Bonfante, A., Giordano, L. G., López-Pérez, D., Garcia-Rodriguez, A., Geraci, G.,
Baracca, P., Butt, M. M., Dzaferagic, M. and Marchetti, N. (2019) Performance of
Massive MIMO Self-Backhauling for Ultra-Dense Small Cell Deployments. In: IEEE
GLOBECOM 2018, Abu Dhabi, United Arab Emirates, 09-13 Dec 2018, ISBN
9781538647271.
There may be differences between this version and the published version. You are
advised to consult the publisher’s version if you wish to cite from it.
http://eprints.gla.ac.uk/166355/
Deposited on: 9 August 2018
Enlighten – Research publications by members of the University of Glasgow
http://eprints.gla.ac.uk
Performance of Massive MIMO Self-Backhauling
for Ultra-Dense Small Cell Deployments
Andrea Bonfante
∗†
, Lorenzo Galati Giordano
∗
, David L
´
opez-P
´
erez
∗
, Adrian Garcia-Rodriguez
∗
, Giovanni Geraci
∗
,
Paolo Baracca
§
, M. Majid Butt
‡
, Merim Dzaferagic
†
, and Nicola Marchetti
†
∗
Nokia Bell Labs Ireland and
§
Nokia Bell Labs Germany
†
CONNECT Centre, Trinity College Dublin, Ireland
‡
University of Glasgow, United Kingdom
Abstract—A key aspect of the fifth-generation wireless com-
munication network will be the integration of different services
and technologies to provide seamless connectivity. In this pa-
per, we consider using massive multiple-input multiple-output
(mMIMO) to provide backhaul links to a dense deployment of
self-backhauling (s-BH) small cells (SCs) that provide cellular
access within the same spectrum resources of the backhaul.
Through a comprehensive system-level simulation study, we
evaluate the interplay between access and backhaul and the
resulting end-to-end user rates. Moreover, we analyze the impact
of different SCs deployment strategies, while varying the time
resource allocation between radio access and backhaul links.
We finally compare the above mMIMO-based s-BH approach
to a mMIMO direct access (DA) architecture accounting for
the effects of pilot reuse schemes, together with their associated
overhead and contamination mitigation effects. The results show
that dense SCs deployments supported by mMIMO s-BH provide
significant rate improvements for cell-edge users (UEs) in ultra-
dense deployments with respect to mMIMO DA, while the latter
outperforms mMIMO s-BH from the median UEs’ standpoint.
Index Terms—Integrated access and backhaul, heterogeneous
network, massive MIMO-based backhaul, network capacity.
I. INTRODUCTION
Fifth-generation (5G) wireless communication systems are
expected to support a 1000x increase in capacity compared
to existing networks [1]. Meeting this gargantuan target will
require mobile network operators (MNOs) to leverage new
technologies such as massive multiple-input multiple-output
(mMIMO), and deploy a large number of additional small
cell base stations [2], [3]. Wireless self-backhauling (s-BH),
achieved through the tight integration of these two comple-
mentary means, lures MNOs with the potential of achieving
the desired capacity boost at a contained investment [4].
Indeed, exploiting the large number of spatial degrees-of-
freedom available with mMIMO to provide sub-6 GHz in-
band wireless backhauling to small cells (SCs) offers multiple
advantages to MNOs: avoiding deployment of an expensive
wired backhaul infrastructure, availing of more flexibility in
A. Bonfante was funded by the Irish Research Council and Nokia Ireland
Ltd under the grant EPSPG/2016/106. This publication has emanated from
research supported in part by a grant from Science Foundation Ireland (SFI)
and is co-funded under the European Regional Development Fund under
Grant Number 13/RC/2077.
the deployment of SCs, and not having to purchase additional
licensed spectrum as in the case of out-of-band wireless
backhauling [5].
The integration of radio access and backhauling – advocated
in s-BH solutions – has been addressed by the Third Gener-
ation Partnership Project (3GPP) with a list of requirements
detailed in [6]. At the same time, various research efforts have
tackled the problem of resource allocation and management
in s-BH networks in the time and frequency domains [7], [8].
Finally, the combination of mMIMO spatial multiplexing and
s-BH has been considered in [9]–[11] and also in [12]–[14],
albeit for full-duplex scenarios.
In this paper, we analyze the end-to-end user equipment
(UE) performance of mMIMO s-BH networks. In particular,
we consider a realistic multi-cell setup where mMIMO base
stations (mMIMO-BSs) provide sub-6 GHz backhauling to a
plurality of half-duplex SCs overlaying the macro cellular
area. We evaluate the UE data rates achieved through s-BH
in two ultra-dense deployment scenarios, namely a random
deployment – where SCs are uniformly distributed over a
geographical area –, and an ad-hoc deployment – where
SCs are purposely positioned close to UEs to achieve line-
of-sight (LOS) access links. In these half-duplex systems, a
s-BH approach entails sharing time-and-frequency resources
between radio access and backhaul links. To the best of the
authors knowledge, in this paper, we also compare for the first
time the performance of the mMIMO s-BH approach and that
of a direct access (DA) approach, where mMIMO-BSs are
solely dedicated to serving UEs in the absence of SCs [15].
Our study provides a number of key takeaways:
• Adding more randomly deployed SCs – where SCs are
randomly placed – provides limited gains for the end-to-
end UE rates. The UEs benefit from more radio resources
allocated and from the SCs proximity, given by the higher
probability of the UEs to be close to – and in LOS with
– the serving SCs. However, the UEs are affected by a
significantly higher inter-cell interference because they
see a growing number of interfering links in LOS con-
ditions. Overall, the detrimental impact of interference
outweighs the combination of the gains provided by more
mMIMO-BS
UE
s-BH
small cell
mMIMO-BS
UE
UE
mMIMO-BS
UE
UE
UE
(a) Random deployment
mMIMO-BS
UE
d
θ
s-BH
small cell
mMIMO-BS
UE
UE
mMIMO-BS
UE
UE
UE
(b) Ad-hoc deployment
Fig. 1: Examples of two different SCs deployments considered in the paper.
radio resources and proximity, preventing the UEs to take
the advantages of the SCs dense deployment.
• Adding ad-hoc deployed SCs – where SCs are placed in
proximity to UEs – provides higher data rates, thanks to
a high signal-to-interference-plus-noise ratio (SINR) on
the access link, given by the higher proximity gains with
respect to the random deployment.
• Partitioning resources between wireless access and back-
haul links is of paramount importance. Indeed, the end-
to-end performance is sensitive to said partition, and
optimal rates can only be achieved through a carefully
designed tradeoff.
• Unlike mMIMO s-BH – where mMIMO-to-SC links
are static, and thus channel acquisition is facilitated
– mMIMO DA suffers more from pilot overhead and
contamination. Indeed, when compared to mMIMO DA
solutions with pilot reuse 3 and reuse 1, ultra-dense SCs
deployments supported by mMIMO s-BH provide rate
improvements for cell-edge UEs that amount to 30% and
a tenfold gain, respectively. On the other hand, mMIMO
DA outperforms s-BH from the median UEs’ standpoint.
Notation: Capital and lower-case bold letters denote matri-
ces and vectors, respectively, while [·]
∗
, [·]
T
and [·]
H
denote
conjugate, transpose, and conjugate transpose, respectively.
II. SYSTEM MODEL
As shown in Fig. 1, we focus on the study of the down-
link (DL) performance for a two-tier heterogeneous network
formed by mMIMO-BSs overlaying a layer of self-backhauled
SCs. The mMIMO-BSs are connected to the core network
through a high-capacity wired connection, while all SCs
receive backhaul traffic through mMIMO-BSs and function
as access points for UEs. We consider a self-backhaul con-
figuration, where mMIMO-BSs are solely dedicated for the
backhaul, while SCs are solely dedicated for the access. For
comparison purposes, we also consider the conventional DA
approach where each mMIMO-BS directly serves the UEs.
A. Macro cell and user topologies
We denote by I the set of mMIMO-BSs placed in a uniform
hexagonal grid with three sectors per site. Each mMIMO-BS
i, is equipped with a large number of antennas M, and serves
L
i
single-antenna SCs. Furthermore, we denote by K
i
the
number of UEs randomly and uniformly distributed over the
sector’s area, and let k denotes single-antenna UE. We assume
that each UE is connected with the SC (in the s-BH approach)
or with the mMIMO-BS (in the DA approach) that provides
the largest reference signal received power (RSRP) [16].
B. Small cell topologies
We denote by L
i
the set of SCs deployed per sector
and connected to the i-th mMIMO-BS that provides the
largest RSRP. Each SC connects K
l
UEs. Two different SCs
deployments are presented in the following:
(a) Random deployment: Self-backhauled SCs are ran-
domly and uniformly distributed over the mMIMO-BS
geographical area as shown in Fig. 1a. This scenario is
used as a baseline and follows the set of parameters
specified by the 3GPP in [16] to evaluate the relay
scenario.
(b) Ad-hoc deployment: Self-backhauled SCs are posi-
tioned targeting nearby UE locations. This scenario is
used as an example of ultra-dense network deployment.
We assume the possibility to realize this target of network
deployment, for example by means of drone-BSs, where
the drone-BSs can reposition following the locations
of UEs [17].
1
As shown in Fig. 1b, we model this
scenario by considering SCs deployed within a 2-D
(two-dimensional) distance d of the UEs, and an angle
θ measured from the straight segment that links UEs
and their closest mMIMO-BS. θ is chosen uniformly at
random from −π/2 and π/2. It is worth noting that even
when the 2-D distance
d
= 0
, UEs and SCs are still
separated in space because the antennas are positioned
at different heights. More precisely, they are assumed
located at 1.5 meters and 5 meters above the ground, for
the UEs and the SCs, respectively [16].
With a dense deployment of SCs, the UE SINRs are
severely affected by the strong inter-cell interference among
1
Although mentioned, the drone-BSs use-case is not the focus of this paper
and it is left for future investigation, since the height of the outdoor SC
antenna is fixed to 5 meters, and we use the channel models adopted for the
relay study [16].
SCs. In addition, to limit the effect of the inter-cell interfer-
ence, with the ad-hoc deployment, we propose to replace at
the SC the Patch antenna with a more directive Yagi antenna
pointing downwards to the ground (as shown by the green
radiation cone in Fig. 1b), and therefore only illuminating the
closest UEs: details about this modeling can be found in Table
I.
C. Frame structure
As shown in Fig. 2a, we consider the time-slot T as a
single scheduling unit in the time domain, and we partition the
access and backhauling resources through the parameter α ∈
[0, 1]. Therefore, α time-slots are allocated to the backhaul
links, while 1 −α time-slots are allocated to the access links.
In the frequency domain, we divide the system bandwidth
B into Q
t
resource blocks (RBs), and we allocate all the
RBs to the backhaul links or the access links. We make the
following assumptions in considering the partition of backhaul
and access time-slots among the SCs and UEs:
• During the backhaul time-slots, all the associated SCs are
served by the mMIMO-BS i, and we use the same value
of α for all the SCs. In this approach, the mMIMO-BSs
precode the backhaul signals towards the single-antenna
SCs, which are spatially multiplexed in the same time-
frequency resources by allocating all the RBs in T to
each SC.
• During the access time-slots, the SCs schedule their
connected UEs by using a Round Robin (RR) mechanism
as frequency domain scheduler. This approach entails that
each SC equally shares the system bandwidth B with its
UEs.
In the Fig. 2b, it is shown the frame structure used for
the DA setup, where all the time-slots are allocated to the
access links. In each time slot, the mMIMO BSs precode
the access signals, and the UEs are spatially multiplexed on
the entire system bandwidth. Figs. 2a and 2b also show the
fraction τ of the time-slots dedicated for the transmission
of the pilot sequences, used to estimate the massive MIMO
channel. Details about the channel training procedure will be
discussed in Section III.
D. Channel model
We define as h
il
= [h
il1
, . . . , h
ilM
]
T
∈ C
M
the propaga-
tion channel between the l -th single-antenna receiver (SC in
the s-BH architecture and UE in the mMIMO DA) and the
M antennas of the i-th mMIMO-BS. The composite channel
matrix between the i-th mMIMO-BS and the devices in the
i
′
-th cell is represented by H
i,i
′
= [h
i1
···h
iL
i
′
] ∈ C
M×L
i
′
.
Since all the RBs are assigned to each SC, we removed the
RB index q from the massive MIMO channel notation.
Furthermore, we define as g
lkq
∈ C the single-input single-
output (SISO) channel between the l-th SC and the k-th UE
in the q-th RB. Each channel coefficient h
ilm
=
√
β
il
˜
h
ilm
,
and g
lkq
=
√
β
lk
˜g
lkq
accounts for both the effects of a large
scale fading and a small scale fading components:
DL Data
UL
Pilot
DL Data
Access
Backhaul
macro
layer
small cell
layer
UL
Pilot
Self-Backhaul Frame
τ
DL DataDL Data
Access
time-slot (T)
Backhaul
time-slot (T)
(a)
small cell
layer
macro
layer
Direct Access Frame
DL Data
UL
Pilot
DL Data
UL
Pilot
DL Data
UL
Pilot
DL Data
UL
Pilot
Access
time-slot (T)
τ
Backhaul
(b)
Fig. 2: DL frame structure for mMIMO s-BH with α = 0.5 (Fig.
2a) and for mMIMO DA (Fig. 2b).
• The large fading components β
il
, β
lk
∈ R
+
have been
modeled by using a combined LOS/Non-LOS (NLOS)
path loss model, which accounts for the shadowing effect,
set to be log-normal distributed with different standard
deviations [16]. Because of its slow-varying characteris-
tic, it does not change rapidly with time, and it can be
assumed constant over the observation time-scale of the
network.
• The small scale fading components
˜
h
ilm
, ˜g
lkq
∈ C,
which results from multi-path, have been modeled as a
Rician fast-fading, which rapidly changes over time and
frequency. For the LOS channels, we characterize the
Rician K factor with the model: K[dB] = 13 − 0.03r
in dB, where r is the distance between transmitter and
receiver in meters [18].
Throughout the paper, we assume a composite fading (i.e.
large scale fading and small scale fading together) for the
SC-UE and the mMIMO-BS-UE links (in the DA approach),
which changes between successive time-slots and between
different RBs. Moreover, because of the static position of the
SCs, we consider that the backhaul channel SC-mMIMO-BS
remains constant for a period T
BH
≫ T .
III. END-TO-END UE RATES
In this section, we provide the detailed description of the
operations required for the DL transmission in the mMIMO s-
BH approach and in the mMIMO DA approach. We describe
the channel training procedure, the mMIMO DL backhaul
transmission, and the DL access transmission, which is treated
separately below for both the s-BH and the mMIMO DA
setups.
A. Massive MIMO channel training
To calculate the DL precoder, we consider that the channel
is estimated through uplink (UL) pilots, assuming UL/DL
channel reciprocity [2]. We also assume that the SCs or UEs
associated to the same mMIMO-BS have orthogonal pilot
sequences, and define the pilot code-book with the matrix
Φ
i
= [ϕ
i1
···ϕ
iL
i
]
T
∈ C
L
i
×S
, which satisfies Φ
i
Φ
H
i
= I
L
i
.
Here, the l-th sequence is given by ϕ
il
= [ϕ
il1
, . . . , ϕ
ilS
]
T
∈
C
S
, and S denotes the pilot code-book length. Note that
L
i
≤ S, i.e., the maximum number of SCs or UEs served by
the mMIMO-BSs in a time-slot is limited by the number of
orthogonal pilot sequences. The matrix Y
i
∈ C
M×S
of pilot
sequences received at the i-th mMIMO-BS can be expressed
as [19]
Y
i
=
P
ul
il
i
′
∈I
H
i,i
′
Φ
i
′
+ N
i
, (1)
where P
ul
il
is the power used by the l-th device located in
the i-th sector for UL pilot transmission, and N
i
∈ C
M×S
represents an additive noise, and is modeled with independent
and identically distributed complex Gaussian random variable.
Let H
i
denote the channel between the i-th mMIMO-
BS and the UEs located in the same sector. During the UL
training phase, the mMIMO-BS obtains an estimate of H
i
by correlating the received signal with a known pilot matrix
Φ
i
. Let us define P ⊆ I as the subset of sectors, whose
UEs share identical pilot sequences with the UEs served by
the i-th mMIMO-BS. The resulting estimated channel can be
expressed as
H
i
=
1
P
ul
il
Y
i
Φ
H
i
= H
i
+
i
′
∈P
H
i,i
′
+
1
P
ul
il
N
i
Φ
H
i
. (2)
The first, second and third terms on the right hand side of (2)
represent the estimated channel, a residual pilot contamination
component and the noise after the correlation, respectively.
The use of the same set of orthogonal pilot sequences among
different sectors leads to the well-known pilot contamination
problem, which can severely degrade the performance of
mMIMO systems [2], [20]. In this paper, we assume that no
pilot contamination occurs for the mMIMO s-BH system. Due
to the longer coherence time of the static backhaul channel,
T
BH
, with respect to the system time-slot, T, mMIMO pilots
do not need to be transmitted in every time-slot dedicated
to backhauling, thus allowing higher reuse factors with fully
orthogonality over the entire network. In contrast, for mMIMO
DA this assumption does not hold and, in this paper, we
consider that a maximum of 16 orthogonal pilot sequences
can be multiplexed in a single orthogonal frequency division
multiplexing (OFDM) symbol [20]. In both mMIMO s-BH
and mMIMO DA, the overhead associated to the UL training
phase are considered and measured in terms of number of
OFDM symbols τ. Two pilot allocation schemes are here
compared:
• Pilot reuse 1 scheme (R1): All K
i
UEs per sector are
trained in τ = 1 OFDM symbol.
• Pilot reuse 3 scheme (R3): The sectors of the same site
use orthogonal pilot sequences. This scheme avoids pilot
contamination from co-sited sectors, but requires τ =
3 OFDM symbols, resulting in a higher pilot overhead
when compared to the R1 scheme.
B. Massive MIMO s-BH DL transmission
The i-th mMIMO-BS uses the precoding matrix W
i
=
[w
i1
···w
iL
i
] ∈ C
M×L
i
to serve its connected UEs during
the DL data transmission phase. In this paper, we consider
that W
i
is computed based on the zero-forcing (ZF) criterion
as
W
i
= D
i
1
2
H
i
H
H
i
H
i
−1
. (3)
Here, the diagonal matrix D
i
= diag (ρ
i1
, ρ
i2
, . . . , ρ
iL
i
) is
chosen to equally distribute the total DL power P
dl
i
among
the L
i
receivers. In the previous expression, ρ
il
represents the
power allocated to the l-th receiver located in the i-th sector,
and Tr{D
i
} = P
dl
i
, where Tr{D
i
} is the trace of matrix D
i
.
The SINR of the l-th DL stream transmitted by the i-th
mMIMO-BS can be expressed as
SINR
il
=
ρ
il
|h
H
il
w
il
|
2
j∈L
i
j=l
ρ
ij
|h
H
il
w
ij
|
2
+
i
′
∈I
i
′
=i
j∈L
i
′
ρ
i
′
j
|h
H
i
′
l
w
i
′
j
|
2
+ σ
2
n
.
(4)
The numerator of (4) contains the power of unit-variance
signal intended for the l-th receiver, while the denominator
includes the co-channel interference from the serving i-th
mMIMO-BS, the inter-cell interference from other mMIMO-
BSs, and the power of the thermal noise at the receiver
σ
2
n
.
The corresponding DL backhauling rate at the l-th SC
receiver can therefore be expressed as
R
BH
il
=
1 −
τ
T
B log
2
(1 + SINR
il
) . (5)
C. Small cell DL transmission
We recall from the channel model that g
lkq
denotes the
SISO channel between the l-th SC and the k-th UE corre-
sponding to the q-th RB. The SINR of the k-th UE served by
the l-th SC in RB q can be expressed as
SINR
lkq
=
P
dl
l
|g
lkq
|
2
i∈I
l
′
∈L
i
l
′
=l
P
dl
l
′
|g
l
′
kq
|
2
+ σ
2
n
2
, (6)
where P
dl
l
and P
dl
l
′
are the transmit powers on the RB of the
l-th and l
′
-th SCs, respectively, and σ
2
n
2
denotes the thermal
noise power at the UE receiver. The DL access rate for UE k
served by SC l can be therefore expressed as
R
AC
lk
=
B
Q
t
Q
t
q =1
x
k
q
log
2
(1 + SINR
lkq
) , (7)
where x
k
q
= 1 if the q-th resource block is assigned to the
k-th user, and x
k
q
= 0 otherwise. The aggregated DL access
rate provided by the l-th SC is R
AC
l
=
K
l
k=1
R
AC
lk
. The actual
aggregated DL access rate provided by the l-th SC depends on
the backhaul DL rate, which entails that R
AC
l
≤ R
BH
il
, ∀l ∈
L
i
, and ∀i ∈ I. In this paper, we assume that the backhaul
capacity is equally divided between the K
l
UEs served by the