster number with of ρ = 0.6788. Similarly,
Fig. 5 shows the latency and DL speed distribution among the
selected two clusters. Based on these results, MNO-2’s slices
are divided based on the latency distributions. Therefore, it can
provide both high and low latency services to its subscribers.
Finally, Fig. 6 shows the average score distribution based
on number of clusters where K3 = 3 is selected to be highest
scoring number of clusters of ρ = 0.4325. Fig. 7 plots the
corresponding latency and DL speed distribution among the
selected K3 = 3 clusters. Based on the data analysis results
of Fig. 7, MNO-3 can provide three network slices with 1-) Fig. 3: K-means clustering on MNO-1 data where K1 = 5.
low throughput -low latency, 2-) high throughput-low latency
and 3-) high latency-low throughput demanding applications
and services. of major MNOs in Turkey. Our evaluation results are utilized
to answer a key observation of how many number of network
V. C ONCLUSIONS slices are needed for a specific MNO based on its network
In this paper, we have performed a data analysis over a traffic characteristics and end-to-end network performance.
large sample of collected data indicating the KPI performance Our evaluation results indicate that based on the existing 4G
Fig. 4: Determining the number of clusters using silhouette Fig. 6: Determining the number of clusters using silhouette
analysis for K-means clustering on MNO-2 data. analysis for K-means clustering on MNO-3 data.

Fig. 5: K-means clustering on MNO-2 data where K2 = 2. Fig. 7: K-means clustering on MNO-3 data where K3 = 3.

[4] M-CORD, “M-CORD Open Reference Solution Paves the Way for 5G
KPI data, each MNOs need to launch different number of In-novation.” http://opencord.org/tag/m-cord/., 2016. [Online; accessed
network slices nationwide in order to accommodate various 30-Oct.-2017].
services for their subscribers. [5] E. Zeydan, E. Bastug, M. Bennis, M. A. Kader, I. A. Karatepe, A. S. Er,
and M. Debbah, “Big data caching for networking: Moving from cloud to
edge,” IEEE Communications Magazine, vol. 54, no. 9, pp. 36–42, 2016.
R EFERENCES [6] N. Baldo, L. Giupponi, and J. Mangues-Bafalluy, “Big data empowered
self organized networks,” in European Wireless 2014; 20th European
[1] 5GPP, “5G-PPP use cases and performance evaluation modeling v1.0 Wireless Conference; Proceedings of, pp. 1–8, VDE, 2014.
(white paper).” https://goo.gl/uhv4bC, 2016. [Online; accessed 30-Oct.- [7] 5GPP, “View on 5G Architecture (Version 2.0) (white paper).”
2017]. https://5g-ppp.eu/white-papers/, 2017. [Online; accessed 04-Sep.-2017].
[8] P. J. Rousseeuw, “Silhouettes: a graphical aid to the interpretation and
[2] O. Narmanlioglu and E. Zeydan, “Software-defined networking based
validation of cluster analysis,” Journal of computational and applied
network virtualization for mobile operators,” Computers & Electrical
mathematics, vol. 20, pp. 53–65, 1987.
Engineering, vol. 57, pp. 134–146, 2017.
[3] O. Narmanlioglu and E. Zeydan, “New Era in shared cellular networks:
Moving into open and virtualized platform,” International Journal of
Network Management, pp. e1986–n/a, 2017. e1986 nem.1986.