Automatic Network Slicing for IoT in Smart City

Abstract

Internet of things (IoT) in smart city consists of a diversity of public utility and vertical industry services with extremely different performance requirements. As a key technology of 5G networks, network slicing (NS) is featured to provide distinct virtual networks and differentiated QoS guarantees in a shared infrastructure. Therefore, it becomes necessary to employ intelligent NS management for IoT in smart city. This work proposes a machine learning (ML) driven automatic NS framework which can intelligently scale slice according to network state. The resource preservation and slicing implementation scheme is given to trade off robustness and resource efficiency. Finally, the present study provides preliminary results via simulation and experimentation to justify the effectiveness and efficiency of the presented design.

Full text

1536-1284/20/$25.00 © 2020 IEEE IEEE Wireless Communications • December 2020
108
AbstrAct
Internet of things (IoT) in smart city consists
of a diversity of public utility and vertical indus-
try services with extremely different performance
requirements. As a key technology of 5G net-
works, network slicing (NS) is featured to provide
distinct virtual networks and differentiated QoS
guarantees in a shared infrastructure. Therefore, it
becomes necessary to employ intelligent NS man-
agement for IoT in smart city. This work proposes a
machine learning (ML) driven automatic NS frame-
work which can intelligently scale slice according
to network state. The resource preservation and
slicing implementation scheme is given to trade
off robustness and resource efficiency. Finally, the
present study provides preliminary results via sim-
ulation and experimentation to justify the effective-
ness and efficiency of the presented design.
IntroductIon
Smart city is a prominent application of the
Internet of Things (IoT) to offer citizens innova-
tive services for improved quality of life. Nations
worldwide are carrying out their plans to build
smart cities, where utility infrastructure and public
assets in every corner of the cities are connected
through IoT infrastructure, enabling governors
and utility service providers to make comprehen-
sive decisions for improved citizen services [1].
IoT devices with various functions, such as meter-
ing, monitoring, actuating, from different ser-
vice providers are thus widely deployed, posing
remarkably disparate requirements on network
performance. The effective and efficient connec-
tivity of the widely deployed devices has become
the basis for building smart cities.
Conventionally, utility companies prefer to build
private networks to connect their IoT devices for its
advantages of privacy preservation and self-design-
ability. However, such IoT solutions face challenges
in smart city scenarios, where the utility service pro-
viders have to deploy sensors densely and widely
to gain more detailed information and computing
facilities as close to the scene as possible, gaining
instant insights from gathered information to intel-
ligentize their services. Private networks are usual-
ly built according to the peak demand, leading to
redundant network construction and the low utili-
zation of network infrastructure and budget. Their
limited network operation and maintenance skills
worsen the situation. Budget constraints, technical
skill limitations, and performance concerns keep
utility service providers from further expanding their
private networks. In the era of 5G, other than build-
ing private networks, service providers in smart cities
can subscribe network slices from operators to meet
their practical needs in such network services.
Unlike its predecessors, 5G has high perfor-
mance indicators that allow it to expand the service
scope from people to things [2]. Designers of 5G
have envisioned the challenge in providing a univer-
sal communication service that fulfills the need for
applications with disparate network requirements
on the same network infrastructure. The concept of
network slicing (NS) is thus introduced. Through slic-
ing, the same network infrastructure can be divided
into logically independent networks, called network
slice instances (NSIs). Data flows of similar service
characteristics can be aggregated in the same slice
instance with a designated set of network resources
to reduce resource-scheduling complexity. More-
over, NS ensures isolation between slice instanc-
es, which is a prerequisite for some service holders
in smart city. Besides isolation for privacy, another
driving factor of NS lies in improving the efficacy of
network management. NS grants a new granularity
to manage a group of services within the network,
so IoT services that pay no attention to isolation can
share slices. Slices are initially designed to achieve
isolation among eMBB, URLLC, and mMTC net-
work services in 5G and grow to be mature enough
in the 5G beyond (B5G) stage. The two key chal-
lenges are AI-enabled management and network
orchestration (MANO) and robust slicing accord-
ing to different requirement. MANO is the basis to
achieve management for NFV based networks. The
robust automatic slicing guarantees the network to
better suit user’s requirements with stable network
resource usage.
To serve smart city and meet the broad
demands of utility companies on customized IoT
networks, it is foreseeable that many slice instanc-
es will be subscribed to and simultaneously run-
ning on the same network infrastructure with an
appointed service level agreement (SLA) attached
to each of them. Then, one of the key problems
in network slicing is how to optimally allocate var-
ious network resources to network slice instances.
The process should be agile and flexible to instant
Fanqin Zhou, Peng Yu, Lei Feng, Xuesong Qiu, Zhili Wang, Luoming Meng, Michel Kadoch,
Liang Gong, and Xianjiong Yao
A N S 
IT  S C
ACCEPTED FROM OPEN CALL
Fanqin Zhou, Peng Yu (corresponding author), Lei Feng, Xuesong Qiu, Zhili Wang, and Luoming Meng (corresponding author) are with
BUPT, Beijing; Michel Kadoch is with Ecole de Technologie Superieure; Liang Gong is with Academy of Broadcast Planning NRTA of China;
Xiantong Yao is with State Grid Shanghai Municipal Electric Power Company.
Digital Object Identifier:
10.1109/MWC.001.2000069
ZHOU_LAYOUT.indd 108ZHOU_LAYOUT.indd 108 12/23/20 11:17 AM12/23/20 11:17 AM
Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on February 07,2021 at 07:28:57 UTC from IEEE Xplore. Restrictions apply.

IEEE Wireless Communications • December 2020 109
ACCEPTED FROM OPEN CALL requests and dynamic traffic demands from ten-
ants, which by no means can be manually accom-
plished. At the same time, machine learning (ML)
approaches are widely used to intelligentize the
behavior of practical systems.
The above facts motivate the authors to inves-
tigate the ML-driven automation of NS from the
resource allocation perspective. As illustrated in
Fig. 1, network services of emergency communi-
cation, hotspots capacity enhancement, and so on
in smart city cause network environment chang-
es, which will be refl ected in the spatial and tem-
poral traffi c distribution of slice instances and the
network performance indicators. By learning from
these observations, the automatic network slicing
(ANS) agent, following the deep reinforcement
learning (DRL) framework, will gain experience in
taking better network slicing actions.
The contributions of this article are as follows:
• An online ML-driven framework of ANS is pro-
posed. To the best of our knowledge, it is the
first time that deep learning-based slice state
prediction, unsupervised learning-based slicing
anomaly detection, and reinforcement learn-
ing-based slicing scaling are considered togeth-
er to achieve ANS.
• A robustness-oriented resource preservation
and slicing implementation scheme is given.
Suffi ciently considering the delay-sensitive and
emergency service, it is necessary to preserve
a portion of resources to generate new slice
instances or scale the existing ones. By intro-
ducing robustness optimization and reward
in online ANS, the exact portion of these pre-
served resources can be determined to trade
off the robustness and resource efficiency.
The rest of this article is organized as follows.
The following section introduces the basic knowl-
edge of NS. We then develop an ML-driven NS
framework to support the automation of NS. We
justify the effectiveness of the proposed frame-
work by presenting an illustrative case and numeric
results. Finally,we conclude the article.
network slIcIng In 5g
This section introduces the general network slic-
ing management architecture, the resource man-
agement issues during slicing, and some existing
automatic slicing related work.
the concept of network slIcIng
The concept of NS in 5G networks is extended from
“slicing” in network virtualization by particularly
stressing the end-to-end (E2E) performance. Through
network slicing, the same network infrastructure
can be sliced into independent logical networks,
referred to as network slice instances with preserved
resources for diff erent utilization preferences, such
as aggregating similar fl ows into the same slice for
improved network resource utilization, or logically
isolating network services for the tenants who pay
special attention to privacy and security.
Massive IoT applications with diff erent proper-
ties can be observed in smart city from different
public utilities, vertical industries and over-the-top
(OTT) companies. Even in a residential communi-
ty, massive-connection but low-data-rate NSIs can
serve water and gas companies for meter reading,
a high-bandwidth slice instance can serve a security
company for video surveillance, and ultra-low-la-
tency slice instance can be established for facility
maintenance when wearable augmented reality
devices are used by the maintainers.
The implementation of an NSI can be either
static or dynamic [3]. For static slice instances,
network resources are exclusively reserved and
will not be affected by the variation of network
status within their life cycle, so isolation to other
network services can be achieved. Resource pres-
ervation also means easy implementation in prac-
tice, whereas it usually leads to low utilization of
resources. For dynamic slice instances, a quota of
resources is assigned, but the resources are dynam-
ically allocated according to the actual demand.
This on-demand property improves the utilization
of network resources, though it requires addition-
al QoS-guaranteeing mechanisms to maintain the
E2E performance. Both types of slice instances will
likely exist in 5G networks, but the dynamic ones
cause the main complexity.
network slIcIng MAnAgeMent ArchItecture
In the 5G mobile system, the full-fledged NS will
be built on SDN and NFV enabled network infra-
structure and a NS management architecture is
necessary to coordinate a variety of enabling tech-
nologies and network resources. A typical NS man-
agement architecture generally consists of slice
management functions and slice implementation
functions. The former ones, which are directly
responsible for the life-cycle management of net-
work slice instances, are comprised of the network
slice management (NSM) function for the manage-
ment of E2E slice instances and network slice sub-
net management (NSSM) functions for managing
slice subnet instances in the access network (AN),
transport network (TN) and/or core network (CN)
domains, respectively. When it comes to the imple-
mentation of slice instances, a variety of functional
components will be generated and orchestrated
and corresponding resources will be allocated and
coordinated, under the control of the latter ones
which basically include NFV management and
orchestration (MANO), SDN controller, and legacy
entity management (EM) functions.
When a communication service request comes
from a tenant, such as OTT companies, vertical indus-
FIGURE 1. Network slicing for IoT applications in smart city.
slice 2
2
RAN Edge TN
uRLLC Slice
eMBB Slice
mMTC Slice
CN
Emergency
Communication
Hotspot
Capacity
Unmanned
Emergency Control
Environmental Change
Space
Time
ANS
Agent
Environmental Information
….
RAN Edge TN
eMBB Slice New
mMTC Slice
CN
MTC
MTC
Slic
Slic
eMBB Slice
uRLLC Slice
uRLLC Slice New
ZHOU_LAYOUT.indd 109ZHOU_LAYOUT.indd 109 12/23/20 11:17 AM12/23/20 11:17 AM
Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on February 07,2021 at 07:28:57 UTC from IEEE Xplore. Restrictions apply.

IEEE Wireless Communications • December 2020
110
tries, and mobile virtual network operators (MVNOs),
communication service management (CSM) turns
the request according to the chosen service type and
its attached service level agreements into a network
slice request which is mainly comprised of user panel
functions and a set of performance indicators. The
network service management (NSM) translates the
network slice request into a service function chain
and divides the service function chain (SFC) into seg-
ments for different subnet domains for implementa-
tion convenience. The NSM does so by being able
to understand how a sequence of network functions
will logically compose an SFC to serve a typical data
flow in the required communication service in the
form of network slice templates. Then NFV-MA-
NO cooperates with the SDN controller and EM to
implement NS by instantiating the SFC. It should be
noted that a slice will possibly be comprised of sev-
eral SFC instances to serve geographically-different
data flows.
To implement NS for smart city IoT, the flexi-
ble management of cross-domain and multi-level
network resources forms a key issue. In the fol-
lowing section, a brief description of the managed
resources in NS will be provided.
network slIcIng resource MAnAgeMent
As smart city IoT applications may cover various
network scopes, the managed network resources
in different network domains will be reviewed.
Radio Access Network Resources: The man-
aged resources in the radio access network (RAN)
mainly include the radio carriers, physical resource
blocks, available spectrum, transmit power, and
RAN facilities, such as antennae and cell sites. New
radio technologies also grant new dimensions of
resources, such as air-interface numerology options
and cooperative transmission modes RAN resourc-
es can be dedicated or shared, depending on the
required level of resource isolation and the cost
to implement the isolation. Dedicated resources
are preferred for mission-critical services, as it per-
mits ideal isolation and stable access performance.
However, to improve the overall utilization of ded-
icated resources, flexibly sharing RAN resources,
like spectrum and bearer, is a major direction [4].
Non-Radio Network Resources: In the SDN/
NFV context, non-radio networks can be viewed as
linked network functions running in different kinds
of cloud infrastructure. This means that the resourc-
es for slicing in the non-radio domain are resources
in edge and center clouds, and the links between
them, which typically include computing, network-
ing, and storage or cache resources. Link resources
typically form an important set of constraints decid-
ing the embedding of VNFs, while the problem of
deciding bandwidths of links between VNFs itself
turns out to be NP-hard, requiring a highly-efficient
solution to expedite the slicing process [5].
As shown above, the sliced network consists
of various types of resources in different network
domains to construct network slice instances. To
make the slice instances adaptive to the real-time
demand of various smart city IoT services, dynamic
slicing with autonomous features is in great need.
To this end, an ML-driven automatic NS manage-
ment architecture is proposed, showing its main
framework and stressing designs for robustness.
exIstIng work on AutoMAtIc slIcIng
A few researchers have realized the necessity of
automating network slicing processes (Table 1).
Velasco et al. in [6] demonstrated an architec-
ture to support autonomic slice networking in the
optical network based on SDN. In [7], a microser-
vice-based NS architecture was implemented to
design and create slices automatically according
to service requirements. Song et al. in [8] applied
machine learning in a traffic-aware dynamic slic-
ing framework to dynamically allocate transport
network resources for flexible resource sharing.
A. Perveen et al. in [9] proposes a deep reinforce-
ment based autonomous radio resource slicing
framework with dynamic resource preservation to
TABLE 1. Comparison of existing automatic network slicing studies.
Ref. Main objectives Focused elements Automation level Validation approach Scenario Key description
[6] Slice networking VNF, bandwidth Low Prototype CN Supporting automatic slice network-
ing in SDN-based optical network.
[7] Slice creation VNF Low – CN A microservice-based network slicing
architecture to design and create slices
automatically according to service re-
quirements.
[8] Resource sharing Bandwidth Low Simulation CN Traffic-aware dynamic slicing frame-
work fortranspor t networkslicing with
machine learning.
[9] Slice tuning Spectrum Medium Simulation RAN Resource slicing framework, quick
convergence and improved perfor-
mance in the context of network re-
source utility and satisfaction of net-
work service.
[10] Resource sharing Power Medium Simulation RAN Dynamic inter-slicing resource reser-
vation and scheduling in LoRa net-
works.
[11] Slice tuning Abstract resource
portion
Medium Experiment RAN/CN Making decision of network resource
orchestration via a deep learning ar-
chitecture.
[12] Slice tuning Flow High Experiment E2E Effect of ML approaches in the au-
tomation of network management and
slicing.
ZHOU_LAYOUT.indd 110ZHOU_LAYOUT.indd 110 12/23/20 11:17 AM12/23/20 11:17 AM
Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on February 07,2021 at 07:28:57 UTC from IEEE Xplore. Restrictions apply.

IEEE Wireless Communications • December 2020 111
serve diff erent types of services. S. Dawaliby et al.
in [10] proposed an adaptive dynamic inter-slic-
ing resource reservation algorithm to schedule
resources in LoRa networks. D. Bega et al. in [11]
implemented data-driven strategies concerning
the decision of network resource orchestration via
a deep learning architecture. These works imple-
ment automatic slicing in the low-to-medium level,
because most of them focus on low-level mechan-
ical processes, such as networking VNFs, design-
ing and creating slices. However, they form the
basis of higher level automatic slicing. L. Le et al.
[12] presented some considerations on the eff ects
of ML in automation of 5G network slicing, show-
ing a high level of automation with full life cycle
self-management; however, it merely considered
link resources and proactive resource tuning.
In addition to academic research, standard
development organizations also initiated work
on automatic slicing. The autonomic networking
integrated model and approach (ANIMA) work-
ing group in IETF has scheduled automatic slicing
as the main part of its long-term goal of network
automation. Study Group 13 in ITU-T is working on
new recommendations on network slice confi gura-
tion to support producing NSIs.
Ml-drIven AutoMAtIc network slIcIng
Different from existing studies on enabling NS,
this article goes further to explore the ML-driven
automation of NS from the resource allocation
perspective, as well as to investigate the robust-
ness issues during these processes.
AutoMAtIc network slIcIng frAMework
The high-level ML-driven automatic network slic-
ing framework follows a deep reinforcement
learning (DRL) paradigm as illustrated in Fig. 2.
The slice traffic analyzer module continuous-
ly monitors the E2E slices running in the network
infrastructure. Data statistics and ML methods are
applied to process multi-dimensional traffic data
according to historical statistics. For instance,
long-short term memory (LSTM) can be used to
predicate traffi c hotspots, while the unsupervised
learning to analyze the abnormality in the situation
of communication break. Then, the observed or
predicated slice traffi c is characterized to become
the input for the ANS agent. The static information
for the tenant, for example, priority or experience,
is considered in the feature engineering process to
form an effective but low-dimensional set of fea-
tures. After online real-time computing is conduct-
ed to decide whether to generate new slices and
how to coordinate resources, the slice broker and
the slice scaling coordinator implement the actions.
At last, the reward module generates a designed
reward according to a function to the observed
networks and slices performance indicators.
drl-bAsed Ans Agent And rewArd Module
DRL is an improved version of reinforcement learn-
ing whose key feature is that agents learn good
behavior through trials and errors under the infl u-
ence of properly set rewards. An agent selects
actions based on a policy determined by Q(s,a)
which is the expected long-term cumulative reward
of an action a in state s. Inspired by [13], a typical
DRL framework called DQN is illustrated in Fig. 2,
featured by the dual deep neural networks with the
same neural structure, denoted as MainQNet and
TargetQNet. The Q(s,a; ) is functional approximat-
ed by MainQNet where  represents the weight
parameters of MainQNet. Similarly, TargetQNet’s
weight parameter is denoted as –. In the learning
process, the MainQNet keeps updating  accord-
ing to ← arg min(Q(s, a; ) – Q(s’, a’; –) – r)2
in each step by training the deep neural network
with a set of state-action-reward-state (SARS)
tuples sampled from the replay memory, while Tar-
getQNet copies the parameters from MainQNet
every several steps for algorithmic stability. In the
execution process, the ANS utilizes MainQNet to
produce an action in each state it encounters, and
the -greedy scheme was utilized to enhance the
exploration. Benefiting from the replay memory,
online training can be achieved to adapt the DRL
model to shifted environments.
The ANS agent may have to deal with diff erent
optimization goals with the variation of network
FIGURE 2. Automatic network slicing framework.
Stac
Features
Network
Historical
Informaon
Slice Feature
Engineering
Slice State Analyzer
Observer Predicter
Slice Broker
Slice Scaling
Coordinator
End-to-End Slices in Mul-Tenants Shared Network Infrastructure
Replay memory D
−
…
, , ,
Random acon
= ( − − )2
Opmizer(Loss)
Weights update
DQN Loss Funcon
TargetQNet
=, =
,
if (memory_size > batch){
memory =
random.sample(memory)}
Update
−every
Tsteps
LSTM
−1
ℎ−1ℎ
( , )
= ( , ; ) =′( , ; −)
MainQNet
Reward
Generaon
−
High-level
preference
SARS Tuple
Generator
A er online real-time
computing is conduct-
ed to decide whether
to generate new slices
and how to coordinate
resources, the slice bro-
ker and the slice scaling
coordinator implement
the actions. At last, the
reward module gener-
ates a designed reward
according to a function
to the observed
networks and slices
performance indicators.
ZHOU_LAYOUT.indd 111ZHOU_LAYOUT.indd 111 12/23/20 11:17 AM12/23/20 11:17 AM
Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on February 07,2021 at 07:28:57 UTC from IEEE Xplore. Restrictions apply.

IEEE Wireless Communications • December 2020
112
state, and at the same time, slice instances have
diverse key performance indicators, which usually
consist of indicators for resource efficiency and
quality of service. For robustness considerations,
robustness indicators, such as the ratio of slices fail-
ing SLA and the affected number of slices, can be
set as key factors of the reward. The reward mod-
ule is thus introduced to generate proper reward
from different kinds of observations. Mathematical-
ly, the reward module is denoted as a function, r(s,
a; y), with respect to the state s and chosen action
a, where y represents the high-level preferences of
operators. The weighted sum average of normal-
ized indicators is often used in the reward design.
slIce stAte AnAlyzer
The slice state analyzer is mainly comprised of
the observer and predictor to gather the newest
values of performance indicators and the predic-
tions, respectively.
Robust Observation: In practical systems, the
observed data should be robust to outliers which
forms the basic requirement of robust observation
for automatic slicing. Robust statistics is a kind of
such techniques used to identify the presence of
outliers. In [14], the median absolute deviation and
Mahalanobis distance metrics in robust statistics
were utilized to identify outliers with univariate and
multivariate information, respectively. The medi-
an absolute deviation is a robust alternative to the
standard deviation and holds advantages in com-
puting, while Mahalanobis distance is one form
of the generalized Euclidean distance the squared
of which approximately follows Chi-squared distri-
bution, and a threshold can be numerically deter-
mined for an observation being an outlier.
Another aspect of robust observation lies in the
geographical initialization of new traffic flows of
a slice instance. To fully gain the automatic slic-
ing paradigm of the on-demand feature, a branch
of a slice instance will not necessarily be allocat-
ed resources when there is no traffic in the corre-
sponding geographical area. However, enabling
this feature relies on a robust observation of traffic
flow arrivals in the network. When a new traffic
flow is observed, an access management function
entity should be able to identify the correspond-
ing slice instance as soon as possible, and the slice
resource coordinator should allocate resources to
the slice within the area.
Robust Prediction: To proactively address
network and slice issues, one prerequisite is a
prediction system that gives predicted values of
performance indicators and should be robust to
data outliers. Traditional prediction models, such as
auto-regressive moving average (ARMA), auto-re-
gressive integrated moving average (ARIMA),
and Holt-Winters, cannot solve the problem well.
Furthermore, the entire data set needs to be han-
dled, which requires a lot of training time for huge
data sets. One possible solution, categorized into
the robust adaptive online gradient learning, uses
LSTM for an accurate prediction in the presence
of exceptional value and introduces modification
for robustness by keeping track of the adaptive
moment estimation of predictions. According to
the relative prediction error of the loss function,
the learning efficiency is adjusted. The smaller the
relative error, the lower the learning efficiency.
slIce broker And slIce scAlIng coordInAtor
The slice broker and slice scaling coordinator
are responsible for the practical implementation
of NS. Two aspects are considered in the robust
implementation of network resource slicing. One
is the preservation-based proactive resource alloca-
tion mechanism, which divides network resources
into a pre-allocatable part and a preserved part.
The other is from implementation aspects of opti-
mizing resource allocations. Three schemes are dis-
cussed to minimize slice adjustment and avoid the
effect on network services caused by the adjust-
ment.
Robustness-Oriented Resource Preservation:
The newest data usually does not accurately reflect
the actual state of slices and networks because of
the lag in information gathering and processing.
During the lagging interval, there might be some
emergent cases where some services call for the
instant scaling up of existing slices instances or the
generation of new slices. The quality of services
will be significantly affected if there are not enough
resources available. Therefore, it is reasonable to
preserve a portion of resources in case of an emer-
gency. However, the exact portion of preserved
resources should be carefully determined to trade
off the robustness and resource efficiency.
Slicing Implementing: In automatic slicing,
adjusting the operating slice instances is a basic
requirement. However, dramatically changing
the allocated resources of slice instances on a
large scale is not expected. For robustness con-
siderations, three dynamic resource management
approaches for ANS are identified, which are
referred to as scaling, remapping, and appending,
respectively. The scaling mechanism is designed to
help VNFs dynamically share a pool of resources,
allowing the capacity to be expanded when nec-
essary. Remapping changes the mapping part of
an SFC instance to an alternative route of VNFs to
TABLE 2. Simulation settings.
(a) Simulation parameters.
Parameter [type] (unit)
Value
Number of data centers [Edge Net, Aggregation Net, Core Net] [5~20, 3~8, 1~5]
Resource capacity of data centers [Edge Net, Aggregation Net, Core Net] (units) [20, 30, 80]
Link capacity among data centers [Edge Net, Aggregation Net, Core Net] (Gbps)
[40, 80, 80]
Number of base-stations connected per edge data center 10
Latency threshold [Delay-sensitive, Non-delay-sensitive]
[20, 80
Traffic demand [Delay-sensitive, Non-delay-sensitive] [2, 50]
Cloud resource demand for NSI [Forward, Terminate] (units/Gbps) [1~7, 1]
Connectivity probability between data centers
0.4
Spectrum bandwidth for each BS (MHz) 50
Process capacity for each BS (Gbps/RRU)
5
(b) State space of ANS agent.
Type
State element
Topology
ØLink connectivity, status, and distributions
ØLocation of data center and base stations
ØUser distributions
Radio
ØRadio spectrum
ØUser traffic flows
ØScheduling priority，
ØLoad of BS
Cloud or edge
date center
ØAvailability of computation resources
ØDelay to process service
Ø
Capacity and load
Transport Net
ØTransport delay
ØTransmission delay
ØTraffic flow distribution
ØUtilization of link
ZHOU_LAYOUT.indd 112ZHOU_LAYOUT.indd 112 12/23/20 11:17 AM12/23/20 11:17 AM
Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on February 07,2021 at 07:28:57 UTC from IEEE Xplore. Restrictions apply.

IEEE Wireless Communications • December 2020 113
avoid influence on the entire instance. Appending
attaches a new SFC or slice instance to an existing
network slice instance for traffic split so that dra-
matically changed traffic flows will not impact the
performance of other flows sharing the slice.
sIMulAtIon And experIMent evAluAtIons
sIMulAtIon And nuMerIcAl results
To evaluate the performance of the scheme,
simulations were conducted in several network
scenarios, where the numbers of data centers in
edge, aggregation, and core networks were A
(5∼20), B (3∼8), and C (1∼5), respectively. The
resource capacities of the three types of data cen-
ters are 20, 30, and 100 units, while the network
link capacities are 40, 80, and 80 Gb/s and the
possibility of links between data centers are 0.4.
In the network, two types of NSIs (100 in total)
are deployed, 20 for delay-sensitive (DS) and 80
for non-delay-sensitive (NDS). Each NSI initially
produces 1 Gb/s traffic in total, and each one
Gpbs traffic consumes two unit resources in the
data center that terminate the services and one
unit resource in the data center that services pass
through. The parameters are summarized in Table
2a, and the utilized network information for state
construction is listed in Table 2b.
First, the delay-sensitive traffic was randomly set
in half of the delay-sensitive NSIs increasing from
1 to 7 Gb/s, while the non-delay-sensitive traffic
remained unchanged to evaluate the effect of auto-
matic resource coordination. The results are present-
ed in Fig. 3, based on a specific scenario, where A,
B, and C correspondingly equal 10, 5, and 3. From
Fig. 3a, it can be determined that as delay-sensi-
tive traffic increases, additional node resources in
the edge network domain will be refarmed from
non-delay-sensitive NSIs to serve the delay-sensitive
service. Therefore, delay-sensitive services take up
more node resources in the edge domain while
non-delay-sensitive services take up more resources
in the aggregation and CN domain.
Then a particular delay-sensitive NSI, with the
traffic curve as depicted in Fig. 3b. was examined.
The predicted traffic data curve is also presented in
Fig. 3b. At the marked time point, the automation
optimization of the NSI is triggered, and a scheme
for scaling should be given in a firm deadline. The
average delay of services within the NSI is depict-
ed in Fig. 3c. From the results, the proposed solu-
tion achieves the lowest delay after optimization
compared with two classical approaches, namely
dynamic reservation autonomous resource slic-
ing (DRARS) [9] and robust network slicing design
programming (NSDP) [15]. The average ratios of
NSIs with fulfilled SLA at the corresponding time
points are presented in Fig. 3(d), showing the
proposed approach has the highest success rate.
These phenomena are probably attributable to the
fast response enabled by the prediction module
and the DRL method, and the consideration of
FIGURE 3. Performance evaluation via simulation: a) resource utilization of both DS and NDS services with
the increase of DS traffic in different network segments. In each group, the bars represent aggregation,
transport, and core networks, respectively; b) the traffic curve and predicted traffic curve of a specific
NSI; c) the average delay performance obtained from different methods; d) the average ratio of slice
instances with fulfilled SLA.
1234567
Increase of delay-sensitive traffic (Gbps)
0
10
20
30
40
50
60
70
80
90
100
Resources utilization (%)
DS(Access) NDS(Access)
DS(Aggregation) NDS(Aggregation)
DS(Core) NDS(Core)
(a)
456789 10
Time (hours)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Unified Data Traffic
predicted
actual
trigger point
(b)
0.4 0.5 0.6 0.7 0.8 0.9
Unified Traffic
0
0.5
1
1.5
2
2.5
Average latency (ms)
Proposed
NSDP
RFPA
(c)
0.4 0.5 0.6 0.7 0.8 0.9
Unified Traffic
80
85
90
95
100
105
Ratio of slices with fulfilled SLA (%)
Proposed
DRARS
NSDP
(d)
Two aspects are con-
sidered in the robust
implementation of
network resource
slicing. One is the
preservation-based
proactive resource
allocation mechanism,
which divides network
resources into a pre-al-
locatable part and a
preserved part. e
other is from imple-
mentation aspects of
optimizing resource
allocations.
ZHOU_LAYOUT.indd 113ZHOU_LAYOUT.indd 113 12/23/20 11:17 AM12/23/20 11:17 AM
Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on February 07,2021 at 07:28:57 UTC from IEEE Xplore. Restrictions apply.

IEEE Wireless Communications • December 2020
114
system robustness in the reward design. A better
solution for global optimization is found before a
given deadline, allowing a better performance to
be achieved.
experIMent And AnAlyses
The proposed scheme is tested in an experimental
5G network for power grid industrial applications,
which is built by State Grid Cooperation of China
in the Shanghai Lingang area. The experimental
network has 4 gNBs with a function-virtualized
CN that has the project-oriented opening interface
to manage network slicing. This 5G experimental
network can cover three 35 KV power distribution
stations and one 110 KV power transformer sta-
tion. To support the proposed ANS mechanism,
the online-learning software stacks are packed to
be deployed in the servers of the CN. It currently
supports two types of slices: the eMBB and URLLC
slices. The eMBB slice is used to bear the opera-
tion-maintenance information monitoring services
while the uRLLC slice is used to bear the grid pro-
duction control services. A brief schematic digram
of the platform is depicted in Fig. 4a. Specifically,
we deploy a VLAN-ID pool for the convenience
of slicing management, which reserve some cer-
tain VLAN for eMBB and URLLC slices. This work
guarantees the timelines when creating new slices
since the control of transmission quality is based
on the SLA with VLAN ID.
To validate the proposed mechanism, one
uRLLC instance for distribution automation and
one eMBB instance for distribution station video
monitoring are initially run. A temporary video
robot inspection service then initiates and requests
an eMBB slice. In a video robot inspection (VRI)
task, the patrol robot continuously captures video
of the observed scenes of the power grid system
for potential issues. The video is then delivered to
the cloud for further instant analyses. In the estab-
lished settings, the data rate of the video changes
and the ANS agents will decide whether to create
a new eMBB slice to bear the service or to use the
existing eMBB slice.
The results are as depicted in Fig. 4b. It can be
observed that when the VRI service is initiated,
FIGURE 4. Performance evaluation via test bed: a) the functional architecture of the test bed; b) the aver-
age delay and resource utilization performance of slices, where A to C indicate the moments when the
VRI service initiates, quadruples, and stops without ANS, while D to F indicate the moments with ANS
enabled.
Project - oriented App. Enabling platform
Embedded ANS
Software Stacks
NS management system (telelcom operator)
5G communication
terminal 5G gNB
Power automation
terminal
Video monitoring
terminal Openstack Virtualization
X86 Infrastructure
…
VNF2
VNF1
Monitoring
Power control
Grid App. System
5G CN
eMBB slice uRLLC slice
(a)
(b)
As the VRI data rate
quadruples, the ANS
agent decides to gen-
erate a new eMBB
slice to serve the VRI.
Aer the VRI flow
is shied to a new
slice, the delay per-
formance of the old
eMBB slice returns to
normal. Similar cases
can be observed in
the resource utilization
degree of slices.
ZHOU_LAYOUT.indd 114ZHOU_LAYOUT.indd 114 12/23/20 11:17 AM12/23/20 11:17 AM
Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on February 07,2021 at 07:28:57 UTC from IEEE Xplore. Restrictions apply.

IEEE Wireless Communications • December 2020 115
the average delay of the eMBB slice experience
increased because the ANS let the VRI service
share the existing eMBB slice, while the URLLC
slice exhibited no obvious changes. However,
as the VRI data rate quadruples, the ANS agent
decides to generate a new eMBB slice to serve the
VRI. After the VRI flow is shifted to a new slice, the
delay performance of the old eMBB slice returns
to normal. Similar cases can be observed in the
resource utilization degree of slices.
conclusIon
This article investigates the end-to-end network
slicing technology for IoT services in smart city.
To improve the efficiency of slicing, an online
machine learning-driven framework of automat-
ic network slicing is specifically proposed to
automate the process, jointly employing deep
learning-based slicing prediction, unsupervised
learning-based anomaly detection and reinforce-
ment learning-based slicing scaling. Some robust-
ness-oriented slicing implementation schemes are
also identified for the performance guarantee of
delay-sensitive and emergency IoT services. The
simulation and experiment results verify the effec-
tiveness of the proposed approach for service-dif-
ferentiated and traffic-varied IoT scenarios. To deal
with a large-scale network, the ANS agent can be
implemented in a distributed way and keep each
of the managed domain within a reasonable scale,
where federated and tiered learning tricks, as well
as more efficient exploration schemes, would assist
the distributed ANS agent training. Different from
5G/B5G, 6G is envisioned to be the infrastructure
for understanding the user’s intent and make the
right decision for these intents. The challenge for
robust auto-slicing may be how to interpret the
user’s intent to be the auto-slicing inputs according
to multi-dimension information, such as human
reputation and emotion, network resource, envi-
ronment and so on. These topics will be covered
in the authors’ future work.
AcknowledgMent
This work was supported by the National Key
R&D Program of China (2018YFE0205502).
references
[1] H. Sun, C. Wang, and B. I. Ahmad, From Internet of Things
to Smart Cities — Enabling Technologies, CRC Press, Taylor &
Francis Group, 2017.
[2] S. Li, L. D. Xu, and S. Zhao, “5G Internet of Things: A Sur-
vey,” J. Ind. Inf. Integr., vol. 10, June 2018, pp. 1–9.
[3] M. R. Raza et al., “Dynamic Slicing Approach for Multi-
Tenant 5G Transport Networks,” J. Opt. Commun. and
Netw., vol. 10, no. 1, Jan. 2018, pp. A77–90.
[4] Y. Jia et al., “Bankruptcy Game Based Resource Allocation
Algorithm for 5G Cloud-RAN Slicing,” Proc. WCNC, Apr.
2018, pp. 1–6.
[5] S. Vassilaras et al., “The Algorithmic Aspects of Network Slic-
ing,” IEEE Commun. Mag., vol. 55, no. 8, 2017, pp. 112–19.
[6] L. Velasco et al., “An Architecture to Support Autonomic
Slice Networking,” J. Lightwave Tech., vol. 36, no. 1, Jan.
2018, pp. 135–41.
[7] Y. Minami et al., “An Architecture and Implementation of
Automatic Network Slicing for Microservices,” Proc. NOMS
2018, Apr. 2018, pp. 1–4.
[8] C. Song et al., “Machine Learning Enabling Traffic-Aware
Dynamic Slicing for 5G Optical Transport Networks,” Proc.
Conf. Lasers and Electro-Optics (2018), Optical Society of
America, May 2018.
[9] G. Sun et al., “Dynamic Reservation and Deep Reinforce-
ment Learning Based Autonomous Resource Slicing for Vir-
tualized Radio Access Networks,” IEEE Access, vol. 7, 2019,
pp. 45 758–72.
[10] S. Dawaliby, A. Bradai, and Y. Pousset, “Adaptive Dynamic
Network Slicing in LoRa Networks,” Future Gener. Comp.
Sys., vol. 98, Sept. 2019, pp. 697–707.
[11] D. Bega et al., “a-OMC: Cost-Aware Deep Learning for
Mobile Network Resource Orchestration,” Proc. INFOCOM
2019 (WKSHPS), Paris, France, Apr. 2019, pp. 423–28.
[12] L. Le et al., “SDN/NFV, Machine Learning, and Big Data
Driven Network Slicing for 5G,” Proc. 2018 IEEE 5G World
Forum, 2018, pp. 20–25.
[13] L. Xiao et al., “Reinforcement Learning-Based Downlink
Interference Control for Ultra-Dense Small Cells,” IEEE Trans.
Wireless Commun., vol. 19, no. 1, 2020, pp. 423–34.
[14] S. Xu, Y. Qian, and R. Q. Hu, “Data-Driven Edge Intelli-
gence for Robust Network Anomaly Detection,” IEEE Trans.
Netw. Sci. Eng., 2019.
[15] A. Baumgartner et al., “Optimisation Models for Robust
and Survivable Network Slice Design: A Comparative Analy-
sis,” Proc. GLOBECOM 2017, Dec. 2017, pp. 1–7.
bIogrAphIes
Fanqin Z hou received his Ph.D. degree in automation from
the Beijing University of Posts and Telecommunications (BUPT),
China, in 2019. He is currently a postdoctoral fellow with the
State Key Laboratory of Networking and Switching Technology
in BUPT. His current research interests include intelligent network
slicing and resource management of mobile edge networks.
Peng Yu received his B.Eng. and Ph.D. degrees from BUPT in
2008 and 2013, respectively. He is currently an associate pro-
fessor at the State Key Laboratory of Networking and Switching
Technology, BUPT. His research interests are autonomic man-
agement and hybrid energy allocation in GreenNet.
Lei Feng received his B.Eng. and Ph.D. degrees in communication
and information systems from BUPT in 2009 and 2015. He is cur-
rently an associate pro fessor in the State Key Laboratory of Net-
working and Switching Technology, BUPT. His research interests
are resource management in wireless networks and smart grid.
Xuesong qiu is a professor and a Ph.D. supervisor with the State
Key Laboratory of Networking and Switching Technology, Beijing
University of Posts and Telecommunications. He has authored
about 100 SCI/EI indexed articles. He presided over a series
of key research projects on network and service management,
including the projects supported by the National Natural Science
Foundation and the National High-Tech R&D Program of China.
ZhiLi Wang is currently an associate professor with BUPT. His
main research directions are network management, communica-
tions software, and interface testing. He has won one National
Science and Technology Progress Awards and contributed more
than eight ITU-T international standards, and served as the Work-
ing Party Chair for ITU-T Study Group 2 and Working Party 2.
Luoming meng is a professor and Ph.D. supervisor with BUPT.
He is the Director of the Communications Software Technical
Committee of China Institute of Communications and the Chair-
man of the National Network Management Standards Study
Group. He is the project chief scientist of the China 973 Pro-
gram, the winner of the Yangtse River Scholar award, and the
Outstanding Youth Science Found Receiver award.
micheL K adoch received his Ph.D. degree from Concordia
University in 1992. He is currently a full professor with Ecole de
Technologie Superieure (ETS), University of Quebec, Montreal,
Canada. He is the director of the research laboratory LAGRIT
at ETS. As the principal investigator, he has managed and par-
ticipated actively in a research program on QoS for multicast
in high-speed networks sponsored by Bell Canada and NSERC.
He is presently working on reliable multicast in wireless ad hoc
networks and 5G heterogeneous networks.
Liang gong received his Ph.D. degree in communication and
information systems from Shang Jiao Tong University in 2011.
He is now working with the Academy of Broadcast Planning
NRTA of China. His research interests include physical layer
technologies such as resource alloction and massvie MIMO, and
network layer technolgies such as SDN and NFV technologies
for future mobile and core networks.
Xianjiong Yao is a senior engineer with the State Grid Shanghai
Electric Power Company. He has long been engaged in the pro-
fessional and technical management of electric power commu-
nication, and is responsible for the application, organization, and
implementation of 5G channel test in the national science and
technology special project on the PMU technology research in
power distribution network.
e challenge for
robust auto-slicing may
be how to interpret
the user’s intent to be
the auto-slicing inputs
according to multi-di-
mension information,
such as human rep-
utation and emotion,
network resource,
environment and so
on. ese topics will be
covered in the authors’
future work.
ZHOU_LAYOUT.indd 115ZHOU_LAYOUT.indd 115 12/23/20 11:17 AM12/23/20 11:17 AM
Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on February 07,2021 at 07:28:57 UTC from IEEE Xplore. Restrictions apply.