Received February 22, 2018, accepted March 29, 2018, date of publication April 2, 2018, date of current version

April 25, 2018.

Digital Object Identifier 10.1109/ACCESS.2018.2822398

A Service-Oriented Deployment Policy of

End-to-End Network Slicing Based
on Complex Network Theory
WANQING GUAN , XIANGMING WEN, LUHAN WANG , ZHAOMING LU, AND YIDI SHEN
Beijing Laboratory of Advanced Information Networks, Beijing University of Posts and Telecommunications, Beijing 100876, China
Beijing Key Laboratory of Network System Architecture and Convergence, Beijing University of Posts and Telecommunications, Beijing 100876, China
Beijing Advanced Innovation Center for Future Internet Technology, Beijing University of Posts and Telecommunications, Beijing 100876, China
Corresponding author: Wanqing Guan ([email protected])
This work was supported in part by the National Science and Technology Major Project under Grant 2018ZX03001016, in part by the
China Ministry of Education-CMCC Research under Grant MCM20160104, in part by the 111 Project B17007, in part by the Beijing
Municipal Science and Technology Commission Research under Project Z171100005217001, and in part by the Fundamental Research
Funds for the Central Universities.

ABSTRACT For fifth-generation wireless communication systems, network slicing has emerged as a key
concept to meet the diverse requirements of various use cases. By slicing an infrastructure network into
multiple dedicated logical networks, wireless networks can support a wide range of services. However, how
to fast deploy the end-to-end slices is the main issue in a multi-domain wireless network infrastructure. In this
paper, a mathematical model is used to construct network slice requests and map them to the infrastructure
network. The mapping process consists of two steps: the placement of virtual network functions and the
selection of link paths chaining them. To efficiently utilize the limited physical resources, we pay attention
to the service-oriented deployment by offering different deployment policies for three typical slices: eMBB
slices, mMTC slices, and uRLLC slices. Furthermore, we adopt complex network theory to obtain the
topological information of slices and infrastructure network. With the topological information, we define
a node importance metric to rank the nodes in node mapping. To evaluate the performance of deployment
policy we proposed, extensive simulations have been conducted. The results have shown that our algorithm
performed better in terms of resource efficiency and acceptance ratio. In addition, the average execute time
of our algorithm is in a linear growth with the increase of infrastructure network size.

INDEX TERMS 5G, network slicing, complex network theory, service-oriented deployment, end-to-end
slices.

I. INTRODUCTION isolated, managed independently and created on demand [3].

With the explosive growth of mobile data traffic, the mas- In NS based 5G system, resources of multi-domain infras-
sive terminal connection and the rise of various kinds of tructure network can be efficiently allocated to multiple net-
new applications, future wireless network needs to be agile, work slices according to the requirements of use cases [4].
programmable and open. Moreover, fifth-generation (5G) As shown in the Fig. 1, NS aims to logically separate the set of
networks are nowadays expected to satisfy different require- virtual network functions (VNFs) within the physical infras-
ments of numerous new services and support vertical mar- tructure to build dedicated and customized logical networks.
kets such as automotive, energy, food and agriculture, The International Telecommunication Union (ITU) [5]
healthcare, etc [1]. A wide range of verticals with diverse has identified three broad use case families: enhanced
requirements spur 5G networks to be flexible, scalable, mobile broadband (eMBB), massive machine-type commu-
manageable, customized, and allow multi-tenancy and multi- nications (mMTC), and ultra-reliable and low-latency com-
service support [2]. In order to realize the above vision, munication (uRLLC).
network slicing (NS) has been proposed as a concept for Although the concept of NS is still nascent, the poten-
slicing a common underlying physical network into multi- tial technologies for achieving it, software-defined network-
ple end-to-end (E2E) logical networks which are mutually ing (SDN) and network functions virtualization (NFV), have

2169-3536
2018 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 6, 2018 Personal use is also permitted, but republication/redistribution requires IEEE permission. 19691
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 W. Guan et al.: Service-Oriented Deployment Policy of E2E NS Based on Complex Network Theory

In order to address these challenges, we propose a service-

oriented deployment policy of E2E network slicing based on
complex network (CN) theory. CN theory aims at analyzing
topological characteristics and predict dynamical behaviors
of networked system [12]. Empirical studies have demon-
strated that many real-life communication networks exhibit
small-world and scale-free topological properties [13], [14].
Moreover, [15] demonstrates that wireless networks are
scale-free with a proper algorithm. Using CN theory analyze
the impact of topology on E2E NS deployment is reasonable
and reliable. The contributions of this work can be concluded
as following:
(1) A mathematical model is created to describe E2E NS
deployment. The mathematical model consists of infrastruc-
ture network model, network slice request (NSR) model and
slice deployment model. These models are through three
FIGURE 1. 5G network slicing architecture.
domains: access network, transport network and core net-
work.
(2) For meeting different service requirements, three ded-
many mature researches with concrete solutions and readily icated deployment policies are provided for three use case
available platforms. SDN architecture is an appropriate tech- families respectively. By analyzing different features of
nology for the configuration and control of the forwarding three use cases, three service-oriented deployment algorithms
planes of the underlying resources, while NFV can manage and three optimization objectives are used for eMBB slice,
the life cycle of network slices and orchestrate VNFs effi- mMTC slice and uRLLC slice accordingly.
ciently. Actually, Ordonez-Lucena et al. [6] have presented a (3) In order to improve the utilization of underlying
deployment example of NS by integrating SDN and NFV. [7] resource and increase revenues of service providers, the topo-
first propose an E2E 5G system framework and illustrate the logical properties of infrastructure are analyzed based on CN
techniques in the radio access network and the core network. theory. The topological properties network are combined with
However, as far as we know, few researches have provided the local resources of nodes in the deployment of NS.
efficient and practical deployment policies for E2E slices in (4) Extensive simulations are carried out to evaluate
wireless networks, and the technology development on NS is the performance of our service-oriented deployment policy.
mostly focused on slicing the core network. The results show that our deployent policy performs better
Reference [8] defines that network slices are chains in terms of resource efficiency and acceptance ratio, and
of VNFs and logical/physical resources meeting the ser- requires less time to execute.
vice requirements. Therefore, the realization of NS in The remainder of this paper is organized as follows.
essence is the deployment and orchestration of VNFs, and In Section II we discuss the related work. Infrastructure net-
there are plentiful related researches on VNF placement. work Model, NSR model and slice deployment model are pre-
References [9], [10] propose the algorithms of VNFs place- sented in Section III. Section IV introduces the details of three
ment, which only for a single type of VNF chain request deployment algorithms based on CN theory. In Section V,
without considering diverse service requirements. Besides, we present the simulation results and analyses. Finally,
the mapping from slices to the infrastructure networks is Section VI draws the conclusions of the paper.
a typical virtual network embedding (VNE) problem [11].
As these two problems have been studied maturely in recent II. RELATED WORK
years, deployment of VNFs brings several advantages in In this section, we briefly introduce the structural character-
terms of the resource utilization and cost saving. Currently, istics in CN theory which are useful enablers for the purpose
little works in existing literature have been done on the of obtaining the topological information. We also give a
deployment of E2E network slices although it is necessary short summary of works in E2E network slicing for 5G net-
for the realization of NS. On one hand, E2E slices need works. In addition, we review some existing works on VNF
to be instantiated rapidly and should support cross-domain placement and VNE problem, showing their contribution to
deployment. On the other hand, the resources of infrastruc- deployment of NS.
ture network need to be allocated dynamically in order to
support a diverse set of use cases. Owing that different use A. CN THEORY
case families have different demands on the multi-domain According to the analysis of many real-world networks,
resources, NS deployment should be service-oriented. The several topological properties have been proposed to cap-
above-mentioned two respects raise extreme challenges for ture structural characteristics of various networks in the past
NS deployment in wireless communication systems. decades. For instance,

19692 VOLUME 6, 2018

W. Guan et al.: Service-Oriented Deployment Policy of E2E NS Based on Complex Network Theory

1) DEGREE resources, these works do not consider the characteristics

The degree of a node measures the number of edges that of infrastructure network, and ignore the investigation of
connect to it, which reflects the level of influence. The node the placement algorithm matching the resource demands of
degree can be formulated as providers.
X
di = δij . (1) D. VIRTUAL NETWORK EMBEDDING
j∈N
The topic of VNE problem has received significant attention
The parameter δij takes the value 1 if node i and node j are of many researchers, and the amount of literature on VNE
directly connected, otherwise it takes the value 0. topic is considerable. Fischer et al. [31] provide a compre-
hensive survey on VNE algorithms. Dealing with the virtual
2) BETWEENNESS CENTRALITY resources allocation both in nodes and links, VNE can be
The betweenness centrality quantifies how much a node is divided in two sub-problems: Virtual Node Mapping (VNoM)
found between the path linking other pair of nodes. The and Virtual Link Mapping (VLiM). Solving VNoM and
betweenness centrality is defined as the fraction of shortest VLiM in an isolated way or a coordinated way divides VNE
paths between any pair of nodes that travel through the node, methods into uncoordinated VNE method and coordinated
which can be denoted by VNE method. The example of uncoordinated VNE method
X σst (i) was proposed in [32] while coordination of VNE was first
bi = . (2) achieved in [33]. In addition, M. Chowdhury et al. introduce
σst ViNEYard [34] for single domain and PolyViNE [35] for
s6=i6=t
multi-domain VNE. Considering that network slices are the
In this equation, σst is the total number of shortest paths from chains of VNFs, the large number of VNE methods provide
node s to node t and σst (i) is the number of those paths that some valuable algorithms that can be used in deployment of
pass through node i. network slices.
B. E2E NETWORK SLICING
III. MATHEMATICAL MODEL OF NETWORK SLICING
The concept of network slicing has captured many In this section, we describe the mathematical model of net-
attentions. Many researches focus on radio resource virtu- work slicing including infrastructure network model, net-
alization [16], [17], management and orchestration of net- work slice request (NSR) model and slice deployment
work functions [18], [19], mapping of service and service model.
description [16]. Since NS allows operators to customize
networks according to various service demands, both industry A. INFRASTRUCTURE NETWORK MODEL
and academia introduce many realization models of NS.
The deployment of NS requires topological information
Although NS has been implemented in fixed network, E2E
of infrastructure network including the structural charac-
NS of the mobile network still lags behind. There is still
teristics of physical nodes (e.g. base stations (BS), optical
lacking efficient deployment and management of E2E slices.
switches (OS), core nodes (CN)). The infrastructure network
Zhou et al. [20] illustrate NS as a service (NSaaS) and intro-
can be abstracted as undirected weighted graph, which can
duce the business model of NSaaS. Reference [21] introduces
be denoted as GI = (NI , EI , CI , BI ). Similar to some previ-
PERMIT slice orchestration system which is first to consider
ous literatures, we only take into consideration the capacity
E2E slicing.
of nodes and bandwidth of links. NI stands for the set of
physical nodes, which can be partitioned into the set of base
C. VNF PLACEMENT
stations NIBS , the set of optical switches NIOS and the set of
Recently, with the rise of NFV concept [22], [23] and open-
core nodes NICN , NI = NIBS ∪ NIOS ∪ NICN . EI stands for the
source platform, such as OPNFV [24], OpenMANO [25]
set of physical links including the wireless wave links EIwl
and OpenBATON [26], there are significant efforts to VNF
and wired optical links EIol , EI = EIwl ∪ EIol .
management and orchestration [9]. In order to dynamically
allocate virtualization resources to VNFs, there are also [  
many algorithms for the deployment of the service function EIwl = EIwl nBS
I (3)
chain (SFC). These algorithms [27], [28] provide solutions nBS BS
I ∈NI

to optimize VNF placement problem. Clayman et al. [29]

where EIwl nBS is the subset of wireless link ewl


describe an architecture based on a distributed orches- I I responsible
BS
of connecting nI and other nodes. CI stands for the capac-
trator which enables automated placement of virtual ele-
ments. Similar with the dynamic VNF placement algorithm, ity of physical nodes, which includes BS wireless channel
Ghaznavi et al. [30] introduce Elastic Virtual Network Func- capacity CIBS and computing resource of core cloud CICN ,
tion Placement (EVNFP) problem and present a model for CI = CIBS ∪ CICN . BI stands for the bandwidth set of physical
minimizing operational costs in providing VNF services. Tar- links, including the available bandwidth set of wireless wave
geting to achieve optimal resource utilization of infrastructure links Bwl ol wl
I and wired optical links BI , BI = BI ∪ BI .
ol

VOLUME 6, 2018 19693

 W. Guan et al.: Service-Oriented Deployment Policy of E2E NS Based on Complex Network Theory

B. NETWORK SLICE REQUEST MODEL 2) VARIABLES

CIwl ewl wl


In our model, the set of NSRs consists of three types of slices • I : Channel capacity of the wireless link eI ,
for three use case families, which can be denoted by RNS , ewl wl
I ∈ EI
;
RNS = Re ∪ Rm ∪ Ru . Re represents eMBB slice, Rm rep- BI eI : Bandwidth assigned to the wireless link ewl
wl wl
• I ;
resents mMTC slice and Ru represents uRLLC slice. Each • µnBS ,nBS : Binary variable, if nBS of G is mapped to nBS
R
request is regarded as GR = (NR , ER , CR , BR , TR ) where R I R I

NR represents nodes of network slice, ER represents links, of GI , µnBS ,nBS = 1; Otherwise, µnBS ,nBS = 0;
R I R I
CR denotes capacity, BR denotes bandwidth and TR is the • µnOS ,nOS : Binary variable, if nOS OS
R of GR is mapped to nI
R I
duration of the NSR remaining in the infrastructure network. of GI , µnOS ,nOS = 1; Otherwise, µnOS ,nOS = 0;
Thus, GRe = NRe , ERe , CRe , BRe , TRe is for request Re , and


R I R I
• µnCN ,nCN : Binary variable, if nCN
R of GR is mapped to
similarly GRm and GRu is for requests Rm and Ru respectively. R I
nCN
I of GI , µnCN ,nCN = 1; Otherwise, µnCN ,nCN = 0;
R I R I
• νewl ,ewl : Binary variable, if a virtual link ewlI of GR
R I
traverse the physical wireless link eI , νewl ,ewl = 1;
wl
R I
Otherwise, ewl I , νewl wl = 0;
R ,eI
• νeol ,eol : Binary variable, if a virtual link eolI of GR tra-
R I
verse the physical optical link eol I , ν e ,e = 1; Other-
ol ol
R I
wise, eolI , νeol ol = 0;
R ,eI
• ξnBS ,ewl Binary variable, if nBS wl
R of GR is served by eI ∈
R I
EIwl nBS R , ξnBS ,ewl = 1; Otherwise, ξnBS ,ewl = 0;


R I R I
FIGURE 2. The schematic diagram of deploying NSRs.
IV. DEPLOYMENT POLICY OF NS BASED ON CN
The main objective of deployment process is minimizing
the deployment cost on the premise of meeting NS require-
C. SLICE DEPLOYMENT MODEL
ments. In addition to the main objective, three types of slices
Slice deployment is a process in which nodes of slice requests have their peculiar objectives because of different service
are mapped onto substrate nodes and links are mapped onto demands. Besides, we propose an VNFs placement algorithm
substrate paths on the premise of meeting service demands of based on CN theory which was used to analyze the topologi-
slices. The mapping process consists of two stages, the node cal properties of nodes in the underlying network.
mapping and the link mapping. The schematic diagram of
deploying NSRs is demonstrated in Fig. 2. As shown in Fig. 2, A. MAPPING ALGORITHM
VNFs of eMBB slice are mapped in nodes B1 , C1 , D3 and
1) PLACING VNFS
the links chaining VNFs are mapped in path B1 → B2 →
Placing VNFs means to select the physical nodes of substrate
C1 → C3 → D3 . Similarly, VNFs of uRLLC slice are
network as host for the virtual nodes of NSRs under the
mapped in nodes A3 , C2 , D2 and the links chaining VNFs are
condition of satisfying the capacity requirements. According
mapped in path A3 → A4 → C2 → C4 → D2 . A node
to the literatures of VNE problem, the local resources for
of NSR can only be mapped on a node of infrastructure
nodes are measured by
network, and a node of infrastructure network can only host
a node from the same of NSR. Below, there is a summary of
X
NR (i) = C (i) · BW (l) (4)
parameters and variables that are used for the formulation of l∈s(i)
the mathematical model and the introduction of the decision
variables: where C (i) represents the capacity of node i, s (i) represents
the set of links that directly connected to the node i, BW (l)
1) PARAMETERS represents the current available bandwidth of link l. The main
shortcoming of this measurement is ignoring the topological
CICN nCN : Computing resource of core node nCN


• I I , characteristics of nodes. Hence, we combine the degree and
CN CN CICN nCN CN ;
P

nI ∈ NI , I = CI betweenness centrality of nodes to measure node importance
nCN CN
I ∈NI in the step of placing VNFs.
• CRCN nCN


R : Computing resource requirement of the vir-
First, the degree and betweenness centrality of nodes are
tual core
node nCN CN
R ∈ NR ; normalized. Considering that the degree of node is not exceed
• wl wl
BR eR : Bandwidth requirement of the virtual wireless N − 1 when the total number of nodes is N , the normalization
link ewl
R
∈ ER ;
wl
of the degree can be expressed by
• Bol ol
R eR : Bandwidth requirement of the virtual optical di
link eR ∈ ERol .
ol di0 = . (5)
N −1
19694 VOLUME 6, 2018
W. Guan et al.: Service-Oriented Deployment Policy of E2E NS Based on Complex Network Theory

Similarly, the betweenness centrality of node can be normal- Algorithm 2 The Node Mapping Algorithm Based on BFS
ized by using Input: RNS : the arrived NSR
2bi Output: Mnode : the results of node mapping
b0i = (6) 1: Sort virtual nodes with Algorithm 1.
(N − 1) (N − 2)
2: Sort physical nodes according to their NI values in non-
because the maximum of bi is (N − 1) (N − 2)/2. In the increasing order.
case of reaching maximum, each node pair of the network 3: for each virtual node do
has at least one shortest path that travel through the node. 4: if it is root R then
Based on these normalized metrics of nodes, the weighting 5: it is mapped into the physical node with the great-
0 0
parameters of node i can be given by d i +b i
2 . Therefore, est value of NI .
combining the local resource and weighting of each node, 6: else
the node importance of i can be given by 7: find the parent node P of it.
 0
d i +b0 i find the mapped physical node I for P.

8:
NI (i) = NR (i) × . (7) 9: find the neighbor nodes of I as the candidate
2
physical nodes C.
According to the node importance, we use graphical 10: choose one of C which owns the greatest value of
breadth-first-search (BFS) algorithm to sort nodes, and map NI in the premise of satisfying the capacity requirements.
the virtual nodes to physical nodes based on BFS. The sorting 11: end if
algorithm of virtual nodes are listed in Algorithm 1. Based on 12: return Mnode .
the sorting algorithm, the node mapping algorithm is intro- 13: end for
duced in Algorithm 2. The node degree describes the number
of its neighborhood nodes and node betweenness centrality
describes the importance of a node with respect to the shortest Algorithm 3 The Link Mapping Algorithm Based on KSP
path. Nodes with higher degree and betweenness centrality Input: RNS : the arrived NSR
mean that they are frequently used and many shortest paths Output: Mlink : the results of link mapping
pass through them. Hence, with BFS algorithm, giving prior- 1: Sort the virtual links according to bandwidth in non-
ity to the nodes with higher node importance will reduce link increasing order.
resources usage in the infrastructure network. 2: for each virtual link l. do
3: calculate the bandwidth requirement BW (l).
4: remove the physical links that can not meet the band-
Algorithm 1 The Sorting Algorithm for Virtual Nodes
width requirement.
in NSRs
Input: NR : the set of virtual nodes in NSR 5: according to Mnode , find the mapped physical nodes
Output: NR0 : the sequence of sorted virtual nodes of l.
1: Calculating NI value of each virtual node.
6: find the physical shortest path between these two
2: Sorting the virtual nodes by NI value in non-increasing
physical nodes by using Floyd algorithm.
order. 7: return Mlink .
8: end for
3: Selecting the virtual node with highest NI value as R.
4: Using R as the root node, traverse the graph of NSR using
BFS algorithm, and get the BFS tree T .
5: Sorting the virtual nodes in each layer of T according to the ultimate goal is to take advantage of infrastructure
NI value in non-increasing order. resources efficiently. Hence, the main objective can be
6: Return NR0 . expressed by
 
X X
2) CHAINING VNFS min  CR (nR ) · µnR ,nI + BR (eR ) · νeR ,eI  . (8)
nR ∈NR eR ∈ER
The procedure of creating paths that interconnect the VNFs
placed nodes would be achieved on the basis of k-shortest Furthermore, the objectives of different kinds of slices are
paths (KSP) algorithm. KSP algorithm is used to select suit- also determined by their service demands. These objectives
able physical paths in the premise of satisfying the bandwidth are listed below.
resource requirements. After removing the link paths that do
not satisfy the requirements, Floyd algorithm is used to calcu- 1) EMBB SLICE
late the shortest path. More details are shown in Algorithm 3. The eMBB usage scenario covers a range of cases, including
wide-area coverage and hotspot. For the hotspot case, i.e. for
B. OBJECTIVES OF THREE TYPES OF SLICES an area with high user density, very high traffic capacity is
In this section, we introduce the objectives of NSRs deploy- needed, while the requirement for mobility is low and user
ment problem. No matter which kind of slice is required, data rate is higher. This kind of slice does not require strict

VOLUME 6, 2018 19695

 W. Guan et al.: Service-Oriented Deployment Policy of E2E NS Based on Complex Network Theory

X X  
delay and plentiful resources. Hence, the deployment objec- µnBS ,nBS · ξnBS ,ewl ·CIwl ewl
I
R I R I
tive of eMBB slices should be maximizing the remaining nBS BS wl wl BS
I ∈NI eI ∈EI nI

resources of physical nodes, which can be represented by  

R ,
≥ CRBS nBS ∀nBS BS
  R ∈ NR (17)
X X
max  CI (nI ) − CR (nR ) · µnR ,nI  (9) Eq. 14 ensures that the bandwidth occupied by all the wireless
nI ∈NI nR ∈NR channels should not exceed the total available bandwidth BwlI
for each BS. Eq. 15 ensures that the channel capacity sum
2) MMTC SLICE of wireless links in each BS should not exceed capacity of
The mMTC usage scenario is characterized by a very large this BS allocated by the control plane. Eq. 16 ensures that
number of connected devices typically transmitting a rela- the capacity sum of all virtual BS undertaken in this BS
tively low volume of non-delay sensitive data. This use case should not exceed its allocated capacity. Eq. 17 ensures that
has plenty of connections, which results in the requirement of the allocated capacity for each virtual BS should not be less
high computing resources and low congestion rate. Therefore, than its capacity requirement.
the deployment objective should be minimizing the usage of X
bandwidth on physical links. In other words, the remaining µnOS ,nOS = 1, ∀nOSI ∈ NI
OS
(18)
R I
bandwidth on physical links should be maximized. Thus, nOS OS
R ∈NR
the deployment objective of mMTC slice can be denoted as
X
µnOS ,nOS ≤ 1, ∀nOS OS
R ∈ NR (19)
  R I
X X nOS OS
I ∈NI
max  BI (eI ) − BR (eR ) · νeR ,eI  (10)
Eq. 18 ensures that each virtual OS only should be mapped
eI ∈EI eR ∈ER
to a physical OS. Eq. 19 ensures that each physical OS only
3) URLLC SLICE can undertake a virtual OS for each NSR.
X
The uRLLC usage scenario has stringent requirements for µnCN ,nCN = 1, ∀nCN ∈ NICN (20)
R I I
capabilities such as throughput, latency and availability. nCN CN
R ∈NR
Some examples include wireless control of industrial manu- X
facturing, remote medical surgery, transportation safety, etc. µnCN ,nCN ≤ 1, ∀nCN CN
R ∈ NR (21)
R I
The QoS guarantee of this use case is low latency, which nCN CN
I ∈NI
causes that the deployment objective should be minimizing X    
µnCN ,nCN ·CRCN nCN
R ≤ CICN nCN
I , ∀nCN CN
I ∈ NI
the delay of slices. We transfer delay time into the number R I
of hops, so minimizing the delay means minimizing each nCN CN
R ∈NR

physical path length. Hence, deployment objective of uRLLC (22)

X    
slices is µnCN ,nCN ·CICN nCN
I R ,
≥ CRCN nCN ∀nCN
R ∈ NRCN
X X R I
νewl ,ewl + νeol ,eol nCN CN
min (11) I ∈NI
R I R I
ewl wl eol ol (23)
R ∈ER R ∈ER

These objectives are subject to Eq. 20 ensures that each virtual CN only should be mapped to
X a physical CN. Eq. 21 ensures that each physical CN only can
µnBS ,nBS = 1, ∀nBS BS
I ∈ NI (12) undertake a virtual CN for each NSR. Eq. 22 ensures that the
R I
nBS BS
R ∈NR computing resource of each physical CN can satisfy the total
requirement of all virtual CNs mapped in it. Eq. 23 ensures
X
µnBS ,nBS ≤ 1, ∀nBS BS
R ∈ NR (13) that the computing resource of selected CN should not be less
R I
nBS BS than the computing resource requirement of virtual CNs.
I ∈NI

Eq. 12 ensures that each virtual BS only should be mapped C. DEPLOYMENT STRATEGIES OF THREE TYPES OF SLICES
to a physical BS. Eq. 13 ensures that each physical BS only As a functional block in the management framework of 5G
should undertake a virtual BS for each NSR. network slicing [36], the deployment strategies we proposed
 
are conducted under the assumption that the arrived NSR has
X
Bwl
I ewl
I ≤ Bwl
I (14)
been identified as one of those three types. In our proposed
ewl wl
I ∈EI
X     deployment policy of NS, we give three different strate-
CIwl ewl
I ≤ CI
BS
nBS
I , ∀nBS BS
R ∈ NR (15) gies to three types of slices. After the arriving of NSRs,
ewl wl
I ∈EI nBS
I


these requests are classified, then implemented by different
X     mapping algorithms respectively. Meanwhile, the resource
µnBS ,nBS · CRBS nBS
R ≤ CIBS nBS
I , ∀nBS BS
I ∈ NI efficiency (RE) and acceptance ratio (AR) of NSRs are cal-
R I
nBS BS
R ∈NR culated. Resource efficiency is defined as the revenues and
(16) cost ratio. The achieved revenues of accepting a NSR by the

19696 VOLUME 6, 2018

W. Guan et al.: Service-Oriented Deployment Policy of E2E NS Based on Complex Network Theory

infrastructure network can be defined as the sum of nodes 1) ALGORITHM A FOR EMBB SLICE
capacity and link bandwidth requirements of a NSR. And the In the deployment algorithm for eMBB slice, the virtual BSs
cost can be defined as the sum of nodes capacity and link are first sorted and mapped according to the node mapping
bandwidth resources of the infrastructure network. Hence, algorithm (Algorithm 2). After mapping the virtual BSs,
the resource efficiency can be formulated as follows: the virtual CNs are mapped similarly considering the com-
P
CR (n) +
P
BR (l) puting resource requirements. Then, the mapping of virtual
n∈NR l∈ER OSs are finished when searching the shortest paths between
RE = P (24) each BS-CN pair. Finally, the virtual links are mapped with
CR (n) + BR (l) × hop (l)
P
n∈NR l∈ER the link mapping algorithm (Algorithm 3). Details are shown
in Algorithm 5.
Where CR (n) represents the capacity of node n and BR (l)
represents the bandwidth of link l, hop (l) represents the map-
ping path length of link l. Furthermore, acceptance ratio is the Algorithm 5 Deployment Algorithm A for eMBB Slice
ratio of the number of NSRs which have been successfully Input: GI and Re
mapped and the total number. Hence, it can be formulated as Output: Mnode and Mlink
PT 1: sort the virtual nodes by Algorithm 1.
NU Macc 2: do node mapping of BSs by Algorithm 2.
AR = Pt=0 T
(25)
3: do node mapping of CNs by Algorithm 2.
t=0 NU Marr
4: do link mapping by Algorithm 3.
In the above formula, NU Macc represents the number of 5: do node mapping of OSs according to the link mapping.
NSRs that have been accepted while NU Marr denotes the 6: return the mapping results.
number of NSRs that have been arrived. Details are presented
in Algorithm 4. For the deployment of NSRs, we consider
two conditions which are the static deployment and dynamic
2) ALGORITHM B FOR MMTC SLICE
deployment. The former means that slices are permanent
In the deployment algorithm for mMTC slice, the virtual CNs
once they are deployed successfully. However, in the latter
are mapped firstly. Next, take these CNs as the source end-
condition, NSRs have life time and the resource allocated to
points and find out the candidate BSs as the target endpoints.
NSRs will be recycled at the end of life time. For dynamic
Then, we select the shortest path from the set of candidate
deployment, step 2 in Algorithm 4 will check if there are any
paths between CN and candidate BSs to map the virtual link.
NSRs which need to be recycled and calculate the resources
Details are shown in Algorithm 6.
after the recycle of NSRs. However, for static deployment,
there are no any NSRs which need to be recycled during the
deployment of the sequence of NSRs. Algorithm 6 Deployment Algorithm B for mMTC Slice
Input: GI and Rm
Output: Mnode and Mlink
Algorithm 4 The NSRs Implementation Algorithm 1: sort the virtual nodes by Algorithm 1.
Input: GI = (NI , EI , CI , BI ) and RNS = Re ∪ Rm ∪ Ru 2: do node mapping of CNs by Algorithm 2.
Output: AR, RE, Mnode and Mlink 3: select available BSs as candidate BSs.
1: while RNS 6 = ∅ do 4: search the set of candidate paths between CN and candi-
2: Calculating the resources of infrastructure net- date BSs.
work GI . 5: do link mapping based on candidate paths by
3: if NSR is eMBB slice Re then Algorithm 3.
4: deploy it using Algorithm 5. 6: do node mapping of OSs according to the link mapping.
5: else if NSR is mMTC slice Rm then 7: do node mapping of BSs.
6: deploy it using Algorithm 6. 8: return the mapping results.
7: else
8: deploy it using Algorithm 7.
9: end if 3) ALGORITHM C FOR URLLC SLICE
10: if Mnode and Mlink are not null then
In the deployment algorithm for uRLLC slice, we first find
11: update the resources of infrastructure
out the set of candidate BS-CN pairs and search all the
network GI .
possible routing paths between these candidate pairs as the
12: calculating resource efficiency RE.
candidate path set. According to the number of virtual links
13: else if Mnode and Mlink are null then
of NSR, we select the shortest paths from the set of candidate
14: calculating acceptance ratio AR.
paths to map the virtual links. Then we map the virtual BSs
15: end if
and CNs into the physical endpoints of the selected paths.
16: end while
Details are shown in Algorithm 7.

VOLUME 6, 2018 19697

 W. Guan et al.: Service-Oriented Deployment Policy of E2E NS Based on Complex Network Theory

Algorithm 7 Deployment Algorithm C for uRLLC Slice

Input: GI and Ru
Output: Mnode and Mlink
1: sort the virtual nodes by Algorithm 1.
2: select available BSs as candidate BSs.
3: select available CNs as candidate CNs.
4: search the set of candidate paths between candidate CN
and candidate BS.
5: do link mapping based on candidate paths by FIGURE 3. The generative process of BA scale-free network.
Algorithm 3.
6: do node mapping of OSs according to the link mapping. TABLE 1. The settings of parameters.
7: do node mapping of BSs.
8: do node mapping of CNs.
9: return the mapping results.

V. SIMULATION RESULTS AND ANALYSIS

In this section, the simulation environment settings and
the simulation results are discussed. The integral linear
programming (ILP) model presented in previous section
is implemented using YALMIP platform embedded in
MATLAB which integrates the commercial programming
solver CPLEX 12.7. The Algorithm 1-7 are achieved using
MATLAB 2015b and simulations are carried out on a laptop
with four 2.4GHz CPU cores and 12GB memories.

A. EXPERIMENTAL ENVIRONMENT SETTINGS time duration of slice is ranged from 15 units to 45 units and
In the growth of communication network, a forthcoming node the arrive time of each NSR is between 1 and 50.
has the tendency to connect itself to the nodes with large
degrees, and the node-degree distribution follows a power- B. EXPERIMENTAL RESULTS AND ANALYSIS
law form. Hence, we apply the algorithms of BarabÃąsi- We use the resource efficiency and accptance ratio which
Albert (BA) scale-free networks [37] to generate the topology were introduced in the previous section to evaluate the per-
of NSRs and infrastructure network. The generating principle formance of the NSRs implementation algorithm. As only
is as follows: a few deployment algorithm of E2E slices can be found
in current literatures, our deployment algorithms are com-
• Growth: Starting from a connected network of small
pared with a simulated annealing (SA) [38] based approach,
size with n0 ≥ 1 nodes, introduce one new node to the
the VNE algorithm LAVA [39] and the VNF placement algo-
existing network each time, and this incoming new node
rithm GLL [40]. The SA-based approach here is similar with
is connected to n existing nodes, where 1 ≤ n ≤ no .
SA-based middel box placement algorithm in [27]. LAVA
• Preferential Attachment: The incoming new node is con-
is a service-oriented substrate resource slicing strategy, and
nected to each of the n Qexisting nodes, according to the
GLL is a greedy algorithm which is based on functions
following probability: i = PNkci , where Nc denotes
j=1 kj being mapped to the node with highest available capacity.
the total number of current existing nodes and ki is the Moreover, we analyze the execution time of three different
degree of node i. deployment strategies while the number of NSRs increases.
Obviously, after t steps of adding new nodes, the BA network To eliminate the randomicity and improve the accuracy,
will have N = n0 + t nodes. Fig. 3 illustrates the growing the simulations have been operated 50 times independently
process of a BA scale-free network. and took the average.
The specific parameter settings are shown in Table.1.
We assume that the total number of physical nodes in infras- 1) RESOURCE EFFICIENCY
tructure network can be N = 100, 200, 300. These nodes In Fig. 4, we compare the RE results of eMBB slice,
are divided into three sets: BS nodes, OS nodes and CN mMTC slice, and uRLLC slice sequences respectively. Each
nodes. For the static deployment, the sequences of NSRs sequence of NSRs contains a single type of slices under the
consist of three different types of slices and the numbers of condition of static deployment. We calculate the results at
eMBB slices, mMTC slices and uRLLC slices are randomly N = 300 to guarantee sufficient physical resource. It can be
decided. For dynamic deployment, the difference is that the seen that RE gradually decreases along with the increasing of

19698 VOLUME 6, 2018

W. Guan et al.: Service-Oriented Deployment Policy of E2E NS Based on Complex Network Theory

FIGURE 4. Resource efficiency of different types of slices. (a) eMBB slices. (b) mMTC slices. (c) uRLLC slices.

FIGURE 5. Acceptance ratio of different deployment algorithms. (a) Static deployment. (b) Dynamic deployment.

the number of NSRs. The main reason is that the bigger the two conditions are plotted respectively in Fig. 5. As illustrated
number of NSRs, the longer the average mapping path length in Fig. 5(a), the AR of four algorithms decrease gradually
and the more link bandwidth resources are consumed. The with the increasing number of NSRs and time duration of
results show that NSRs implementation algorithm and LAVA each NSR. And the AR of our algorithm is higher than others
perform better than two heuristic algorithms, SA-based and when the number of NSRs is lower than 28 approximately.
GLL. Thus, it can be concluded that service-oriented resource However the AR of our algorithm is lower than SA-based
slicing approach can improve the efficiency of resource using. algorithm when the number of NSRs is more than 28. That is
Since we have improved the node’s ranking method with the because, our algorithm focuses on increasing the RE and pro-
topological characteristics based on CN theory, our algorithm vide different objectives for three types of slices. This more
performs even better than LAVA. In addition, we can see that strict deployment policy leads to lower AR when the number
the RE of uRLLC slices is highest compared to mMTC slices of NSRs is higher. It also can be observed that the AR of GLL
and eMBB slices. This is due to the fact that the average is higher than LAVA. The reason is that LAVA allocates the
mapping path length of links in the sequence of uRLLC slices substrate resources according to the service class, which is
is lowest, which is the result of the deployment objective of more strict than GLL. In Fig. 5(b), the infrastructure network
uRLLC slices. recycles resources with the ending of slices, which causes
more available resources. Hence, the number of NSRs that are
accepted successfully is higher than static deployment. When
2) ACCEPTANCE RATIO the time duration is long, the resources are not frequently
According to equation Eq. 22, the AR is proportional to updated and it causes the decreasing of AR.
the number of successfully deployed slices. We analyze the
relationship between the number of NSRs and the AR in the
condition of static deployment. Furthermore, we calculate the 3) EXECUTION TIME
AR when the time duration of each NSR is changed in the To evaluate the execution time, we use the proposed algo-
condition of dynamic deployment. The AR results of these rithms and SA-based approach to deploy three types of slices

VOLUME 6, 2018 19699

 W. Guan et al.: Service-Oriented Deployment Policy of E2E NS Based on Complex Network Theory

FIGURE 6. Execution time performance of different deployment algorithms. (a) Algorithm A for eMBB slices. (b) Algorithm B for mMTC slices.
(c) Algorithm C for uRLLC slices.

seperately in the condition of static deployment. In Fig. 6, simulations have shown that our proposed algorithm achieve
with the increasing number of slices, both average execution higher resource efficiency and acceptance ratio. Analysis of
time of our algorithms and SA-based algorithm increase. execution time prove that E2E network slices can be deployed
And the SA-based algorithm takes much more time than rapidly with our proposed service-oriented deployment
the proposed algorithm for three types of slices. It can also policy.
be noticed that average execution time of our algorithm is
different for three types of slices. For eMBB slices, it requires REFERENCES
quite less time and the execution time for uRLLC slices is [1] 5G Empowering Vertical Industries, 5GPPP and ERTICO,
highest. That is because, algorithm B and algorithm C need 2016. [Online]. Available: https://5g-ppp.eu/wp-content/
uploads/2016/02/BROCHURE_5PPP_BAT2_PL.pdf
to search the set of candidate paths. In addition, algorithm [2] View on 5G Architecture, 5GPPP, 2016. [Online]. Available: https://5g-
C searches the paths between candidate BS nodes and can- ppp.eu/wp-content/uploads/2014/02/5G-PPP-5G-Architecture-WP-July-
didate CN nodes, which cause more time consumption than 2016.pdf
[3] Description of Network Slicing Concept, NGMN-Alliance, 2016, vol. 1.
algorithm B. For the proposed algorithm, the search of paths [Online]. Available: https://www.ngmn.org/fileadmin/user_upload/
between BS domain and CN domain becomes the main factor 160113_Network_Slicing_v1_0.pdf
of execution time. However, for SA-based algorithm, it takes [4] H. Zhang, N. Liu, X. Chu, K. Long, A.-H. Aghvami, and V. C. M. Leung,
‘‘Network slicing based 5g and future mobile networks: Mobility, resource
too much time to get an optimal location for each node and management, and challenges,’’ IEEE Commun. Mag., vol. 55, no. 8,
an optimal path for each link. Moreover, as the total number pp. 138–145, Aug. 2017.
of physical nodes increases, the execution time increments of [5] ‘‘Imt vision–framework and overall objectives of the future develop-
ment of IMT for 2020 and beyond,’’ Tech. Rep., 2015. [Online].
our deployment algorithm are less than the SA-based algo- Available: https://www.itu.int/dms_pubrec/itu-r/rec/m/R-REC-M.2083-0-
rithm. As a heuristic method, time to find the optimal solution 201509-I!!PDF-E.pdf
for SA-based algorithm increases exponentially when the [6] J. Ordonez-Lucena, P. Ameigeiras, D. Lopez, J. J. Ramos-Munoz, J. Lorca,
and J. Folgueira, ‘‘Network slicing for 5G with SDN/NFV: Concepts,
underlying infrastructure network scales up. Oppositely, there architectures, and challenges,’’ IEEE Commun. Mag., vol. 55, no. 5,
is an approximately liner increase in the execution time of our pp. 80–87, May 2017.
algorithm. Hence, our algorithm performs better in terms of [7] Q. Li, G. Wu, A. Papathanassiou, and U. Mukherjee. (2016). ‘‘An end-to-
end network slicing framework for 5G wireless communication systems.’’
the execution time. [Online]. Available: https://arxiv.org/abs/1608.00572
[8] M. R. Sama, X. An, Q. Wei, and S. Beker, ‘‘Reshaping the Mobile core
network via function decomposition and network slicing for the 5G era,’’
VI. CONCLUSION in Proc. IEEE Wireless Commun. Netw. Conf. Workshops (WCNCW),
Apr. 2016, pp. 90–96.
The 5G network is envisioned to support a range of verticals [9] F. Bari, S. R. Chowdhury, R. Ahmed, R. Boutaba, and O. C. M. B. Duarte,
and use cases, which causes that network slicing draws a lot of ‘‘Orchestrating virtualized network functions,’’ IEEE Trans. Netw. Service
attention. In 5G network slicing, the deployment of different Manag., vol. 13, no. 4, pp. 725–739, Dec. 2016.
[10] R. Riggio, A. Bradai, D. Harutyunyan, T. Rasheed, and T. Ahmed,
type of slices is one of the most challenging tasks aimed at ‘‘Scheduling wireless virtual networks functions,’’ IEEE Trans. Netw.
enabling 5G architecture to accommodate diverse use cases. Service Manag., vol. 13, no. 2, pp. 240–252, Jun. 2016.
In this paper, we introduce the metric of node importance [11] S. Herker, A. Khan, and X. An, ‘‘Survey on survivable virtual network
embedding problem and solutions,’’ in Proc. Int. Conf. Netw. Services,
based on complex network theory. Based on topological 2013, pp. 1–6.
information, we present the mathematical model of deploying [12] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and D.-U. Hwang, ‘‘Com-
end-to-end slices. And the network slice request implementa- plex networks: Structure and dynamics,’’ Phys. Rep., vol. 424, nos. 4–5,
pp. 175–308, 2006.
tion algorithm was proposed, which includes three different [13] B. Zhang, R. Liu, D. Massey, and L. Zhang, ‘‘Collecting the internet
deployment algorithms for eMBB slices, mMTC slices and AS-level topology,’’ ACM SIGCOMM Comput. Commun. Rev., vol. 35,
uRLLC slices respectively. Extensive simulations was con- no. 1, pp. 53–61, 2005.
[14] J. Wu, C. K. Tse, and F. C. M. Lau, ‘‘Optimizing performance of communi-
ducted to validate the performance of our proposed algorithm cation networks: An application of network science,’’ IEEE Trans. Circuits
and analyze the average execution time. Results analysis of Syst. II, Exp. Briefs, vol. 62, no. 1, pp. 95–99, Jan. 2015.

19700 VOLUME 6, 2018

W. Guan et al.: Service-Oriented Deployment Policy of E2E NS Based on Complex Network Theory

[15] Y. B. Kim, B. Hong, and W. Choi, ‘‘Scale-free wireless networks with [40] R. Mijumbi, J. Serrat, J.-L. Gorricho, N. Bouten, F. de Turck, and S. Davy,
limited degree information,’’ IEEE Wireless Commun. Lett., vol. 1, no. 5, ‘‘Design and evaluation of algorithms for mapping and scheduling of
pp. 428–431, Oct. 2012. virtual network functions,’’ in Proc. 1st IEEE Conf. Netw. Softwariza-
[16] N. Nikaein et al., ‘‘Network store: Exploring slicing in future 5G net- tion (NetSoft), Apr. 2015, pp. 1–9.
works,’’ in Proc. 10th Int. Workshop Mobility Evol. Internet Archit., 2015,
pp. 8–13.
[17] X. Costa-Pèrez, J. Swetina, T. Guo, R. Mahindra, and S. Rangarajan, WANQING GUAN received the B.S. degree in
‘‘Radio access network virtualization for future mobile carrier networks,’’
electronic information science and technology
IEEE Commun. Mag., vol. 51, no. 7, pp. 27–35, Jul. 2013.
from Beijing Forestry University, Beijing, China.
[18] P. Rost et al., ‘‘Mobile network architecture evolution toward 5G,’’ IEEE
Commun. Mag., vol. 54, no. 5, pp. 84–91, May 2016. She is currently pursuing the Ph.D. degree in com-
[19] K. Samdanis, X. C. Perez, and V. Sciancalepore, ‘‘From network sharing munications engineering at the Beijing Univer-
to multi-tenancy: The 5G network slice broker,’’ IEEE Commun. Mag., sity of Posts and Telecommunications, China. Her
vol. 54, no. 7, pp. 32–39, Jul. 2016. current research interests include network slicing,
[20] X. Zhou, R. Li, T. Chen, and H. Zhang, ‘‘Network slicing as a ser- network architecture, and complex network theory.
vice: Enabling enterprises’ own software-defined cellular networks,’’ IEEE
Commun. Mag., vol. 54, no. 7, pp. 146–153, Jul. 2016.
[21] T. Taleb, B. Mada, M. Corici, A. Nakao, and H. Flinck, ‘‘PERMIT:
Network slicing for personalized 5G mobile telecommunications,’’ IEEE XIANGMING WEN received the B.E., M.S., and
Commun. Mag., vol. 55, no. 5, pp. 88–93, May 2017. Ph.D. degrees from the Beijing University of Posts
[22] ‘‘Network Functions Virtualisation (NFV); architectural framework,’’
and Telecommunications (BUPT), Beijing, China,
ETSI and GSNFV, Tech. Rep., 2013. [Online]. Available:
all in electrical engineering. He is currently the
http://www.etsi.org/deliver/etsi_gs/NFV/001_099/002/01.01.01_60/gs_
NFV002v010101p.pdf
Vice President of BUPT, where he is also a Pro-
[23] R. Mijumbi et al., ‘‘Network function virtualization: State-of-the-art fessor with the Communication Network Center
and research challenges,’’ IEEE Commun. Surveys Tuts., vol. 18, no. 1, and the Director of the Beijing Key Laboratory of
pp. 236–262, 1st Quart., 2016. Network System Architecture and Convergence.
[24] LinuxFoundation. Opnfv. Accessed: Jun. 7, 2017. [Online]. Available: He is the Vice Director of the Organization Com-
https://www.opnfv.org/ mittee of the China Telecommunication Associ-
[25] Telefonica. Openmano. Accessed: Jun. 7, 2017. [Online]. Available: ation. In the last five years, he has been the author of over 100 papers
https://github.com/nfvlabs/ published. He is the Principle Investigator of over 18 projects, including
[26] FraunhoferFOKUS. Openbaton. Accessed: Jun. 7, 2017. [Online]. Avail- the National Key Project of Hi-Tech Research and Development Program of
able: http://openbaton.github.io/ China (863 Program) and the National Natural Science Foundation of China.
[27] J. Liu, Y. Li, Y. Zhang, L. Su, and D. Jin, ‘‘Improve service chaining His current research interests include broadband mobile communication
performance with optimized middlebox placement,’’ IEEE Trans. Services theory, multimedia communications, and information processing.
Comput., vol. 10, no. 4, pp. 560–573, Jul./Aug. 2017.
[28] L. Wang, Z. Lu, X. Wen, R. Knopp, and R. Gupta, ‘‘Joint opti-
mization of service function chaining and resource allocation in net-
LUHAN WANG received the B.S. degree in com-
work function virtualization,’’ IEEE Access, vol. 4, pp. 8084–8094,
2016. munication engineering from Shandong Univer-
[29] S. Clayman, E. Maini, A. Galis, A. Manzalini, and N. Mazzocca, ‘‘The sity, Jinan, China, and the Ph.D. degree from the
dynamic placement of virtual network functions,’’ in Proc. IEEE Netw. Beijing University of Posts and Telecommunica-
Oper. Manage. Symp. (NOMS), May 2014, pp. 1–9. tions in 2017. He joined the School of Information
[30] M. Ghaznavi, A. Khan, N. Shahriar, K. Alsubhi, R. Ahmed, and and Communication Engineering, Beijing Univer-
R. Boutaba, ‘‘Elastic virtual network function placement,’’ in Proc. IEEE sity of Posts and Telecommunications, in 2017.
4th Int. Conf. Cloud Netw. (CloudNet), Oct. 2015, pp. 255–260. His current research interests include network
[31] A. Fischer, J. F. Botero, M. T. Beck, H. de Meer, and X. Hesselbach, architecture, network function virtualization, and
‘‘Virtual network embedding: A survey,’’ IEEE Commun. Surveys Tuts., software-defined networks.
vol. 15, no. 4, pp. 1888–1906, 4th Quart., 2013.
[32] M. Yu, Y. Yi, J. Rexford, and M. Chiang, ‘‘Rethinking virtual net-
work embedding: Substrate support for path splitting and migration,’’
ZHAOMING LU received the Ph.D. degree from
ACM SIGCOMM Comput. Commun. Rev., vol. 38, no. 2, pp. 17–29,
Apr. 2008. the Beijing University of Posts and Telecom-
[33] N. M. M. K. Chowdhury, M. R. Rahman, and R. Boutaba, ‘‘Virtual network munications in 2012. He joined the School of
embedding with coordinated node and link mapping,’’ in Proc. IEEE Information and Communication Engineering,
INFOCOM, Apr. 2009, pp. 783–791. Beijing University of Posts and Telecommunica-
[34] M. Chowdhury, M. R. Rahman, and R. Boutaba, ‘‘ViNEYard: Virtual tions, in 2012. His research interests include open
network embedding algorithms with coordinated node and link mapping,’’ wireless networks, QoE management in wireless
IEEE/ACM Trans. Netw., vol. 20, no. 1, pp. 206–219, Feb. 2012. networks, software-defined wireless networks, and
[35] M. Chowdhury, F. Samuel, and R. Boutaba, ‘‘PolyViNE: Policy-based cross-layer design for mobile video applications.
virtual network embedding across multiple domains,’’ in Proc. 2nd
ACM SIGCOMM Workshop Virtualized Infrastruct. Syst. Archit., 2010,
pp. 49–56.
[36] W. Guan, X. Wen, L. Wang, and Z. Lu, ‘‘Network slicing management of YIDI SHEN received the B.S. degree in com-
5G ultra-dense networks based on complex network theory,’’ in Proc. IEEE
munication engineering from Jilin University,
Globecom Workshops (GC Wkshps), Dec. 2017, pp. 1–6.
Changchun, China. She is currently pursuing the
[37] A.-L. Barabási and R. Albert, ‘‘Emergence of scaling in random net-
works,’’ Science, vol. 286, no. 5439, pp. 509–512, 1999. M.S. degree in communications engineering at the
[38] E. Aarts and J. Korst, ‘‘Simulated annealing and boltz- Beijing University of Posts and Telecommunica-
mann machines,’’ Tech. Rep., 1988. [Online]. Available: tions, China. Her current research interests include
https://www.osti.gov/biblio/5311236 network slicing and complex network theory.
[39] C. Yu, W. Hou, Y. Guan, Y. Zong, and P. Guo, ‘‘Virtual 5G network
embedding in a heterogeneous and multi-domain network infrastructure,’’
China Commun., vol. 13, no. 10, pp. 29–43, Oct. 2016.

VOLUME 6, 2018 19701