Received 24 February 2023, accepted 7 March 2023, date of publication 10 March 2023, date of current version 21 March 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3255828

Study of Clustering Solutions for Scalable

Cell-Free Massive MIMO
DANAISY PRADO-ALVAREZ 1 , DANIEL CALABUIG 1 , (Member, IEEE),
JOSE F. MONSERRAT 1 , (Senior Member, IEEE), SAMER BAZZI 2 ,
AND WEN XU2 , (Senior Member, IEEE)
1 Instituto de Telecomunicaciones y Aplicaciones Multimedia (iTEAM), Universitat Politècnica de València, 46022 Valencia, Spain
2 Huawei Technologies Duesseldorf GmbH, 80992 Munich, Germany
Corresponding author: Danaisy Prado-Alvarez ([email protected])

ABSTRACT In response to the requirements of 5G and beyond, cell-free systems have emerged to ensure
uniform service throughout the area. However, the idea of having no cells leads to a considerably large
monolithic system, which is not scalable. In this context, common sense suggests the creation of clusters
of Access Points (APs) and User Equipments (UEs) that allow the system to be managed locally to some
extent. The clustering procedure has to ensure certain performance for the global system and, at the same
time, ensure its scalability. In this paper, first, scalability is analyzed. Subsequently, we study and compare
two clustering approaches for uplink and downlink. Finally, we propose a combination of clustering and
resource allocation techniques that outperforms the rest of the analyzed state-of-the-art solutions.

INDEX TERMS Cell-free massive multiple-input multiple-output (MIMO), scalability, clustering.

I. INTRODUCTION base-band (BB) unit. On the other hand, a Central Processing

The performance of systems based on a cell-centric design Unit (CPU) is in charge of the coordination among all APs.
has always been limited by the interference experienced by Some challenges arise when implementing the canon-
UEs at cell borders. New applications for Fifth Generation ical form of cell-free massive MIMO described before.
(5G) and beyond require seamless deployments to ensure First, since all APs must be connected to the same CPU
uniform service over the entire area. A natural solution is the via front-haul, the system has to withstand a considerable
use of user-centric cell-free systems. In this context and con- amount of traffic and signaling, which increases the compu-
sidering massive Multiple-Input Multiple-Output (MIMO) as tational requirements of the CPU and the front-haul capacity
one of the enabling technologies of next-generation systems, demands. This problem can be tackled via a user-centric
cell-free massive MIMO [1] was proposed to overcome the approach, so that a given UE is served by a subset of APs
high interference levels at the cell borders. close to it that share a CPU, helping to relax the front-
Normally, when talking about cell-free massive MIMO, haul performance requirements significantly [2]. In [3], the
and due to the large number of antennas at the network authors described four network deployments going from the
side, the APs acquire the Channel State Information (CSI) conventional cellular network to a cell-free massive MIMO
through the use of uplink pilot signals sent by the UEs, network considering a user-centric approach. In this work,
and then this CSI is used in the design of precoders. To be several practical aspects are analyzed such as the cost and
able to use this CSI in the downlink, Time Division Duplex- complexity of the deployment, limited capacity of back/front-
ing (TDD) mode is considered. In general, at the APs, the haul connections, and network synchronization. It is impor-
signal is multiplexed/de-multiplexed, converted by the ana- tant to remember that the objective of cell-free massive
log/digital and digital/analog converters, and processed by the MIMO is to reduce the interference at the cell borders of
traditional cellular systems. Therefore, the subset of APs for
The associate editor coordinating the review of this manuscript and each UE has to be selected taking this objective into account.
approving it for publication was Pietro Savazzi . In this line, it is possible to identify two different clustering
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 11, 2023 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 26703
 D. Prado-Alvarez et al.: Study of Clustering Solutions for Scalable Cell-Free Massive MIMO

approaches depending on how interference is reduced. The (i) decoding the signal of the first UE considering the signal
first approach creates a different cluster for each UE, locating of the rest as interference, (ii) subtracting the signal decoded
the UEs at the center of their clusters. In this case, the cluster from the first UE from the received signal, (iii) decoding
border is distanced as much as possible. The second approach the signal of the second UE considering the signal from the
includes many UEs in the same cluster and uses interference third UE onwards as interference, and so on. This reception
cancellation techniques to reduce the interference. technique is known to be capacity-achieving if all the sources
In this work, we will analyze these two approaches for of interference have a known covariance. Specifically, since
uplink and downlink under different scenarios in order to the decoding order does not affect the sum rate [5], we assume
figure out which would be the most suitable in terms of that the first signal to be decoded is the one corresponding
achievable data rate under certain circumstances. Based on to the UE with the highest index, i.e., the UE k̃(m) of the
the obtained results, we will propose a combination of clus- m-th cluster, and then we follow the decreasing order of the
tering and resource allocation techniques that outperforms the indices. Therefore, the achievable data rate of the k-th UE in
current solutions in the literature. the uplink with an identity noise covariance is as in (1), shown
The reminder of this paper is organized as follows. at the bottom of the page, where P is the available power at the
In Section II, the system model is presented. Afterward, the UEs. In (1), the interference has been divided into two terms
scalability concept is analyzed in Section III. Section IV according to its origin. In the numerator, the first term is the
describes the two previously mentioned clustering approa- noise covariance; the second term includes the interference
ches. The performance of particular clustering solutions from from UEs in clusters with indices m < m̃(k), the interference
the two approaches with the corresponding resource alloca- from the UEs in the same cluster, and the signal of the
tion techniques is evaluated in Section V for the uplink, and k-th UE; and the third term includes the interference from the
in Section VI for the downlink. Finally, Section VII draws the UEs in clusters with indices m > m̃(k). In the denominator,
conclusion. the terms are similar except for the fact that the signal of the
k-th UE is not included.
II. SYSTEM MODEL Analogously, in the case of the downlink, Dirty Paper Cod-
In this work, a cell-free massive MIMO system with clus- ing (DPC) [6] is considered. This coding technique consists
ters is considered. In particular, we define the system S = of adapting the codebooks of the transmitter to a known
{{Lm }Mm=1 , {Km }m=1 } as a system with M clusters of single-
M
interference in such a way that the interference does not affect
antenna APs and UEs. The APs and UEs in the m-th cluster the achievable rate. Therefore, the CPU of the AP cluster can
are those in Lm and Km , respectively. We assume that the UEs (i) encode the signal of the first UE, (ii) treat this signal as a
are exactly in one of those clusters, although APs can be in known interference for the second UE, and so on. In this case,
several clusters, i.e., Km ∩ Kn = ∅ for all m ̸ = n, but this the first UE sees the signal from the other UEs as interference,
is not necessarily true for any pair of sets in {Lm }M m=1 . The the second UE sees the signal from the third UE onwards
amount of APs and UEs in the system is L = | ∪M m=1 Lm | and as interference, etc. As in the case of SIC, this encoding
K = | ∪M m=1 mK |, respectively. technique is known to be capacity-achieving if all the sources
Let m̃(k) be the cluster the k-th UE belongs to. Without of interference have a known covariance. Some works found
loss of generality, we index the UEs in the same cluster an uplink-downlink duality when SIC and DPC are used [7].
continuously, i.e., the UEs in the m-th cluster are those with In particular, the achievable rates are the same. In [7], it is
indices k̃(m − 1) + 1, k̃(m − 1) + 2, . . . , k̃(m), where k̃(m) is shown that the DPC encoding order that achieves certain
the largest index of the UEs in the m-th cluster, and k̃(0) = 0. achievable rates is the reversed SIC decoding order of the dual
The channel between the k-th UE and the APs of the m-th uplink system. Following this principle, we assume that the
cluster is denoted by Hkm ∈ C|Lm | , where |Lm | is the amount encoding order is the reversed order considered for the uplink,
of APs in the m-th cluster. i.e., the encoding follows the increasing order of the UEs’
Interference cancellation techniques are used for the set indices. Therefore, the achievable data rate of the k-th UE
of UEs in each cluster. In particular, some interference can- in the downlink with an identity noise covariance is as in (2),
cellation is performed between the UEs 1, 2, . . . , k̃(1), also shown at the bottom of the next page, where Qk is the transmit
between the UEs k̃(1) + 1, . . . , k̃(2), etc. covariance matrix of the k-th UE. In (2), the interference has
As a performance measure of the system, the total sum data also been divided into two terms according to its origin, and
rate is calculated for both uplink and downlink. In the case in a similar manner as in (1). Since single-antenna UEs are
of the uplink, Successive Interference Cancellation (SIC) [4] considered, it is not needed to include determinants in the
is considered in the decoding. This algorithm consists of equation.

Pk ∗
PK
|I + P i=1 Him̃(k) Him̃(k) + P H H∗ |
i=k̃(m̃(k))+1 im̃(k) im̃(k)
RUL
k = log . (1)
+ P i=1 Him̃(k) Hi∗m̃(k) + P K
Pk−1
H∗ |
P
|I H
i=k̃(m̃(k))+1 im̃(k) im̃(k)

26704 VOLUME 11, 2023

D. Prado-Alvarez et al.: Study of Clustering Solutions for Scalable Cell-Free Massive MIMO

III. ANALYSIS OF THE SCALABILITY the amount of UEs, specially considering deployments with
Cell-free massive MIMO systems in their canonical form rely more APs than UEs like in cell-free massive MIMO systems.
on the cooperation of the L APs to serve the K UEs, i.e., Second, real systems have access control mechanisms to
M = 1. This implies that, as the network grows, so do the avoid this situation. In this sense, we wonder whether it is
computational requirements of the network and the volume reasonable to criticize the scalability of the solution in [3]
of data and signaling that has to be exchanged between each because the amount of UEs served by an AP is not explicitly
AP and each UE. Due to this, these systems are said to be limited.
non-scalable. It is therefore important to find solutions with
superior performance and good scalability with the network B. PROPOSED DEFINITION
size. To do this, it is important to agree on the main features From our point of view, the definition of scalability should
that a scalable system should satisfy. Then, we could design take into account the system viability in terms of computation
systems with such features. In this section, we analyze a scal- and resource requirements as the system increases in size.
ability definition from the state of the art, and subsequently, To do this, we have to discuss first when a system is viable.
we propose a new definition. In this context, we can define viability using the system denial
of service probability, i.e., the probability that a UE cannot be
A. STATE-OF-THE-ART DEFINITION served by the system due to the lack of computational, front-
In [8], the authors formally defined that the network is scal- haul or radio resources. Let Dk be the denial of service prob-
able if the following tasks have finite complexity and resource ability for the k-th UE. Then, the denial of service probability
requirements for each AP as the number of UEs, K , tends of the system S is defined as
to ∞: ′
K
1) Signal processing for channel estimation; 1 X
2) Signal processing for data reception and transmission; PD (S) = lim Dk . (3)
′
K →K K′
k=1
3) Front-haul signaling for data and CSI sharing;
4) Power control optimization. This definition is valid for both the case K = ∞, and the
In [3], a solution was proposed and considered to be fully case K < ∞.PIn the latter, this definition can be rewritten as
scalable. The solution is based on the definition of fixed PD (S) = K1 K k=1 Dk . We say that a system is viable if the
clusters of APs, each of them connected to a different CPU. denial of service probability is below a maximum threshold,
Each UE selects the best-server AP and a set of serving i.e., PD (S) ≤ Pmax < 1.
APs, up to a maximum quantity of APs. The CPU of the Let Sm = {{Lm }, {Km }} be a subsystem of the system S
best-server is in charge of the channel estimation and the composed of the m-th cluster alone, whose denial of service
reception/transmission processing. This processing and chan- probability is PD (Sm ). Considering that the m-th cluster is
nel estimation is done independently for each UE and takes viable, i.e., PD (Sm ) ≤ Pmax , more clusters can be added with-
into account only the serving APs. Therefore, even in the out affecting the viability of the system if they are designed to
worst case (all UEs require the maximum amount of serving meet the same criterion. This is explained by the fact that the
APs), the total resource requirements of this solution increase availability of computational, front-haul, and radio resources
linearly with the amount of UEs in the network. When the of a cluster does not depend on the rest of the system. It is
network grows, all, the amount of UEs, APs, and CPUs grow worth noticing that with the increase in UEs, these may
with the same proportion and, hence, both the amount of UEs experience a deterioration of the received signal but this is
and the available resources grow with the same proportion. caused not by a reduction in the radio resources available, but
As a consequence, the system is scalable. However, this by interference. Considering that for the k-th UE the denial
scalability claimed in [3] was questioned by [8], since [3] of service probability is Dk = PD (Sm̃(k) ), we say that the
does not explicitly limit the number of UEs that can connect system is viable since following (3), we get, PD (S) ≤ Pmax .
to one AP. In other words, if we design viable clusters, the overall system
This raises the question of whether all the criteria in the composed of those clusters will also be viable, and certainly,
scalability definition in [8] are necessary in practice. For scalable.
instance, although one can argue the necessity of limiting the We now look at the canonical cell-free massive MIMO
amount of UEs that can be served by one AP, the reality is that case, where a unique CPU manages the reception and trans-
many works in the literature dealing with resource allocation mission to all UEs, and collects CSI information from all APs.
problems do not consider this limitation. This is due to two Taking into account that the CPU computational capabilities,
main reasons. First, it is unlikely to saturate the APs by as well as the front-haul resources, are limited, the number

Pk̃(m̃(k)−1) ∗
Hk m̃(i) Qi Hk m̃(i) + K ∗
P
1 + i=1 i=k Hk m̃(i) Qi Hk m̃(i)
RDL
k = log Pk̃(m̃(k)−1) ∗ . (2)
Hk m̃(i) Qi Hk m̃(i) + K ∗
P
1 + i=1 H Q H
i=k+1 k m̃(i) i k m̃(i)

VOLUME 11, 2023 26705

 D. Prado-Alvarez et al.: Study of Clustering Solutions for Scalable Cell-Free Massive MIMO

of UEs with Dk < 1 is finite. This means that if K = ∞,

the result of (3) is 1. In other words, the canonical cell-
free massive MIMO system is not viable, and of course, not
scalable. Note that this conclusion is valid even if we provide
the CPU with infinite computational capabilities, i.e., the
front-haul still makes the system not scalable.
Therefore, using our (less restrictive) definition of scalabi-
lity, clustering ensures the scalability of the system, if coop-
eration between clusters is not allowed.

IV. CLUSTERING APPROACHES

As shown in Section III, the solution to guarantee scalability
is self-presenting: divide and conquer. However, the division
FIGURE 1. Representation of overlapping clusters.
of the network into clusters and the posterior allocation of
resources should be done in a smart way, trying to take full
advantage of the flexibility of cell-free deployments and user-
centric solutions. In this sense, our work focuses on the
study of the strengths and weaknesses of current clustering
solutions with the aim of proposing more advanced solutions
to achieve better performance.
In what follows, two clustering approaches are analyzed.
The first approach focuses on ensuring that the UEs are closed
to the center of the cluster, thus avoiding edge effects. In the
extreme case of this approach, the UEs are placed at the
center creating a cluster of APs for each individual UE. Due
to this, each cluster of APs serves only one UE as shown
in Figure 1. This approach requires the highest number of
CPUs in a particular area. However, in this case, each CPU FIGURE 2. Representation of disjoint clusters clusters.
can have lower computational capabilities than in other cases
where a CPU has to serve several UEs. These clusters can
different clustering techniques are analyzed for both uplink
be overlapping. The second approach includes several UEs
and downlink. Then, conclusions are reached on the suitabil-
in each cluster of APs and focuses on using interference can-
ity of using one clustering solution or another, depending on
cellation techniques within each cluster to reduce the effect
the characteristics of the scenario and the inherent needs of
of interference. In the extreme case of this approach, the AP
each of the transmission modes.
clusters are enlarged as much as possible including all the
UEs in their coverage area in order to maximize the amount V. ANALYSIS OF THE CLUSTERING SOLUTIONS
of canceled interference in the clusters. This produces disjoint FOR THE UPLINK
clusters as shown in Figure 2. For the same average amount In this section, we present a clustering solution for each of
of APs in each cluster, the disjoint cluster approach is the the two clustering approaches described in Section IV for
one that requires the lowest number of clusters in a particular the uplink. Subsequently, we study the performance of the
area, hence, the lowest number of CPUs. However, in this two solutions in a particular scenario to gain insights into the
case, each CPU needs to serve the highest amount of UEs. performance of the two clustering approaches.
Therefore, this solution demands more powerful CPUs but,
at the same time, a smaller number of them. A. USER-AT-CENTER
It is worth noticing that the channel estimation and sig- This solution is part of the first clustering approach in
naling overheads do not depend directly on the clustering Section IV. Therefore, each UE is at the center of a custom-
approach but on the number of APs in each cluster and is, designed cluster and therefore does not experience edge
at least, linearly proportional to the total number of UEs effects. The clusters of different UEs can overlap. In [8],
served. Synchronization is also a critical aspect. With this the clusters are formed according to the following proce-
respect, in [3], a synchronization mechanism is proposed, dure. The UE appoints the AP with the strongest large-scale
which could be used in any of the clustering approaches. With fading channel coefficient as its master AP. The maximum
this synchronization mechanism, the overhead is proportional amount of UEs that one AP can serve is limited. So, when
to the number of APs each AP is collaborating with and to the the UE requests the service, the AP will assign the channel
number of clusters that contain each pair of AP. it considers to be the least affected by interference from the
In the following sections, these two approaches are com- other UEs it serves. The master AP informs a limited set of
pared in terms of sum data rate. For this purpose, first, neighboring APs that it is going to serve this UE. In particular,

26706 VOLUME 11, 2023

D. Prado-Alvarez et al.: Study of Clustering Solutions for Scalable Cell-Free Massive MIMO

the informed APs are those whose channel with the UE is at

most a ‘‘threshold’’ weaker than the channel of the master AP.
Then, the informed APs decide whether they serve the UE or
not. To make this decision, the APs have to take into account
other UEs they are serving. Using this clustering solution,
as many clusters as UEs are created, with the UEs being
relatively centered in their respective clusters. This implies
that the last UE of the m-th cluster is m, i.e., k̃(m) = m and
the cluster where the k-th UE belongs to is k, i.e., m̃(k) = k.
Considering this, the achievable data rate of the k-th UE for
uplink and downlink results in, respectively,
|I + P K ∗
P
i=1 Hik Hik |
UL-UaC
Rk = log PK , (4) FIGURE 3. Achievable data rate in the uplink for the UaC solution.
|I + P i=1,i̸=k Hik Hik∗ |
1+ K ∗
P
i=1 Hik Qi Hik
DL-UaC
Rk = log PK . (5)
1 + i=1,i̸=k Hik∗ Qi Hik
Since this clustering solution creates a different AP cluster
for each UE, it is not possible to cancel interference between
any of the UEs. Therefore, all the UEs in this scenario are
considered interferers by the k-th UE.
B. DISJOINT CLUSTERS
This solution is part of the second clustering approach in
Section IV. For this solution, the APs are grouped into non-
overlapping clusters of a specific size, i.e., |Lm | = N for
all m. Regarding the UE clusters, the clustering process is as FIGURE 4. Achievable data rate in the uplink fixing the average cluster
follows. First, the UEs select their master APs as in the previ- size. The lines for canonical cell-free massive MIMO (CF) are obtained
with only one cluster with and without IC.
ous solution, i.e., the APs with the strongest large-scale fading
channel coefficient. Then, the m-th UE cluster is composed
of the UEs whose master AP is in the m-th AP cluster. It is As shown in Figure 3, the average data rate per UE decreases
clear that, following this clustering process, the UE clusters as the number of UEs in the scenario increases. This decre-
are composed of several UEs in general. Therefore, since the ment can be softened by allowing each AP to serve more UEs,
signals of the UEs in the same cluster are processed by the i.e. by increasing α.
same CPU, some interference cancellation technique could In the case of the UaC solution, the size of the AP clusters
be used for the decoding/encoding. However, UEs are not is variable. Assuming that the master AP of each UE informs
prevented from suffering cluster border effects. all APs in the scenario to create the AP cluster of this UE, then
all APs serve min(K , α) UEs. Using this assumption, we can
C. COMPARISON OF SOLUTIONS FOR UPLINK compute the average AP cluster size as L min(K , α)/K . For
In this section, we study the performance of the two previous instance, in the case of 256 UEs and α = 16, the average
clustering solutions in a particular scenario. The scenario is cluster size is 1024 × 16/256 = 64. In order to facilitate the
a square of 32 by 32 ceiling-mounted APs in a squared grid comparison between the UaC solution and the DC solution
with an inter-AP distance of 10 m. In order to avoid scenario for different cluster sizes, we will compare the data rate of
border effects, a wrap-around technique is implemented. The certain average cluster size of the UaC solution with the data
UEs are randomly distributed in the scenario. The heights of rate of the same cluster size of the DC solution.
APs and UEs are 6 m and 1.5 m respectively. The channel Figure 4 shows the average data rate per UE versus the total
model used is the ‘‘Industrial indoor scenario’’ presented number of UEs in the scenario for UaC and DC solutions for
in [9]. The UEs have 10 mW of available power, and the cluster sizes of 16, 64 and 256 APs. The figure also shows
noise power at the receivers is 1.5887 · 10−10 mW, which the average data rate for canonical cell-free massive MIMO
corresponds to the noise power at 15 ◦ C and a transmission with and without interference cancellation as a benchmark.
bandwidth of 20 MHz. In this scenario, the AP clusters of the Note that canonical cell-free massive MIMO without inter-
‘‘disjoint clusters’’ (DC) solution are squared groups of APs ference cancellation is a particularization of the UaC solution,
located in a squared grid over the scenario. whereas canonical cell-free massive MIMO with interference
First, the results for several configurations of the ‘‘user-at- cancellation is a particularization of the DC solution, both
center’’ (UaC) solution are obtained. From one configuration for N = 1024. In Figure 4, when cluster sizes of 16 APs
to the other the only parameter that changes is the maximum and 64 APs or less than 64 UEs are considered, the compared
number of UEs that can be served by one AP, denoted by α. solutions have similar behaviour. However, when considering

VOLUME 11, 2023 26707

 D. Prado-Alvarez et al.: Study of Clustering Solutions for Scalable Cell-Free Massive MIMO

clusters of 256 APs and more than 64 UEs, the DC solution

clearly outperforms the UaC solution. Objectively speaking,
according to what it is shown, there is no solution whose per-
formance prevails over the other for any number of UEs. This
motivates the definition of a mixed solution in the following
section.

D. MIXED SOLUTION FOR THE UPLINK

This solution is inspired by the idea of merging the best
of the previous solutions: the avoidance of border effects
and the possibility of applying interference cancellation. The
UE and AP clustering is performed as follows. First, the
APs are grouped into non-overlapping cluster cores as for
FIGURE 5. Representation of the mixed solution.
the disjoint clusters approach. Then, more APs are added
to those clusters. The criterion to follow for such cluster
addition could be, e.g., proximity. As a result, the cores will
be surrounded by other APs of their clusters. Second, each
UE is linked to its best-serving AP, as previously. The cluster
of UEs served by each AP cluster is the set of UEs linked
to the APs in the corresponding core. The latter implies that
the cluster of UEs is centered with respect to the cluster
of APs preventing them from experimenting edge effects.
Figure 5 illustrates this solution. In the figure, four clusters
are delimited by colored solid lines. The core of each cluster is
formed by the APs that have the same color as the mentioned
lines. The UEs are represented in the color of the cluster
they are served by. For the sake of simplicity, for the results
in this section and for the same scenario described in the FIGURE 6. Achievable data rate comparison in the uplink for N = 256.

previous
√ section,
√ the AP clusters are assumed to be squares of
√ |L m |× |L
√ m | APs, whereas the cluster cores are squares of preserve the scalability of the system, suboptimal strategies
|Lm |/2 × |Lm |/2 APs at the center of the corresponding should be considered. One of the options is to allocate power
cluster, for m = 1, . . . , M . equally as will be detailed later in Section VI-A for the UaC
In Figure 6 the achievable data rate per UE for a cluster solution. Another option would be to design a scheme to
average size of 256 APs is shown for the UaC, the DC and the make a fixed pre-allocation of power and then optimize the
mixed solutions. The curve of the UaC solution was obtained power allocation in each cluster. However, there is no clear
following the procedure explained in Figure 4. As can be criterion for such pre-allocation. Due to this and in order
noticed, the performance with the mixed solution is better to simplify the subsequent analysis, we will not consider a
than the UaC and the DC solutions. Therefore, in this case, mixed solution, which requires the APs to pre-allocate power
the best performance is not obtained by only (i) avoiding the to all the clusters they belong to.
cluster border effects or (ii) canceling all the interference of
UEs with the master AP in the cluster. In this case, we need to A. USER-AT-CENTER
use a solution that mixes the two clustering approaches. In any According to [8], in the downlink, the AP clusters are created
case, the DC solution is the most similar to the mixed solution. as explained in Section V-A. We recall that this clustering
This highlights the importance of canceling interference. solution produces overlapping AP clusters that serve only one
UE each. Moreover, since the APs serve a maximum of α
VI. ANALYSIS OF THE CLUSTERING SOLUTIONS UEs (or belong to a maximum of α clusters), they split their
FOR THE DOWNLINK available power into this amount of UEs. This implies that the
In the uplink, the APs can be considered passive receivers, APs pre-allocate a certain amount of power for the different
since they just need to send the signal received to the clusters they belong to. Letting P be the available power at the
corresponding CPUs. However, in the downlink, the APs are APs, they use a maximum of P/α for each UE or cluster [8].
actively transmitting so the available power needs to be allo- The actual power used to transmit to the UEs is obtained by
cated to each of those transmissions. This power allocation multiplying P/α by a normalized precoding vector, which,
becomes more challenging if the clusters are overlapping. in the case of the k-th UE, is computed as
In that case, as each AP could be connected to more than
one CPU, all the CPUs involved would have to cooperate Hkk
wk = . (6)
and this could lead to a non-scalable system. In order to |Hkk |

26708 VOLUME 11, 2023

D. Prado-Alvarez et al.: Study of Clustering Solutions for Scalable Cell-Free Massive MIMO

Then, the power used by the l-th AP of the k-th cluster to limited available power in the APs. Mathematically, for the
transmit to the k-th UE is cluster m = 1, we want to solve
P P k̃(1)
|wkl |2 ≤ , (7) X k̃(1)
α α max RDL-2
k ({Qi }i=1 ),
k̃(1)
which indicates that the total power used to transmit to the {Qk }k=1 k=1
UEs by all APs is, indeed, P/α. This power allocation, which s.t. Qk ⪰ 0, k = 1, . . . , k̃(1),
is proposed in [8], ensures the APs do not use more power
k̃(1)
than P, although in a very conservative way. In particular, X
on average, the APs use significantly less power than P qkl ≤ P, l = 1, . . . , N , (11)
to transmit to all the UEs. In order to illustrate this fact, k=1

we assume that each AP serves exactly α UEs. Then, the where qkl is the l-th element of the main diagonal of Qk .
average amount of APs serving each UE is To find a solution for (11), we can use the algorithm proposed
in [10], which is based on [11]. With the optimum transmit
Lα
Lserving = . (8) covariance matrices in (11), the APs use all their available
K power. It is worth recalling that this solution is optimum
As mentioned before, the total power used to transmit to
locally, i.e., for one isolated cluster. The use of all the avail-
each UE is P/α, which is split into the APs serving this
able power can be dangerous in terms of the interference
UE. Therefore, the average power used per AP and UE is
caused to other clusters. Due to this, we also consider a
P/αLserving . Since the APs serve α UEs, the average total
different resource allocation.
power used per AP is
For the alternative resource allocation, we consider that
P PK the available power can be shared among all the APs in
α= . (9)
αLserving Lα the cluster. By doing this, the optimum covariance matrices
We consider here two numerical examples. If α = 4, contain, in the eigenvectors, the optimum precoding matri-
K = 256, and L = 1024, the average total power used per ces without the limitation of the power constraints. These
AP would be P/16, and if α = 8, K = 64, and L = 1024, precoding matrices define the optimum power distribution
the average drops to P/128. This conclusion is quite striking, to maximize the data rate. Due to this, we propose to use a
as it shows that a large part of the available power is unused. scaled version of these covariance matrices. The scaling is
necessary to ensure that the power constraints are met. More
B. DISJOINT CLUSTERS specifically, we solve
As in the case of the uplink, we can also consider the creation
k̃(1)
of disjoint clusters in the downlink with size |Lm | = N for X k̃(1)
all m. As discussed before, the resource allocation in each max RDL-2
k ({Qi }i=1 ),
k̃(1)
{Qk }k=1 k=1
cluster should not be conditioned to the particular allocation
performed in neighboring clusters. If not, the effect of each s.t. Qk ⪰ 0, k = 1, . . . , k̃(1),
cluster would propagate throughout the scenario, making the k̃(1)
X
resource allocation not scalable. In order to avoid this scala- Qk ≤ NP. (12)
bility problem, we propose to perform the resource allocation k=1
in each cluster as if they were isolated from the rest, i.e., k̃(1)
assuming that the interference from the other clusters do not Let {Q̄k }k=1 be the optimum solution of (12). In general,
affect their UEs. Without loss of generality, we are going to these covariance matrices do not satisfy the individual power
particularize this resource allocation for the first cluster, i.e., constraints in each AP. Due to that, we compute
m = 1. In this case, and taking into account that we neglect P
the interference from other clusters, the achievable rate of the
Qk = P  Q̄k , k = 1, . . . , k̃(1), (13)
k̃(1)
maxl k=1 q̄kl
k-th UE is, using (2),
P 
1 + Hk1∗ k̃(1) where q̄kl is the l-th element of the main diagonal of Q̄k .
k̃(1) i=k i Hk1
Q
k̃(1)
RDL-2
k ({Qi }i=1 ) = log Pk̃(1)  , (10) The covariance matrices {Qk }k=1 in (13) satisfy the power
∗
1 + Hk1 i=k+1 Qi Hk1 constraints in each AP and, since they are scaled versions
k̃(1)
of {Q̄k }k=1 , they have the same eigenvectors and, hence,
for k = 1, . . . k̃(1). The rate expression in (10) is the one used
they define the same precoding matrices. The optimization
for the resource allocation, i.e., the computation of the trans-
k̃(1) problem in (12) can be solved with the algorithm presented
mit covariances {Qi }i=1 . However, the actual achievable rate in [10], or with other techniques that make use of the uplink-
is shown in (2), which takes all the interference into account. downlink duality like in [12] and [7].
The expression in (10) is the one used in the following section
to compute the achievable rate in the figures. C. COMPARISON OF SOLUTIONS FOR DOWNLINK
To obtain the transmit covariances, we propose to maxi- In this section, we study the performance of the two pre-
mize the sum data rate in the cluster taking into account the vious clustering solutions in the same scenario described in

VOLUME 11, 2023 26709

 D. Prado-Alvarez et al.: Study of Clustering Solutions for Scalable Cell-Free Massive MIMO

FIGURE 9. Achievable data rate comparison in the downlink.

FIGURE 7. Achievable data rate in the downlink for the UaC solution.

with N = 1024. As can be observed, the DC solution shows

a better performance than the UaC solution. Specifically, the
performance could improve by 2.5 times for N = 256 and
256 UEs in the scenario. Note also that the rate improvement
from one cluster size to another is almost constant, e.g., the
absolute rate improvement from the N = 4 case to the
N = 16 case is similar to that from the N = 16 case to the
N = 64 case. An exception is the rate improvement from the
N = 256 case to the N = 1024 case. As mentioned before,
the latter case corresponds to canonical cell-free massive
MIMO. Since, in this case, there is only one cluster, this
cluster does not suffer from external interference, and this is
FIGURE 8. Achievable data rate in the downlink for the DC solution.
the cause of the extra rate gain between the N = 256 and the
Section V-C, with the same available power at the APs and N = 1024 cases.
the same noise power at the receivers. We start showing the
achievable sum data rate with different configurations of VII. CONCLUSION
the UaC solution. Figure 7 shows the sum data rate versus the In this work, we studied scalability issues of cell-free massive
number of UEs for the UaC solution for different values of α, MIMO. After considering a scalability definition presented in
i.e., the maximum number of UEs that each AP can serve. the literature, we proposed a new definition that is less restric-
As it can be observed, the difference between the curves is tive and includes more systems that are intuitively scalable.
not significant, even considering cases where the average We also studied two clustering approaches to design scal-
total power used per AP is very low (see Section VI-A). This able cell-free massive MIMO systems. The first approach is
suggests that the system is limited by interference. to design clusters for each UE in such a way that the UEs
Regarding the DC solution, the two power allocation are at the center of their clusters, hence avoiding cluster-edge
approaches presented are compared for N = 16 and N = 64 effects. The second approach is to include several UEs in each
in Figure 8. In the legend, the results corresponding to the cluster to be able to use interference cancellation techniques
cluster-optimal resource allocation are labeled as ‘‘Local inside each cluster. In the case of the uplink, none of the
Opt.’’, whereas the results of the suboptimal resource allo- two approaches seem to be the best for all cluster sizes and
cation are labeled as ‘‘Subopt.’’. The results highlight that, amounts of UEs. This motivated the development of a clus-
although the optimal resource allocation per cluster (neglect- tering solution that mixes the two approaches. This solution
ing inter-cluster interference) allows for achieving higher per- showed the best performances in all tested configurations.
cluster rate values, the effect of interference makes the total In the case of the downlink, whether or not the clustering
data rates worse than those obtained with the suboptimal solution allows overlapping AP clusters has a major impact
resource allocation algorithm. In other words, by using all on the performance. In particular, in order to avoid a global
the available power in the APs, the sum data rate is more optimization to make the system scalable, if an AP belongs
affected by the interference than benefiting from the trans- to several clusters, pre-allocating certain amount of power
mitted power. to each cluster is needed. This is an additional resource
Figure 9 shows the sum data rate versus the number of management procedure not present in the uplink. This power
UEs for the best result obtained for UaC, i.e. α = 1, the DC pre-allocation makes APs underutilize their available power.
solution with suboptimal resource allocation for cluster sizes, Due to this, a disjoint clustering solution that follows the
N , equal to 4, 16, 64, 256, and the canonical cell-free massive second approach, i.e., interference cancellation, provided the
MIMO with DPC. The latter corresponds to the DC solution best performance.

26710 VOLUME 11, 2023

D. Prado-Alvarez et al.: Study of Clustering Solutions for Scalable Cell-Free Massive MIMO

REFERENCES JOSE F. MONSERRAT (Senior Member, IEEE)

[1] H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta, received the M.Sc. and Ph.D. degrees (Hons.) in
‘‘Cell-free massive MIMO versus small cells,’’ IEEE Trans. Wireless Com- telecommunications engineering from Universi-
mun., vol. 16, no. 3, pp. 1834–1850, Mar. 2017. tat Politècnica de València (UPV), in 2003 and
[2] S. Buzzi and C. D’Andrea, ‘‘Cell-free massive MIMO: User-centric 2007, respectively. He is currently a Full Professor
approach,’’ IEEE Wireless Commun. Lett., vol. 6, no. 6, pp. 706–709, with the Communications Department, UPV, and
Dec. 2017. the Vice-President of UPV. He has been involved
[3] G. Interdonato, P. Frenger, and E. G. Larsson, ‘‘Scalability aspects of cell- in several European projects, especially his sig-
free massive MIMO,’’ in Proc. IEEE Int. Conf. Commun. (ICC), May 2019, nificant participation in NEWCOM, PROSIMOS,
pp. 1–6. WINNER+, and METIS/METIS-II, where he led
[4] S. Verdu, Multiuser Detection. Cambridge, U.K.: Cambridge Univ. Press, the simulation activities, or 5G-CARMEN on autonomous driving and
1998. 5G-SMART on industry 4.0. He also participated in one External Eval-
[5] D. Calabuig, R. H. Gohary, and H. Yanikomeroglu, ‘‘Optimum trans-
uation Group, ITU-R, on the performance assessment of the candidates
mission through the multiple-antenna Gaussian multiple access channel,’’
for the future family of standards IMT-Advanced, in 2010. He co-edited
IEEE Trans. Inf. Theory, vol. 62, no. 1, pp. 230–243, Jan. 2016.
two Special Issues in IEEE Communications Magazine on IMT-advanced
[6] M. H. M. Costa, ‘‘Writing on dirty paper (Corresp.),’’ IEEE Trans. Inf.
Theory, vol. IT-29, no. 5, pp. 439–441, May 1983. and 5G technologies and is a co-editor of the book Mobile and Wireless
[7] S. Vishwanath, N. Jindal, and A. Goldsmith, ‘‘Duality, achievable rates, Communications for IMT-Advanced and Beyond (Wiley) and 5G Mobile
and sum-rate capacity of Gaussian MIMO broadcast channels,’’ IEEE and Wireless Communications Technology (Cambridge). He manages around
Trans. Inf. Theory, vol. 49, no. 10, pp. 2658–2668, Oct. 2003. e0.5 million yearly budget, holds nine patents, and has published more than
[8] E. Björnson and L. Sanguinetti, ‘‘Scalable cell-free massive MIMO sys- 100 journal articles. He is a group head of five postdoctoral fellows, eight
tems,’’ IEEE Trans. Commun., vol. 68, no. 7, pp. 4247–4261, Jul. 2020. Ph.D. students, and two master’s students. He has been an advisor to the
[9] G. Interdonato, E. Björnson, H. Quoc Ngo, P. Frenger, and E. G. Larsson, European Parliament and the World Bank. His current research interests
‘‘Ubiquitous cell-free massive MIMO communications,’’ EURASIP include the design of future 5G wireless systems and their performance
J. Wireless Commun. Netw., vol. 2019, no. 1, pp. 1–3, Aug. 2019. assessment. He was a recipient of the First Regional Prize of Engineering
[10] H. Huh, H. Papadopoulos, and G. Caire, ‘‘MIMO broadcast channel opti- Studies for his outstanding student record, in 2003, and the Best Thesis Prize
mization under general linear constraints,’’ in Proc. IEEE Int. Symp. Inf. from UPV, in 2008. In 2009, he was a recipient of the Best Young Researcher
Theory, Jun. 2009, pp. 2664–2668. Prize of Valencia. In 2016, he received the Merit Medal from the Spanish
[11] L. Zhang, R. Zhang, Y.-C. Liang, Y. Xin, and H. V. Poor, ‘‘On Gaussian Royal Academy of Engineering, in the young researcher category.
MIMO BC-MAC duality with multiple transmit covariance constraints,’’
IEEE Trans. Inf. Theory, vol. 58, no. 4, pp. 2064–2078, Apr. 2012.
[12] W. Yu, ‘‘Uplink-downlink duality via minimax duality,’’ IEEE Trans. Inf. SAMER BAZZI received the B.E. degree in com-
Theory, vol. 52, no. 2, pp. 361–374, Feb. 2006. puter and communications engineering from the
American University of Beirut, in 2008, and
the M.Sc. and Dr.-Ing. degrees in communica-
tions engineering from Technische Universität
DANAISY PRADO-ALVAREZ graduated in elec- München (TUM), in 2010 and 2016, respec-
tronics and telecommunication engineering from tively. From 2010 to 2014, he was a member
the Technological University of Havana ‘‘José of DOCOMO Euro-Labs, Munich, Germany, and
Antonio Echeverría,’’ Cuba, in 2015. She received an External Dr.-Ing. Candidate with the Chair
the M.Sc. degree in telecommunication tech- of Signal Processing Methods, TUM, where he
nologies, systems, and networks and the Ph.D. worked on coordinated multi-point techniques and precoding for interference
degree in telecommunications from Universitat channels. From 2015 to 2022, he was a member of the research staff at
Politècnica de València (UPV), Valencia, Spain, in the European Research Center, Huawei Technologies Duesseldorf GmbH,
2017 and 2022, respectively. In 2018, she joined Munich. His research interests include multiple-input-multiple-output sys-
Instituto de Telecomunicaciones y Aplicaciones tems, parameter estimation, interference management, and general signal
Multimedia (iTEAM), UPV. Her current research interests include 5G&6G processing techniques for wireless communications.
wireless technologies design, channel modeling, network planning, and
resource allocation. WEN XU (Senior Member, IEEE) received the
B.Sc. and M.Sc. degrees in electrical engineering
from Dalian University of Technology (DUT),
China, in 1982 and 1985, respectively, and the
DANIEL CALABUIG (Member, IEEE) received Dr.-Ing. degree in electrical engineering from
the M.Sc. and Ph.D. degrees in telecommuni- Technische Universität München (TUM), Germany,
cations from Universitat Politècnica de València in 1996. From 1995 to 2006, he was with Siemens
(UPV), Valencia, Spain, in 2005 and 2010, respec- Mobile (later BenQ Mobile), Munich, where he
tively. In 2005, he joined Instituto de Telecomu- was the Head of the Algorithms and Standard-
nicaciones y Aplicaciones Multimedia (iTEAM), ization Laboratory. As a competence center, the
UPV. During the Ph.D. degree, he participated in laboratory was responsible for physical layer and multimedia signal process-
some European projects and activities, such as ing, and partly protocol stack aspects of 2G, 3G, and 4G mobile terminals,
NEWCOM, COST2100, and ICARUS, where he and was actively involved in standardization activities of ETSI, 3GPP,
was working on radio resource management in DVB, and ITU. From 2007 to 2014, he was with Infineon Technologies
heterogeneous wireless systems and Hopfield neural network optimization. AG (later Intel Mobile Communications GmbH), Neubiberg, focusing on
In 2010, he received the Marie Curie Fellowship from the European Com- wireline and wireless system concepts/architectures and software/hardware
mission to research cooperative multipoint transmissions. He visited the implementations. In 2014, he joined the European Research Center, Huawei
Department of Systems and Computer Engineering, Carleton University, Technologies Duesseldorf GmbH, Munich, where he is currently the Head of
Ottawa, Canada, from 2010 to 2012. In 2012, he returned to the iTEAM the Radio Access Technologies Department. His research interests include
and worked with the European projects Mobile and wireless communications signal processing, source/channel coding, and wireless communications
Enablers for the Twenty-twenty Information Society (METIS), METIS-II, systems in general. He is a member of the Verband der Elektrotechnik,
and 5G for Connected and Automated Road Mobility in the European Union Elektronik, Informationstechnik (VDE), Germany.
(5G-CARMEN).

VOLUME 11, 2023 26711