This is a postprint version of the following published document:
Sciancalepore, V., Samdanis, K., Costa-Perez, X., Bega, D.,
Gramaglia, M. y Banchs, A. (2017). Mobile Traffic Forecasting
for Maximizing 5G Network Slicing Resource Utilization. In
IEEE INFOCOM 2017 Conference on Computer
Communications.
DOI: https://doi.org/10.1109/INFOCOM.2017.8057230
© 2017 IEEE. Personal use of this material is permitted. Permission from IEEE
must be obtained for all other uses, in any current or future media, including
reprinting/republishing this material for advertising or promotional purposes,
creating new collective works, for resale or redistribution to servers or lists, or
reuse of any copyrighted component of this work in other works.
Mobile Traffic Forecasting for Maximizing
5G Network Slicing Resource Utilization
Vincenzo Sciancalepore
∗
, Konstantinos Samdanis
†
, Xavier Costa-Perez
∗
,
Dario Bega
‡§
, Marco Gramaglia
‡§
, Albert Banchs
‡§
∗
NEC Europe Ltd.
†
Huawei Europe
‡
IMDEA Networks Institute
§
Universidad Carlos III de Madrid
Abstract—The emerging network slicing paradigm for 5G
provides new business opportunities by enabling multi-tenancy
support. At the same time, new technical challenges are intro-
duced, as novel resource allocation algorithms are required to
accommodate different business models. In particular, infras-
tructure providers need to implement radically new admission
control policies to decide on network slices requests depending
on their Service Level Agreements (SLA). When implementing
such admission control policies, infrastructure providers may
apply forecasting techniques in order to adjust the allocated
slice resources so as to optimize the network utilization while
meeting network slices’ SLAs. This paper focuses on the design
of three key network slicing building blocks responsible for (i)
traffic analysis and prediction per network slice, (ii) admission
control decisions for network slice requests, and (iii) adaptive
correction of the forecasted load based on measured deviations.
Our results show very substantial potential gains in terms of
system utilization as well as a trade-off between conservative
forecasting configurations versus more aggressive ones (higher
gains, SLA risk).
I. INTRODUCTION
In addition to the clear advantages in terms of, among
others, enhanced bandwidth, reduced latency or extended
coverage, the introduction of future 5G networks will have
a significant impact on how operators manage their infrastruc-
ture. In contrast to the relatively monolithic architectures of
3G and 4G, by building on the recent advances in network
softwarization, 5G networks will be highly modular and de-
signed to be future-proof.
5G networks will hence allow higher flexibility: network
virtualization can boost the introduction of very diverse ser-
vices to be deployed on-demand using shared infrastructure.
This feature enables new business opportunities for Mobile
Network Operators (MNO); indeed, hosting different services
with possibly conflicting requirements on the same infras-
tructure is currently not achievable with the current one-size-
fits-all architectures. However, it also introduces new critical
challenges; the network slicing concept [1] is expected to be
one of the technical solutions to these challenges.
Network slicing allows MNOs to open their physical net-
work infrastructure platform to the concurrent instantiation of
multiple logical self-contained networks, orchestrated in dif-
ferent ways depending on their specific service requirements;
such network slices are (temporarily) owned by the respec-
tive tenants. The availability of this vertical market provides
new monetization opportunities of the network infrastructure,
since (i) new players may come into play (e.g., automotive
This work has been partially funded by the European Union’s Horizon 2020
research and innovation programme under the grant agreement No. 671584
5GNORMA.
industry, e-health,...), and (ii) a higher infrastructure capacity
utilization can be achieved by admitting network slice requests
and exploiting multiplexing gains. In the above context, the
technical enablers for network slicing admission control need
to be investigated.
The 5G Network Slice Broker [2] is a novel network
element that builds on the capacity broker functional block
considered by 3GPP for advanced RAN sharing [3]. It maps
incoming Service Level Agreement (SLA) requirements as-
sociated to network slice requests into physical resources.
Tenants hence obtain a “slice” of the appropriate Radio Access
Network (RAN) elements. The architectural specifications for
this new network paradigm are currently under definition and
the necessary algorithms yet to be devised.
Although very conservative mappings may be considered
for mission critical services that need ultra-high availability,
enhanced admission control algorithms that leverage mul-
tiplexing gains of traffic among slices are the key to the
optimization of network utilization and monetization. To this
end, the ability to predict the actual footprint of a particular
network slice is essential to increase the maximum number of
slices that might be run on the same infrastructure.
Building on this idea, in this paper we design three key
network slicing building blocks: (i)aforecasting module that
predicts network slices’ traffic based on past traffic and user
mobility, (ii) a network slicing admission control algorithm
and (iii) a network slicing scheduler algorithm in charge
of meeting the agreed SLAs and report deviations to the
forecasting module.
The remaining of the paper is organized as follows. In
Section II we review the state-of-the-art solutions, before
presenting our framework building blocks in Section III. In
Section IV we establish the basis of our slice forecasting
model, whereas in Section V we formulate the admission
control problem as a geometric knapsack, providing that this
problem is NP-Hard. In Section VI we explain the slice
scheduling process and how its feedback is used to adjust the
forecasting process. In Section VII we discuss the simulation
results and, finally, we conclude the paper in Section VIII.
II. R
ELATED WORK
The support for multi-tenancy in 3GPP LTE networks is
related to early proposals on active RAN sharing, which
enables network sharing based on contractual agreements.
A study on virtualization for wireless and mobile networks
considering preliminary proposals such as the GENI project
as well as early LTE base station virtualization is elaborated
in [4]. Two active network sharing architectures are specified
in 3GPP, the Multi-Operator Core Network (MOCN), allowing
each operator to share eNBs connected on a separate core
network, and the Gateway Core Network (GWCN), where
operators share additionally the Mobility Management Entity
(MME) [5]. A complementary network sharing management,
which enables MVNOs to control the allocated resources,
is designed in [6]. Our proposal exploits the experience of
early deployments, while being compatible with the 3GPP
specifications.
A RAN sharing solution applying proportional fairness
criterion is proposed in [7]. To share resources among different
operators under diverse radio conditions, [8] introduces the
Network Virtualization Substrate (NVS), a two-step process
where the infrastructure first allocates resources to the virtual
instances of eNBs and then each tenant customizes scheduling
within its eNB instance [9]. In our work we adopt a similar
two-step process, allocating slices via a broker entity that
performs admission control based on the requested SLAs.
Our approach builds on the concept of a signaling-based
network slicing broker solution by implementing a capacity
forecasting algorithm that considers guaranteed and best-effort
traffic in addition to user mobility. A study that explores
different options of network sharing based on a centralized
broker is provided in [10], considering mobility means, spec-
trum transfer policies and resource virtualization to optimize
the usage of MNO’s limited resources. Unlike our proposal,
such a study introduces new 3GPP interfaces to accommodate
the broker functionality. A scheme that integrates the capacity
broker with a minimum set of enhancements on the 3GPP
architecture is documented in [11]. Such capacity broker
forecasts the network capacity when allocating guaranteed and
best-effort slices, considering their respective SLAs. Our ap-
proach enhances previous solutions by introducing algorithms
that dynamically evaluate network slices SLA requests, while
maximizing the infrastructure resources utilization.
III. S
YSTEM DESIGN
This paper builds on the concept of a 5G network slice
broker in the context of the 3GPP network sharing manage-
ment architecture [6] for establishing network slices through
signaling. The 5G network slice broker is introduced in the
network management system of the infrastructure provider to
exploit 3GPP conventional monitoring procedures for gather-
ing global network load measurements. Such information can
assist the forecasting process, facilitating admission control
while considering the specified network slice SLAs. To support
a signaling-based slice allocation, certain 3GPP interfaces
need to be enhanced (Type 5 and Itf-N) to enable the in-
stantiation and configuration of network slices, indicating the
time duration, the required resource amount, and additional
requirements, such as, e.g., the slice SLA. We refer the reader
to [2] for further architectural details.
Fig. 1 depicts the 5G Network Slice Broker building blocks
addressed in this paper. Network slice requests are collected
within a fixed negotiation time window. When the time win-
dow is closed, network slice requests are processed and eval-
uated. A key aspect for an efficient network slice admission
control mechanism is to accurately predict the tenants’ traffic
Training phase
(Legacy solution)
Network
Slices Packer
Forecasting-aware
Network Slicer
ADMISSION CONTROL
Granted Slice Requests
(X
i
(k)
)
Penalty History Function
(H
i
(k)
)
Forecasted Information
(ZǺ
i,z
(k)
)
Traffic Patterns
(R
i
(k)
)
SLICE FORECASTING
HoltWinters
Core
Prediction
Intervals
HW Params {ɲ,ɴ,ɶ}
(
i,z
)
i,z
SLICE SCHEDULING
Scheduler s
i,k
Monitoring P
i,k
Slice Requests
ɇ={ʍ
i
(k)
}
5G Network Slice Broker
Fig. 1: Block diagram of the 5G Network Slice Broker.
evolution in the near future. This is achieved through a Slice
Forecasting module in charge of analyzing the network slices
traffic patterns and providing forecasting information to the
Admission Control module, as explained in Section IV. When
no forecasting solution is applied (w/o forecasting) or during
the training period (for adjusting the forecasting algorithm
parameters), the only information used are the SLA requests.
Based on this information, Admission Control policies are
applied in order to select which network slice requests will
be granted for the next time window. To this end, two
different algorithms are devised, with different performance
and complexity features, as explained in Section V. The list
of granted slice requests is sent to the Slice Scheduling module,
which allocates network slice physical resources and monitors
(with a penalty history function) the served traffic levels and
potential SLA violations. Such a function is used to provide
feedback to the forecasting module and thus adaptively adjust
the system, as explained in Section VI.
IV. S
LICE FORECASTING
Information on forecasted traffic patterns is used to predict
future slice load and thus maximize the system resource
utilization. The effectiveness highly depends on the accuracy
of the forecasting algorithm: the more accurate, the more
aggressive we can be in leasing available resources while
keeping a small probability of violating slice SLAs. While
the first aspect is deeply analyzed in this section, we refer the
reader to Section VI for more details on SLA violations and
dynamic forecasting parameters adjustments.
A. Tenant traffic analysis: characterization and forecasting
Traffic predictions are performed on an aggregate basis for
every tenant. Each tenant i might ask for a different network
slice request tailored to its specific service requirements.
Indeed, the forecasting process can easily categorize the traffic
requests based on the associated service requirements, thereby
performing a prediction separately per slice. In our analysis,
we first assume that traffic requests are uniformly distributed
within the whole network. However, in Section IV-B we extend
this assumption by considering multi-cellular environments
where tenant traffic requests are significantly affected by the
user mobility.
We assume different classes of traffic based on specific
SLAs, as shown in Table I. We let the traffic volumes
of tenant i for traffic class k (e.g., satisfying particular
service requirements) be a realization of a point process,
ζ
(k)
i
=
T
t=0
δ
t
r
(k)
i
(t), where δ
t
denotes the Dirac measure
for sample t. We express traffic requests r
(k)
i
(t) in terms
of required resources (note that these resources could be
easily translated into different metrics, such as latency or
throughput demands). Given the periodic nature of traffic
requests, the traffic forecasting is based on an observed
time window T
OBS
, and is given by the vector r
(k)
i
=
(r
(k)
i
(t − T
OBS
),r
(k)
i
(t − (T
OBS
+1)), ··· ,r
(k)
i
(t)). Then, the
forecasting function f
HW
provides forecasted traffic volumes
for time period [t +1,t + T
WINDOW
], denoted as ˆr
(k)
i
=
(ˆr
(k)
i
(t +1), ˆr
(k)
i
(t +2), ··· , ˆr
(k)
i
(t + T
WINDOW
)). For fixed
traffic patterns, the system exhibits a periodic behavior, which
translates into seasons of length W
S
that are repeated over
time. Within a single season we assume that process ζ
(k)
i
is
stationary and ergodic. Thus, we can use the Holt-Winters
(HW) forecasting procedure to analyze and predict future
traffic requests associated to a particular network slice. We
denote a specific predicted traffic request ˆr
(k)
i
(t) by ˆr
(k)
i,t
.We
rely on the additive version of the HW forecasting problem as
the seasonal effect does not depend on the mean traffic level of
the observed time window but instead it is added considering
values predicted through level and trend effects. Following
HW standard procedure, we can predict such requests based
on the level l
t
, trend b
t
and seasonal s
t
factors, as follows:
ˆr
(k)
i,t+T
WINDOW
= l
t
+ b
t
h + s
t+T
WINDOW
−W
where
l
t
= α(r
(k)
i,t
− s
t−W
)+(1 − α)(l
t−1
+ b
t−1
),
b
t
= β(L
t
− l
t−1
)+(1− β)b
t−1
,
s
t
= γ(r
(k)
i,t
− l
t−1
− b
t−1
)+(1− γ)s
t−W
.
(1)
While the set of optimal HW parameters α, β and γ
can be obtained during a training period employing existing
techniques [12], we focus on the forecasting errors and how the
forecasting inaccuracy may affect our network slicing solution.
We define the one-step training forecasting error e
(k)
i,t
as follows
e
(k)
i,t
= r
(k)
i,t
− ˆr
(k)
i,t
= r
(k)
i,t
− (l
t−1
+ b
t−1
+ s
t−1
), (2)
which is computed during the training period of our fore-
casting algorithm (when predicted values are compared with
the observed ones). Given that our process ζ
(k)
i
is ergodic
and assuming an optimal HW parameter set, for any pre-
dicted value at time z we can derive the prediction interval
ˆ
ll
(k,χ)
i,z
,
ˆ
hh
(k,χ)
i,z
wherein future traffic requests lie for that
particular network slice with a certain probability χ
(k)
i
. Then,
it holds that
Pr
ˆ
ll
(k,χ)
i,z
≤ ˆr
(k)
i,z
≤
ˆ
hh
(k,χ)
i,z
= χ
(k)
i
, ∀z ∈ [t+1,t+T
WINDOW
]
(3)
where
ˆ
hh
(k,χ)
i,z
(or
ˆ
ll
(k,χ)
i,z
)=ˆr
(k)
i,z
+(−)Ω
χ
Var(e
(k)
i,z
) and
Var(e
(k)
i,z
) ≈
(1 + (z − 1)α
2
[1 + zβ +
z(2z − 1)
6
β
2
]
σ
2
e
.
In the above equation, Ω
χ
denotes the one-tailed value of a
standard normal distribution such that we obtain χ
(k)
i
proba-
bility and σ
2
e
is the variance of one-step training forecasting
error, i.e., σ
2
e
= Var(e
(k)
i,t
), over the observed time window.
TABLE I: Network slice traffic requirements [13]
k T
(k)
Type and QCI
010ms GBR - 65
150ms GBR - 3
2 100 ms GBR - 1
3 150 ms GBR - 2
4 300 ms non-GBR - 6
5 1000 ms non-GBR - 9
Due to the penalties imposed by traffic SLAs, we focus only
on the upper bound of the prediction interval as it provides
the “worst-case” of a forecasted traffic level. From Eq. (3), a
larger prediction time window T
WINDOW
, e.g., a higher number
of predicted values z, leads to a lower accuracy and behaves
closer to the real network slice demand (limited network slice
resources utilization). Conversely, an accurate forecasting with
a lower error probability can result in severe penalties in case
it does not guarantee the desired slice SLAs. Therefore, we
adjust the forecasting error probability χ
(k)
i
according to the
service requirements and to the number of prediction points the
forecasting process needs to perform. For instance, best-effort
traffic requests having no stringent requirements can tolerate
a prediction with a longer time pace that results in imprecise
values. This makes the upper bound
ˆ
hh
(k,χ)
i,z
very close to
the real (future) values r
(k,χ)
i,z
regardless the error probability
χ
(k)
i
as the number of z values to predict is limited. Hence,
we might select a low forecasting error probability χ
(k)
i
for
this service type. On the other hand, when guaranteed bit rate
traffic is considered, the corresponding SLA must be fulfilled
in a shorter time basis, which makes our forecasting process
much more complex, requiring significantly more predicted
values z. To achieve this, our system models such a type of
traffic with a higher forecasting error probability χ
(k)
i
.
We implement the above mathematically as follows. Ac-
cording to the traffic classes defined in Table I, traffic class
k =0provides a forecasted horizon shorter than the other
traffic classes, and hence a larger number of values z must
be predicted. To achieve this, we can derive an upper bound
for the forecasting probability error per tenant i for this
traffic class. We calculate the maximum potential gain be-
tween the slice request and the forecasted traffic requests
as
ˆ
d
(k)
i
=max
z∈T
WINDOW
R
(k)
i
− ˆr
(k)
i,z
. We then compute the
forecasting error probability as follows
χ
(k=0)
i
:Ω
χ
Var(e
(k=0)
i,z
)=
ˆ
d
(k=0)
i
. (4)
As soon as the potential gain
ˆ
d
(k=0)
i
becomes very large, we
cap the one-tailed value Ω
χ
to 3.49, resulting in χ
(k=0)
i
=
99.9%. Conversely, for the best-effort traffic (k =5)we
compute the forecasting error probability χ
(k=|K|)
i
= 50%,
due to its more relaxed service. For the other traffic classes k,
intermediate forecasting error probabilities χ
(k)
i
are calculated
from (4) by deriving
ˆ
d
(k)
i
values from the upper and the lower
bound values. However, note that forecasting error probability
values are dynamically evaluated and adjusted based on the
SLA violations experienced during the slice scheduling pro-
cess, as explained in detail in Section VI-B.
B. User mobility and traffic model periodicity
We next extend our forecasting model to dynamic scenarios
where user mobility is considered and the traffic periodicity
assumption no longer holds. We consider a multi-cellular
environment covering the whole area. In order to design
forecasting algorithms that are accurate under realistic settings,
we rely on human-based mobility patterns. Specifically, we
employ the well-accepted SLAW mobility model [14] for
user motions. According to this model, users move among a
number of waypoints, which are distributed over the network
area according to self-similarity rules forming a given number
of clusters. Clusters with more waypoints can be seen as
hotspots attracting more users. While performing a flight (a
movement from one waypoint to the other within the same
trip), based on some given probabilities users choose a set
of clusters which are dynamically and randomly replaced
during the flight. Then, users start moving between a subset
of waypoints residing within the selected clusters according to
a least-action trip planning (LATP) with α
SLAW
=3. Traffic
requests come randomly during the user trip. Assuming that
users stop when reaching a waypoint for a pause-time, we can
model the value of the flight-time (x
L
) and pause-time (x
P
)as
a random value drawn from a heavy-tailed distribution function
defined in terms of Fourier transformations as
f
L
(x)=f
P
(x)=
1
2π
∞
−∞
e
−iu x−|ρu|
α
DISTR
du (5)
where ρ is the scale factor and α
DISTR
depends on the
distribution considered (pause-time or flight-time).
Given a uniform user speed distribution, the traffic model
of the considered users is dominated by a heavy-tailed dis-
tribution whose components can be decoupled, as showed
in Eq. (5). Under these conditions, the system exhibits a
periodic behaviour (like in the previous section). Without
loss of generality, we can obtain a periodic traffic vector
as follows. Let M denote the period and r
(k)
i
= {r
t
} a
generic traffic vector. Then, the forecasting process applies a
Discrete Fourier Transform (DFT) to retrieve the M-periodic
samples R
w
=
M−1
n=0
r
t
e
−iw
2π
N
t
, where w =0, ··· ,M − 1.
Note that R
w
is a complex number translating the sinusoidal
component of r
t
. Then, the forecasting process can obtain
all single time-series components derived by each of those
frequency samples by applying the Inverse Discrete Fourier
Transform (IDFT), e.g., r
n
=
1
N
M−1
w=0
R
w
e
2πi
Nwn
, where
n =0, ··· ,N − 1, which provides a periodic traffic vector
r
(k)
i
=(r
(k)
i
(n),r
(k)
i
(n +1), ··· ,r
(k)
i
(n + M )).
V. A
DMISSION CONTROL:DESIGN AND VALIDATION
A 5G Network Slice Broker might decide on the network
slice requests to be granted for the subsequent time window
T
WINDOW
based solely on the current resource availability.
However, if forecasting information is taken into account,
the resources consumed by network slice requests might be
accurately reshaped to fit additional slice requests into the
system (see Fig. 2).
A mathematical approach is proposed next address the
admission control problem for both cases. First, we prove its
NP-Hardness and suggest a baseline algorithm for allocating
Resources
Time
T
WINDOW
Ⱥ
L
1
R
1
i = 1
i = 5
i = 2 i = 3
R
3,1
Forecasted Traffic Level SLA Slice Request Traffic Offered
i = 4
R
3,3
R
3,2
Ǻ
Ǻ
Ǻ
r
4,t
i = 6
Fig. 2: Admission control problem as geometric knapsack
problem.
network slice requests when no forecasting information is
available. Then, we design the Forecasting-aware Network
Slicer algorithm to efficiently perform the admission control
phase exploiting accurate traffic pattern predictions.
A. Problem Formulation
In our problem formulation, we first assume that network
slice instantiations require a constant amount of resources,
and then we show that, when relaxing such an assumption
by considering different forecasted traffic levels, the problem
becomes more complex but still tractable for our admission
control process.
Let us define a network slice request as σ
(k)
i
= {R, L, i, k}
where i identifies the tenant, R is the amount of resources
required, L is the time duration of the slice and k is the
traffic class. Hereafter, we simply refer to a tenant request as
R
(k)
i
(L
i
). Recalling that our main objective is to accommodate
network slice requests within a fixed time window T
WINDOW
while maximizing the network resource utilization, we next
derive our model.
Let us assume a rectangular box with fixed width W and
height H representing the resource availability within a fixed
time window. In particular, the box width corresponds to
T
WINDOW
and box height corresponds to the total amount of
resources Θ. Let us assume a set of items I, where each
item i ∈Icorresponds to a network slice request having
width w
i
(corresponding to slice duration L
i
) and height h
i
(corresponding to the amount of resources R
i
). In addition,
each item provides a profit c
i
that corresponds to the amount
of resources needed (here we are assuming that a slice request
pays an amount of money proportional to the number of
resources granted
1
). Then, the objective of our admission
control problem is to find a subset of items I
⊆Ithat
maximizes the total profit
i∈I
c
i
, i.e., the total amount of
used resources, as shown in Fig. 2.
Lemma 1. Let the overall system resource availability be a
box with height Θ and width T , and let each item i ∈Ibe the
network slice request σ
i
with height R
i
and width L
i
. Then,
the admission control problem is mapped into a Geometric
Two-dimensional knapsack problem with the objective of filling
1
This assumption could be relaxed to reflect a different economic model
within the multi-tenancy framework, which is out of the scope of the paper.
![TABLE I: Network slice traffic requirements [13]](/figures/table-i-network-slice-traffic-requirements-13-1f3kp48l.webp)
![TABLE II: System parameters ([19])](/figures/table-ii-system-parameters-19-2e8ne0gy.webp)