Scheduling wireless virtual networks functions

240 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 13, NO. 2, JUNE 2016
Scheduling Wireless Virtual Networks Functions
Roberto Riggio, Abbas Bradai, Davit Harutyunyan, Tinku Rasheed, Member, IEEE, and Toufik Ahmed
Abstract—Network function virtualization (NFV) sits firmly on
the networking evolutionary path. By migrating network functions
from dedicated devices to general purpose computing platforms,
NFV can help reduce the cost to deploy and operate large IT
infrastructures. In particular, NFV is expected to play a pivotal
role in mobile networks where significant cost reductions can be
obtained by dynamically deploying and scaling virtual network
functions (VNFs) in the core network. However, in order to achieve
its full potential, NFV needs to extend its reach also to the radio
access segment. Here, mobile virtual network operators shall be
allowed to request radio access VNFs with custom resource allo-
cation solutions. Such a requirement raises several challenges in
terms of performance isolation and resource provisioning. In this
work, we formalize the wireless VNF placement problem in the
radio access network as an integer linear programming prob-
lem and we propose a VNF placement heuristic, named wireless
network embedding (WiNE), to solve the problem. Moreover, we
present a proof-of-concept implementation of an NFV manage-
ment and orchestration framework for enterprise WLANs. The
proposed architecture builds on a programmable network fabric
where pure forwarding nodes are mixed with radio and packet
processing capable nodes.
Index Terms—Network Management, Resource allocation,
Virtual Network Embedding, Virtual Network Functions
Placement, Network Function Virtualization, Wireless Networks.
I. INTRODUCTION
N ETWORK Function Virtualization (NFV) promises to
reduce the cost to deploy and operate large networks by
migrating network functions from dedicated hardware appli-
ances to software instances running on general purpose vir-
tualized networking and computing infrastructures. This, in
time, shall improve the flexibility and the scalability of mobile
networks in that the deployment of new applications and ser-
vices will be quicker (software Vs. hardware development
life–cycles) and different network functions can share the same
resources paving the way to further economies of scale. This
progressive process of network softwarization is set to play a
pivotal role in fifth generation mobile networks. In this context,
the Network–as–a–Service business model shall allow opera-
tors to tap into new revenue streams by further abstracting the
Manuscript received December 9, 2015; revised March 6, 2016; accepted
March 13, 2016. Date of publication April 1, 2016; date of current version
June 8, 2016. This work was supported by the European Union’s H2020
Research and Innovation Action under Grant H2020-ICT-644843 (VITAL) and
Grant H2020-ICT-671639 (COHERENT). The associate editor coordinating
the review of this paper and approving it for publication was A. Pras.
R. Riggio, D. Harutyunyan, and T. Rasheed are with the CREATE-NET,
Trento 38123, Italy (e-mail: rriggio@create-net.org; dharutyunyan@create-
net.org; trasheed@create-net.org).
A. Bradai is with the XLIM Institute, University of Poitiers, Poitiers 86130,
France (e-mail: abbas.bradai@univ-poitiers.fr).
T. Ahmed is with the CNRS-LaBRI, University of Bordeaux/Bordeaux-INP,
33400, France (e-mail: tad@labri.fr).
Digital Object Identifier 10.1109/TNSM.2016.2549563
physical network into service specific slices possibly operated
by different mobile virtual network operators (MVNOs). The
envisioned vertical applications range from high–definition
video delivery to machine–to–machine applications.
In order to cope with the diverse range of requirements that
sprout for such use cases, future wireless and mobile networks
will further rely on virtualized resources and on dynamic ser-
vice orchestration. Although a rich body of literature exists
on VNF placement [1], virtual network embedding [2], and
component placement [3], most of these works focus on the
problem of mapping an input virtual network request (often in
the form of a VNF Forwarding Graph) onto a physical virtual-
ized network substrate (often offering computational as well as
networking resources). However, these works implicitly assume
that once a VNF is mapped on a node, the virtualization layer
(i.e. the hypervisor) will take care of scheduling the various
VNFs ensuring both logical isolation and an efficient use of the
substrate resources [4]. Such an assumption does not hold any-
more if radio nodes are added to the set of virtualized resources
available in the substrate network (alongside computational
and networking resources). In this case, in fact, the amount of
resources available at each substrate radio node is a stochastic
quantity depending on both channel fluctuations and end–users
distribution.
In this work, we investigate the VNF placement and schedul-
ing problems in the Radio Access Network (RAN) domain. In
this scenario we expect MVNOs to specify their requests in
terms of a VNF Forwarding Graph. Such VNFs can include
functions such as load–balancing and firewall, as well as virtual
radio nodes. Moreover, in order to satisfy the diverse require-
ments imposed by future applications and services, MVNOs
shall be allowed to deploy custom resource allocation schemes
within their network slice. At the same time, the underlying
system shall both enforce strict performance isolation between
MVNOs and ensure efficient resource utilization across the net-
work in spite of the non–deterministic nature of the wireless
medium.
The contribution of this paper is twofold: (i) we formal-
ize the VNF placement problem for radio access networks,
and (ii) we propose a slice scheduling mechanism that ensures
resource and performance isolation between different slices.
The proposed solutions work jointly, i.e. performance isolation
is ensured if slices are accepted under the constraints imposed
by our VNF placement problem formulation. This paper builds
upon our previous work [5] by refining the VNF placement
heuristic WiNE (Wireless Network Embedding), extending the
simulation study to additional types of VNF requests, and ana-
lyzing in deeper detail how the type of VNFs impacts on
the substrate network utilization. Numerical simulations show
that the proposed heuristic can approximate the performance
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/
www.redpel.com +917620593389

RIGGIO et al.: SCHEDULING WIRELESS VIRTUAL NETWORKS FUNCTIONS 241
of the optimal ILP–based placement algorithm. Moreover, the
efficiency of the heuristic increases with the size of the substrate
network hinting at the fact that WiNE may be capable of closing
the gap with the ILP–based placement algorithm for realistic
substrate networks. Finally, we also report on a updated proof–
of–concept implementation of the proposed solution and on its
field evaluation. The programmable data–path, the controller
and the SDK have been released under a permissive license for
academic use1.
The rest of this paper is structured as follows. In Sec. II we
discuss the related work. The physical network model, the VNF
request model, and ILP problem formulation are presented in
Sec. III. The VNF placement heuristics and its evaluation are
presented in Sec IV and in Sec. V. The proof–of–concept is
presented in Sec VI while some illustrative VNFs and their
evaluation are presented in Sec. VII. Finally, Sec. VIII draws
the conclusions pointing out the future work.
II. RELATED WORK
The recent advances in general purpose computing plat-
forms paved the way to a new generation of software routers.
However, many of these solutions focus on improving the pure
raw packet processing speed [6]–[8] but do not tackle the prob-
lem of deploying and orchestrating VNFs. In parallel, there
are significant efforts toward VNF management and orchestra-
tion. In particular the European Telecommunications Standards
Institute (ETSI) has recently tackled the NFV concept [9] while
OPNFV [10], OpenMANO [11], and OpenBATON [12] are
working toward an open–source carrier grade platform for NFV.
A. Virtual Network Embedding
The amount of literature on virtual network embedding
(VNE) topic is considerable. Seminal works in this domain
include VINEYard [13] for single domain VNE and PolyVINE
[14] for multi–domain VNE. For a comprehensive survey
on VNE algorithms we point the reader to [2]. In [15] the
authors put forward a novel model that reflects the time–varying
resource requirements of a virtual network request. The authors
also consider VNE in the context of opportunistic resource shar-
ing at the level of the entire substrate network. In [16], the
authors present the SiMPLE VNE algorithm. SiMPLE exploits
path diversity in order to protect virtual networks from single
link failures. However, to the best of the authors’ knowledge,
none of these works formulate the VNE problem for hybrid
wired/wireless networks with the goal of ensuring performance
isolation between tenants.
B. VNF Placement
The VNF placement problem is conceptually similar to the
component placement problem in data–centers and cloud com-
puting environments. The amount of literature in this domain
is thus humbling [4], [17]–[19]. A survey on resource manage-
ment in cloud computing environments can be found in [3]. In
[17], the authors study the problem of placing virtual machine
1On–line resources available at: http://empower.create-net.org/
instances on physical containers in such a way to reduce
communication overhead and latency. In [18], the authors pro-
pose a novel design for a scalable hierarchical application
components placement in cloud computing environments. The
proposed solution operates in a distributed fashion ensuring
scalability while at the same time providing performances very
close to that of a centralized algorithm. This work is extended
in [19] where several algorithms for efficient management of
component–based applications in cloud environments are pro-
posed. In [4] the elasticity overhead and the trade–off between
bandwidth and host resource consumption are jointly consid-
ered by the authors when formulating the VNF placement
problem. In [20], a joint node and link mapping algorithm
is proposed. Finally, the authors of [1], [21], [22] tackle the
problem of dynamic VNF placement. Targeting resource allo-
cation in data–centers, these works do not tackle the problem of
virtualized radio function placement.
C. Wireless & Mobile Networks
The topic of radio resources virtualization has received
significant attention in the literature. In [23], a WLAN virtu-
alization approach named Virtual WiFi is proposed extending
the VNE problem from the wired to the wireless domain.
Kernel–based virtual machines are used as a virtual wireless
LAN devices while time domain multiplexing is used in order
to provide isolation between the virtual wireless devices. In
[24], [25], wireless network virtualization is applied to wireless
mesh networks. A virtual network traffic shaper is introduced
in [26], [27] for air time fairness in 802.16e networks. In [28],
[29] the problem of virtualizing OFDMA–based wireless net-
works (i.e. WiMAX and LTE) is studied. The authors tackle the
problem both at the radio and the core network level opening
the way to interesting infrastructure sharing scenarios. Similar
consideration can be also made for [30] where a framework
for sharing a single WiMAX base station is proposed. Wireless
Virtualization of 802.11 devices is the focus of [31]. In all the
cases above, however, the channel–aware placement of VNFs
over radio and wired resources is not formulated nor is the
performance isolation challenge between multiple MVNOs
tackled.
D. Middlebox Management
Systems like OpenNF [32] and its derivatives [33], [34]
focus on providing a platform for consistent VNF management.
However their focus is on maintaining backward compatibility
with currently available VNFs such as Bro [35] for IDS and
Squid [36] for caching web proxies. Conversely, in this work we
set to explore the possibilities opened by a fully programmable
networking substrate where also radio access is treated as a
standard VNF. Similar considerations can be made also for
Split/Merge [37], CoMB [38], and XoMB [39].
III. NETWORK MODEL
In the VNF placement problem, the input consists of Service
Function Chains (SFC) composed by a variable number of

TABLE I
SUBSTRATE NETWORK PARAMETERS
VNFs, whereas the substrate network, called Network Function
Virtualization Infrastructure (NFVI), provides the physical con-
straints in terms of bandwidth and capacity [9]. In this context,
the term capacity is not related only to pure computational
resources, such as number of CPU cores, memory, and/or
storage. Instead it refers also to packet forwarding and radio
processing capabilities. Before introducing the proposed, solu-
tion we need to detail specific notations for the NFVI and the
SFC requests.
A. Network Function Virtualization Infrastructure Model
Let Gnf vi = (Nnf vi , Enf vi ) be an undirected graph model-
ing the physical network, where Nnf vi is the set of n = |Nnf vi |
physical nodes that compose the substrate network and Enf vi
is the set of edges or links. An edge enm ∈ Enf vi if and only
if a point–to–point connection exists between n ∈ Nnf vi and
m ∈ Nnf vi . With respect to the physical network, links are
actual wiring media, e.g., an Ethernet cable interconnecting
the two nodes. Four weights, ωs
c(n), ωs
m(n), ωs
s (n), ωs
r (n), are
assigned to each node n ∈ Nnf vi : ωs
c,m,s(n) ∈ N+ and ωs
r (n) ∈
R+, 0 ≤ ωs
r (n) ≤ 1 representing the packet and radio process-
ing resources available on that node. Nodes with all weights
equal to 0 (zero) are assumed to be pure packet forwarding
nodes. Nodes with ωs
c > 0, ωs
m > 0, ωs
s > 0, and ωs
r = 0 are
assumed to be pure packet processing nodes. Finally, nodes
with ωs
c = ωs
m = ωs
s = 0 and ωs
r > 0 are assumed to be pure
radio access nodes.
Another weight ωs
e(enm) is assigned to each link enm ∈
Enf vi : ωs
e(enm) ∈ N+ representing the capacity of the link con-
necting two nodes. In order to avoid exceeding the nominal
capacity of the substrate links, traffic shaping is implemented
at the nodes with packet and/or radio processing capabili-
ties. Finally, let Pnf vi be the set of all substrate paths and
Pnf vi (s, t) the shortest path between nodes s, t ∈ Nnf vi . Table I
summarizes the NFVI parameters.
The weights ωs
c,m,s(n) associated with the packet process-
ing nodes represent, respectively, the amount of CPU, memory,
and storage resources available on that node, while the weights
ωs
r (n) are specific to the radio access nodes and represent the
normalized amount of wireless resources available at that node.
Notice how with the term radio access nodes we refer to the
generic nodes providing end–users terminals with wireless con-
nectivity and how this model makes no assumptions on the type
of resources that can be available at radio nodes. For example,
Fig. 1. NFVI network model. The figure shows the three basic virtual
resources: forwarding, packet processing, and radio access.
in a 802.11–based network ωs
r (n) could model the amount of
airtime available at a certain Access Point (AP). Similarly, in
an OFDMA–based network (e.g., LTE), ωs
r (n) could be used to
model the available radio resources in time and frequency at a
certain eNodeB (eNB). For the sake of simplicity and without
any loss of generality in this work we assume that all radio–
enabled nodes initially have the same amount of resources
ωs
r (n) = 1, ∀n ∈ Nnf vi .
A sample substrate network is sketched in Fig. 1. The net-
work is composed by 8 nodes interconnected together. In order
to improve readability link weights have been omitted. The
substrate network in this example consists of 4 radio access
nodes (at the bottom of the picture), and 4 switches, 2 of which
supporting just basic forwarding capabilities.
B. Service Function Chain Requests
Users are allowed to request SFCs as a directed graphs
Gs f c = (Ns f c, Es f c). Where Ns f c denotes the set of nodes
(i.e. VNFs) and Es f c ⊆ Ns f c × Ns f c denotes the set of virtual
links. An edge enm ∈ Es f c if and only if the packets from VNF
n ∈ Ns f c must be forwarded to the VNF m ∈ Ns f v. Notice that
as opposed to the previous NFVI model, nodes in SFC requests
represent virtual network functions through which packets must
undergo before leaving the network. Packet processing nodes
and links in the SFC request shares the same weights as for
the NFVI substrate network (ωv
c, ωv
m, ωv
s, ωv
e). On the other
hand, provisioning of radio resources can be made either in
terms of fraction of available radio resources (ωv
r ) or in terms
of bandwidth (ωv
b). In the former case, the users will specify
the percentage of radio resources they want to be assigned to
a certain node, while in the latter case the users will spec-
ify the amount of bandwidth assigned to the node. A single
SFC request can mix bandwidth–based and resource–based
provisioning models.
Due to their stochastic nature, available bandwidth is a time–
varying quantity in wireless networks. Channel fading, but also
to the distribution of end–users, can greatly influence the net-
work performance. For example, users at the center of the cell
can, in general, use more efficient modulations and coding
schemes thus achieving higher throughput for a fixed amount of
radio resources than users at the edges of the cell. As a result,
when the bandwidth–based provisioning model is employed,

TABLE II
SERVICE FUNCTION CHAIN REQUEST PARAMETERS
also the actual channel conditions experienced by the end–users
must be taken into account.
Let us call b(n) the actual aggregated throughput of the vir-
tual radio node n ∈ Ns f c in the fraction of resources currently
assigned the node. We can then introduce an additional parame-
ter named reference throughput v
b(n) ≥ ωv
b(n) upon which the
bandwidth reservation ωv
b(n) is enforced. We can then define
the effective target bandwidth ˜ωv
b(n) for the virtual radio node
n as follows:
˜ωv
b(n) =
ωv
b(n) if b(n) ≥ v
b(n)
ωv
b(n) b(n)
v
b(n) if b(n) < v
b(n)
(1)
The parameter v
b(n) represents a threshold. When the actual
throughput b(n) of the virtual radio node n is above the thresh-
old, then the bandwidth reservation is respected. Conversely,
when b(n) < v
b(n) the requested bandwidth ωv
b(n) is linearly
scaled down. Choosing a small value for v
b(n) means that
the network can utilize more resources in order to satisfy the
bandwidth requirements, which in time could result in a higher
pricing. Conversely, a high value for v
b(n) means that the net-
work will try to satisfy the bandwidth requirements only for
the tenants that are making a better use of the wireless spec-
trum. Notice how, as it will be more clear in Sec. VI, tenants
can make a poor use of the wireless spectrum also due to cus-
tom scheduling disciplines that favor fairness at the expense of
the aggregated throughput. This could be a reasonable trade–
off in network slices aimed at emergency response applications
where it is critical that all mobile clients get the same share
of resources even if this could lead to a sub–optimal spectrum
utilization.
If we define ωv
r (n) =
˜ωv
b(n)
b(n) , then, in order for the SFC request
to be feasible, it must hold:
n∈Nb
s f c
˜ωv
b(n)
b(n)
+
n∈Nr
s f c
ωv
r (n) ≤ ωv
r (m) = 1 (2)
Which means that the sum of the fractions of radio resources
allocated to virtual nodes must be less than or equal to the
resources available at the substrate node m. Table II summarizes
the SFC request parameters.
For example, consider the case depicted in Fig. 2a. Here two
tenants (A and B) requested two slices with aggregated capacity
of, respectively, 6 Mb/s and 4 Mb/s, while the reference band-
width ( v
b) for both tenants has been set to 10 Mb/s. Moreover,
for simplicity and without losing generality, let us assume that
only one client is active in either slice. If the cell capacity b(n)
experienced by both clients is equal to, or higher than, 10 Mb/s
it is easy to see how both slices can be accommodated by the
system, thus in this case b(n) ≥ v
b(n) and ˜ωv
b(n) = ωv
b(n) for
both slices.
However, if the cell capacity experienced by the wireless
client in the tenant B’s slice is reduced to 5 Mb/s, then the
same request can not be accommodated anymore. This is due
to the fact that serving the same amount of data to the wire-
less client in the tenant B’s slice now requires twice the amount
of radio resources it did before (see Fig 2b). In this case since
b(B) < v
b(B), thus:
˜ωv
b(B) = ωv
b(B)
b(B)
v
b(B)
= 4
5
10
= 2
This results in the resource allocation sketched in Fig 2c which
penalizes only the performance of the tenant B without affect-
ing tenant A. It is worth noticing that this example can be easily
generalized to an OFDMA system were resources are assigned
across a time/frequency matrix.
A few sample SFC requests are sketched in Fig. 3. Notice that
wireless terminals are not represented since they are outside the
control of the orchestration framework. The linear SFC request
in Fig. 3a consists of three VNFs including a WiFi hotspot, a
firewall, and a load balancer. The WiFi hotspot request will
also include parameters such as the name of the network and
the authentication parameters (e.g. type of encryption, RADIUS
server to be used, etc.) however these kind of information are
purely functional and are thus omitted in this section. The
SFC request in Fig. 3b implements a performance enhancing
network service, namely duplicates filtering in dense urban sce-
narios. Finally, the SFC request in Fig. 3c accounts for an access
enforcement scenario whereby multiple security related VNFs
are deployed in parallel. Notice how all radio access VNF use
the bandwidth–based resource provisioning model.
C. Virtual Network Function Placement
In this section we shall provide the optimal ILP formulation
for the SFC embedding problem, while in the next section we
will present a scalable heuristic. The overall objective is to com-
pute the optimal VNF placement based on the available radio
resources. The chosen objective function is:
min
⎛
⎝
n∈Nnf vi n ∈Ns f c
wv
c(n ) c(n) + wv
m(n ) m(n)
+wv
s (n ) s(n) + wv
r (n ) r (n) n
n
+
e∈Enf vi e ∈Es f c
ωv
e(e ) e(e) e
e
⎞
⎠
where, n
n , e
e ∈ 0, 1 are two binary variables indicating
respectively if the VNF n ∈ Ns f v has been mapped to node
n ∈ Nnf vi and if the virtual link e ∈ Es f c has been mapped to
the substrate link e ∈ Enf vi .

Fig. 2. Impact of channel conditions on slice isolation. Bandwidth demands can be met when channel conditions are good for both slices (a). Conversely,
bandwidth targets can not be achieved when one of the tenants is experiencing poor channel conditions (b). Isolation can be preserved by linearly scaling down
the bandwidth target for the tenant with adverse channel conditions (c).
Fig. 3. Sample SFC Requests. Notice how resource requests for all radio VNFs
are bandwidth–based.
A valid solution is the one where the resources utilized by
the SFC request are at most equal to the available resources on
the substrate network nodes:
n ∈Ns f c
ωv
c(n ) n
n ≤ ωs
c(n) ∀n ∈ Nnf vi (3)
n ∈Ns f c
ωv
m(n ) n
n ≤ ωs
m(n) ∀n ∈ Nnf vi (4)
n ∈Ns f c
ωv
s(n ) n
n ≤ ωs
s (n) ∀n ∈ Nnf vi (5)
n ∈Ns f c
ωv
r (n ) n
n ≤ ωs
r (n) ∀n ∈ Nnf vi (6)
and links:
e ∈Es f c
ωv
e(e ) e
e ≤ ωs
e(e) ∀e ∈ Enf vi (7)
In this work we also assume that every VNF in the SFC request
shall be mapped to a different substrate node:
n ∈Ns f c
n
n ≤ 1 ∀n ∈ Nnf vi (8)
Every VNF in the SFC request shall be mapped only once:
n∈Nnf vi
n
n = 1 ∀n ∈ Ns f c (9)
In terms of radio resources requirements, the following con-
strain, deriving from (1) and (2), enforces that for every radio
processing node n ∈ Nnf vi a feasible request has been made:
n ∈Nb
s f c
ωv
b(n )
v
b(n )
n
n +
n ∈Nr
s f c
ωv
r (n ) n
n ≤ 1 ∀n ∈ Nnf vi
(10)
Finally, the following constraint enforces that for each link
enm ∈ Es f c there must be a continuous path allocated between
the pair of physical nodes on top of which the VNFs n, m ∈
Ns f c have been mapped.
j>i
j∈Nnf vi
enm
ei j −
j<i
j∈Nnf vi
enm
e ji = n
i − m
i
∀i ∈ Nnf vi ∀enm
∈ Es f c (11)
IV. HEURISTIC
The ILP formulation described in the previous sections can
not be applied to realistic scenarios due to its limited scalability.
For example, embedding a 6–nodes linear SFC request over a
k = 8 fat–tree substrate topology can take several hours on Intel
Core i7 laptop (3.0 GHz CPU, 16 Gb RAM) using the Matlab®
ILP solver (intlinprog). Notice also that even a k = 8 fat–tree
substrate topology is rather small if compared with realistic

Algorithm 1. Compute list of candidate substrate nodes
1: procedure FindCandidates(Nnf vi , Ns f c)
2: for n ∈ Ns f c do
3: for p ∈ Nnf vi do
4: If ωs
c,m,s,r (p) >= ωv
c,m,s,r (n) then
5: n.candidates.add(p)
6: end if
7: end for
8: end for
9: end procedure
Algorithm 2. Sort list of candidate substrate nodes
1: procedure SortCandidates (Ns f c)
2: sort(Ns f c)
3: end procedure
deployments where it is common to find k = 24 fat–tree net-
works, making the ILP problem formulation intractable. In this
section we present a heuristic, named WiNE, that can handle
similar (and more complex) requests in tens of minutes.
The proposed heuristic is composed of three steps. In the
first step, for each virtual node n ∈ Ns f c, the heuristic com-
putes the list of candidate substrate nodes (see Alg. 1). These
are the substrate nodes that can support the virtual nodes in
the SFC request given the input capacity constraints. In the
second step, the virtual nodes n ∈ Ns f c are sorted in decreas-
ing order according to the number of candidate substrate nodes
(see Alg. 2). In the third step, the sorted list of virtual nodes is
traversed starting with the virtual node with more embedding
opportunities. For each candidate substrate node p ∈ Nnf vi
the heuristic considers all the neighboring nodes m ∈ Ns f c of
the virtual node n. The heuristic, then, assigns the node n to the
substrate node m ∈ Nnf vi with the lowest virtual edge mapping
cost (see Alg. 3).
We define the virtual edge mapping cost W : Es f c × Nnf vi ×
Nnf vi → R between a virtual edge enm ∈ Es f c and a pair of
substrate nodes p, q ∈ Nnf vi as follows:
W(enm
, p, q) = c,m,s,r (p)ωv
c,m,s,r (n)
+ c,m,s,r (q)ωv
c,m,s,r (m)
+
e∈Pnf vi (p,q)
e(e)ωv
e(enm
)
This represents the cost of embedding the virtual edge enm ∈
Es f c over the path Pnf vi (p, q) between the substrate nodes
p, q ∈ Nnf vi given that virtual nodes n, m are mapped on,
respectively, the substrate nodes p, and q. Minimizing the
virtual edge mapping cost for node m essentially means that
substrate nodes that are far away from node’s m embedding
opportunities are penalized. This results in virtual nodes in an
SFC request to be placed close to each other over the substrate
network, which in time means that less substrate resources are
needed to support a given number of requests.
Algorithm 3. Nodes and links assignment
1: procedure NodeAndLinkAssignment (Gnf vi , Gs f c)
2: for n ∈ Ns f c do
3: for p ∈ n.candidates do
4: ff p.used then
5: continue
6: end if
7: for m ∈ n.neighbors do
8: if m.mapped then
9: cost = W(enm, p, m.mapped)
10: else
11: cost = +∞
12: for q ∈ m.candidates do
13: cost = min(cost, W(enm, p, q))
14: end for
15: end if
16: mapping_cost(p)+ = cost
17: end for
18: end for
19: p ← argmin(mapping_cost(p))
20: n.mapped ← p
21: p.used ← True
22: for m ∈ n.neighbors do
23: if m.mapped then
24: Allocate path Pnf vi (n.mapped,
m.mapped)
25: end if
26: end for
27: end for
28: end procedure
V. EVALUATION
The goal of this section is to compare the relative perfor-
mance of the ILP–based placement algorithm with the per-
formance of our placement heuristic using different synthetic
substrate networks and different SFC requests. In this section,
we shall first describe the simulation environment and then the
performance metrics. Simulations are carried out in a discrete
event simulator implemented in Matlab®.
A. Simulation Environment
The ILP–based placement algorithm and the proposed place-
ment heuristic are evaluated in three different scenarios. In the
first scenario, linear VNF requests, similar to the one depicted
in Fig. 3a, are considered. In the second scenario, branched
VNF requests, similar to the one depicted in Fig. 3b, are con-
sidered. Finally, in the third scenario, VNF requests with loops,
similar to the one depicted in Fig. 3c, are used. The number
of VNFs in each SFC request as well as the actual amount of
radio, computational, memory, storage, and link resources are
randomly generated for each request.
The reference substrate network is a k–ary fat–tree with
k = 4, 6, 8, where leaf nodes are WiFi APs rather than servers.
This results in a total of, respectively, 16, 54, and 128 WiFi APs.
The computational, memory, storage, radio, and link resources

Fig. 4. Acceptance ratio using the ILP–based algorithm and the heuristics with different virtual and substrate topologies.
Fig. 5. Average embedding cost using the ILP–based algorithm and the heuristics with different virtual and substrate topologies.
Fig. 6. Average computational nodes utilization using the ILP–based algorithm and the heuristics with different virtual and substrate topologies.
for the substrate network are initially all set to 100. The cost of
using each unit of node c,m,s,r (n) and link e(e) resources is
set to 1. The number of VNFs in each SFC request depends on
the SFC type. In the case of linear and branched SFC requests
the number is randomly picked in the set {3, 6}, while in the
case of cyclic SFC requests the number is randomly picked in
the set {4, 6}. The computational, memory, and storage require-
ments for each SFC requests are uniformly distributed between
[5, 30], while the radio and link requirements are uniformly
distributed between [5, 60].
The metrics used in this study are the standard ones adopted
in several other related works [13], [40], [41]. For each scenario
the number of accepted requests, the average embedding cost,
the average node and link utilization, and the execution time
using either the ILP–based placement or the proposed heuristic
are considered. In this study we assume that a fixed number
of SFC requests are embedded sequentially onto the substrate
network. In particular in each run, the simulator tries to embed
30 randomly generated SFC requests. Reported results are the
average of 10 simulations.
B. Simulation Results
Figures 4 and 5 shows the percentage of accepted SFC
requests for different substrate networks and the average
embedding cost. As expected the ILP–based placement algo-
rithm is more efficient than WiNE in mapping the incoming
requests. This can be seen in terms of both a higher number of
accepted requests as well as a lower average embedding cost.
Notice however that, at least for linear SFCs, WiNE actually
has a lower embedding cost. This does not mean that WiNE is
more efficient than the ILP–based algorithm but rather that, by
accepting a lower number of SFC requests, WiNE also utilizes,
on average, less substrate resources. It is also worth noticing
that the efficiency of the proposed heuristic increases with the
size of the substrate network. This hints toward the fact that
WiNE may be capable of closing the gap with the ILP–based
placement algorithm for realistic substrate networks.
Figure 6, 7 and 8 summarize the substrate resource utiliza-
tion. As it can be noticed the ILP–based placement algorithm
is characterized by a higher utilization ratio for both radio
and computational nodes. This results in fewer substrate nodes

Fig. 7. Average radio nodes utilization using the ILP–based algorithm and the heuristics with different virtual and substrate topologies.
Fig. 8. Average link utilization using the ILP–based algorithm and the heuristics with different virtual and substrate topologies.
Fig. 9. Average execution time using the ILP–based algorithm and the heuristic with different virtual and substrate topologies.
being used to support the same number of SFC requests which
in time could result in a more energy efficient operation if
unused nodes can be turned–off. Notice that this is further sup-
ported by Fig. 8 where the substrate links utilization is reported.
As it can be seen, WiNE is characterized by a higher link utiliza-
tion, which means that the proposed heuristic is less efficient in
finding shorter paths between VNFs. However, it is also worth
noticing that the gap between the ILP–based placement algo-
rithm and WiNE gets smaller as the size of the substrate network
increase.
Figure 9 shows that the average amount of time required
to embed a single SFC request using the ILP–based place-
ment algorithm is significantly higher than the time required
to embed the same request using WiNE. The ILP problem
becomes essentially intractable for substrate networks with
more than a few tens of nodes (irrespective of the number
of VNFs in the request), while WiNE can effectively embed
complex SFC requests on substrate networks with hundreds of
nodes in a limited amount of time.
VI. IMPLEMENTATION
A. Overview
We implemented the VNF placement and scheduling solution
presented in this work in a proof–of–concept NFV management
and orchestration framework, named EmPOWER. Notice that,
the prototype currently supports only wireless access networks
based on the 802.11 family of standards and, as a consequence,
the applications described in the next section target Enterprise
WLANs and Campus networks scenarios. Nevertheless, as seen
in the previous sections, the provisioning model does not make
any assumption about the particular link–layer technology and
can be as well applied to any kind of radio access network
including OFDMA networks such as LTE and LTE–Advanced.
Our proof–of–concept is loosely modeled after the ETSI ref-
erence NFV Architecture [9]. As it can be seen in Fig. 10, the
architecture is conceptually divided into three layers. The bot-
tom layer consists of the physical as well as the virtualized
resources composing the NFVI. In the second layer we have

Fig. 10. Reference network function virtualization architecture [9].
the actual VNFs which are the software implementation of a
particular network function capable of being executed over the
NFVI. We remind the reader that in this work also radio access
is treated as a VNF. Finally, in the third layer we have the
Operational Support System (OSS) and the Business Support
System (BSS) used by the network administrators to oper-
ate and manage their virtual networks. The Management and
Orchestration plane covers the orchestration and the manage-
ment of physical and/or virtual resources that support the NFVI
as well as the life–cycle management of the VNFs, i.e. creation,
configuration, monitoring, and destruction.
B. Network Function Virtualization Infrastructure
Our architecture currently accounts for three kinds of NFVI
resources, namely: basic forwarding nodes (i.e. OpenFlow
switches), packet processing nodes, and radio access nodes.
The latter, in addition to the features supported by the packet
processing node, also embed specialized hardware in the form
of one or more 802.11 Wireless NICs. Figure 11 sketches the
system architecture.
We name Wireless Termination Points (WTPs) the phys-
ical points of attachment in the RAN (e.g. WiFi APs or
LTE eNBs) supporting virtualized radio processing capabili-
ties. Conversely, the Click Packet Processors (CPPs) are the
forwarding nodes with computational capacity. These nodes
are essentially programmable switches running an embedded
version of Linux and capable of performing arbitrary opera-
tions on the traffic, e.g. load–balancing, firewalling, deep packet
inspection, etc.
As the name suggests the actual packet processing is per-
formed by multiple instances of the Click Modular Router [42].
Click allows to build complex VNFs using simple and reusable
components, called elements. Click includes over 300 elements
supporting functions such as packet classification, access con-
trol, deep packet inspection. Elements can be composed in order
to realize complex VNFs. Finally, Click can be easily extended
with custom processing elements making it possible to support
features that are not provided by the standard elements.
Each WTP/CPP includes an OpenVSwitch instance, one or
more VNFs, and one Agent. The latter is in charge of monitor-
ing the status of each VNF as well as handling requests coming
from the controller. In the current implementation the monitor-
ing features include: number of packets/bytes transmitted and
received as well as the amount of resources (cpu time, memory,
storage) utilized by each VNF.
In the current prototype CPPs are built upon the Soekris
6501-70 platform consisting in single 1.6 GHz Intel Atom CPU,
2 Gbyte if SDRAM, and 12 Gigabit Ethernet Ports. CPPs
run Ubuntu 15.04 Server as operating system. WTPs exploit
the PCEngines ALIX (x86) embedded platform and run the
OpenWRT operating system.
C. Virtual Infrastructure Managers
As Virtual Infrastructure Managers (VIMs) we use a combi-
nation of frameworks. Ryu [43] is used to configure the switch-
ing fabric while for the NFVI we extended the EmPOWER
controller presented by the authors in [44] in order to support
also the generalized packet processing nodes. The controller
supports multiple slices on top of the same physical infrastruc-
ture. A slice is a virtual network with its own set of WTPs/CPPs.
The controller is responsible for the deploying the VNFs on the
network devices.
D. Orchestrator
From an architectural standpoint, the VNF placement algo-
rithm resides in the Orchestrator which is in charge of deciding
whether a particular request can be accepted or if it must be
refused. If a request is accepted, then the Orchestrator is in
charge of mapping the request onto the substrate network, i.e.,
network resources must be allocated and configured on both the
substrate nodes and the substrate links and the VNFs must be
instantiated on the selected nodes (see Fig. 11). Notice how, our
proof–of–concept does not impose a specific Orchestrator leav-
ing network administrators free to use other currently available
solutions, such as OpenBaton [12].
VII. APPLICATIONS
In this section we shall describe two SFCs implemented
and tested over a small scale testbed deployed at CREATE–
NET premises. The testbed consists of 2 OpenFlow–switches,
2 CPPs, and 20 WTPs.
A. Radio Access Network Slicing
In this use case we aim at demonstrating the performance
isolation features enabled by our joint VNF placement and
scheduling solution. Such use case is supported by introducing
a MAC Hypervisor implementing the slice scheduling tech-
niques described in the previous sections. Figure 12 sketches

Fig. 11. The EmPOWER network architecture.
Fig. 12. Radio Access Network Slicing Application.
the internal architecture of a radio access node. Notice how, not
the entire WiFi MAC is virtualized. More precisely, WiFi con-
trol frames generation (e.g. ACK and RTS/CTS) as well as the
actual physical layer are not virtualized due to their tight tim-
ing/computational constraints. On the other hand each tenant
can have its own Radio Resource Management (RRM) instance
as well as its own management place. In this work with RRM
we refer to wireless clients scheduling and to rate adaptation,
which can be defined on a per–tenant basis2. For example, one
tenant can use a wireless station scheduler aimed at ensuring
fairness while another tenant could opt for a scheduler aimed at
improving the aggregated slice bandwidth.
2Notice how, the hypervisor implementation details are out of the scope of
this work and are thus omitted.
The hypervisor monitors the channel utilization as well as the
amount of traffic generated by each Tenant. Such information is
used to compute the effective aggregated goodput of each slice
(b(n)). It is worth stressing that both channel activity and good-
put are two quantities that can be monitored by the hypervisor
with low overhead and without affecting the operation of the
RRM policy implemented by Tenants. For example a Tenant
with wireless clients experiencing poor channel conditions will
see its aggregated goodput reduced due to re–transmissions.
Similarly a Tenant implementing wireless clients scheduling
policies aimed at maximizing fairness will also see a reduction
in goodput if some of its wireless clients require less efficient
modulation and coding schemes.
The network setup used for this set of measurements con-
sists of two tenants A and B. The bandwidth requests coming
from Tenant A and B are, respectively, 4 and 2 Mb/s, while the
reference bandwidth ( v
b) is 6 Mb/s for both tenants. A total
of three wireless clients is active in the networks. Clients 1
and 2 belongs to Tenant A while Client 3 belongs to Tenant
B. Traffic is generated from a server sharing the same back-
haul with the APs and consists of three UDP streams (one
for each client). Each stream has constant inter–departure time
and packet size resulting in a transmission rate of 10 Mb/s
for each stream. In order to simulate a hotspot with limited
capacity the rate control algorithm used by the AP has been
modified in order to always use the lowest transmission rate (6
Mb/s) . Measurements have been carried in two different sce-
narios differentiated by the channel conditions experienced by
client number 1 which is positioned in such a way to experience
channel conditions raging from Good to Poor.
As it can be seen from Fig. 13a, when Client 1 is experi-
encing good channel conditions the proposed VNF scheduling

Fig. 13. Isolation across clients in different networks. Client 1 and 2 are in
Tenant 1’s network, while Client 3 is in Tenant 2’s network.
method can satisfy the bandwidth request for both tenants
(notice that Tenant A request of 4 Mb/s has been equal-
ity partitioned among the two clients). On the other hand,
when Client 1 starts experiencing poor channel conditions (see
Fig. 13b) the legacy resource provisioning mechanism allocates
the same bandwidth to all the clients. This behavior, known as
IEEE 802.11 performance anomaly [45], allows a node which
experiences poor channel conditions to monopolize the wire-
less medium lowering the performance of the whole system.
Conversely, the proposed resource provisioning mechanism can
meet the bandwidth reservations made by Tenant A by linearly
scaling down the amount of resources allocated to Tenant B that
is experiencing poor channel conditions.
B. Data/Management Plane Offloading
Wireless, and in particular mobile networks, have been so
far designed around the requirements of the downlink, i.e. cell
or AP selection is performed using downlink signal strength
measurements. In the recent years, however, we have wit-
nessed a mushrooming of new uplink–centric applications such
as Machine Type Communications (MTC), Internet of Things
(IoT), and Vehicle to Infrastructure (V2I) as well as of symmet-
ric mobile applications. This calls for a paradigm shift where
the traffic originated from a wireless client is received by one
node while the traffic destined to the same client is transmit-
ted by another node. This kind of network setup is usually
referred to as uplink/downlink decoupling and, in its most gen-
eral form, can consists of two possibly non overlapping sets of
transmitting and receiving nodes.
The EmPOWER platform [44] allows for network configu-
rations where a wireless client can be attached to one AP in
the downlink direction and to one or more APs in the uplink
direction. This feature allows to exploit the broadcast nature of
the wireless medium and to opportunistically receive the same
transmission at multiple in–range APs. However, if not properly
controlled such a feature can lead to an overload in the network
core. For example, a wireless client scheduled on N APs in the
uplink direction could increase the load on the network core
by a factor of N. Moreover, a straightforward implementation
of such a mechanism could generate a significant increase in
the number of duplicate packets which in time could trigger
anomalous behaviors at the transport layer.
As a result of the fact that the WiFi MAC is virtualized
it is possible to offload parts of it to dedicated processing
Fig. 14. Data/Management Plane Offloading application. Data–plane (dupli-
cates filtering) and Management–plane (probe response generation) operations
are offloaded to two different CPPs.
nodes as VNFs that can be then shared among several APs. In
this use case we implemented a VNF which filters out dupli-
cate 802.11 frames based on their sequence number. Figure 14
sketches a slightly more generic case were both data–plane
(duplicates filtering) and management (probe response genera-
tion) functionalities are offloaded to two different CPPs. In this
section however we shall evaluate only the performances of the
duplicate filtering VNF.
Traffic originated from clients is received at one or more
WTPs where it is encapsulated (802.11 over Ethernet) and then
forwarded to a CPP where the duplicates filtering VNF is
deployed. This VNF is also responsible for decapsulating the
802.11 frame and converting it into an Ethernet frame before
forwarding it to its intended destination.
In order to evaluate this VNF we exploited a network setup
composed of a single client and three APs. Traffic is injected
from the wireless client as a single UDP stream. Packet trans-
mission rate and payload are kept fixed at, respectively, 100
packets/s and 1472 bytes. Impairments on the link between
client and APs are simulated by randomly dropping received
frames with probability PL at all receiving APs. For all mea-
surements a total of 6000 frames were generated. Confidence
intervals were very small for all the data points and have there-
fore be omitted in order to improve readability. Measurements
have been taken using three packet dropping probabilities,
namely: 0.05 (good channel conditions), 0.2 (medium channel
conditions), and 0.8 (poor channel conditions)
As it can be seen from Fig. 15a, the end–to–end packet deliv-
ery ratio increases with the number of available uplinks. The
proposed uplink/downlink decoupling solution can provide a
small performance improvement even when the channel condi-
tions are good (PL = 0.05). On the other hand the performance
improvements are significant when the channel conditions get
worse. In particular this solution allows to turn an essentially
broken channel (4 out of 5 dropped packets) into an usable
channel (1 out of 2 dropped packets). Finally, in Fig. 15b and
Fig. 15c we can see the impact of the duplicate filtering VNF
on the bandwidth utilization. As expected without the VNF
the bandwidth utilization increases with the number of uplinks,
while using the VNF filtering the bandwidth utilization does not
exceed the nominal goodput.

Fig. 15. Packet loss (a) and bandwidth utilization (b and c) for a single client scheduled at multiple APs in the uplink direction. The duplicate filtering VNF can
reduce the load on the backhaul by selectively dropping duplicate packets.
VIII. CONCLUSIONS
NFV is rapidly emerging as a flexible solution to deploy
and operate future mobile networks. However, in order to fully
deliver on its promises, the concept of NFV must be extended
also to the radio access segment of the mobile network. In this
paper, we tackled this challenge by presenting a novel formu-
lation of the VNF placement problem encompassing also radio
access VNFs. We then introduced a ILP–based algorithm for
small networks and a scalable heuristic, named WiNE, for larger
deployments. Numerical simulations showed that the proposed
heuristic can approximate the performance of the ILP–based
placement algorithms for realistic substrate networks. Finally,
we reported on a preliminary proof–of–concept implementa-
tion of a NFV Management and Orchestration framework for
Enterprise WLANs.
As future work, we plan to investigate the resiliency proper-
ties of WiNE in case of nodes and link failures and to study how
VNF placement can be optimized by taking into account wire-
less clients distribution and user mobility. We also plan to verify
the applicability of our problem formulation to OFDMA–based
radio access networks like LTE and LTE–Advanced and to
extend the prototype (both the hypervisor and the controller)
adding support for VNF migration and scaling as well as for
Mobile Edge Computing capabilities.
REFERENCES
[1] H. Moens and F. De Turck, “VNF-P: A model for efficient placement of
virtualized network functions,” in Proc. IEEE 10th Int. Conf. Netw. Serv.
Manage. (CNSM), Rio de Janeiro, Brazil, 2014, pp. 418–423.
[2] A. Fischer, J. Botero, M. Till Beck, H. de Meer, and X. Hesselbach,
“Virtual network embedding: A survey,” IEEE Commun. Surveys Tuts.,
vol. 15, no. 4, pp. 1888–1906, Apr. 2013.
[3] B. Jennings and R. Stadler, “Resource management in clouds: Survey and
research challenges,” J. Netw. Syst. Manage., vol. 23, no. 3, pp. 567–619,
Jul. 2015.
[4] M. Ghaznavi, A. Khan, N. Shahriar, K. Alsubhi, R. Ahmed, and
R. Boutaba, “Elastic virtual network function placement,” in Proc. IEEE
CloudNet, Niagara Falls, ON, Canada, 2014, pp. 255–260.
[5] R. Riggio, A. Bradai, T. Rasheed, J. Schulz-Zander, S. Kuklinski, and
T. Ahmed, “Virtual network functions orchestration in wireless net-
works,” in Proc. IEEE 11th Int. Conf. Netw. Serv. Manage. (CNSM),
Barcelona, Spain, 2015, pp. 108–116.
[6] M. Dobrescu et al., “RouteBricks: Exploiting parallelism to scale soft-
ware routers,” in Proc. ACM Symp. Oper. Syst. Principles (SOSP), Big
Sky, MT, USA, 2009, pp. 15–28.
[7] S. Han, K. Jang, K. Park, and S. Moon, “PacketShader: A GPU-
accelerated software router,” SIGCOMM Comput. Commun. Rev., vol. 40,
no. 4, pp. 195–206, Aug. 2010.
[8] K. K. Ram, A. L. Cox, M. Chadha, and S. Rixner, “Hyper-switch: A
scalable software virtual switching architecture,” in Proc. USENIX Annu.
Tech. Conf. (ATC), San Jose, CA, USA, 2013, pp. 13–24.
[9] European Telecommunications Standards Institute (ETSI), “ETSI GS
NFV 002 network functions virtualisation (NFV); Architectural frame-
work,” Dec. 2014.
[10] Linux Foundation. OPNFV [Online]. Available: https://www.opnfv.org/
[11] Telefonica. OpenMANO [Online]. Available: https://github.com/
nfvlabs/
[12] Fraunhofer FOKUS. OpenBATON [Online]. Available: http://openbaton.
github.io/
[13] M. Chowdhury, M. R. Rahman, and R. Boutaba, “ViNEYard: Virtual net-
work embedding algorithms with coordinated node and link mapping,”
IEEE/ACM Trans. Netw., vol. 20, no. 1, pp. 206–219, Feb. 2012.
[14] M. Chowdhury, F. Samuel, and R. Boutaba, “PolyViNE: Policy-based
virtual network embedding across multiple domains,” in Proc. ACM
Workshop Virtualized Infrastruct. Syst. Archit. (VISA), New Delhi, India,
2010, pp. 49–56.
[15] S. Zhang, Z. Qian, J. Wu, S. Lu, and L. Epstein, “Virtual network embed-
ding with opportunistic resource sharing,” IEEE Trans. Parallel Distrib.
Syst., vol. 25, no. 3, pp. 816–827, Mar. 2014.
[16] M. A. Khan, N. Shahriar, R. Ahmed, and R. Boutaba, “Simple:
Survivability in multi-path link embedding,” in Proc. IEEE 11th Int. Conf.
Netw. Serv. Manage. (CNSM), Barcelona, Spain, 2015, pp. 210–218.
[17] D. Breitgand, A. Epstein, A. Glikson, A. Israel, and D. Raz, “Network
aware virtual machine and image placement in a cloud,” in Proc. IEEE
9th Int. Conf. Netw. Serv. Manage. (CNSM), Zurich, Switzerland, 2013,
pp. 9–17.
[18] M. Barshan, H. Moens, and F. De Turck, “Design and evaluation of a scal-
able hierarchical application component placement algorithm for cloud
resource allocation,” in Proc. IEEE 10th Int. Conf. Netw. Serv. Manage.
(CNSM), Rio de Janeiro, Brazil, 2014, pp. 175–180.
[19] M. Barshan, H. Moens, S. Latre, and F. De Turck, “Algorithms for
efficient data management of component-based applications in cloud
environments,” in Proc. IEEE Netw. Oper. Manage. Symp. (NOMS),
Krakow, Poland, 2014, pp. 1–8.
[20] R. Guerzoni et al., “A novel approach to virtual networks embedding
for SDN management and orchestration,” in Proc. IEEE Netw. Oper.
Manage. Symp. (NOMS), Krakow, Poland, 2014, pp. 1–7.
[21] S. Clayman, E. Maini, A. Galis, A. Manzalini, and N. Mazzocca, “The
dynamic placement of virtual network functions,” in Proc. IEEE Netw.
Oper. Manage. Symp. (NOMS), Krakow, Poland, 2014, pp. 1–9.
[22] F. Bari, S. R. Chowdhury, R. Ahmed, and R. Boutaba, “On orchestrat-
ing virtual network functions,” in Proc. IEEE 11th Int. Conf. Netw. Serv.
Manage. (CNSM), Barcelona, Spain, 2015, pp. 50–56.
[23] L. Xia et al., “Virtual WiFi: Bring virtualization from wired to wireless,”
SIGPLAN Not., vol. 46, no. 7, pp. 181–192, Mar. 2011.
[24] P. Lv, X. Wang, and M. Xu, “Virtual access network embedding in wire-
less mesh networks,” Ad Hoc Netw., vol. 10, no. 7, pp. 1362–1378, 2012.
[25] P. Lv, Z. Cai, J. Xu, and M. Xu, “Multicast service-oriented virtual
network embedding in wireless mesh networks,” IEEE Commun. Lett.,
vol. 16, no. 3, pp. 375–377, Mar. 2012.
[26] G. Bhanage, R. Daya, I. Seskar, and D. Raychaudhuri, “VNTS: A vir-
tual network traffic shaper for air time fairness in 802.16e systems,” in
Proc. IEEE Int. Conf. Commun. (ICC), Cape Town, South Africa, 2010,
pp. 1–6.
[27] G. Bhanage, D. Vete, I. Seskar, and D. Raychaudhuri, “SplitAP:
Leveraging wireless network virtualization for flexible sharing of
WLANs,” in Proc. IEEE GLOBECOM, Miami, FL, USA, 2010, pp. 1–6.

[28] R. Kokku, R. Mahindra, H. Zhang, and S. Rangarajan, “NVS: A substrate
for virtualizing wireless resources in cellular networks,” IEEE/ACM
Trans. Netw., vol. 20, no. 5, pp. 1333–1346, Oct. 2012.
[29] R. Kokku, R. Mahindra, H. Zhang, and S. Rangarajan, “CellSlice:
Cellular wireless resource slicing for active RAN sharing,” in Proc. IEEE
5th Int. Conf. Commun. Syst. Netw. (COMSNETS), Bangalore, India,
2013, pp. 1–10.
[30] G. Bhanage, I. Seskar, R. Mahindra, and D. Raychaudhuri, “Virtual
basestation: Architecture for an open shared WiMAX framework,” in
Proc. ACM Workshop Virtualized Infrastruct. Syst. Archit (VISA), New
Delhi, India, 2010, pp. 1–8.
[31] G. Smith, A. Chaturvedi, A. Mishra, and S. Banerjee, “Wireless vir-
tualization on commodity 802.11 hardware,” in Proc. ACM Workshop
Wireless Netw. Testbeds Exp. Eval. Charact. (WinTECH), Montreal, ON,
Canada, 2007, pp. 75–82.
[32] A. Gember-Jacobson et al., “OpenNF: Enabling innovation in network
function control,” in Proc. ACM SIGCOMM, Chicago, IL, USA, 2014,
pp. 163–174.
[33] B. Kothandaraman, M. Du, and P. Sköldström, “Centrally controlled dis-
tributed VNF state management,” in Proc. ACM HotMiddlebox, London,
U.K., 2015, pp. 37–42.
[34] A. Gember-Jacobson and A. Akella, “Improving the safety, scalabil-
ity, and efficiency of network function state transfers,” in Proc. ACM
HotMiddlebox, London, U.K., 2015, pp. 43–48.
[35] The Bro Network Security Monitor [Online]. Available:
https://www.bro.org/
[36] The Squid Caching Proxy [Online]. Available: http://www.squid-
cache.org/
[37] S. Rajagopalan, D. Williams, H. Jamjoom, and A. Warfield, “Split/merge:
System support for elastic execution in virtual middleboxes,” in Proc.
USENIX Netw. Syst. Des. Implement. (NSDI), Lombard, IL, USA, 2013,
pp. 227–240.
[38] V. Sekar, N. Egi, S. Ratnasamy, M. K. Reiter, and G. Shi, “Design
and implementation of a consolidated middlebox architecture,” in Proc.
USENIX Netw. Syst. Des. Implement. (NSDI), San Jose, CA, USA, 2012,
pp. 323–336.
[39] J. W. Anderson, R. Braud, R. Kapoor, G. Porter, and A. Vahdat,
“xOMB: Extensible open middleboxes with commodity servers,” in Proc.
ACM/IEEE ABCS, Austin, TX, USA, 2012, pp. 49–60.
[40] Y. Minlan, Y. Yi, J. Rexford, and M. Chiang, “Rethinking virtual net-
work embedding: Substrate support for path splitting and migration,”
SIGCOMM Comput. Commun. Rev., vol. 38, no. 2, pp. 17–29, Mar. 2008.
[41] Y. Zhu and M. Ammar, “Algorithms for assigning substrate network
resources to virtual network components,” in Proc. IEEE INFOCOM,
Barcelona, Spain, Apr. 2006, pp. 1–12.
[42] E. Kohler, R. Morris, B. Chen, J. Jannotti, and M. F. Kaashoek, “The click
modular router,” ACM Trans. Comput. Syst., vol. 18, no. 3, pp. 263–297,
Aug. 2000.
[43] Ryu SDN Framework [Online]. Available: http://osrg.github.io/ryu/
[44] R. Riggio, M. Marina, J. Schulz-Zander, S. Kuklinski, and T. Rasheed,
“Programming abstractions for software-defined wireless networks,”
IEEE Trans. Netw. Serv. Manage., vol. 12, no. 2, pp. 146–162, Jun. 2015.
[45] M. Heusse, F. Rousseau, G. Berger-Sabbatel, and A. Duda, “Performance
anomaly of 802.11b,” in Proc. IEEE INFOCOM, San Francisco, CA,
USA, 2003, pp. 836–843.
Roberto Riggio received the Ph.D. degree in commu-
nications engineering from the University of Trento,
Trento, Italy, in 2008. He is currently a Senior
Researcher Staff Member with the Future Networks
Team, CREATE-NET, Trento, Italy, and a Guest
Lecturer with the University of Trento. His research
interests include performance isolation in multitenant
mobile networks; software defined mobile network-
ing and deep programmable mobile networks. He has
been involved in several European projects includ-
ing three 5G-PPP projects and has directly generated
more than 1 MC in competitive funding. He has authored more than 60 papers
published in internationally refereed journals and conferences. He has served
on the TPC of leading conferences in the networking field and is an Associate
Editor for the International Journal of Network Management (Wiley) and for
the Wireless Networks Journal (Springer).
Abbas Bradai received the M.S. degree in com-
puter science from the National Institute of Computer
Science (ESI ex-INI), Algiers, Algeria, and the
University of Rennes 1, Rennes, France, in 2009,
and the Ph.D. degree from the LaBRI/University of
Bordeaux, Talence, France, in 2012. He is actually
an Associate Professor with University of Poitiers,
Poitiers, France, and a Research Fellow with XLIM
Laboratory, Poitiers, France. His research interests
include multimedia communications over wired and
wireless networks, cognitive radio, software defined
network and virtualization. He is/was involved in many French and European
projects (FP7, H2020) such as ENVISION and VITAL.
Davit Harutyunyan received the first level degree
in telecommunication engineering and the master’s
degree in telecommunication engineering (both with
hons.) from the National Polytechnic University of
Armenia, Yerevan, Armenia, in 2011 and 2015,
respectively. He is currently pursuing the Ph.D.
degree at the University of Trento, Trento, Italy.
Since 2015, he has been a member of the Future
Networks Team, CREATE-NET, Trento, Italy. His
research interests include cellular networks and soft-
ware defined mobile networking.
Tinku Rasheed (M’05) received the bachelor’s
degree in telecommunications from University of
Kerala, Thiruvananthapuram, India, in 2002, the M.S.
degree from Aston University, Birmingham, U.K.,
in 2003, and the Ph.D. degree from the Department
of Computer Science, University of Paris-Sud XI.,
Orsay, France, in 2007. He is a Senior Research Staff
Member and the Head of the Future Networks Area
within CREATE-NET, Trento, Italy. Before join-
ing CREATE-NET, he was Research Engineer with
Orange Labs for four years. He has extensive indus-
trial and academic research experience in the mobile wireless communication
area, end-to-end network architectures and services. He has several granted
patents and has authored more than 60 articles published in major journals and
conferences. He is a member of the ACM.
Toufik Ahmed is a Full Professor with ENSEIRB-
MATMECA School of Engineers, Institut
Polytechnique de Bordeaux (Bordeaux-INP),
Talence, France, and is involved with research
activities in CNRS-LaBRI Lab-UMR 5800 at the
University Bordeaux, Bordeaux, France. His research
interests include end-to-end Quality of Service
(QoS) management and provisioning for multimedia
wired and wireless networks, media streaming, and
cross-layer optimization. He has also worked on a
number of national and international projects. He has
served as a TPC member for international conferences and journals.

Scheduling wireless virtual networks functions

More Related Content

What's hot (19)

Viewers also liked (6)

Similar to Scheduling wireless virtual networks functions (20)

More from redpel dot com (20)

Recently uploaded (20)

Scheduling wireless virtual networks functions