Received October 1, 2014, accepted October 21, 2014, date of publication December 18, 2014, date of current version

January 13, 2015.

Digital Object Identifier 10.1109/ACCESS.2014.2381215

Energy Efficiency Benefits of RAN-as-a-Service

Concept for a Cloud-Based 5G Mobile
Network Infrastructure
DARIO SABELLA1 , ANTONIO DE DOMENICO2 , EFSTATHIOS KATRANARAS3 , (Member, IEEE),
MUHAMMAD ALI IMRAN3 , (Senior Member, IEEE), MARCO DI GIROLAMO4 , UMER SALIM5 ,
MASSINISSA LALAM6 , KONSTANTINOS SAMDANIS7 , AND ANDREAS MAEDER7
1 Telecom Italia, Turin 10148, Italy
2 Laboratory for Electronics and Information Technology, Atomic Energy Commission, Grenoble 38054, France
3 University of Surrey, Guildford GU2 7XH, U.K.
4 HP Italy Innovation Center, Milan 20003, Italy
5 Intel Mobile Communications, Paris 75017, France
6 Sagemcom Broadband, Paris 92848, France
7 NEC Laboratories Europe, Heidelberg 69115, Germany

Corresponding author: D. Sabella ([email protected])

This work was supported by the European Community’s 7th Framework Programme (FP7/2007-2013) under
Grant 317941 through the iJOIN Project.

ABSTRACT This paper focuses on energy efficiency aspects and related benefits of radio-access-
network-as-a-service (RANaaS) implementation (using commodity hardware) as architectural evolution of
LTE-advanced networks toward 5G infrastructure. RANaaS is a novel concept introduced recently, which
enables the partial centralization of RAN functionalities depending on the actual needs as well as on network
characteristics. In the view of future definition of 5G systems, this cloud-based design is an important
solution in terms of efficient usage of network resources. The aim of this paper is to give a vision of the
advantages of the RANaaS, to present its benefits in terms of energy efficiency and to propose a consistent
system-level power model as a reference for assessing innovative functionalities toward 5G systems. The
incremental benefits through the years are also discussed in perspective, by considering technological
evolution of IT platforms and the increasing matching between their capabilities and the need for progressive
virtualization of RAN functionalities. The description is complemented by an exemplary evaluation in terms
of energy efficiency, analyzing the achievable gains associated with the RANaaS paradigm.

INDEX TERMS Energy efficiency, wireless communication, radio access networks, RAN-as-a-service,
power model, cloud-RAN, LTE-advanced, 5G.

I. INTRODUCTION of Operational Expenditure (OPEX) and its relevance in

The current vision towards 5G is often driven by traffic upcoming systems will be somehow proportional to network
forecasts that suggest increasing data volumes, number of growth. In fact, energy consumption at network level depends
more intelligent terminals and an ever growing capacity and on large part on the number of installed radio base stations.
service-aware demand. This discussion is also accompanied In addition to this essential densification of network nodes,
by the fundamental question related to the need to define a integrating new systems upon the existing ones unavoidably
new air interface or consider 5G as an evolution of current increases the energy consumption, even if new systems are
systems. more efficient than the old ones (this happened by adding
Launching a higher number of base stations, i.e. intro- LTE on top of 2G/3G). Increased energy consumption means
ducing traditional network densification, may provide a higher costs and a greater carbon footprint, since today mobile
solution for such a tremendous traffic increase, but would systems are present everywhere in the world. The European
result in increased network infrastructure costs, which are Commission (EC) recognized the need for further actions
expected to see a consequent explosion through the years. towards energy efficiency and green communications and
In particular, energy consumption is an important part introduced the Code-of-Conduct [1] to provide a policy

2169-3536 2014 IEEE. Translations and content mining are permitted for academic research only.
1586 Personal use is also permitted, but republication/redistribution requires IEEE permission. VOLUME 2, 2014
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
D. Sabella et al.: Energy Efficiency Benefits of RANaaS Concept

that regulates energy consumption and carbon dioxide network EE is still an open research field with GreenTouch
emissions. recently announcing that the potential of reducing the net
As a consequence, Energy Efficiency (EE) and sustainabil- energy consumption may reach up to 90% by 2020 [6].
ity of 5G networks have recently received significant attention However, a main challenge is to elaborate realistic and
from mobile operators, vendors and research projects [2]. complete models of innovative solutions for new generation
Figure 1 shows our vision of the EE evolution in mobile networks, in order to correctly influence the definition of
networks toward a sustainable 5G, where the exponential affordable and sustainable 5G systems.
mobile traffic growth toward 2020 (blue curve) goes with In [7], the EE benefits of a centralized RAN are analyzed in
a stable network energy consumption (red curve), resulting terms of hardware design, i.e., cooling and soft technologies
in an increasing EE of the system through the years including cooperative processing, virtualization and dynamic
(green curve). cell re-configuration. In addition, signaling and control opti-
mizations are explored to move away from the conventional
connection-oriented paradigm, i.e., the use of ‘‘virtual’’ eNBs
with irregular antenna arrays and new interference mitigation
schemes.
A more flexible cloud-based RAN architecture is pro-
posed in [8], where the front-haul is logically re-adapted
to the corresponding requirements. Such cloud-RAN archi-
tecture enables energy saving on the cloud-RAN platform,
by re-arranging the number of active Base Band processing
Units (BBUs), when the traffic is low.
More recently, the concept of RAN-as-a-Service (RANaaS)
has been introduced in [9], where some RAN functionalities
are partially and flexibly centralized, depending on the actual
load and network characteristics. In the view of future defini-
tion of 5G systems, these cloud-based designs together with
RAN sharing mechanisms appear to be the most promising
solutions in terms of efficient usage of network resources.
The aim of the paper is to present a vision of the advantages
of the RANaaS paradigm and its benefits in terms of EE.
FIGURE 1. Energy efficiency evolution in mobile networks toward a
sustainable 5G. To do so, we will introduce a consistent system-level power
consumption model, which is intended to be a useful refer-
From the standardization point of view, EE in wireless ence for the EE performance evaluation of innovative RAN
systems mainly concentrates on the efforts of 3GPP for LTE mechanisms. The rest of the paper is organized as follows.
and LTE-Advanced (LTE-A) [3]. The 3GPP Radio Access Section II describes the proposed mobile access architecture.
Network (RAN) groups were the first to deal with EE in Section III gives an overview of current status and future
Release 9 by investigating energy–aware network manage- trends of IT platforms, with a description of related per-
ment with small cells, allowing LTE base stations (called formances offered for hosting RANaaS entities. Section IV
eNBs) to switch-off, based on local load information and con- provides a detailed analysis of the proposed architecture from
figuration data or via the Operation and Management (OAM). an EE perspective, with a derivation of a comprehensive
The RAN groups further introduced the ‘‘cell DTX’’ mech- power model for each involved entity. Section V gives an
anism, which deactivates the transmitter of an eNB based on exemplary description of small cell management using the
certain patterns, conserving energy on the fraction of inactiv- RANaaS paradigm, as a promising solution for future sus-
ity time [4]. Energy saving management was later suggested tainable 5G systems. Finally, Section VI concludes the paper.
by the Service Architecture Group 5 (SA5) that introduced
Self-Organizing Network (SON) operations in relation with II. REFERENCE SCENARIO AND ARCHITECTURE
overlaid and dense urban networks, wherein certain eNBs In mobile communications, the trend has always been to push
may compensate in terms of coverage and service allowing the computation burden toward the last miles, to reduce the
selected eNBs to be powered-off [5]. round-trip-time and improve the system reactivity. With dense
Besides 3GPP, the effort of the research community in small cells, coordination is needed again to deal with the
this topic is also notable. Current activities mainly focus high level of interference introduced due to the proximity of
on evolving flexible hardware for enhancing eNBs, novel base stations. To cope with this requirement, we consider in
architectures based on small cells deployment and adaptive Figure 2 an evolutionary architecture of the LTE-A one [10],
schemes that adjust network capacity with respect to service where RAN functions associated with small cells (iSCs)
loads [2]. Nevertheless, while research on green communi- can be centralized in a flexible manner. Such function
cations has produced notable results, improving the wireless centralization reduces the processing on radio access, where

VOLUME 2, 2014 1587

 D. Sabella et al.: Energy Efficiency Benefits of RANaaS Concept

FIGURE 2. Mobile network architecture implementing the RANaaS concept [9].

iSCs with low transmission power are used to allow high III. EVOLUTION OF CLOUD INFRASTRUCTURES
data-rates and enhanced EE compared to macro cells. AND RELATED PERFORMANCE
Moreover, to jointly optimize RAN and backhaul, The cloud computing model is the computational paradigm
a Network Controller (NC), i.e., a software defined network- which is substantially marking the latest and future era in
based controller, configures the routing among the Transport the IT realm. According to one of the most recognized
Nodes (TNs) based on the associated constraints such as definitions, the one elaborated by NIST [11], cloud
RAN/backhaul load, user density and mobility. computing is:
The purpose of the RANaaS, based on generic data centres, ‘‘. . . a model for enabling convenient, on-demand
is to centrally execute part of the RAN functionalities, thus network access to a shared pool of configurable comput-
benefitting from centralization gains, which are fundamental ing resources (e.g., networks, servers, storage, applica-
in ultra-dense deployments. In practice, the RANaaS and iSCs tions, and services) that can be rapidly provisioned and
entities appear as classical eNBs to the existing network. released with minimal management effort or service provider
Therefore, such ‘‘virtual eNB’’ (veNB) entity can be seam- interaction. . . ’’
lessly integrated in the legacy architecture. Although standard cloud computing is not the only possible
Under such setup, the power consumption of the veNB way to implement the RANaaS model described in chapter II,
will take into account the iSCs, the backhaul network, and it is the most obvious ‘‘first option’’ to take into account,
the RANaaS. At a first glance, it appears that this power since its characteristics (like full transparency of underly-
consumption will be much higher than those required by ing physical resources, potential portability and dynamic
a classical deployment that covers the same area, but this scalability) are optimally fitting the main drivers behind RAN
will have to be evaluated relatively to the effective gain on virtualization and centralization. On the other hand, sticking
the system performance as well. For instance, this architec- to an industry standard model like cloud computing imposes
ture is potentially enabling advanced RAN sharing solutions some ‘‘overheads’’ requested to preserve generality, which
that may dramatically improve the system performance. can have an impact on the overall performance and ultimately
Accordingly, EE evaluations are traditionally performed by on the feasibility of RAN centralization. Assessing such
considering two kinds of metrics (and variations of them) [2]: constraints and limitations is among the key objectives of
Energy per Information Bit [J/bit] and Power per Area research initiatives underlying what is discussed in the current
Unit [W/m2 ]. paper.

1588 VOLUME 2, 2014

D. Sabella et al.: Energy Efficiency Benefits of RANaaS Concept

FIGURE 3. Evolution of computational power (GFLOPS) (left) and Energy Efficiency (Watts/GFLOPS) (right).

The RANaaS concept sets challenging requirements right-side of Figure 3 depicts energy cost per billion of
to the hardware (and software) infrastructure on which FLOPS (FLoating point Operations Per Second) and shows
the RAN functions have to be outplaced from standard how deployment of huge servers, realizing trillions of
LTE (LTE-A) base stations. On one hand, the infrastruc- operations, is more and more becoming financially and
ture should be based on industry standard servers (ISS) environmentally sustainable. These aspects are also at some
and computational equipment (such as storage and network extent interrelated, e.g., a reduced footprint can also help to
appliances), to keep a reasonable cost balance with respect save power.
to standard base stations: paying a financial penalty would The current standard generation mostly employed in large
override one of the key reasons to go after centralization, data centres is based on X86 servers, in rackable or blade
i.e. CAPEX reduction. On the other hand, commoditization form factor. However, the ISS technology is undergoing
must not hinder the minimal performance requested to sustain a breakthrough transformation, which is bringing towards
the performance requirements of the processing and man- much more powerful servers with a far smaller form factor
agement functions swapped to the RANaaS layer. Finally, and a previously unknown reconfigurability level: Hyperscale
centralization must be kept under control from an EE stand- servers have larger computational power, along with a high
point, to make sure that the centralization gains of functional degree of scalability, vertical or horizontal (this latter typical
split do not introduce a total excessive energy consumption enabler of the cloud computing model). Also, new processor
increase at overall system level. families like ARM or Atom are starting to be employed inside
Accordingly, one of the key objectives of our research ISS, hugely increasing the computational power density per
is finding out a set of minimal conditions that guarantee square meter and, thus, reducing the facility costs of a data-
performance and efficiency goals at once in acceptable extent. centre.
One of such assessments concerns the RANaaS hardware Finally, these new systems implement the software defined
and software infrastructure: ISS have gone through a relevant server concept to optimize their configuration to the peculiar-
evolution over the last years, affecting their performance, ities of the workload to be executed on them. For instance,
the performance/footprint ratio and the EE. Figure 3 gives HP Moonshot chassis are 7.5’’ tall and can host 45 hot-plug
a statistical visual idea of such trend, considering the top independent server cartridges, sharing power, cooling and
500 supercomputers. Supercomputers are not the actual management resources [12]. Server cartridges are based on
target of RAN centralization, for the reasons exposed Intel Atom processors and are design-optimized for specific
above orienting the research towards ISS equipment. classes of applications. A 45-server chassis shows a power
However, ISS servers are not benchmarked versus low-level consumption of around 850W, able to power 180 × 2.0 GHz
operations/second Key Performance Indicators (KPIs); threads, with 2GB of RAM for each thread, at under 5W per
their standard industry benchmarks are executed through thread. Cartridges can be optimized by adjusting local mass
specific applications, not easily relatable to RAN func- storage capacity on each module and by experimenting with
tions. Nevertheless, the figure trend can very well chassis-level network switch configurations, for instance to
approximate a similar evolution undergone in the ISS focus on front-end performance. With similar server archi-
domain. tectures, it is easy to envision an upcoming availability of
In detail, the left-side of Figure 3 fully captures standard servers optimized on a specific problem domain like
the growth in computational power. On the other hand, RAN processing.

VOLUME 2, 2014 1589

 D. Sabella et al.: Energy Efficiency Benefits of RANaaS Concept

Clearly, this breakthrough passage in the ISS technology this study, the maximum power consumption of a single
has relevant impacts on the applicability and actual viability eNB sector can be given by [14]:
of the RANaaS concept. The above aspects drive a bulky
cost reduction of computational power in data centres, on one PBB + PRF + PPA
PeNB = (2)
hand filling the performance gap existing between custom (1 − σDC )(1 − σMS )(1 − σCO )
components and ISS servers, on the other cutting costs and
where PBB , PRF , PPA and σDC , σMS , σCO denote the power
making an ISS based infrastructure really competitive versus
consumption and loss factors, respectively, of the differ-
legacy eNBs.
ent components. Furthermore, it was found that an affine
function of the transmission power (comprising a static
IV. POWER CONSUMPTION MODEL FOR
and a linearly increasing load-dependent share) can approx-
CLOUD-BASED MOBILE NETWORKS
imate very well the generalized model of eNBs [15],
While the traditional architecture considers several sophis-
i.e.,:
ticated small cells, our reference system is composed
by several iSCs of low processing capability and a 
P0 + y · 1p · Pmax if 0 < y < 1
RANaaS platform where pooling of complex (e.g. baseband) PeNB = (3)
Psleep if y = 0
processing can be performed. In order to investigate the
convenience (from an energy performance perspective) of where P0 , Pmax and Psleep stand for the eNB power con-
the proposed architecture compared to the traditional one, we sumption at zero load, full load and sleep mode (considering
need to introduce a generalized holistic power model; in fact, that eNBs may enter a low consumption sleep mode where
power consumption at system level should be evaluated by some of their main units are turned off when no data is
considering the sum of all power consumption contributions received or transmitted), respectively. Furthermore, y and 1p
in the network. This will help us to perform a quantitative denote the cell load and the slope of the load-dependent linear
analysis on the power consumption of the RANaaS concept in model.
cloud-based mobile networks and evaluate the potential bene- To approximate iSC power consumption, we have adopted
fits in terms of energy efficiency in different load conditions. and combined the aforementioned models and we have taken
Considering the system architecture as introduced in into account that: 1) the BB and RF power consumptions scale
Section 2, comprising NiSC iSCs in total, its overall power linearly with the system bandwidth (W ) and the number of
consumption can be modelled by: antennas (Nant ) used [16], 2) PA power consumption can be
NiSC
approximated as a linear function of the PA output power,
X
PTotal = PRANaaS + PBh + PiSC-n (1) and 3) no active cooling is needed in that case. Moreover,
n=1 for simplicity, we have considered that only the PA is turned
off when an iSC goes to sleep mode (i.e. no BB engine
where PRANaaS , PBh , and PiSC-n stand for the power
reductions due to sleep mode are considered). Therefore, the
consumed at the RANaaS platform, the power needs for
power consumption of an iSC n with cell load yn can be given
backhauling network, and the power usage at any iSCn n,
by:
respectively.
In the following, the power consumption of each individual W
Nant 10[MHz] P0BB + P0RF + yn PPA-max


network element is discussed. Furthermore, some examples PiSC-n = (4)
of measures are provided to correlate and obtainan idea on (1 − σDC ) (1 − σMC )
the order of magnitude of each element’s power consumption,
where P0BB and P0RF denote the BB and RF base consumptions
depending on the cells’ load (which is interrelated to the cells’
(i.e. using 10MHz and one antenna) while PPA-max is the
RF output power).1
PA maximum transmission power.
In general, we consider that an iSC may implement only
A. iSC POWER CONSUMPTION
a portion of the eNB protocol stack. Therefore, its power
The FP7 EARTH has investigated how the power consump- consumption will be lower and upper bounded by the two
tion of distinct components of several eNBs, such as power extreme cases: a) Radio Remote Head (RRH) and b) baseline
amplifier (PA), baseband engine (BB), Radio Frequency (RF) Small Cell, respectively. RRHs are considered as low com-
small-signal transceiver, direct-current (DC)-DC converter, plexity and processing nodes that solely perform RF opera-
main supply (MS), and active cooling (CO), depends on tions and rely on self-backhauling (i.e. PBB = 0 in that case).
the transmission bandwidth, the transmission power, and On the other hand, baseline small cells perform all the BB
the number of radio chains/antennas [13]. According to operations.
1 In addition, this paper has supplementary downloadable material Table 1 provides exemplary realistic values for the param-
available at http://ieeexplore.ieee.org, provided by the authors. This includes: eters associated to the iSC power model, considering two
a multimedia video endorsing the importance of energy efficiency assessment (per-antenna) maximum transmit power, i.e., 24 dBm (PTx,1 )
applied to architectures based on the RANaaS concept, the Excel implemen-
tation of the overall power consumption model for RANaaS-enabled mobile and 30 dBm (PTx,2 ), and Figure 4 depicts the estimate power
networks, and a readme file. This material is about 2 MB in size. consumption of iSCs.

1590 VOLUME 2, 2014

D. Sabella et al.: Energy Efficiency Benefits of RANaaS Concept

TABLE 1. iSC power consumption - exemplary realistic parameter values [13].

where Psrv srv

0 and Pmax denote the power consumption of the
server when in idle mode and maximum usage, respec-
tively, while 1srv
p stands for the slope of the equivalent
linear power model which depends on the specific server
considered.
Considering the RANaaS BBU as an enclosure hosting
several identical interconnected ISS Blade servers (each con-
sidered as a set of multiple processors) equally sharing the
requested workload, the servers’ processing capacity (XCap ),
in Giga-Operations-per-Second3 (GOPS), will define the total
number of servers (Nsrv ) required to process the system
BB-related workload (X in GOPS) moved to RANaaS
BBU, i.e.:
 
X
Nserver = (6)
FIGURE 4. Complete small cell and RRH power consumption with respect XCap
to different RF output power and constraints.
and the percentage CPU usage at each server will be:
X
B. RANaaS PLATFORM POWER CONSUMPTION xserver = · 100% (7)
XCap Nserver
In order to obtain an accurate estimation on the power con-
sumption of the RANaaS platform due to BB processing2 The question that arises next is how the extra RANaaS BBU
moved from iSCs, we make use of the commodity hard- workload can be related to cell load. In that regard, the work
ware consumption coming from the IT world; Fit4Green has in [16] considered the functionalities of various base station
investigated the power consumption for IT resources of data types and examined how the GOPS per function block scale
centres [17]. In particular, results for the various computing with cell load for a specific reference system. Targeting to a
style servers are provided using a monitoring tool and a more generalized view, Werthman et al. defined recently the
generic power consumption prediction model. Considering resource effort required to serve a UE at a specific time as
the measurements results for the cloud computing testbed a function of the number of used antennas, the modulation
hardware equipment as the closest paradigm for the BBUs at bits, the code rate, the number of spatial MIMO-layers and
a RANaaS platform, it can be observed that a linear model can the number of allocated frequency/time resources each as
approximate well a server’s power consumption (Psrv ) versus allocated to the UE at that time [18]. Based on this work,
its percentage CPU usage (xsrv %) we introduce an average sum to approximate the total extra
RANaaS BBU workload required to serve all UEs when an
average βBB % of BB processing is assumed to be moved
0 + 1p Pmax xsrv
Psrv = Psrv srv srv
(5)
towards RANaaS platform from each iSC. To this end, the
GOPS required at RANaaS platform will depend on the total
2 Note that for fair comparison, we consider as power consumption of number of iSCs in the system, the load of each iSC, the
the RANaaS platform only the extra power needed for computations due to
BB processing moved to RANaaS from iSCs (e.g at a RANaaS BBU). 3 It is noted that the processing capacity of the server is usually expressed
Other functionalities at RANaaS platform can be assumed that exist also at in GFLOPS [19]; however, it can be converted in GOPS, and in this work we
the core network of a conventional system. use a 1:1 ratio as a conservative estimation.

VOLUME 2, 2014 1591

 D. Sabella et al.: Energy Efficiency Benefits of RANaaS Concept

TABLE 2. RANaaS platform power consumption - exemplary realistic parameter values.

FIGURE 5. RANaaS power consumption with respect to the small cell RF output power for different BB shift options.

system bandwidth, the number of antennas used to serve a due to BB processing moved towards RANaaS platform will
UE per iSC (NTx ), the average number of data bits per symbol be the sum of the power consumption at each of the required
per user (eMSC ) and the average number of MIMO layers per servers (9), as shown at the bottom of this page.
user (eMIMO ): To obtain an exemplary view on the power consumption
βBB W  eMSC  of RANaaS platform we consider a cloud computing server
2
X = · 30NTx + 10NTx + 20 eMIMO comprising Intel Xeon E5540 processors as presented in [17]
100 10 [MHz] 6
under the high efficiency configuration policy (i.e. CPU
NiSC
X frequency always set to maximum value). Table 2 intro-
× yn (8)
duces the respective realistic values and Figure 5 illustrates
n=1
the power consumption of this specific RANaaS BBU with
Therefore, for any RANaaS BBU comprising servers with respect to the small cell RF output power for different
specific processing capacity, its overall power consumption BB shift.

 
NiSC W 2 + 20 eMSC e
X (y)  srv X


· 30NTx + 10NTx
 
MIMO
PRANaaS = · P0 + yn · 10MHz 6
1srv
p Pmax .
srv 
(9)
X (y)
l m
XCap XCap XCap
n=1

1592 VOLUME 2, 2014

D. Sabella et al.: Energy Efficiency Benefits of RANaaS Concept

TABLE 3. Backhauling network power consumption - exemplary realistic parameter values.

As we can see in Figure 5, RANaaS power consumption maximum capacity (Ymax in Gbps):
trend depends on the particular functional split considered  n
(and the consequent percentage of BB processing moved in  0, Nmw = 1

the RANaaS). Linear increase of the power consumption is Pnswitch = l or yn = 0 m (11)
 Ps yn Ymax fcell-Bh , otherwise

obtained (with 10% BB in RANaaS) when the processing Cswitch
capacity of a single server is sufficient to handle all load where Ps stands for the switch basic power consumption and
in the small cell. When more computational effort is moved fcell-Bh is a factor denoting the relationship between backhaul
in the RANaaS (e.g. 50%, 100%), the power consumption traffic load and cell load.
trend shows some jumps due to the fact that more servers Moreover, the power consumption for transmitting and
are required to handle the processing workload moved at the receiving the aggregate backhaul traffic will generally depend
RANaaS as the cell load and % shift increases. on the traffic conditions. In this work, we consider that back-
haul links can have an idle mode (i.e. when no data needs
C. BACKHAUL POWER CONSUMPTION to be transported through backhaul) and a two-step function
The last important element that we have modelled is the back- (low/high capacity traffic), where the two capacity regions are
haul network. In general, centralised systems have notable distinguished by a single threshold (Cthr ):
backhaul load; therefore, power consumption due to data
 Pidle , yn = 0

transport and switching can become a significant percentage

C
of the total system power consumption [2]. Plink = Plow-traffic , 0 < yn ≤ Ymax fthr
n
cell-Bh
(12)
Cthr
high-traffic , yn ≥ Ymax fcell-Bh

P
Monti et al. provided some fundamental power consump-
tion models for data transport through various backhaul Table 3 presents the backhaul relevant parameters with
technologies and topologies in small cells [20]. Considering exemplary realistic values. Note that backhaul traffic load
self-backhauling iSCs with microwave links and omitting is translated into cell load considering current LTE-based
TNs for simplicity, the backhaul power consumption can be RAN. For this, we consider the iSC maximum bits-per-second
estimated by adjusting the respective model in [20] to our capacity (evaluated assuming a single carrier with 10MHz
system architecture paradigm. In our case, backhaul power bandwidth, 2 × 2 MIMO, 64QAM, and 28% control over-
consumption shall scale with the number of iSCs in the head) and the non-negligible overheads from X2 U- and
system, the load of each iSC, the number of microwave C-plane (∼4%), the transport protocol (∼10%), and the IPsec
n ), the power for transmitting and
antennas at each iSC n (Nmw (∼14%) [21]. Finally, Figure 6 depicts the estimate backhaul
receiving the aggregate backhaul traffic at each iSC (Pnlink ) power consumption where we can see that for generic small
and the power consumption of switches at each iSC (Pnswitch ) cells, the backhaul always operates in low capacity region,
aggregating traffic from other iSCs in case more than one which results in flat power consumption for medium/high cell
backhaul link originates at the reference iSC, as: RF output power.

N
XiSC V. RANaaS-BASED OPTIMAL CONTROLLER FOR
PBh = Pnswitch (yn ) + Nmw
n n
Plink (yn ) (10) ENERGY EFFICIENT SMALL CELLS
n=1 In this section, the proposed power model is used to evaluate a
mechanism that manages the activity of NiSC iSCs according
Note that the power consumption at any switch will depend to access characteristics and the Quality of Service (QoS).
on the aggregated traffic at the associated iSC and its We consider that part of the small cell functionalities, i.e.

VOLUME 2, 2014 1593

 D. Sabella et al.: Energy Efficiency Benefits of RANaaS Concept

decided action changes the current status from st to st+1 ,

with a state transition probability T(st+1 |st ; at ), and incurs
in a cost Ct = C(st , at ). Our goal is to find an optimal
policy π that associates an action at (st |π ) to the state st .
This policy aims to minimize the energy consumption and
satisfying the QoS constraints while avoiding concurrent
access of multiple iSCs. This problem can be represented as
follows:
• a set of states S;
• a set of actions A;
• a state transition probability T ;
• a cost function C.

FIGURE 6. Backhauling Network Power consumption with respect to the

The statistics of the data rate and the incoming traffic are mod-
small cell RF output power. elled as Markov chains, independent of the time steps, buffer
state and actions. Hence, the model is Markovian because the
state transition depends only on the current status/actions and
is independent of any previous states/actions. Furthermore,
the Radio Resource Control, is shifted to the RANaaS.
we can define the system cost function as:
Accordingly, the RANaaS manages the activity of the iSCs
to match available resources and required services. Ct = P̃(st , at ) + α · d̃(st , at ), (13)
Markov decision process is used to model this opti-
mization problem [22]. In the considered scenario, the where α is a weighting factor that prioritizes between EE
RANaaS receives data from thecore network and stores it in and QoSwhile P̃(st , at ) and d̃(st , at ) denote the sum of the
a dedicated buffer. When required, RANaaS activates a given NiSC length vectors indicating the power consumption at iSCs
iSC through the backhaul and forwards the associated data and packet losses, respectively. Eq. (13) models the trade-
to it. Thereafter, the iSC will autonomously manage available off between the packet loss and the power consumption; to
resources to transmit received data according to a first-in- limit packet drop the RANaaS should keep the iSCs activated
first-out policy. When a given iSC is idle, energy saving and transmit packets stored in its buffer. On the contrary,
is realised due to the reduced running functionalities at the minimizing power consumption requires to keep iSCs idle,
small cell and at the backhaul network. On the contrary, which in turn may lead to packet loss.
as described in Section IV, when an iSC is activated, its To provide performance evaluation, we consider a hotspot
contribution to the energy consumption of the RAN (that composed by four iSCs (NiSC = 4) coordinated by the nearby
includes the backhaul) depends on its load, i.e., the transmit- RANaaS. Without loss of generality, we consider that in a
ted data on the backhaul and access resources. Finally, when transmission time interval (set equal to 1ms), at most one
an iSC is deactivated, basic coverage and synchronization is packet is received at each buffer and that an activated iSC
provided to the nearby users by the macro cell through dual can transmit up to two packets. Packet length is considered
connectivity [23]. Accordingly, the iSCs can rely on new equal to 1Kbits. The model presented in Section IV is used to
carrier type, which avoids systematic activation to broadcast compute the network power consumption, due to the RANaaS
cell-specific reference signals [23]. In this case, deploying platform, the backhaul and the iSCs. Accordingly, we assume
the RANaaS at the macro cell site may enable to manage that iSC and backhaul equipment in idle mode consumes the
the status of a limited number of neighboring ISCs, which 60% of its zero-load power. Moreover, we consider that 5%
seems to be the optimal choice in terms of complexity and of the BB load is transferred by the iSCs to the RANaaS to
overhead. manage their activation/deactivation.
In our model, the RANaaS is equipped with NiSC buffers Figure 7 shows the performance of the optimal stationary
of size M packets, each one dedicated to an iSC. The number solution (obtained through value iteration [22]) with respect
of packets received at an iSC at time step t( ft ) is finite and to a random policy and greedy approach. By using the random
depends on its load. Moreover, when activated, an iSC i policy, at each time step, the RANaaS randomly decides
transmits at most Ri packets, according to the available whether to keep iSCs idle or activated; on the contrary,
bandwidth and the transmission spectral efficiency. when using the greedy method, action is taken to minimize
Let S be a set describing the network state space and the instantaneous cost, i.e., without considering the total
defined as S = Q × R × F, where Q, R, and F are the (over time) expected cost. Solid, dotted and dashed lines,
composite spaces modelling the buffer, the data rate and the respectively, correspond to the optimal, greedy and random
incoming traffic, respectively. At each time step, the RANaaS solutions. Performance is presented in terms of cumulative
observes the current state of the network st ∈ S and selects an network EE [bit/J] computed as the ratio of the cumulative
action from the action space A, where each action corresponds number of transmitted bits and the associated network power
to activating at most one iSC and keep the others idle. The consumption measured over 1 second.

1594 VOLUME 2, 2014

D. Sabella et al.: Energy Efficiency Benefits of RANaaS Concept

REFERENCES

[1] Code of Conduct on Energy Consumption of Broadband Equipment,

Version 5.0, European Commission, Brussels, Belgium, Dec. 2013.
[2] A. D. Domenico, E. C. Strinati, and A. Capone, ‘‘Enabling green cellular
networks: A survey and outlook,’’ Comput. Commun., vol. 37, pp. 5–24,
Jan. 2014.
[3] 3GPP, ‘‘Potential solutions for energy saving for E-UTRAN (release
11),’’ Radio Access Netw. Tech. Specification Group, Tech. Rep. 36.927,
Sep. 2012. [Online]. Available: http://www.3gpp.org
[4] 3GPP, ‘‘Solutions for energy saving within UTRA node B (release
11),’’ Radio Access Netw. Tech. Specification Group, Tech. Rep. 25.927,
Sep. 2012. [Online]. Available: http://www.3gpp.org
[5] 3GPP, ‘‘Telecommunication management; energy saving management
(ESM); concepts and requirements (release 11),’’ Tech. Specification
Group Services Syst. Aspects, Tech. Rep. TS 32.551, Jun. 2013. [Online].
Available: http://www.3gpp.org
[6] Green Meter Research Study: Reducing the Net Energy Consumption in
Communications Networks by up to 90% by 2020, GreenTouch, Kolkata,
India, Jun. 2013.
[7] I. Chih-Lin, C. Rowell, S. Han, Z. Xu, G. Li, and Z. Pan, ‘‘Toward green and
soft: A 5G perspective,’’ IEEE Commun. Mag., vol. 52, no. 2, pp. 66–73,
FIGURE 7. Cumulative Network EE with respect to different small cell Feb. 2014.
management schemes. [8] K. Sundaresan, M. Y. Arslan, S. Singh, S. Rangarajan, and
S. V. Krishnamurthy, ‘‘FluidNet: A flexible cloud-based radio access
network for small cells,’’ in Proc. 19th Annu. Int. Conf. Mobile Comput.
Netw. (MobiCom), Sep. 2013, pp. 99–110.
The optimal solution leads up to 96% of gain with respect [9] P. Rost et al., ‘‘Cloud technologies for flexible 5G radio access networks,’’
to the greedy solution; however, it gains only 16.4% with IEEE Commun. Mag., vol. 52, no. 5, pp. 68–76, May 2014.
respect to the random policy. This surprising result is due to [10] ‘‘Evolved universal terrestrial radio access (E-UTRA) and evolved uni-
versal terrestrial radio access network (E-UTRAN); overall descrip-
the limited energy consumption associated with the random tion; (release 11),’’ Radio Access Netw. Tech. Specification Group,
solution, which in turns results in an unacceptable number of Tech. Rep. TS 36.300, Sep. 2014. [Online]. Available: http://www.3gpp.org
dropped data. [11] National Industry of Standards and Technologies. Cloud
Computing Synopsis and Recommendations. [Online]. Available:
http://csrc.nist.gov/publications/nistpubs/800-146/sp800-146.pdf,
VI. CONCLUSIONS accessed May 2012.
This article discusses EE aspects and related benefits [12] HP Moonshot System. [Online]. Available: http://h17007.www1.hp.com/
of RANaaS implementation (using commodity hard- us/en/enterprise/servers/products/moonshot, accessed Apr. 2014.
[13] EARTH Project. D2.3—Energy Efficiency Analysis of the Reference Sys-
ware) as evolution of LTE-Advanced networks toward
tems, Areas of Improvements and Target Breakdown. [Online]. Available:
5G infrastructure. In fact, in the view of future definition of http://www.ict-earth.eu/publications/publications.html
5G systems, this cloud-based design is an important archi- [14] G. Auer et al., ‘‘How much energy is needed to run a wireless network?’’
tectural solution, especially in terms of efficient usage of IEEE Commun. Mag., vol. 18, no. 5, pp. 40–49, Oct. 2011.
[15] H. Holtkamp, G. Auer, V. Giannini, and H. Haas, ‘‘A parameterized
network resources. After providing a vision of the advantages base station power model,’’ IEEE Commun. Lett., vol. 17, no. 11,
of RANaaS and its benefits in terms of EE, we presented pp. 2033–2035, Nov. 2013.
a consistent power model as a useful reference for further [16] C. Desset et al., ‘‘Flexible power modeling of LTE base stations,’’ in Proc.
evaluations of flexible centralization of RAN functionalities IEEE Wireless Commun. Netw. Conf. (WCNC), Apr. 2012, pp. 2858–2862.
[17] CT FP7 FIT4GREEN. (Mar. 2012). Presentation of the Full-
in terms of EE. The incremental benefits through the years Featured Federated Energy-Consumption Models. [Online]. Available:
have been also discussed in perspective, by considering http://www.fit4green.eu/sites/default/files/attachments/documents/D3.3_
technological evolution of IT platforms and the increasing final.pdf
matching between their capabilities and the need for progres- [18] T. Werthmann, H. Grob-Lipski, and P. Proebster, ‘‘Multiplexing gains
achieved in pools of baseband computation units in 4G cellular net-
sive virtualization of RAN functionalities. The description is works,’’ in Proc. IEEE 24th Int. Symp. Pers. Indoor Mobile Radio
finally complemented by an exemplary evaluation, in terms of Commun. (PIMRC), Sep. 2013, pp. 3328–3333.
EE, of a small cell activity controller based on the RANaaS [19] Intel Xeon Processor 5500 Series. [Online]. Available:
http://download.intel.com/support/processors/xeon/sb/xeon_5500.pdf,
paradigm, as a promising technology for future sustainable accessed Apr. 2014.
5G radio. [20] P. Monti, S. Tombaz, L. Wosinska, and J. Zander, ‘‘Mobile backhaul
in heterogeneous network deployments: Technology options and power
consumption,’’ in Proc. 14th Int. Conf. Transparent Opt. Netw. (ICTON),
ACKNOWLEDGMENT
Jul. 2012, pp. 1–7.
The European Union and its agencies are not liable or [21] NGMN Alliance. Small Cell Backhaul Requirements. [Online]. Available:
otherwise responsible for the contents of this document; http://www.ngmn.org/uploads/media/NGMN_Whitepaper_Small_Cell_
its content reflects the view of its authors only. The Backhaul_Requirements.pdf, accessed Jun. 2012.
[22] R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction.
authors gratefully recognize the great contributions of many Cambridge, U.K.: Cambridge Univ. Press, 1998.
colleagues from iJOIN, who in fruitful cooperation, [23] C. Hoymann, D. Larsson, H. Koorapaty, and J.-F. Cheng, ‘‘A lean carrier
contributed with valuable insight, surveys and vision. for LTE,’’ IEEE Commun. Mag., vol. 51, no. 2, pp. 74–80, Feb. 2013.

VOLUME 2, 2014 1595

 D. Sabella et al.: Energy Efficiency Benefits of RANaaS Concept

DARIO SABELLA received his Dr. Eng. Degree MARCO DI GIROLAMO Graduated in Electron-
in Electronics Engineering from Politecnico di ics Engineering (summa cum laude) in November
Torino, Italy, in 2002. After graduation, he joined 1986. More than 25 years of working experience
the research centre of Telecom Italia, called TILab, in R&D, IT and Consulting environments, in dif-
being involved in several R&D projects target- ferent Companies, both Italian and international.
ing future generation communications. In 2004 he In HP since 1993, covered positions in R&D as
received an additional post-degree specialization Development Engineer, in Customer Care and IT
Master in Telecommunications. Currently working Department. Since 2000, he is working in the Italy
in the Wireless Innovation Department of Telecom Innovation Centre, where served as senior tech-
Italia as responsible for some international projects nology specialist, solution architect and project
and innovative activities on Energy Efficiency for the mobile network. Hes manager in different engagements, both internal and external.
leading Telecom Italia activities for 5GrEEn project, on green guidelines for Since 2002, he is leading HP Italy activities into European collaborative
5G mobile system. From 2012 he is involved in Carbon Footprint project with research and innovation. In this position, hes coordinated ad led HP Italy
the Italian Environment Ministry, and hes leading Telecom Italia activities initiatives related to European innovation programs, including IST FP5, FP6,
related to energy efficiency of the mobile network in the framework of the FP7, eContent and CIP project participations. He co-authored different pub-
LTE auction in Italy. He is co-author of several publications and patents in lications within the projects he contributed to. In particular, he contributed
the field of wireless communications and energy efficiency. as WP leader to the FIT4Green and All4Green projects, which investigated
ANTONIO DE DOMENICO received the methods, tools and metrics to improve energy efficiency of datacentres in
M.Sc. degree in telecommunication engineering integrated ecosystems. He is currently engaged in the DC4Cities project,
from the University of Rome La Sapienza, Rome, investigating technologies to maximize the share of renewable energy used
Italy, in 2008, and the Ph.D. degree from the in supplying datacentres.
University of Grenoble, Grenoble, France, in 2012.
Since 2009, he has been with the Laboratory for
Electronics and Information Technology, Micro
and Nanotechnology Innovation Centre, Atomic
Energy Commission, Grenoble, as a Research UMER SALIM received the Ph.D. and M.S. degrees
Engineer. His current research topics are cloud- in electrical engineering with a specialization
enabled heterogeneous wireless networks, millimetric-wave-based commu- in communication theory and signal processing
nications, machine learning, and green communications. He is the main from EURECOM, Biot, France, and Suplec,
Inventor or Co-Inventor of seven patents. Cesson-Svign, France. He has several years
of research experience in digital communications
EFSTATHIOS KATRANARAS (M’06) received and signal processing, and has authored several
the B.Sc. degree in electrical and computer engi- papers in well-known conferences and journals.
neering from the National Technical Univer- His main areas of interest include signal processing
sity of Athens, Athens, Greece, in 2005, and techniques for multicell multiuser multiple-input
the M.Sc. (Hons.) and Ph.D. degrees in mobile multiple-output systems, novel and practical CSI feedback design techniques
communications from University of Surrey, and analysis, information theoretic analysis of cognitive radio, and multiuser
Guildford, U.K., in 2006 and 2009, respectively. information theory in general. He has been serving as a reviewer of the
He is currently a Research Fellow with the Institute IEEE TRANSACTIONS ON INFORMATION THEORY, the IEEE TRANSACTIONS ON
for Communication Systems, University of Surrey, WIRELESS COMMUNICATIONS, the IEEE TRANSACTIONS ON SIGNAL PROCESSING, and
where he is an Active Research Member of the numerous well-known conferences. He co-authored a paper that received the
5G Innovation Centre. His main research interests include the evaluation Best Paper Award at the European Wireless Conference in 2011.
of fundamental capacity limits and energy efficiency in cellular systems He is currently with the Department of Systems Engineering, Intel Mobile
and advanced multicell cooperation techniques for future dense small cell Communications, Paris, France, where he is involved in the design of
networks. advanced receivers for future wireless standards and has designed sophis-
ticated interference cancellation algorithms, which are in use in the modern
MUHAMMAD ALI IMRAN (M’03–SM’12) high-end smart-phones and tablets.
received his M.Sc. (Distinction) and Ph.D. degrees
from Imperial College London, UK, in 2002 and
2007, respectively. He is currently a Reader in the
Institute for Communication Systems, home of 5G
Innovation Centre, at the University of Surrey, UK.
In this role, he is leading a number of multimillion MASSINISSA LALAM received the master’s
international research projects encompassing the degree from the Institut National Polytechnique
areas of energy efficiency, fundamental perfor- de Grenoble, cole Publique D’ingnieurs
mance limits, sensor networks and self-organising en Informatique, Mathmatiques Appliques
cellular networks. He is also leading the new physical layer work area et Tlcommunications de Grenoble, Greno-
for 5G innovation centre at Surrey. He has a global collaborative research ble, France, in 2002, and the Ph.D. degree
network spanning both academia and key industrial players in the field of from the Ecole Nationale Suprieure de
wireless communications. He has supervised 18 successful PhD graduates Tlcommunications de Bretagne, Telecom
and published over 150 peer-reviewed research papers including more than Bretagne, Plouzan, France, in 2006, where he
20 IEEE Journals. His research interests include the derivation of information held a post-doctoral position, in 2007. He joined
theoretic performance limits, energy efficient design of cellular system and Orange Labs, Paris, France, from 2008 to 2009. He is currently an Expert
learning/self-organising techniques for optimisation of cellular system oper- with Sagemcom Broadband, Paris, France, where he works on wireless (Wi-
ation. He is an editor of IET Communications and a guest editor of special Fi) and cellular (3G/LTE) technologies. He has been involved in several
issues in IEEE COMMUNICATIONS MAGAZINE and IEEE WIRELESS COMMUNICATIONS European projects, such as BeFEMTO, iJOIN, and LEXNET, as a Work
MAGAZINE. He is a senior member of IEEE and a Fellow of Higher Education Package Leader and Task Leader or Contributor. His expertise includes
Academy (FHEA), UK. He has won the Faculty Learning and Teaching in link and system-level performance evaluation, heterogeneous network,
Award 2014 and has been awarded the 2014 IEEE Communications Society network modeling, and radio resource management. He has authored or co-
Fred W. Ellersick Prize. authored over 15 international publications.

1596 VOLUME 2, 2014

D. Sabella et al.: Energy Efficiency Benefits of RANaaS Concept

KONSTANTINOS SAMDANIS received the ANDREAS MAEDER received the Ph.D. degree
Ph.D. and M.Sc. degrees in mobile communica- from the University of Wrzburg, Wrzburg,
tions from Kings College London, London, U.K. Germany, in 2008, with his thesis on radio resource
He is currently a Senior Researcher and Backhaul management in integrated 3GPP UMTS/HSPA
Standardization Specialist with NEC Laboratories systems. Since 2008, he has been with NEC
Europe, Heidelberg, Germany. He leads a research Laboratories Europe, Heidelberg, Germany, where
project on long-term evolution (LTE) network he is currently a Senior Researcher working
management and multitenancy, while at the same on 5G mobile networks, including RAN and
time, he is involved in the Marie Curie Initial Train- core network virtualization, energy efficiency, and
ing Network through the CROSSFIRE project as a the convergence of IT and telecommunication
Work Package Leader on LTE network virtualization. He is currently active in technologies. He has contributed to the standardization of mobile networks
5G, while in the past, he participated in the iJOIN and BeFemto FP7 projects in IEEE and 3GPP since 2008. He has authored numerous scientific articles
and IST EVEREST project. He is also active in the Broadband Forum in and conference papers, holds patents, and is an Editor of a book on green
Software-Defined Networking/Network Functions Virtualization and is an data communication.
Editor of the TR-293 Energy-Efficient Mobile Backhaul. He has served as a
Guest Editor of the IEEE COMMUNICATION MAGAZINE and the IEEE MULTIMEDIA
COMMUNICATIONS TECHNICAL COMMITTEE E-LETTERS, and the Co-Chair of the Next-
Generation Networking Symposium of the IEEE International Conference on
Communications 2014.

VOLUME 2, 2014 1597