Content uploaded by Xavier Costa-Pérez
Author content
All content in this area was uploaded by Xavier Costa-Pérez
Content may be subject to copyright.
IEEE Communications Magazine • July 2013 27
0163-6804/13/$25.00 © 2013 IEEE
The research leading to
these results has been par-
tially funded by the Euro-
pean Community’s
Seventh Framework Pro-
gramme (FP7/2007-2013)
under grant agreement no.
257263 (FLAVIA).
INTRODUCTION
Future mobile carrier networks need to address
the predicted growth in mobile traffic volume
which is expected to have explosive growth in
the next years mainly driven by video and web
applications [1]. Based on the expected capacity
demand increase, mobile network operators
(MNOs) are required to increase capital
(CAPEX) and operational expenses (OPEX)
accordingly. However, as cellular networks move
from being voice-centric to data-centric, increase
in revenue is not keeping pace with the increase
in traffic volume. Forecasted future mobile net-
works capacity requirements together with the
decreasing operators’ benefits margin have led
to network sharing being considered a key busi-
ness model for reducing future deployment and
operational costs.
Network sharing solutions are already avail-
able, standardized [2], and partially used in some
mobile carrier networks. These solutions can be
divided into passive and active network sharing.
Passive sharing refers to the reuse of compo-
nents such as physical sites, tower masts, cabling,
cabinets, power supply, air-conditioning, and so
on. Active sharing refers to the reuse of back-
haul, base stations, and antenna systems, the
reuse of the latter two labeled as active radio
access network (RAN) sharing. According to a
market survey [3], mobile infrastructure sharing
has already been deployed by over 65 percent of
European operators in different ways, and this
trend is expected to expand in the future. One of
the main motivations for network sharing is that
currently, a considerable number of sites con-
sume energy and computational resources, even
though they carry a negligible level of traffic.
For instance, in [4] it was reported that around
20 percent of all sites carry about 50 percent of
total traffic. Estimations regarding the expected
savings for operators by implementing active
network sharing have been conducted [5]. This
study concluded that operators worldwide could
reduce combined OPEX and CAPEX costs up
to $60 billion over a five-year period, and at
least 40 percent of these cost savings is expected
to come from active network sharing.
The traditional model of single ownership of
all network layers and elements is thus being
challenged: since the most basic sharing con-
cepts related to passive network sharing are easi-
er to implement and have already been partially
exploited, active network sharing is rising in
importance to enable substantial and sustainable
reduction in network expenses to ensure opera-
tors’ future cost competitiveness.
Active RAN sharing involves sharing anten-
nas/base stations across multiple mobile (virtual)
network operators with either separate spectrum
resources for each entity or shared spectrum
resources through spectrum pooling. Current
solutions, however, have limitations in terms of
separating both data and control planes among
operators, flexibility and customizability for
accommodating different requirements per oper-
ator, and the capability to adapt to new or chang-
ing requirements.
We first review 3GPP’s network sharing stan-
dardized functionality. Emerging business mod-
ABSTRACT
This article presents a survey of cellular net-
work sharing, which is a key building block for
virtualizing future mobile carrier networks in
order to address the explosive capacity demand
of mobile traffic, and reduce the CAPEX and
OPEX burden faced by operators to handle this
demand. We start by reviewing the 3GPP net-
work sharing standardized functionality followed
by a discussion on emerging business models
calling for additional features. Then an overview
of the RAN sharing enhancements currently
being considered by the 3GPP RSE Study Item
is presented. Based on the developing network
sharing needs, a summary of the state of the art
of mobile carrier network virtualization is pro-
vided, encompassing RAN sharing as well as a
higher level of base station programmability and
customization for the sharing entities. As an
example of RAN virtualization techniques feasi-
bility, a solution based on spectrum sharing is
presented: the network virtualization substrate
(NVS), which can be natively implemented in
base stations. NVS performance is evaluated in
an LTE network by means of simulation, show-
ing that it can meet the needs of future virtual-
ized mobile carrier networks in terms of
isolation, utilization, and customization.
FUTURE CARRIER NETWORKS
Xavier Costa-Pérez and Joerg Swetina, NEC Laboratories Europe
Tao Guo, NEC Telecom Modus
Rajesh Mahindra, Sampath Rangarajan, NEC Laboratories America
Radio Access Network Virtualization for
Future Mobile Carrier Networks
PEREZ-COSTA LAYOUT_Layout 1 6/26/13 4:39 PM Page 27
IEEE Communications Magazine • July 2013
28
els calling for new sharing capabilities are pro-
vided next, followed by the Third Generation
Partnership Project’s (3GPP’s) ongoing efforts in
the RAN Sharing Enhancements (RSE) Study
Item to address them. Since base station virtual-
ization-based techniques are key enablers for
RAN sharing, along with the required manage-
ment flexibility and independence among the
sharing entities, we provide an overview of the
state of the art of mobile carrier network virtual-
ization. A concrete implementation of a RAN
sharing base station virtualization solution, the
network virtualization substrate (NVS), and
quantitative performance results in a Long Term
Evolution (LTE) system are provided. Finally,
we summarize our contributions and conclude
the article.
3GPP NETWORK SHARING
STANDARDIZATION
3GPP has recognized the importance of support-
ing network sharing since Release 6, and defined
a set of architectural requirements [6] and tech-
nical specifications [2] that have been updated,
at the time of writing this article, through
Releases 10 and 11, respectively.
3GPP NETWORK SHARING
FUNCTIONALITY OVERVIEW
In order to derive the network sharing require-
ments for 3GPP, five main use cases were
defined in [6]:
•Scenario 1 corresponds to multiple core net-
works (CNs) sharing a common RAN in Release
99. For operators that have multiple frequency
allocations, it is possible to share the RAN ele-
ments, but not to share the radio frequencies. In
this case the operators connect directly to their
own dedicated carrier layer in the shared radio
network controller (RNC) in the shared RAN.
•Scenario 2 corresponds to two or more oper-
ators with individual frequency licenses with
which their respective RANs cover different
parts of a country, but together provide coverage
of the entire country.
•Scenario 3 corresponds to one operator
deploying coverage in a specific geographical
area, and other operators being allowed to use
this coverage for their subscribers. Outside this
geographical area, coverage is provided by each
of the operators independently.
•Scenario 4 corresponds to common spec-
trum network sharing when:
–One operator has a frequency license and
shares the allocated spectrum with other
operators.
–A number of operators decide to pool
their allocated spectra and share the total
spectrum.
•Scenario 5 corresponds to multiple RANs
sharing a common CN. The multiple RANs can
belong to different public land mobile networks
(PLMNs) and network operators. Due to opera-
tors’ deployment choices, different nodes or part
of the common CN can be shared.
Based on these scenarios, 3GPP has defined
a network sharing architecture with the objective
of allowing different CN operators to connect to
a shared RAN [2]. Operators may share not only
network elements, but radio resources as well.
Two possible architectural network sharing con-
figurations have been specified: the gateway CN
(GWCN) and the multi-operator CN (MOCN).
• GWCN: In a GWCN network sharing con-
figuration, CN operators share CN in addi-
tion to RAN nodes.
• MOCN: In an MOCN network sharing con-
figuration, multiple CN nodes are connect-
ed to the same RNC, and the CN nodes are
operated by different operators.
In both configurations, the network sharing
agreement between operators is transparent to
the users. Base stations broadcast multiple
PLMN-ids according to the number of mobile
(virtual) network operators. Mobile stations,
user equipment (UE) in 3GPP terminology, that
support network sharing functionality are able to
discriminate among operators in a spectrum net-
work sharing configuration based on the PLMN-
ids. Non-supporting UE devices ignore the
broadcast system information, and the shared
network selects a CN operator from the avail-
able ones. PLMN-ids are used for routing and
handover purposes within the network.
DEVELOPING BUSINESS MODELS FOR
NETWORK SHARING
As mobile carrier operators look for new ways to
reduce costs due to their increasing financial
investment burden, new business models for
operators and infrastructure owners are emerg-
ing, calling for an extension to existing standard-
ized network sharing functionality. See [7] for
use cases under discussion in 3GPP and [8] for
an extensive industry survey on mobile opera-
tors’ views. In the following, we provide several
examples of these disruptive developing business
models.
On-Demand Capacity — There are several
business cases where different service providers
could be interested in buying capacity from
MNOs for a specific period of time and/or pur-
pose. Examples include specialized mobile virtu-
al network operators (MVNOs) offering specific
services (e.g., VoIP, video telephony, live stream-
ing), mobile operators requiring additional
capacity for major events (e.g., sports, concerts,
fairs), and emerging machine-to-machine (M2M)
service providers (e.g., meter measurements,
security surveillance). Dynamic and flexible shar-
ing of the infrastructure would be required in
this case on timescales smaller than the current
times needed for contract agreement and imple-
mentation.
Wholesale-Only Network — In this model
MNOs do not offer services to end customers
but sell network capacity to businesses who do
not have their own wireless network or have lim-
ited geographic coverage/spectrum. This repre-
sents an extreme case for network sharing
requirements where full-fledged flexibility in
terms of dynamic resource allocation, traffic iso-
lation, and customizability is needed to maximize
its potential.
As mobile carrier
operators look for
new ways to reduce
costs due to their
increasing financial
investment burden,
new business models
for operators and
infrastructure owners
are emerging, calling
for an extension to
existing standardized
network sharing
functionality.
PEREZ-COSTA LAYOUT_Layout 1 6/26/13 11:49 AM Page 28
IEEE Communications Magazine • July 2013 29
Over-The-Top Service Providers — In current
mobile networks the biggest capacity share is
consumed by the content of third party service
providers (over-the-top, OTT, services). Such
OTT service providers may wish to pay wireless
network operators in the future to ensure a
satisfactory quality of experience for their
users. In such cases no additional M(V)NO
infrastructure might be involved in the RAN
sharing process.
These new business models together with
additional ones under discussion [7, 8], omitted
here for the sake of brevity, have triggered the
creation of a RAN Sharing Enhancements Study
Item in 3GPP, described next.
3GPP RAN SHARING
ENHANCEMENTS STUDY ITEM OVERVIEW
The 3GPP RAN Sharing Enhancements (RSE)
Study Item [7] of the System Architecture Work-
ing Group 1 (SA1) is defining new scenarios of
multiple mobile carrier operators sharing radio
network resources. Some of the authors of this
article are actively contributing to this effort and
have been appointed as the official 3GPP RSE
Study Item Rapporteurs.
The objective of the RSE Study Item is to
create new requirements that complement exist-
ing system capabilities for sharing common RAN
resources. The new scenarios aim to:
• Provide means to be able to verify that the
shared network elements provide allocated
RAN resources according to the sharing
agreements and/or policies
• Provide means to efficiently share common
RAN resources according to identified
RAN sharing scenarios (e.g., pooling of
unallocated radio resources)
• Provide means to flexibly and dynamically
allocate RAN resources on demand at
smaller timescales than those supported
today
• Provide means to act on overload situations
considering sharing agreements and/or poli-
cies
Based on these objectives, we describe below
some of the main use cases being used for defin-
ing new 3GPP requirements within the SA1 RSE
Study Item.1
RSE Monitoring — This use case describes the
situation of a hosting RAN provider (primary
RAN operator) that allows participating opera-
tors (e.g., MVNOs) to use a share of the RAN
capacity where the amount of resources might
be different for each participating operator. In
such a scenario, the hosting RAN shall allow
participating operators to retrieve operation,
administration, and management (OAM) status
information at the same level of detail as in a
non-shared RAN for the share of their
resources.
RSE Dynamic Capacity — This use case
describes the situation of participating operators
requiring varying network capacities during dif-
ferent time periods (e.g., within a day or a week).
Participating operators might request various
allocations of a portion of the shared RAN in
order to meet the projected variation in network
usage requirements. An example of this use case
could be an MVNO requiring significant RAN
capacity during business hours but lower capaci-
ty during the night or weekends.
RSE On-Demand Automated Capacity Bro-
kering — This use case describes the situation
of a hosting RAN provider that supports on-
demand automated capacity requests from par-
ticipating operators. A typical situation for this
use case would be a hosting RAN provider with
excess capacity during the night that participat-
ing operators could request on demand (e.g.,
M2M services such as smart meter measure-
ments. Another example could be a major event
(e.g., sports, concerts, trade fairs, etc.) requiring
additional capacity from a participating operator
for that event. On-demand requests for capacity
shall indicate the time period in which the capac-
ity is needed, the amount of capacity required,
and the specific service(s) treatment desired, for
example, based on standardized quality of ser-
vice (QoS) class identifiers (QCIs). The hosting
RAN provider then verifies automatically
whether the RAN sharing request can be ful-
filled; if so, the shared RAN is re-configured
accordingly.
RSE Load Balancing — This use case
describes the situation of overlapping cells in a
shared RAN. The hosting RAN shall be able to
support load balancing within a shared RAN
while respecting the agreed shares of RAN
resources based on the whole cell load level
and the load level for each participating opera-
tor. If the load levels of individual cells are
exceeded, the hosting RAN provider shall
enforce agreed maximal usage limits of each
participating operator by handing over UE
devices to neighboring cells.
Based on the described use cases and addi-
tional ones under discussion, a set of new con-
solidated requirements has been derived. In
Table 1 we summarize some of these main new
requirements, which illustrate the new function-
ality expected to be specified based on the RSE
Study Item contributions.
These set of new requirements illustrate the
paradigm change being defined by the RSE
Study Item work. The 3GPP acceptance of
automated means (no human intervention) to
allocate RAN resources dynamically and on
demand deserves special mention. This
enhanced functionality opens the door to new
industry players as well as to an evolution of
operators’ business models which could poten-
tially have a deep impact on the industry. Addi-
tionally, the enhanced sharing flexibility along
the capacity, time, and sector dimensions enable
higher efficiency of RAN infrastructure
exploitation.
The RSE work in TR 22.852 [7] is expected
to be completed and kick off a new Work Item
producing normative text this year. In the next
section, we discuss the current state of the art of
mobile carrier networks’ virtualization technolo-
gies that could potentially address the require-
ments of the previously described emerging
RAN sharing use cases.
1Note that at the time of
writing this article the
Study Item is close to
completion, but changes
might still be made in the
final document release.
In current mobile
networks the biggest
capacity share is con-
sumed by the con-
tent of third party
service providers
(over-the-top, OTT,
services). Such OTT
service providers may
wish to pay wireless
network operators in
the future to ensure
a satisfactory quality
of experience for
their users.
PEREZ-COSTA LAYOUT_Layout 1 6/26/13 4:39 PM Page 29
IEEE Communications Magazine • July 2013
30
MOBILE CARRIER NETWORKS
VIRTUALIZATION
Although network virtualization in the wired
domain has received significant attention in
the past years [9], mobile carrier network vir-
tualization is still in its nascent stage. The
main focus of wired network virtualization has
been on the design of an adaptive network
substrate that can support multiple virtual net-
works running customized services over a
physical network. To enable end-to-end cellu-
lar network virtualization, the mobile CN and
RAN have to be virtualized. While solutions
from the wired network and server domains
can be used to virtualize the mobile CN, RAN
or base station virtualization has to deal with
problems specific to the characteristics of
wireless access links. Wireless access links usu-
ally have more dynamic sets of users, user
mobility, and varying channel conditions that
make it harder to virtualize the wireless
resources across multiple entities.
Base station virtualization may be performed
at either the hardware level (dedicated spectrum)
or the flow level (shared spectrum). Hardware
virtualization solutions exist commercially today
for traditional MNOs to cut operating costs. The
virtual base transceiver station (vBTS) [10] is one
such virtualized base station solution that enables
sharing radio components at the hardware level
and running multiple base station protocol stacks
in software. To support spectrum-sharing-based
models, virtualization at higher levels such as the
flow level at the base station is more appropriate
and leads to better multiplexing of resources.
Furthermore, it can support multiple deployment
scenarios such as MVNOs where the virtual net-
works do not own spectrum. Multiple MNOs can
also pool their spectrum to save costs and accel-
erate network rollouts.
Several GENI design documents describe
proposals and issues of wireless network virtu-
alization [11]. The aim of GENI is to create a
unified nationwide wired-wireless virtualized
testbed that can support multiple concurrent
experiments. In this context, recent work [12]
focuses on virtualizing the wireless resources of
WiMAX base stations remotely from an access
service network (ASN) gateway. Other efforts
[13] propose to virtualize the LTE network by
implementing a hypervisor in the eNodeB.
Each entity runs its LTE stack in a virtual
machine. The hypervisor allocates spectrum to
the different entities in accordance with a spec-
ified guarantee. An entity can request either a
fixed or dynamic guarantee based on its current
load up to a maximum amount of resources.
However, all of the above mentioned schemes
only ensure isolation of resources and do not
guarantee bandwidth across the entities. More-
over, these schemes fail to provide flexibility to
the slices to customize the resource allocation
across their users.
The European FP7 project Flexible Architec-
ture for Virtualizable Future Wireless Internet
Access (FLAVIA) [14] is defining and prototyp-
ing a new base station architecture with the
objective of enabling a higher level of pro-
grammability. FLAVIA promotes the concept of
a wireless medium access control (MAC) proces-
sor, a programmable device that:
• Provides a set of stateless MAC commands
• Embeds a MAC protocol engine in charge
of executing a finite state machine able to
exploit and compose the sequence of com-
mands forming a desired protocol
This enhanced base station programmability
functionality is being explored in the context of
wireless access virtualization, and project partici-
pants contribute to the 3GPP RSE Study Item
efforts.
In parallel, the emerging software-defined
networking (SDN) paradigm is a key enabler to
simplify the network provisioning, management,
(re)configuration, and control of such virtualized
and shared mobile carrier networks. Wireless
SDN requires the identification of abstractions
for wireless primitives and functions, which com-
promise between flexibility and ability of the
abstractions to permit developing innovative
wireless functions and mechanisms. An overview
of ongoing SDN standardization efforts can be
found in [15].
With respect to commercial deployments,
there are several MVNOs that lease spectrum
Table 1. 3GPP SA1 RSE Study Item, consolidated requirements sample, TR
22.852 [7].
Hosting RAN resources allocation
• A hosting RAN provider shall be able to allocate the share of RAN capacity
per participating operator as:
– Fixed, minimum allocation guaranteed
– Fixed for a period of time and/or sectors
– First come first served (i.e., on demand)
• A shared RAN shall be capable of providing differentiated traffic treatment
per participating operator
• A shared RAN shall conduct admission control based on the proportion of
assigned RAN usage per participating operator
On-demand capacity negotiation
• The hosting RAN shall be able to offer by automatic means shareable RAN
resources as on-demand capacity to participating operators
• Participating operators shall be able to automatically request hosting-RAN-
offered on-demand resources
• The hosting RAN provider shall allow a participating operator to request
the cancellation of granted on-demand requests
• The hosting RAN provider shall be able to withdraw a granted request
(within SLA limits)
OAM access to the hosting RAN
• The hosting RAN shall be able to provide and control selective OAM access
to participating operators
• The hosting RAN shall be able to allow participating operators to retrieve
selective OAM status information at the same level of detail as would be
available from a non-shared RAN
Handovers due to RAN sharing agreements
• Participating operators shall be able to direct UE toward the hosting RAN
at the beginning of a RAN sharing period
• A hosting RAN provider shall be able to direct UE away from the hosting
RAN at the end of a RAN sharing period
• Participating operators shall be involved in the handover decisions at the
end of a RAN sharing period
PEREZ-COSTA LAYOUT_Layout 1 6/26/13 11:49 AM Page 30
IEEE Communications Magazine • July 2013 31
from existing MNOs to provide cellular services.
There are also a few instances of MNO-based
RAN sharing in different parts of the world [8].
However, there is very little publicly available
information on the base station virtualization
aspects of such sharing solutions.
BASE STATION
VIRTUALIZATION FEASIBILITY
So far we have considered mobile carrier net-
works virtualization from a conceptual perspec-
tive. In this section, we take the next step and
focus on describing a specific solution to imple-
ment it in real networks. In order to do so, we
build on our work on base station virtualization,
which allows multiple entities to share the same
spectrum, achieving effective virtualization of its
wireless resources: the network virtualization
substrate (NVS) [16]. NVS can be readily
deployed to virtualize orthogonal frequency-divi-
sion multiple access (OFDMA)-based 4G base
stations such as LTE and WiMAX, and aims to
provide key virtualization features as defined by
the 3GPP SA1 RSE consolidated requirements:
•Isolation ensures that traffic, mobility, and
fluctuations in channel conditions of users
of one entity do not affect the reserved
resource allocations of other entities shar-
ing the same base station: RSE hosting
RAN resources allocation.
•Customization provides flexibility to the dif-
ferent entities to program the base station
to optimize their service delivery: RSE
OAM access to the hosting RAN.
•Utilization of base stations is maximized by
allowing usage of resources unused by one
entity by other entities: RSE hosting RAN
resources allocation.
From this point onward we call virtual net-
works slices. A slice is defined as a virtual net-
work consisting of a group of flows that share a
physical base station with multiple other virtual
networks.
BASE STATION VIRTUALIZATION: DEFINITIONS
To meet the three aforementioned requirements
for virtualization, NVS is designed as a hierar-
chical scheduler, as shown in Fig. 1, with two
components:
•A slice scheduler ensuring resource isolation
across the slices
•A flow scheduling framework enabling flexi-
bility of flow scheduling to the different
slices
Separating the slice scheduler from the flow
scheduler(s) enables greater flexibility in cus-
tomizing the slices in both the uplink and down-
link directions, and also makes service level
agreements (SLAs) between the slice owners
and the physical network owner easier to define
and manage.
NVS operates in fine timescales in order to
fill up the resources of a physical layer frame. At
every scheduling instant, NVS selects a slice that
maximizes the overall utility or revenue. Once a
slice is selected, NVS schedules flows within the
slice according to a custom slice policy.
Slice Scheduler — NVS defines two possible
ways in which slices can reserve a part of the
base station resources:
•Resource-based provisioning: Defines
resource allocation to a slice in terms of a
fraction of the total base station resources.
This implies that the slice is guaranteed to
receive at least a fraction of the total spec-
trum resources. For instance, if two MNOs
share RAN resources, MNO A can reserve
60 percent of the total resources while
MNO B gets 40 percent of them.
•Bandwidth-based provisioning: Defines
resource allocation to a slice in terms of
Figure 1. Network virtualization substrate architecture illustration.
Frame scheduler
Downlink:
1. Choose slice
2. Choose flow
Uplink:
1. Choose slice
2. Choose flow
DL packets
NVS
DL flows
UL flows
Bandwidth requests
Slice reservations and utility functions
DL/UL flow scheduling model
Classifier
Slice 1 Slice 3
Slice 2
NVS operates at fine
time-scales in order
to fill up the
resources of a physi-
cal layer frame. At
every scheduling
instant, NVS selects a
slice that maximize
the overall utility or
revenue. Once a slice
is selected, NVS
schedules flows
within the slice
according to a cus-
tom slice policy.
PEREZ-COSTA LAYOUT_Layout 1 6/26/13 11:49 AM Page 31
IEEE Communications Magazine • July 2013
32
aggregate bandwidth (in megabits per sec-
ond) that will be obtained by the flows of
that slice.
For instance, a content provider like youtube
can reserve 2 Mb/s rate on an MNO’s network
so that its users receive good quality video even
during periods of congestion. Note that while
MNOs and MVNOs may prefer to share
resources based on resource provisioning, con-
tent or application service providers may prefer
to reserve bandwidth on the cellular network.
To satisfy both bandwidth- and resource-
based provisioning on the same base station,
NVS defines different utility functions for the
two provisioning schemes while ensuring that
they are comparable.
Assuming that G1and G2are the sets of slices
with bandwidth-based and resource-based provi-
sioning, respectively, NVS defines the following
utility functions:
(1)
(2)
where rg
rsv and rgdenote the minimum reserved
cumulative bandwidth and the achieved cumula-
tive bandwidth (units of megabits per second),
while tg
rsv and tgdenote the minimum reserved
cumulative resources and the achieved cumula-
tive resources (units of resource blocks). rg
eff
denotes the average effective transmission rate
of a slice gin terms of the resource blocks allo-
cated to it.
The slice utilities Ug(rg) and Vg(tg) are based
on SLAs between the slice owners and the physi-
cal network owner. NVS uses concave utility
functions to model the revenue generated from
the different slices. Specifically, NVS uses a log-
arithmic function of resources or bandwidth allo-
cated to the slice. This choice of utility function
assumes that the marginal utility of a slice
decreases as its achieved resources or bandwidth
increases. By employing weighted log-utility
functions, NVS effectively provides fairness
among slices that is similar to proportional fair-
ness after meeting the reserved bandwidth or
resources for each slice.
The slice scheduler is designed to maximize
the overall utility of the base station, while
ensuring the requirements of individual slices.
To achieve maximum utility, the slice scheduler
employs the following weight function and picks
the slice with the largest weight at every time
instant j.
(3)
where rg,j
exp and tg,j
exp are exponential moving aver-
ages of bandwidth and resource blocks that are
allocated to slice gat time j, respectively.
If there are unused resource blocks in the
system after the flow scheduling within the
selected slice, the slice with the second largest
weight at this time instant may be selected for
flow scheduling. This process continues until all
the resources are utilized or no slice remains. In
this way, the utilization of the base station can
be maximized.
In addition to the data traffic generated by
users, note that NVS also accounts for the con-
trol signaling generated by each user of a slice
when updating the cumulative resources tgor the
cumulative bandwidth rgof a slice. This ensures
that NVS provides isolation of resources for a
slice against slices that generate high amounts of
signaling (e.g., due to high user mobility and/or
increased transitions to and from idle mode).
Flow Scheduling Framework — NVS builds a
generic flow scheduling framework that facili-
tates slices to emulate a wide variety of flow
schedulers. NVS defines three modes of per-
forming flow scheduling for slices. Each mode is
designed with trade-offs between complexity and
flexibility.
Scheduler Selection — In this mode, NVS
provides several common flow schedulers already
implemented in the base station. A slice can
choose from one of these schedulers. While this
mode is simple, it is not suitable for evaluating
or employing more innovative schedulers.
Model Specification — In this mode, NVS
provides an interface for slices to specify a
weight distribution function to schedule the
flows of that slice. The weight distribution can
be a function of several parameters like average
rate, and modulation and coding scheme (MCS).
When a slice is scheduled, NVS chooses a flow
within this slice with the smallest weight and
then updates the weights of all flows. This mode
enables arbitrary weight-based schedulers.
Virtual-Time Tagging — In this mode, NVS
provides real-time flow-level feedback to a slice
to perform flow scheduling itself. It exposes
higher flexibility to the slice to implement its
own arbitrary scheduler, but it is also more com-
plicated as slices need to implement their own
scheduler.
Although NVS introduces a hierarchical
scheduler in the base station, the implementa-
tion is lightweight to ensure it is readily deploy-
able on today’s base station hardware. The NVS
scheduler is integrated into the current base sta-
tion flow scheduling framework. The implemen-
tation of NVS simply requires each flow to be
tagged by the slice ID to which it belongs, and at
every scheduling epoch, only those flows that
belong to the slice selected as per Eq. 3 are
scheduled.
EVALUATION
In our previous work [16], NVS was implement-
ed and evaluated on a picochip-based WiMAX
testbed as part of the MAC layer in software.
Several experiments were conducted that demon-
strated the isolation, customization, and utiliza-
tion efficiency of NVS in a real testbed, with a
special focus on bandwidth reservation. Due to
U=∈rr
r
rgG( ) log( ),
gg
g
g
g
rsv
eff 1
V=∈tt tgG( ) log( ),
gg g g
rsv 2
=
∈
∈
⎧
⎨
⎪
⎪
⎪
⎩
⎪
⎪
⎪
w
r
r
gG
t
t
gG
if
if .
gj
g
gj
g
gj
,
rsv
,
exp 1
rsv
,
exp 2
Although NVS intro-
duces a hierarchical
scheduler in the base
station, the imple-
mentation is
lightweight to ensure
it is readily deploy-
able on today’s base
station hardware.
The NVS scheduler is
integrated into the
current base station
flow scheduling
framework.
PEREZ-COSTA LAYOUT_Layout 1 6/26/13 11:49 AM Page 32
IEEE Communications Magazine • July 2013 33
hardware limitations, the experiments included
only a limited number of users, and no admis-
sion control mechanism was considered.
In this article, we evaluate an LTE system,
and implement and analyze NVS on a C++ sys-
tem-level simulator in a typical cellular network
scenario. We focus on the resource reservation
approach with an LTE resource block (RB) as
the minimum allocable resource unit. A hexago-
nal cell layout with seven sites and three sectors
per site is considered. The system bandwidth is
10 MHz consisting of 50 RBs. NVS is compared
with two baseline schemes. The interaction
between NVS and admission control is also eval-
uated.
Full Sharing — In the first set of simulations,
we compare NVS with a full spectrum sharing
(FS) scheme to illustrate the isolation and cus-
tomization capabilities of NVS. We set up each
base station with two slices. In the FS case
there is no resource reservation, and the whole
bandwidth is shared by all users, whereas in
the NVS case, slices 1 and 2 reserve 40 and 60
percent of the total RBs in the system band-
width, respectively. Ten video users and 20
FTP users are placed per sector in each slice.
Video users continuously stream an average
rate of 384 kb/s in the downlink. FTP users
have backlogged traffic also in the downlink. In
the FS case, a MAC scheduler based on con-
ventional proportional fairness without traffic
type prioritization is used. In the NVS case,
slice 1 runs a proportional fair MAC scheduler
with video users having strict priority over FTP
users, while for slice 2 the same scheduler as in
the FS case is used.
Figure 2a shows the mean throughput
received by the end users and the mean RB uti-
lization of both slices in the FS case. As expect-
ed, all users, regardless of their traffic type and
associated slices, receive a similar mean through-
put. Both slices also experience 50 percent RB
utilization on average. As a result, the video rate
requirements are not met for all users in both
slices. On the other hand, Fig. 2b shows that
with NVS, slices 1 and 2 receive 40 and 60 per-
cent of the total RBs, as configured, although
they have the same traffic load. Slice 1 users are
prioritized based on their traffic type according
to the customized slice scheduler satisfying the
video requirements, while in slice 2 the video
and FTP users receive similar throughput,
although the value is larger than the one in the
FS case since slice 2 has 60 percent of the total
RBs allocated.
Static Reservation — In the second set of sim-
ulations, we compare NVS with a static reserva-
tion (SR) scheme to illustrate the benefits of
NVS in terms of resource utilization. In SR each
slice is allowed to use only its reserved RBs. RBs
0–14 (30 percent) are permanently reserved for
slice 1, while RBs 15-49 (70 percent) are perma-
nently reserved for slice 2. In the NVS case, we
set up the base station with two slices such that
slices 1 and 2 are allocated 30 and 70 percent of
the total RBs in the system, respectively. In this
set of experiments, all users in slice 1 download
FTP files with a size of mean 250 kbytes, while
all users in slice 2 stream video at an average
rate of 128 kb/s in the downlink that last for a
duration of 10 s on average. Following common
practice in commercial cellular networks, FTP
requests are always admitted regardless of the
load conditions, while RB-usage-based admis-
sion control (AC) is used in slice 2 to support
the QoS of the video users. For illustration pur-
poses, the AC threshold is set to 80 percent of
the RBs. In both slices a proportional fair MAC
scheduler is used.
In Fig. 3a the mean RB utilization of slices 1
and 2 is shown with respect to an increasing
offered traffic load for both NVS and SR
schemes. As the load increases, slice 1 with SR
reaches saturation first as its mean RB utilization
is upper bounded to 30 percent of the system
capacity. In contrast to this, since NVS can
opportunistically allocate the unused resources in
slice 2 to slice 1, slice 1 reaches an RB utilization
above 30 percent as long as there are unused
resources in slice 2. Once slice 2 reaches satura-
tion at 56 percent, both SR and NVS provide
resources per slice as configured. NVS reaches
higher system throughput than SR due to its
capability to allocate the free resources in slice 2
Figure 2. NVS isolation and customization features: a) full sharing (FS); b)
network virtualization substrate (NVS).
Video (1)
100
0
Mean user throughput (kb/s)
Mean RB utilization (%(
200
300
80
100400
60
40
20
0
Mean user throughput
Mean RB utilization
Video (2) FTP (1) FTP (2)
(a)
SLICE 1 SLICE 2
Video (1)
100
0
Mean user throughput (kb/s)
Mean RB utilization (%)
200
300
80
100400
60
40
20
0
Mean user throughput
Mean RB utilization
Video (2) FTP (1) FTP (2)
(b)
SLICE 1 SLICE 2
PEREZ-COSTA LAYOUT_Layout 1 6/26/13 11:49 AM Page 33
IEEE Communications Magazine • July 2013
34
(20 percent blocked for video traffic due to the
AC mechanism) to FTP traffic from slice 1.
SUMMARY AND CONCLUSIONS
The virtualization of mobile carrier networks is
an appealing opportunity for operators to handle
the forecasted “mobile data apocalypse” while at
the same time increasing the return on invest-
ment in CAPEX and OPEX infrastructure costs.
In this article, we first review 3GPP’s network
sharing standardized functionality followed by an
overview of the RAN sharing enhancements cur-
rently being discussed in the 3GPP RSE Study
Item based on emerging business models. Then
a summary of the state of the art of mobile carri-
er network virtualization research is provided
with a special focus on RAN sharing, base sta-
tion programmability, and customization.
The feasibility of mobile network virtualiza-
tion is analyzed by presenting a RAN sharing
technique meeting the required isolation, net-
work utilization, and customization virtualization
needs, NVS, which can be natively implemented
in base stations. LTE performance results are
presented and benefits of the proposed approach
discussed.
Based on the business trends reported in this
article, the corresponding 3GPP RAN sharing
enhancements efforts and research results in the
area, we conclude that future mobile networks
will increasingly include advanced virtualization
solutions, opening the market to a wide range of
new business models building on the enhanced
capabilities of such virtualized mobile carrier
networks and likely to be exploited by emerging
wireless SDN frameworks.
REFERENCES
[1] Cisco Visual Networking Index: Global Mobile Data Traf-
fic Forecast Update, 2012–2017, http://www.cisco.com/.
[2] 3GPP TS 23.251, “Network Sharing; Architecture and
Functional Description,” v. 11.3.0, 2012.
[3] Mobile Network Sharing Report 2010-2015, Develop-
ment, Analysis & Forecasts, Market Study, Visiongain,
2010.
[4] H. Guan et al., “Discovery of Cloud-RAN,” Cloud-RAN
Workshop, 2010.
[5] “Active RAN Sharing Could Save $60 Billion for Opera-
tors,” http://www.cellular-news.com/story/36831.php.
[6] 3GPP TS 22.951, “Service Aspects and Requirements for
Network Sharing,” v. 11.0.0, 2012.
[7] 3GPP TR 22.852, 3GPP System Architecture Working
Group 1 (SA1) RAN Sharing Enhancements Study Item.
[8] Telecoms.com Intelligence, Industry Survey, 2013.
[9] I. Stoica et al., “A Hierarchical Fair Service Curve Algo-
rithm for Link-Sharing, Real-Time, and Priority Services,”
IEEE/ACM Trans. Net. J., 2000.
[10] Vanu Networks, http://www.vanu.com/.
[11] S. Paul and S. Seshan, “Virtualization and Slicing of
Wireless Networks,” GENI Design Doc. 06-17, 2006.
[12] G. Bhanage et al., “Virtual Base Station: Architecture
for an Open Shared WiMAX Framework,” Proc. ACM
SIGCOMM VISA, 2010.
[13] Y. Zaki et al., “LTE Wireless Virtualization and Spec-
trum Management,” Proc. IFIP WMNC Conf., Oct 2010.
[14] EU FP7 FLAVIA Project, http://www.ict-flavia.eu/.
[15] X. Costa-Pérez et al., “Latest Trends in Telecommuni-
cation Standards,” ACM Comp. and Commun. Review,
Apr. 2013.
[16] R. Kokku et al., “NVS: A Substrate for Virtualizing
Wireless Resources in Cellular Networks,” IEEE/ACM
Trans. Net., vol. 20, issue 5, 2012.
BIOGRAPHIES
XAVIER COSTA-PÉREZ [M] (xavier.costa@ieee.org) is chief
researcher at NEC Laboratories, where he has managed
several projects related to mobile networks. In the wireless
LAN area, he led a project contributing to 3G/WiFi mobile
phones evolution and received NEC’s R&D Award for his
work on N900iL, NEC’s first 3G/WiFi phone. In the 4G area
he managed a team researching base station enhance-
ments and recently received NEC’s R&D Award for success-
ful technology transfers. In 3GPP he is contributing to the
SA1 RSE Study Item where new requirements for future
systems are being defined. He has served on the Program
Committees of several conferences, including IEEE Green-
com, WCNC, and INFOCOM, and holds over 20 patents.
He received both his M.Sc. and Ph.D. degrees in telecom-
munications from Universitat Polite`cnica de Catalunya
(UPC) and was the recipient of the national award for the
best Ph.D. thesis on multimedia convergence in telecom-
munications.
JOERG SWETINA (joerg.swetina@neclab.eu) studied chemistry
and mathematics at the University of Vienna, Austria, and
later conducted research in theoretical chemistry. Moving
from academia to industry, he led a development team
dealing with GSM call processing and software testing at
Siemens Austria. Since the early days of 3GPP he represent-
ed Siemens and later Nokia Siemens Networks in standard-
ization bodies like ETSI SMG, 3GPP, and OMA. In 2008 he
Figure 3. NVS resource utilization features: a) mean RB utilization; b) system
throughput.
Offered load (Mb/s/cell)
(a)
20
10
0
Mean RB utilization (%)
20
30
40
50
60
4 6 8 10 12 14 16
Offered load (Mb/s/cell)
(b)
20
2
0
System throughput (Mb/s/cell)
4
6
8
10
4 6 8 10 12 14 16
SR (SLICE 1)
NVS (SLICE 1)
SR (SLICE 2)
NVS (SLICE 2)
SR
NVS
PEREZ-COSTA LAYOUT_Layout 1 6/26/13 11:49 AM Page 34
IEEE Communications Magazine • July 2013 35
moved to NEC Europe, Heidelberg, Germany, continuing
his work in standards. In addition to 3GPP, his current field
of interest covers machine-to-machine communication. He
is active in ETSI TC M2M and is currently acting as Vice
Chair of the requirements group of the “oneM2M Global
Initiative” organization.
TAO GUO (tao.guo@emea.nec.com) received his B.Sc.
degree in information engineering from Xi’an Jiaotong Uni-
versity, China, in 2004, his M.Sc. degree with distinction in
communications engineering from the University of Birm-
ingham, United Kingdom, in 2005, and his Ph.D. degree in
wireless networking from Newcastle University, United
Kingdom, in 2009. In May 2009 he joined the University of
Surrey, United Kingdom, as a research fellow working on
various EU projects and the Mobile VCE project. Since June
2012 he has been with NEC Telecom Modus Ltd as a senior
systems engineer working on the development of the 3GPP
LTE system. His current research interests focus on radio
resource and mobility management, self-organizing net-
works, and network virtualization.
RAJESH MAHINDRA (rajesh@nec-labs.com) received his M.S.
degree from the School of Electrical and Computer Engi-
neering, Rutgers University, New Jersey, and his B. Tech.
degree from the Department of Electronics and Communi-
cations Engineering, M. S. Ramaiah Institute of Technology,
Bangalore, India, in 2005. He is currently working as a
senior associate research staff member at NEC Labs Ameri-
ca. His research interests include video streaming over
wireless, wireless network virtualization, and wireless
resource management with focus on next-generation cellu-
lar networks.
SAMPATH RANGARAJAN [SM] (sampath@nec-labs.com)
received his M.S. degree in electrical and computer engi-
neering and Ph.D. degree in computer sciences from the
University of Texas at Austin. He heads the Mobile Commu-
nications and Networking Research Department at NEC
Laboratories America in Princeton, New Jersey. Previously,
he was a researcher at Bell Laboratories in New Jersey. He
was also a co-founder and vice president of technology at
Ranch Networks, Morganville, New Jersey, a venture-fund-
ed startup in the IP networking space. Before joining Bell
Laboratories, he was an assistant professor with the Electri-
cal and Computer Engineering Department, Northeastern
University, Boston, Massachusetts. His research interests
span the areas of mobile communications, mobile net-
works, and distributed systems. He has been on the Edito-
rial Boards of IEEE Transactions on Computers, IEEE
Transactions on Parallel and Distributed Systems, and ACM
Mobile Computing and Communications Review.
PEREZ-COSTA LAYOUT_Layout 1 6/26/13 11:49 AM Page 35
