Received July 9, 2018, accepted August 27, 2018, date of publication October 9, 2018, date of current version October

29, 2018.
Digital Object Identifier 10.1109/ACCESS.2018.2874221

Drastic Energy Reduction With gDTX

in Low Cost 5G Networks
FANNY PARZYSZ , (Member, IEEE), AND YVON GOURHANT
Architecture and uRbanism for teleComs Team, Orange Labs, 22300 Lannion, France
Corresponding author: Fanny Parzysz (fanny.parzysz@orange.com)

ABSTRACT Base station (BS) sleep modes handle the progressive deactivation of unused hardware
components and bring a promising solution to the increasing energy bill and environmental concerns. Taking
advantage of the 5G ultra-lean design and spaced synchronization signals, we propose a novel gDTX mode
(for next-generation discontinuous transmission), based on flexible duty cycling. First, an inactive period
enables the BS to enter deep sleep state for some tens of milliseconds, even if users are requesting service.
We show that adapting the inactivity duration to the daily traffic load fluctuations provides significant
energy reduction, reaching half of the time more than 45% gain in suburban environment and 78% in
rural area. Second, an active period maintains some cell capacity and normal signaling, rendering the BS
always available, even if not always on. Questioning the feasibility of gDTX regarding real-time delay-
non-tolerant voice traffic, we analyze its impact on the overall jitter, end-to-end delay and estimated mean
opinion score. We evidence that sleeping 60 ms every 80 ms while sustaining acceptable voice quality is not
contradictory in most network conditions, for light to moderate loads. The proposed gDTX cycle definitely
opens new perspectives on energy reduction, targeting primarily developing countries where powering BSs
is still a priority to deploy reliable mobile services.

INDEX TERMS 5G, energy consumption, emerging countries, base stations, sleep modes, voice, QoS, mean
opinion score.

I. INTRODUCTION
Future 5G architecture is widely being shaped around
enhanced mobile broadband (eMBB), ultra-reliable low
latency communications (URLLC) and massive machine-
type communications (mMTC), promising unprecedented
user experience. Despite the growing environmental concerns
and the push toward green networks, such high-performance
technologies can hardly be sustained in digital ecosystems
where powering base stations is still a major issue and where
many consumers cannot afford a basic Internet connection.
As pointed out in [1] and [2], many emerging and devel-
oping countries are heavily restricted by energy, while the
overall consumption in the RAN is expected to increase by
16% from 2014 to 2020. More than half of towers are off-grid
in Burkina Faso, Nigeria, Madagascar or Chad, where electri-
fication barely reaches 4%. Generators powering remote base
stations (BS) may run out of diesel, by cause of excessive
cost for local communities or delivery issues on poor roads,
notably during the rainy season. Alternative energy sources
have been developed, notably through the ORYX project [2],
but in many places, solar panels still often experience low
efficiency, e.g. due to insufficient cleaning. Even on-grid sites
may undergo regular power cuts or may not be powered more
than a few hours per day.
Contributions: In the light of this background,
we consider a complementary approach to ubiquitous high
capacity technologies and propose a new gDTX cycle (‘‘next-
generation Discontinuous Transmission’’) firstly dedicated
to cellular networks where energy is the uppermost priority.
Taking advantages of the new features of 5G radio, the pro-
posed gDTX cycle, depicted in Figure 1, operates periodical
deactivation of all transmissions and receptions at a gNB
(next-generation NodeB) for some tens of milliseconds, while
maintaining normal signaling and some cell capacity depend-
ing on the traffic load and battery level fluctuations.
First, we detail the concept of gDTX within the 5G frame-
work and highlight how it offers a proximity access for all
consumers at all time, particularly for isolated sites with no
possibility of traffic offloading. Contrary to other approaches,
base stations do not enter sleeping state uniquely at zero load
timespan. Periodical sleeps for energy reduction are set up
based on traffic load, rather than on individual packet arrivals,
while minimizing QoE (Quality of Experience) degradation
as much as permitted by the energy constraints.

2169-3536
2018 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 6, 2018 Personal use is also permitted, but republication/redistribution requires IEEE permission. 58171
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 F. Parzysz, Y. Gourhant: Drastic Energy Reduction With gDTX in Low-Cost 5G Networks
FIGURE 1. Overview of the proposed gDTX cycle.
Second, we provide a benchmark evaluation of its energy
performance and show that adapting the gDTX cycle to the
average traffic load is sufficient to reach half of the time up to
45% energy gain in suburban environments and up to 78% in
rural areas. Contrary to existing strategies, the gDTX cycle
results in a predictable energy consumption, which greatly
facilitates battery management at off-grid base stations.
Third, we evidence that large sleeping periods do not
impede the transmission of real-time traffic like voice in
light to moderate load conditions. We analyze the impact of
inactivity periods on the overall jitter and end-to-end delay,
then use an E-model for Adaptive Multi-Rate Wideband
(AMR-WB) codec to estimate the Mean Opinion Score.
We highlight that a gNB can enter deep sleep mode for some
30ms every 80ms in typical network propagation conditions
without being perceived by users. Scheduling inactivity peri-
ods of 60ms every 80ms is almost transparent to users in good
network conditions (low jitter, short propagation delay) and
still provides acceptable voice quality (MOS≥ 3.5) in most
conditions.
Paper organization: The literature on BS energy reduction
is reviewed in Section II and the proposed gDTX cycle is
detailed in Section III. We show in Section IV that significant
energy gain can be achieved and, in Section V, that gDTX
cycles are also suitable for real-time voice traffic. Concluding
remarks and future investigations are discussed in Section VI.
II. RELATED WORK
This section covers the main approaches on energy reduction
for radio cellular networks, with respect to the considered low
cost 5G context.
A. IDLE ENERGY CONSUMPTION IN LOW LOAD
NETWORKS
The push towards network densification for 5G has spurred
advanced technologies for energy reduction, such as BS
cooperation [3], cloud-RAN optimization [4] or User/Control
planes separation [5]. Nevertheless, harsh technical con-
straints render such technologies hard to operate in rural
areas of developing countries. They encompass long distance
between nodes, long latency backhaul, lower cost architecture
and sometimes, poor equipment interconnections.
58172
Aside from the energy consumed for the transmission
itself, significant energy is dissipated at BS for site cooling,
network maintenance and signaling, whether data is trans-
mitted or not. In low load networks, such idle energy off-
set largely dominates the overall consumption and a typical
3-sector site consumes around 2kW for a transmission-related
power of only 120W [6]. Various solutions have been pro-
posed to approach the ‘‘zero watt at zero load’’ objective
through standalone BS sleep modes. It consists in setting in
idle state part of the BS hardware components to reduce the
load-independent consumption.
B. ON-DEMAND BS ACTIVATION
Switching off unused base stations (BS) and activating them
only when needed is one of the most popular responses to
power issues in low load networks [7]. The coupling existing
between the achievable energy gain and the cell densification
in heterogeneous networks has been analyzed in [8] and [9]
and noticeable reduction in the daily energy consumption is
achieved through cooperation among delay-tolerant commu-
nications in [10] or using traffic prediction in [11].
However, on-demand BS activation is principally designed
for dense heterogeneous networks and presents three main
drawbacks. First, most literature on the subject focuses on
switching off small-cells, while macro base stations are the
most energy-consuming [3], [12]. Second, the radio coverage
and service have to be provided by the remaining active
cells, implying that such technique is not suitable for iso-
lated BS. As a response to this issue, innovative solutions
for re-activating cells have been proposed in [13], using
out-of-band low consumption radio modules for femtocells,
and in [14], through BS waking up using a LoRa interface.
Third, BS switching off implies long system recovery and
slow reactivity to traffic variations or sudden power cuts [6].
Indeed, the delay necessary to wake up, re-establish connec-
tions and transfer traffic is non-negligible and counted in
seconds or minutes. Its maximum acceptable value in low
density networks is still an open research issue.
C. CELL DISCONTINUOUS TRANSMISSION (CELL DTX)
While, at a long time scale (a few tens of minutes), only very
few cells carry no traffic, it is found that, at a much shorter
time resolution (on millisecond level), no user is served in
most of the cells, most of the time [6], [15]. There is thus a
large energy saving potential for fast mechanisms that enable
a BS to enter light sleep mode when no data is scheduled.
Cell Discontinuous Transmission has been proposed in [16]
to clean up unnecessary subframe transmissions and reduce
control signaling overhead. Taking inspiration from the Dis-
continuous Reception (DRX) mode, standardized in LTE
for mobile terminals, cell DTX aims at quickly switching
between active and empty symbol timespans (' 71µs), dur-
ing which the power amplifier and part of RF subsystem are
turned off.
A wide range of algorithms have been proposed to fully
harness the potential of Cell DTX and its performance
VOLUME 6, 2018
F. Parzysz, Y. Gourhant: Drastic Energy Reduction With gDTX in Low-Cost 5G Networks
has been analyzed at country level in [15]. As shown
in [17], significant energy gains are achieved in denser
networks with lightly loaded cells. Regarding rural envi-
ronments, with lower population and high heterogeneity in
the inhabitant distribution, beamforming has been analyzed
in [18] for mmWave networks using cell DTX. Performance
is also subjected to BS scheduling. Energy-efficient tech-
niques encompass adapting the resource blocks placement
within the frame [19], controlling the number of empty
subframes [20], [21], notably depending on the delay toler-
ance of user applications, or coordinating the sleep periods
among cells to reduce inter-cell interference [22]. Whereas
cell DTX offers valuable reactivity to packets arrival in
moderate load conditions, higher energy savings could be
achieved during long periods with low traffic, e.g. at night,
by turning off more BS hardware components.
D. TOWARDS ADVANCED SLEEP MODES
On-demand BS activation and cell DTX can be understood as
two extreme sleep modes (SM), the former relying on switch-
ing off a whole site to save energy, the latter aiming at turn-
ing off radio subframes. In between, several other potential
sleep levels can be implemented for energy reduction. A first
sleep scale has been proposed in [6] for 2G/3G networks,
which includes site, sector, layer and carrier component sleep
modes. This concept has been further investigated by IMEC
within the GreenTouch project [12], [23]. In particular, the BS
component deactivation and activation duration, or transition
latency, has been shown to be a serious hindrance to advanced
sleep mode and energy reduction.
Hardware sleep capabilities have been investigated in [23]
to predict the power consumption of BS up to the year
2020. Four levels of sleep depth have been characterized,
by grouping the sub-components with similar transition
latency:
• SM1: This is the fastest sleep mode as its transition
latency is only one OFDM symbols (71µs). The power
amplifier and some other RF components are deacti-
vated.
• SM2: This medium sleep state, where more components
are disabled, is reached at the cost of 1ms transition time.
• SM3: In this idle mode, all components of the digital
baseband and analog front-end (both Rx and Tx) are
switched off, excepting the clock generator. The transi-
tion time is evaluated to 10ms.
• SM4: In this standby mode, a BS is out of operations
and only wake-up functionalities are kept active. This
refers to on-demand BS activation. The transition time is
measured in seconds (not accounting for the connection
re-establishment time).
The power consumed by components that are still active
for each sleep mode is given in Figure 2. SM2 and
SM3 are of particular interest when assuming non-zero
traffic load, as they allow larger energy reduction with
reasonable latency. The proposed gDTX strategy aims at
VOLUME 6, 2018
FIGURE 2. Power consumption of sleep modes (2x2 MIMO BS).
enabling such advanced sleep modes while insuring user
service.
E. NEW OPPORTUNITIES OF 5G TECHNOLOGIES
Due to their longer transition latency, advanced sleep modes
are first hampered by BS synchronization signaling, which
is transmitted even within an empty cell for network access
and cell (re-)selection. In LTE, this signaling includes the
Primary and Secondary Synchronization Signals (PSS and
SSS), as well as the Physical Broadcast CHannel (PBCH),
and is repeated every 5ms. In addition, only a few subframes
within a frame can be totally emptied due to reference
signals for demodulation, phase-tracking and channel-state
information. The concept of hyper-cellular networks has
been proposed in [5] to decouple the coverage of control
signals from the coverage of data and has been extended
in [24] to enable long sleep periods. Yet, including syn-
chronization and channel estimation within the Control/User
plane separation remains complex to implement in practical
architectures.
The key design principle of ‘‘5G Forward compatibil-
ity’’ aims at minimizing always-on transmissions and opens
new perspectives for BS sleep modes. The ultra-lean design
of 5G radio enables to confine signals and channels for
physical layer functionalities within a configurable time and
frequency resource [25]. More flexibility is provided in the
uplink/downlink frame scheduling and in timing relations for
HARQ and ACK management [26]. Full service, including
data/control management and users connection, can be pro-
vided with limited time-domain resources. Finally, synchro-
nization signals are grouped in SS-blocks and scheduled with
a much longer and configurable periodicity of 5, 10, 20, 40,
80 or 100ms, as standardized in [27].
While spaced SS-bursts primarily targets overhead reduc-
tion for beam sweeping in multi-antenna coverage, it also
holds significant potential for energy savings. It is shown
in [28] that the power consumption of a 2x2 macro-BS with
a 5ms periodicity (as in LTE) is 4.5 times larger than a
similar BS with a 100ms periodicity. Next, strategies have
been proposed in [29] and [30] to enter deep sleep state
during network idle periods and quickly wake up upon traffic
requests, then allowing energy reduction while minimizing
58173
 F. Parzysz, Y. Gourhant: Drastic Energy Reduction With gDTX in Low-Cost 5G Networks
FIGURE 3. Example of use of the guaranteed active period.
the additional latency due to switching from idle to active BS
state.
III. A gDTX CYCLE TOWARD ADVANCED SLEEP MODES
We present in this Section the proposed gDTX strategy within
the context of energy-restricted cellular networks.
A. MOTIVATIONS FOR A DIFFERENT APPROACH ON BS
SLEEPING
The major restriction of aforementioned solutions is that
advanced sleep modes are permitted only as a response
to zero-load conditions, to provide the best QoS. Practical
implementations of Cell-DTX allow full sleep pattern (num-
ber of empty subframes) only if there is no RRC-connected
users, and partial sleep pattern if there are less than 5 RRC-
connected users, with low DL traffic load [31]. In [20], sleep
state is dependent on the buffered packet with the most strin-
gent delay constraint. In [29] and [30], a BS is activated at any
packet or signaling message arrival, even if only a few PRB
(physical resource blocks) are scheduled during a subframe.
Consequently, a BS managing only a few users doing
phone calls may be unceasingly woken up by grant requests
and packet transmissions, and may be prevented from enter-
ing sleep modes, even if the traffic load is almost null.
By tracking the load fluctuations on operated networks,
we have observed that the time occupancy of a typical rural
BS, i.e. the ratio between occupied and available subframes,
is in average 3 to 4 times higher than the effective resource
occupancy in PRB. Thus, significant energy reduction is
expected from approaching sleep management differently.
The proposed gDTX cycle aims at solving this issue. In addi-
tion, we propose a strategy that is applicable for isolated sites,
with no possibility of traffic offloading, and for any type of
base stations, including macro BS which are generally used
for network coverage.
B. OVERVIEW OF gDTX CYCLES WITHIN THE 5G
FRAMEWORK
To exploit the capabilities of BS sleeping in non-zero load
conditions, we propose a novel Discontinuous Transmis-
sion mode for gNB, named gDTX cycle. As illustrated
58174
in Figures 1 and 3, it hinges on:
• A flexible duty cycle of length NT , composed of an
active period, containing SS-blocks, and an inactive
period;
• An inactive period, of length N , during which the gNB
enters sleep mode, even if users are requesting service.
Incoming packets, signaling and requests are buffered;
(max)
• A guaranteed service period every NT to enable
extended DRX, used in particular for IoT devices.
Applying the ultra-lean design of 5G radio to the proposed
gDTX cycle, as introduced in Subsection II-E, all transmis-
sions, including user data, user-specific signaling and ref-
erence signals, are interrupted during the inactive period to
leave blank resources and let the gNB enter advanced sleep
modes. As it operates on periodical basis rather than as a
response to packet arrivals, the proposed strategy minimizes
the number of wakes-up. The control necessary to wake up
and switch to an active mode is limited and no extra energy
is dissipated to listen to the incoming traffic. As full ser-
vice is provided during the active period, no connection re-
establishment or traffic offloading is needed before user data
transmissions. In addition, the proposed gDTX can operate
jointly with any resource scheduler. The active period should
be sufficiently large to allow scheduling optimization, for
example depending on the channel quality, and may be com-
bined with cell DTX.
LONGER CYCLES FOR BETTER PERFORMANCE
For a fixed sleep ratio NNT , extending the cycle duration NT
may limit the interactivity of some low latency real-time
traffic but it also provides numerous benefits. First, deeper
sleep states are achievable, bringing more power gain. For
N
NT = 0.5, the average power consumed during sleep states is
equal to 54.3W when NT = 10ms (i.e. 4ms of SM2 and 1ms
SM1↔SM2) and is reduced to 35.7W when NT = 40ms (i.e.
9ms of SM3, 1ms SM1↔SM2, 10ms SM2↔SM3). Although
further extending NT does not bring significant additional
gain, it offers more granularity for sleep ratio optimization
and lets the possibility to implement new intermediate sleep
states, with a (de)activation latency between 10ms and 50ms.
BASE STATION gDTX AND USER DRX
The joint management of gDTX and user DRX is two-fold.
First, trying to reach a gNB in sleep state is vain and energy-
wasting. In contrast to conventional systems, where each user
can negotiate its own DRX parameters, the proposed strategy
benefits from synchronizing the user DRX to the gDTX cycle
of its associated gNB. To this end, the cycle length NT and
inactivity duration N should be broadcast, e.g. within System
Information Block messages.
Second, the necessity for users to know when the gNB
is awake raises challenges for terminals in extended DRX
(eDRX), for example IoT devices. A such terminal may
stay in dormant mode for a predefined number of hyper
frames that can reach a few hours. Consequently, it may not
VOLUME 6, 2018
F. Parzysz, Y. Gourhant: Drastic Energy Reduction With gDTX in Low-Cost 5G Networks
be informed if the parameters (N , NT ) have been modified
meanwhile, and thus, loose synchronization. This is illus-
trated in Figure 3. To this end, the proposed gDTX design
includes a guaranteed service window, i.e. an active period
(max)
is scheduled every NT , whatever the current values of
(N , NT ). A simple way of implementation is to delay the
modification of (N , NT ) so that the active period of the new
gDTX cycle matches the next scheduled guaranteed service
period.

period, even though less cell capacity is provided to users.
C. gDTX CYCLES AND OPTIMIZATION OF N, NT
Conventional systems are based on the sharing of all time
and frequency resources among users, whether there are
one or hundreds of them. On the contrary, the proposed strat-
egy aims at limiting the time-domain resources and control-
ling the duration NT −N of the active period, depending on the
traffic load and network energy constraints. The optimization
of (N , NT ) is thus a key design parameter for reducing energy
and maintaining QoE.
First, reactive or proactive algorithms can be developed
depending on the ability to track and/or predict the traffic
load fluctuations in the time (e.g. day/night), but also in the
space (e.g. work/residential areas). Solutions may be rather
static, leading to a pair (N , NT ) set for a few hours, or more
dynamic, with (N , NT ) changed after a few frame trans-
missions. Long-term adaptation of gDTX cycles based on
network statistics may not lead to the optimal energy gain
and may cause under-utilization of resources, or reversely,
may excessively degrade the users QoS. However, it is less
sensitive to traffic sudden variations and limits the extra
energy dissipated for computation and broadcast of (N , NT ).
In Section IV and V, we consider a semi-static adaptation,
based on 15-minute periods. Further performance gain is
achievable through the joint optimization of several BS duty
cycles, notably through smart traffic offloading, as in [22] for
Cell DTX.
Next, energy harvesting base stations raise unique chal-
lenges as the energy arrival is uncertain. In addition to traffic
load considerations, the gDTX cycle can be adapted to the BS
battery level and predictions on the harvesting capabilities.
For example, the inactivity period may be shorter during
sunny days when solar panels can operate at full efficiency,
and longer during rainy days or nights, when the BS must rely
on the energy stored in its battery.
Finally, the gDTX cycle should be tailored not only to the
energy savings but also to the resulting quality of service,
as long inactivity period may degrade user experience. The
impact of gDTX cycles on the voice quality is introduced in
Section V but further optimization is left to future work.
D. DISCUSSION
Whereas QoS metrics such as minimal latency or broadband
capacity primarily target markets of developed countries,
insuring, at lower cost, sufficient network availability, both
in the space (coverage) and time (battery lifespan), remains a
major priority for many developing countries.
VOLUME 6, 2018
On the one hand, the proposed gDTX design allows a
refined energy management. The inactivity period for energy
savings is set by the gNB itself as a function of traffic load
and energy harvesting statistics, and independently of users
individual requests and packet arrivals. Thereby, the energy
consumption can be estimated quite accurately in advance,
given the chosen (N , NT ) and becomes predictable. Manag-
ing battery is then greatly facilitated. For example, in case
of low battery, a gNB may decide to increase the inactivity
Such a situation is preferred in networks severely restricted
by energy as it maintains minimal connectivity, whereas pro-
viding full capacity may imply running out of energy and
letting network access down for all users, including in case
of emergency.
On the other hand, gDTX renders gNB always available,
even if not always on. It particularly fits isolated sites, where
traffic cannot be offloaded. Moreover, it offers more oppor-
tunity for an operator to maintain connection at rural or sub-
urban sites, contrary to BS switching off which tends to
favor users in denser areas or cities. Thus, gDTX can offer a
proximity access to all users, whatever their location or time
of day.
IV. ENERGY PERFORMANCE OF gDTX CYCLES
In this Section, first insights are given on the daily energy
savings in a cellular network implementing gDTX cycles.
A. TRAFFIC MODEL AND SIMULATION SETUP
The energy consumption largely varies depending on the BS
type, i.e. its architecture, hardware components and transmis-
sion technology. As in [29], we consider a 2x2 MIMO tri-
sectored base station operating on a 20MHz bandwidth. For
performance evaluation, we assume three cases:
• A BS using gDTX cycles, which can switch
between active state, SM1, SM2 and SM3, depend-
ing on the transition latency constraints detailed in
Subsection II-D.
• A BS implementing Cell DTX only, modeled as a two-
state BS (active or SM1).
• An always-on BS, which does not use sleep modes.
As highlighted in [15] and illustrated in Figure 4, small-
time-scale simulations may result in underestimated energy
gains when considering busy-hour network conditions (worst
case scenario) or conversely, in overestimated results when
considering night traffic pattern (best case scenario). This
motivates us to measure real traffic over several days, to bet-
ter capture the living behavior of users. In particular, we track
the daily load fluctuations of three BSs, respectively located
in urban, suburban and rural environments and currently in
operation on the Orange radio cellular network.
A BS generally manages many users and heterogeneous
traffic types at the same time, for both data and signaling. The
footprint of such traffic on the media resources is subjected to
protocol overheads, channel quality, but also network failures
58175
 F. Parzysz, Y. Gourhant: Drastic Energy Reduction With gDTX in Low-Cost 5G Networks
FIGURE 4. Daily average power consumption with and without sleep modes in a suburban environment.
FIGURE 5. Cumulative Distribution Function of the measured traffic load and simulated power gain over several week days.
(a) CDF. (b) Complementary CDF.
and congestions, which may lead to retransmissions or con-
nection reconfigurations. To account for their joint impact,
we measure the BS load through the bandwidth (BW) occu-
pancy, i.e. the ratio between the number of utilized physical
resource blocks (PRB) and the number of available PRB,
averaged over a 15-minute period. The cumulative distribu-
tion function (CDF) of the BW occupancy observed in rural,
suburban and urban environments is depicted in Figure 5a.
The power consumed by each base station is computed as
the sum of the load-dependent power, assumed proportional
to the number of used PRB, and the power dissipated in
circuitry, which depends on the number of consecutive empty
subframes (1ms each). It accounts for the time spent in each
sleep mode (SM1, SM2 or SM3), as well as for the power
dissipated during transition periods between sleep modes and
active states. Values used for power consumption evaluation
are given in Figure 2.
Remark: To accurately evaluate the gain achieved by Cell
DTX, we also measure the actual number of empty subframes
at the considered base stations.
Next, a wide range of parameters can be considered to
adapt the gDTX cycle to the network conditions. In this
paper, we consider a simple adaptation method for (N , NT ) to
provide a first benchmark on the energy performance. Taking
NT =80ms to extend the inactivity period, we assume that
the gNB knows the average BW occupancy along the day
58176
NT −N
and uses this value to set the activity ratio NT , for each
15-minute period.
B. SIMULATION RESULTS: ENERGY REDUCTION USING
gDTX CYCLES
As illustrated in Figure 4, the power gain achieved by sleep
modes (through Cell DTX or gDTX) considerably varies
depending on the time of day. During empty hours (between
1am and 7am), traffic does not occupy more that 10% of
the available resources, whether considering an urban, sub-
urban or rural environment. The consumption of a BS imple-
menting gDTX cycles is almost 5.9 times lower than for an
always-on BS and 3.5 times lower than for a BS implementing
cell DTX.
But energy reduction is not limited to night hours.
Although gains achieved by sleep modes are generally low
when the BW occupancy reaches 80-85%, such scenario has
been observed only in urban environment during busy hours
(i.e. 12am-2pm, 4-6pm and 8-10pm). As shown in Figure 5a,
a suburban BS does not carry more than 35% load half of
the time, and the load of a rural BS rarely exceeds 20%.
This clearly encourages the use of advanced sleep modes and,
in light of this traffic distribution, we plot in Figure 5b the
complementary CDF of the power gain achieved by gDTX
cycles over always-on BS and cell DTX. As observed, gDTX
cycles provide more than 50% gain over cell DTX for 20%
VOLUME 6, 2018
F. Parzysz, Y. Gourhant: Drastic Energy Reduction With gDTX in Low-Cost 5G Networks
TABLE 1. Simulated delays for voice traffic [32]–[34].
of the time in an urban environment (i.e. P (Gain ≥ 50%) =
0.2), 30% of the time in a suburban area and almost all the
time in rural environment.
A finer analysis of the traffic transmitted and received at
the considered base stations shows that significant energy
gain is provided by gDTX over cell DTX during periods
with low sporadic data traffic (typically smartphone software
updates) or some users doing phone calls (a few PRB every
20ms). In this case, a BS which does not implement gDTX
regularly switches to active state, even if many subframes are
scheduled with a few occupied PRB only. Such scenario is
observed notably in rural areas or at the beginning of night.
The ratio of utilized subframes over available subframes is
then 4 to 5 times higher than the ratio of utilized PRB.
V. VOICE QUALITY USING gDTX CYCLES
Setting up long BS inactivity periods affects the throughput,
delay and jitter experienced throughout the network. While
best-effort traffic can be spread over duty cycles, transmitting
real-time delay-non-tolerant applications, like voice, raises
new challenges to handle the unpredictable behavior of IP
networks and limit the potential QoS degradation. As they
are mostly based on proprietary algorithms that regulate
the packet flow as a function of network conditions, other
real-time applications, e.g. Skype video-conference or online
gaming, are left to future investigations.
A. TRAFFIC MODEL AND SIMULATION SETUP
To analyze the performance of voice traffic under gDTX
cycles, we consider a Matlab-based simulator modeling the
network depicted in Figure 6. We can reasonably assume
that the network behavior is statistically symmetrical in both
directions of the phone call and we focus on one direction
only, from mouth to ear. The delays modeled in our simulator
are listed in Table 1 and further details on each part of the
simulated network are provided in the following.
1) SOURCE ENCODER
Voice is traditionally characterized by an ON-OFF macro-
behavior, consisting of intercalated speech and silent bursts.
According to [32] and [33], this behavior can be modeled as
a two-state Markov model and the probability of being in the
talking state is given by the voice activity factor (equal to
0.5 in our setup). Both talking and silent periods are exponen-
tially distributed, with mean 3sec. We assume an Adaptive
Multi-Rate Wideband (AMR-WB) codec, widely used in
VOLUME 6, 2018
current networks. In the talking state, a voice frame is gener-
ated every 20ms, while a Silence Insertion Descriptor (SID)
frame is sent every 160ms during silence periods. For sim-
plicity, we assume that AMR mode 1 is used throughout the
whole conversation. This baseline model provides a tractable
benchmark for performance comparison and other features,
e.g. TTI bundling, are left to future extensions.
2) ACCESS NETWORK
We consider the following cases:
S0-R0: no gDTX cycle is implemented, neither at the
•
sender side, nor at the receiver side (Reference scheme,
depicted by N = 0ms in following figures).
• S0-R1: gDTX cycles are used at the receiver side only
(named ‘‘downlink gDTX cycle’’ in the following). This
case also models calling from a landline.
• S1-R0: gDTX cycles are used at the sender side only
(named ‘‘uplink gDTX cycle’’). This case also models
receiving a phone call at a landline.
• S1-R1: gDTX cycles are implemented at both sides of
the phone call. In this paper, we consider equal values
for N and NT at both sides.
We further assume no synchronization of the phone call,
the downlink and the uplink gDTX cycles. To this end, ran-
dom call start times and gDTX cycle start times are simulated.
In the considered framework, only one phone call is simulated
and a packet can be sent immediately if it arrives during
the BS active period. Otherwise, it is sent at the beginning
of the next active period. Such assumption allows to isolate
the impact of gDTX on performance. It is valid for voice
with guaranteed QoS in low to moderate load and for non-
prioritized voice in light load (e.g. during nights or in rural
area). Performance analysis at higher load, in particular for
non-prioritized voice, is left to future investigation.
3) CORE NETWORK
Real packet route consists of a mixture of various networks,
using different technologies, passing through different equip-
ments. Each of them adds a certain degree of uncertainty
in the processing and transmitting times, such that packets
experience random delay, can arrive out of sequence, or may
be lost due to traffic congestion and buffer overflows. The
Generalized Pareto distribution is well suited to describe the
resulting overall network delay statistics [32], [35], [36].
For performance simulation, we consider different network
conditions, with a packets loss probability of 0.1%, an overall
propagation delay Dp of {10, 50, 75, 100}ms and a jitter J of
{0, 10, 20, 30}ms. The case J = 0ms allows to evaluate the
extra jitter generated by gDTX cycles solely.
4) RECEIVER DECODING
To obtain a steady stream out of the irregular packet flow
coming from the Internet, a jitter buffer is used at the receive
side where received packets are held and re-ordered before
being played out. A longer holding time, or jitter buffer size,
58177
 F. Parzysz, Y. Gourhant: Drastic Energy Reduction With gDTX in Low-Cost 5G Networks

FIGURE 6. Model for analyzing voice performance.

allows to overcome large network delay variations and to

reduce packet loss due to late arrivals. Yet, this also increases
the end-to-end delay. In this paper, we consider a fixed buffer
policy, for which network conditions are not tracked to adapt
the holding time. This latter is set to insure a packet loss no
larger than 5%.

B. SIMULATION RESULTS: VOICE QUALITY USING

gDTX CYCLES
For performance analysis, we simulate a large number of
phone calls, with random start times for the call and gDTX
cycles. Then, the achieved performance is averaged out.

1) CONSIDERED METRICS
Robust coding and error correction make voice more tolerant
to packet loss, but very sensitive to the end-to-end delay
De2e (or mouth-to-ear) and overall jitter. In addition to these FIGURE 7. Impact of gDTX cycles on the overall jitter.
two metrics, we estimate the perceived speech quality and
user satisfaction [37]. It is usually measured through the
0 for x < 0

Mean Opinion Score (MOS), which ranges from 1 (Bad) where H (x) = (3)
1 otherwise.
to 5 (Excellent). While subjective methods involve humans
listening and rating of a live stream or recorded file, E-models The equipment impairment factor Ie captures the effect of
have been developed to predict voice quality based on the signal distortion due packet loss and codec insufficient bit
measured network parameters and without referring to the rate. It is expressed as [39]:
original signals. Ppl
In this paper, we consider the wideband E-model standard- Ie = Ie,wb + (129 − Ie,wb ) , (4)
Ppl + Bpl
ized by ITU-T [38] to estimate the achieved MOS. The output
of this model is the R-factor, which is computed as in [39]: in which Bpl is a codec-specific factor. Ie,wb is given by
[39, Fig. 1] as a function of the packet loss Ppl . Following
1
R = (129 − Id − Ie ) × (1) an approach similar to [36], Ppl encompasses losses due to
1.29 both network impairments and limited-size jitter buffer.
and then, mapped to the MOS as follows:
for R < 0 2) PERFORMANCE ANALYSIS

 1
for R > 100


4.5 First, gDTX cycles are themselves a non-negligible source
MOS = (2) of jitter, as illustrated in Figure 7 for an ideal network (case

 1 + 0.035R + 7.10−6 R×

(R − 60)(100 − R) otherwise. J=0ms). When the inactivity duration N increases, pack-
ets are more likely received in bursts. Yet, gDTX cycles
In this, the delay impairment factor Id accounts for all
have some beneficial effect with non-zero network jitter.
delays experienced by voice and is given by [39]:
When N ≤ 10ms, the overall jitter is equal to the sum of
Id = 0.0024 De2e + 0.11 (De2e − 177.3) H (De2e − 177.3) the network-generated and gDTX-generated jitters but, for

58178 VOLUME 6, 2018

F. Parzysz, Y. Gourhant: Drastic Energy Reduction With gDTX in Low-Cost 5G Networks

FIGURE 8. Impact of gDTX cycles on the jitter buffer size. FIGURE 9. Mean Opinion score and network jitter.

higher N , the effect of network jitter diminishes and part of it

is absorbed by gDTX cycles.
A direct consequence is that gDTX cycles used in the
downlink act as jitter buffer. This is particularly visible
in Figure 8, for the case S0-R1 with J=30ms. By increasing
the inactivity period, the size of actual jitter buffer (at the
receiver side) can be reduced and tends to NT − 20ms,
reflecting the gDTX-generated jitter only.
The use of gDTX in the uplink raises more challenges,
especially for the case S1-R1 for which the overall jitter
results from the uplink gDTX, core network and downlink
gDTX. As observed in Figure 8 for J=30ms, the downlink
gDTX efficiently absorbs part of the uplink and core jitters
for N ≤ 50ms but, for higher values of N (i.e. both gNB are
sleeping most of the time), too many packets are delayed by
two gDTX cycles and the receiver buffer must be lengthened
FIGURE 10. Mean Opinion score and network propagation delay.
to maintain a packet loss of 5%. Nevertheless, a such increase
is quite reasonable and, compared to the reference scheme
N=0ms with J=30ms, does not exceed 35ms, even for two
gDTX cycles with N=75ms each. In good network conditions (low jitter and short propaga-
This leads us to state that the impact of gDTX on the tion delay), the inactivity period can be increased to 60ms.
overall jitter is generally well-handled by the receiver buffer Whereas the MOS degradation is larger when gDTX is
and that the main QoS impediment for voice is the increased used at both side (S1-R1), the voice quality remains accept-
end-to-end delay De2e . This latter encompasses the holding able (MOS≥ 3.5) in most cases, even for long inactivity
time in (a) the transmitter buffer, when packets arrives dur- (N '50/60ms). Consequently, we deduce that large network
ing the inactivity period of uplink gDTX (cases S1-R0 and propagation delay is the major hindrance for gDTX imple-
S1-R1), (b) the downlink gNB buffer (cases S0-R1 and mentation.
S1-R1), and (c) the receiver jitter buffer (all cases), which all We now verify the ability of gDTX cycles to handle traffic
add up to the network propagation delay. However, we show load and respectively plot in Figures 12 and 13 the Mean
that the degraded jitter and end-to-end delay are not necessar- Opinion Score and the percentage of PRB available for data,
ily perceived by users. with 50, 100 and 150 simultaneous voice users (Dp = 50ms
To this end, we plot in Figures 9 and 10 the simulated and J = 20ms). We observe that good MOS is maintained in
MOS for different values of J and Dp . We deduce from these loaded networks till 45ms of sleep for S1-R1 and till 60ms for
figures that implementing gDTX cycles at one side (either S0-R1, but in this case, the whole cell capacity is occupied
S1-R0 or S0-R1) is transparent to users for N up to 30ms, by voice packets. Given that traffic load is generally low
in typical network conditions (i.e. J ' 20ms and Dp ' in rural areas, as depicted in Figure 5a, slightly reducing
50/60ms), already allowing the gNB to enter SM3 for 20ms. the sleep duration allows to sustain both voice and data.

VOLUME 6, 2018 58179

 F. Parzysz, Y. Gourhant: Drastic Energy Reduction With gDTX in Low-Cost 5G Networks
FIGURE 11. Daily Mean Opinion Score achieved with the same simulation parameters as in Fig. 4.
FIGURE 12. Mean Opinion score with 50 (black), 100 (blue) and 150 (red)
simultaneous voice users.
FIGURE 13. Portion (in %) of PRB left for data, with 50 (black), 100 (blue)
and 150 (red) simultaneous voice users.
For example, 100 simultaneous voice users and 35% load can
be maintained while sleeping 30ms every 80ms.
As concluding observations, we map in Figure 11 the daily
power consumption of Figure 4 to the Mean Opinion Score
for both cases S0-R1 and S1-R1 (worst case), with J=20ms
and Dp =50ms. We observe that the large power gain achieved
by gDTX cycles does not significantly impair the voice
58180
quality as perceived by users for S0-R1. Only phone calls
from 1am to 7am may experience perceptible MOS degra-
dation. It is recommended in this case to reduce the inactivity
period to 60ms at most and use SM1 (i.e. Cell DTX) when
an empty subframe is scheduled during the gDTX activity
period.
VI. CONCLUSION AND FUTURE WORK
We have proposed in this paper a novel 5G Discontinuous
Transmission mode (gDTX), primarily dedicated to energy-
restricted networks as often encountered in developing coun-
tries. Taking advantages of the 5G ultra-lean design and
spaced synchronization signals, it operates on a duty cycle
basis and alternates between an active period, for signal-
ing broadcast and user service, and an inactive period, dur-
ing which the gNB enters idle state. Contrary to existing
approaches on sleep modes, the gNB does not wake up at
each packet or request arrival and does not rely on neigh-
boring gNB to offload traffic. Next, we have shown that
large energy savings can be expected from gDTX cycles and,
quite counter-intuitively, that it is also suitable for real-time
delay-non-tolerant voice traffic, in light to moderate load.
Such encouraging results calls for further investigations to
dynamically configure (N , NT ) for energy minimization and
measure the resulting power gain at users.
REFERENCES
[1] ‘‘MEA’s MNOs and towercos talk business,’’ TowerXchange, Tech. Rep.,
2017. [Online]. Available: www.towerxchange.com
[2] D. Marquet, M. Aubrée, S. Le Masson, A. Ringnet, P. Mesguich, and
M. Kirtz, ‘‘The first thousand optimized solar BTS stations of Orange
group,’’ in Proc. IEEE Int. Telecommun. Energy Conf. (INTELEC),
Oct. 2011, pp. 1–9.
[3] S. Zhang, N. Zhang, S. Zhou, J. Gong, Z. Niu, and X. Shen, ‘‘Energy-aware
traffic offloading for green heterogeneous networks,’’ IEEE J. Sel. Areas
Commun., vol. 34, no. 5, pp. 1116–1129, May 2016.
[4] T. Sigwele, A. S. Alam, P. Pillai, and Y. F. Hu, ‘‘Evaluating energy-efficient
cloud radio access networks for 5G,’’ in Proc. IEEE Int. Conf. Data Sci.
Data Intensive Syst., Dec. 2015, pp. 362–367.
[5] S. Zhou, T. Zhao, Z. Niu, and S. Zhou, ‘‘Software-defined hyper-cellular
architecture for green and elastic wireless access,’’ IEEE Commun. Mag.,
vol. 54, no. 1, pp. 12–19, Jan. 2016.
[6] G. Micallef, L. Saker, S. E. Elayoubi, and H.-O. Sceck, ‘‘Realistic energy
saving potential of sleep mode for existing and future mobile networks,’’
J. Commun., vol. 7, no. 10, pp. 740–748, 2012.
[7] J. Wu, Y. Zhang, M. Zukerman, and E. K. N. Yung, ‘‘Energy-efficient
base-stations sleep-mode techniques in green cellular networks: A sur-
vey,’’ IEEE Commun. Surveys Tuts., vol. 17, no. 2, pp. 803–826,
2nd Quart., 2015.
VOLUME 6, 2018
F. Parzysz, Y. Gourhant: Drastic Energy Reduction With gDTX in Low-Cost 5G Networks
[8] K. Hiltunen, ‘‘Utilizing eNodeB sleep mode to improve the energy-
efficiency of dense LTE networks,’’ in Proc. IEEE Annu. Int. Symp. Pers.,
Indoor, Mobile Radio Commun. (PIMRC), Sep. 2013, pp. 3249–3253.
[9] M. M. Mowla, I. Ahmad, D. Habibi, and Q. V. Phung, ‘‘A green communi-
cation model for 5G systems,’’ IEEE Trans. Green Commun. Netw., vol. 1,
no. 3, pp. 264–280, Sep. 2017.
[10] S. Gamboa, A. Pelov, P. Maillé, X. Lagrange, and N. Montavont, ‘‘Reduc-
ing the energy footprint of cellular networks with delay-tolerant users,’’
IEEE Syst. J., vol. 11, no. 2, pp. 729–739, Jun. 2017.
[11] N. Saxena, A. Roy, and H. Kim, ‘‘Traffic-aware cloud RAN: A key
for green 5G networks,’’ IEEE J. Sel. Areas Commun., vol. 34, no. 4,
pp. 1010–1021, Apr. 2016.
[12] ‘‘GreenTouch green meter research study: Reducing the net energy
consumption in communications networks by up to 90% by 2020,’’
GreenTouch White Paper, Tech. Rep., 2013. [Online]. Available:
https://s3-us-west-2.amazonaws.com/belllabs-microsite-greentouch/
uploads/documents/GreenTouch_Green_Meter_Research_Study_26_
June_2013.pdf
[13] H. Claussen, I. Ashraf, and L. T. W. Ho, ‘‘Dynamic idle mode procedures
for femtocells,’’ Bell Labs Tech. J., vol. 15, no. 2, pp. 95–116, Sep. 2010.
[14] I. Allal, S.-M. Senouci, J. Penhoat, and Y. Gourhant, ‘‘A new sustainable
mechanism to wake-up bast stations in mobile networks,’’ in Proc. Global
Inf. Infrastruct. Netw. Symp. (GIIS), Oct. 2016, pp. 1–6.
[15] S. Tombaz, P. Frenger, M. Olsson, and A. Nilsson, ‘‘Energy performance
of 5G-NX radio access at country level,’’ in Proc. IEEE 12th Int. Conf.
Wireless Mobile Comput., Netw. Commun. (WiMob), Oct. 2016, pp. 1–6.
[16] P. Frenger, P. Moberg, J. Malmodin, Y. Jading, and I. Gódor, ‘‘Reducing
energy consumption in LTE with cell DTX,’’ in Proc. IEEE 73rd Veh.
Technol. Conf. (VTC Spring), May 2011, pp. 1–5.
[17] S. Tombaz, S.-W. Han, K. W. Sung, and J. Zander, ‘‘Energy efficient
network deployment with cell DTX,’’ IEEE Commun. Lett., vol. 18, no. 6,
pp. 977–980, Jun. 2014.
[18] A. Karlsson, O. Al-Saadeh, A. Gusarov, R. V. R. Challa, S. Tombaz, and
K. W. Sung, ‘‘Energy-efficient 5G deployment in rural areas,’’ in Proc.
IEEE Int. Conf. Wireless Mobile Comput., Netw. Commun. (WiMob),
Oct. 2016, pp. 1–7.
[19] H. Holtkamp, G. Auer, S. Bazzi, and H. Haas, ‘‘Minimizing base sta-
tion power consumption,’’ IEEE J. Sel. Areas Commun., vol. 32, no. 2,
pp. 297–306, Feb. 2014.
[20] R. Gupta and E. C. Strinati, ‘‘Base-station duty-cycling and traffic buffer-
ing as a means to achieve green communications,’’ in Proc. IEEE Veh.
Technol. Conf. (VTC Fall), Sep. 2012, pp. 1–6.
[21] R. Wang, J. S. Thompson, and H. Haas, ‘‘A novel time-domain sleep mode
design for energy-efficient LTE,’’ in Proc. Int. Symp. Commun., Control
Signal Process. (ISCCSP), Mar. 2010, pp. 1–4.
[22] K. Abdallah, I. Cerutti, and P. Castoldi, ‘‘Energy-efficient coordinated
sleep of LTE cells,’’ in Proc. IEEE Int. Conf. Commun. (ICC), Jun. 2012,
pp. 5238–5242.
[23] B. Debaillie, C. Desset, and F. Louagie, ‘‘A flexible and future-proof
power model for cellular base stations,’’ in Proc. IEEE Veh. Technol. Conf.
(VTC Spring), May 2015, pp. 1–7.
[24] X. Guo, Z. Niu, S. Zhou, and P. R. Kumar, ‘‘Delay-constrained energy-
optimal base station sleeping control,’’ IEEE J. Sel. Areas Commun.,
vol. 34, no. 5, pp. 1073–1085, May 2016.
[25] Summary of Performance Evaluation Results for Standalone NCT,
document 3GPP R1-133 976, Ericsson, ST-Ericsson, Huawei, Shenzhen,
China, Aug. 2013.
[26] E. Farag, ‘‘5G RAN features and capabilities,’’ Nokia Bell Labs, Murray
Hill, NJ, USA, Tech. Rep. ATIS 3GPP Webinar, 2017.
[27] NR; Physical Layer Procedures for Control (Release 15), document TS
38.213, 3GPP, Dec. 2017.
[28] Long Network DTX in NR, document R1-161900, 3GPP, Nov. 2016.
[29] F. E. Salem, A. Gati, Z. Altman, and T. Chahed, ‘‘Advanced sleep modes
and their impact on flow-level performance of 5G networks,’’ in Proc.
IEEE 86th Veh. Technol. Conf. (VTC-Fall), Sep. 2017, pp. 1–7.
[30] P. Lähdekorpi, M. Hronec, P. Jolma, and J. Moilanen, ‘‘Energy efficiency
of 5G mobile networks with base station sleep modes,’’ in Proc. IEEE Conf.
Standards Commun. Netw. (CSCN), Sep. 2017, pp. 163–168.
VOLUME 6, 2018
[31] ‘‘LTE1891—eNB power saving—Micro DTX,’’ Nokia Academy,
Tech. Rep., 2016.
[32] J. F. M. R. Cardoso, ‘‘Implementation of VoLTE considering different
QoS and QoE overall approaches,’’ M.S. thesis, Instituto Superior Técnico,
Lisboa, Portugal, Nov. 2017.
[33] F. Khan, LTE for 4G Mobile Broadband: Air Interface Technologies and
Performance. Cambridge, U.K.: Cambridge Univ. Press, 2009.
[34] X. Zhang, LTE Optimization Engineering Handbook, Hoboken, NJ, USA:
Wiley, 2018.
[35] L. Ding and R. A. Goubran, ‘‘Speech quality prediction in VoIP using
the extended E-model,’’ in Proc. IEEE Global Telecommun. Conf.
(GLOBECOM), vol. 7, Dec. 2003, pp. 3974–3978.
[36] A. Kovac, M. Halas, M. Orgon, and M. Voznak, ‘‘E-model MOS estimate
improvement through jitter buffer packet loss modelling,’’ Adv. Electr.
Electron. Eng., vol. 9, no. 5, pp. 233–242, 2011.
[37] M. Anehill, M. Larsson, G. Strömberg, and E. Parsons, ‘‘Validat-
ing voice over LTE end-to-end,’’ Ericsson, Stockholm, Sweden, Tech.
Rep., 2012. [Online]. Available: https://pdfs.semanticscholar.org/5b94/
4b6ddd849416efccc9a8bf35f017217257c6.pdf
[38] Wideband E-Model, document G.107.1, ITUT, Geneva, Switzerland, 2011.
[39] D.-H. Nguyen and H. Nguyen, ‘‘A dynamic rate adaptation algorithm
using WB E-model for voice traffic over LTE network,’’ in Proc. Wireless
Days (WD), Mar. 2016, pp. 1–6.
FANNY PARZYSZ received the Ph.D. degree from
the École de Technologie Supérieure, Montreal,
Canada, in 2015, and the M.Sc.Eng. degree from
Télécom ParisTech, France, in 2009. She was
a Post-Doctoral Researcher at the University of
Barcelona, within the Marie Skłodowska-Curie
Program, and at the Universitat Pompeu Fabra,
Barcelona, Spain. She recently joined the
Architecture and uRbanism for teleComs Team,
Orange Labs, Lannion, France. Her interest covers
various aspects of wireless communication, with a particular focus on
energy-efficient technologies for next-generation radio access networks,
including energy harvesting, relaying, and performance modeling using
stochastic geometry tools.
YVON GOURHANT received the Ph.D. degree
in computer science from inria in 1991. After
15 months in the DSG Group, Trinity College
Dublin, he joined the R&D Division, Orange Labs,
Lannion (Brittany in France) in 1993, still involved
in distributed systems. From 1998 to 2009, he
was leading a R&D Team on Programmable and
Adaptable Networks. In 2007 and 2008, he was
also leading and acting in a R&D Program on
Light Infrastructure Networks that encompass ad
hoc/mesh networks, self-organized sensor networks, and vehicular networks.
Since 2009, he has been leading and acting in a research program on low-cost
network infrastructures that take into account emerging countries constraints
(e.g., low ARPU, energy efficiency, and infrastructure less networks). From
2014 to 2017, he has lead also the LCI4D collaborative project on low-cost
cellular infrastructure networks for accelerating broadband penetration in
emerging countries.
58181