=Paper=
{{Paper
|id=Vol-2988/SSN2020_paper_9
|storemode=property
|title=Fronthaul Requirements Analysis for Cell-Free MIMO
|pdfUrl=https://ceur-ws.org/Vol-2988/SSN2020_paper_9.pdf
|volume=Vol-2988
|authors=Andrey Nakamura,Leonardo Ramalho,Aldebaro Klautau
}}
==Fronthaul Requirements Analysis for Cell-Free MIMO==
https://ceur-ws.org/Vol-2988/SSN2020_paper_9.pdf
Fronthaul Requirements Analysis for Cell-Free MIMO
Andrey Nakamura Leonardo Ramalho Aldebaro Klautau
Federal University of Pará Federal University of Pará Federal University of Pará
Belém, Brazil Belém, Brazil Belém, Brazil
andrey.nakamura@itec.ufpa.br leonardolr@ufpa.br aldebaro@ufpa.br
between the APs by a central processing unit (CPU)
coordinating them through a FH link, and allows for
Abstract better resource usage by implementing a power opti-
mization method either in the CPU [Ngu17, Bor19] or
Cell-free massive multiple-input multiple- in the AP [Ngo17, Nay17, Int19].
output (CF-mMIMO) is one of the compo-
nents of the fth-generation mobile communi-
cations, where a large number of distributed
While in cloud-radio access network (C-RAN) archi-
access points (APs) serve many users simulta-
tecture the signal processing is moved from a base sta-
neously, and provides scalability and high ca-
tion (BS) to the C-RAN computer, usually described
pacity data transmission. However, resource
in a star architecture, the CPU in a CF-mMIMO en-
usage increases as the number of APs and user
viroment should not be seen as a physical unit, but
equipment (UEs) grows in the network, and
a set of tasks that must be carried somewhere in the
practical systems need to meet these require-
network. Therefore, dierent C-RAN solutions can be
ments. In this work, we evaluate resource us-
used in the network [Bjo20]. Other works may call the
age of the fronthaul (FH) link capacity using
APs as remote radio units (RRUs), and the CPU as
two precoding methods, zero-forcing and con-
baseband unit (BBU) or distributed unit (DU) when
jugate beamforming, with regards to user data
explaining C-RAN architecture [Li2019, Lar2019].
and channel state information (CSI) transmis-
sion.
Despite the advantage of CF-mMIMO, its practi-
1 Introduction
cal implementation brings a lot of challenges [Int19,
CF-mMIMO systems provide spectral e ciency, relia- Bjo20], such as intensive computational process-
bility and fairness among users, where a large number ing [Zha20] and increased FH tra c among the
of distributed APs simultaneously serve a smaller num- high number of APs and the CPU. The required
ber of UEs using the same time/frequency resources. FH throughput depends on many parameters of
This is achieved by conducting precoding and power CF-mMIMO, such as, the radio signal, as well as
allocation algorithms [Nay17]. Cellular networks have the number of APs and the number of users. One
the drawback of increased inter-cell interference, par- of the contributions of this work is to provide the
ticularly when a UE is located near cell bound- equations to estimate the FH rate, based on many
aries [Ngo17], and the superposition is necessary in parameters of the orthogonal frequency-division mul-
order for the UEto not lose connection when migrating tiplexing (OFDM) signal and the CF-mMIMO sys-
to another cell. CF-mMIMO increases coverage prob- tem. There are consolidated equations to estimate
ability by removing cells and cell boundaries, allowing the FH rate for the IQ data [Li2019, Lar2019], but
all UEs to be served by all APs, reduces interference this work takes into consideration not only IQ data,
Copyright ©
2020 for this paper by its authors. Use permitted
but CSI transmission as well. This is highlighted in
CF-mMIMO because of the dierent C-RAN solutions
under Creative Commons License Attribution 4.0 International
that can be used [Bjo20]. Furthermore, this work ex-
(CC BY 4.0).
plores the dierent throughput requirements on the
In: Proceedings of the IV School of Systems and Networks (SSN
2020), Vitória, Brazil, December 14-15, 2020. Published at FH of CF-mMIMO, when dierent strategies for power
http://ceur-ws.org. allocation and precoding calculation are deployed.
2 Precoding and Power Allocation Fully Centralized
Strategies on Cell-Free Wireless
CPU AP 1 AP M UE 1 UE K
The two types of precoding methods investigated in
Beginning of Coherence Interval
this work are zero-forcing (ZF) and conjugate beam- Uplink Pilots from UE 1
forming (CB). The latter allows for distributed pre-
Uplink Pilots from UE K
coding calculation on the APs and optimal power al-
Channel Channel
location on CPU, where the power allocation with Estimation Estimation
K Estimated Channels
CB typically relies on large-scale CSI. Alternatively, from AP 1
the ZF approach centralizes both tasks on the CPU K Estimated Channels
through a procedure that requires short-term CSI and from AP M
ZF Precoding
therefore poses stronger requirements on uplink (UL) Calculation
FH tra c [Pal19]. However, some works show that Power Allocation
Symbol Precoding
the ZF greatly outperforms CB precoding in terms of
Send Precoded
max-min rate [Nay17]. Symbols to AP 1
The ZF requires the APs to send to the CPU the Send Precoded
Symbols to AP M
short-term CSI, greatly increasing FH bandwidth us-
Synchronous Transmission of Precoded Symbols
age. On the other hand, CB can be implemented
in a distributed manner, where each AP calculates
Symbol Precoding and Precoded Symbol Transmission
the precoding locally, and the power allocation can Continues Until the Next Coherence Interval
be implemented locally or on the CPU, based on the
Next Coherence Interval
long-term CSI, which reduces the FH rate require-
ments [Pal19, Int19]. Figure 1: MSC of the fully centralized method.
The methods referenced above can be categorized as
ZF fully centralized [Nay17, Bor19, Ngu17], CB par- bols, and sends 𝐾 coe cients to every AP. For each
tially distributed [Ngo17], and CB fully distributed OFDM symbol, the CPU sends 𝐾 × 𝑁 𝑠𝑐 QAM symbols
[Int19]. These three approaches are discussed in the to the APs that perform symbol precoding and send
sequel and the respective fronthaul requirements are the precoded symbols to the UEs.
evaluated. The symbol precoding and transmission processes
are repeated until the next coherence interval, how-
2.1 Fully Centralized ever the power allocation is not calculated in every
coherence interval as in the fully centralized strategy,
In the ZF fully centralized method, at the beginning and are only updated when the large-scale coe cient
of the coherence interval, the 𝐾 UEs send orthogonal changes [Pal19]. The MSC of the method is shown in
pilots to the 𝑀 APs, in order to estimate the chan- Fig. 2.
nels. Then, each APs send 𝐾 × 𝑁 𝑠𝑐 /𝐶BW estimated
channels to the CPU, where 𝑁 𝑠𝑐 is the number of sub-
2.3 Fully Distributed
carriers of the OFDM signal and 𝐶BW is the number
of subcarriers in the coherence bandwidth. Then, the In the fully distributed method, at the beginning of the
CPU calculates the precoding coe cients, power al- coherence interval, the UEs send the UL pilots to the
location, performs symbol precoding, and sends the APs, which estimate the large-scale coe cients of the
precoded symbols to every AP. Finally, the APs send channel and send them to the CPU. The CPU broad-
the precoded symbols to the UEs. Symbol precoding, casts 𝐾 × 𝑁 𝑠𝑐 QAM symbols to the APs. The power
FH transport and air transmission is repeated for each allocation, precoding calculation and symbol precod-
OFDM symbol over the coherence interval. The mes- ing are done in the APs [Int19]. In this case, no CSI
sage sequence chart (MSC) of the method is shown in is required on the CPU, and it is only responsible to
Fig. 1. provide the user QAM symbols for the OFDM signal.
The MSC of the method is shown in Fig. 3.
2.2 Partially Distributed
2.4 Fronthaul Link Usage
In the partially distributed method, at the beginning
of the coherence interval, the UEs send the UL pilots The FH rate is estimated for each AP during UL for IQ
to the APs, who estimates the large-scale channel be- samples and CSI samples, and during downlink (DL)
tween them and the UEs. Then, each AP sends 𝐾 for IQ samples. The FH rate during UL IQ data for
channel coe cients to the the CPU. The CPU com- all methods and DL IQ data for the fully centralized
putes the power allocation coe cients of the user sym- method is:
Partially Distributed Fully Distributed
Wireless
Wireless
CPU AP 1 AP M UE 1 UE K
CPU AP 1 AP M UE 1 UE K
Beginning of Coherence Interval
Uplink Pilots from UE 1
Beginning of Coherence Interval
Uplink Pilots from UE K Uplink Pilots from UE 1
Channel Channel
Estimation Estimation Uplink Pilots from UE K
Large-scale Estimated
Channels from AP 1 Channel Channel
Estimation Estimation
Large-scale Estimated
Channels from AP M Power Power
Power Allocation Allocation
Allocation
Precoding Precoding
Power Coefficients
Calculation Calculation
Precoding Precoding
Send IQ Symbols
Calculation Calculation
Send IQ Symbols Symbol Symbol
Symbol Symbol Precoding Precoding
Precoding Precoding
Synchronous Transmission of Precoded Symbols
Synchronous Transmission of Precoded Symbols
Symbol Precoding and Precoded Symbol Transmission Symbol Precoding and Precoded Symbol Transmission
Continues Until the Next Coherence Interval
Continues Until the Next Coherence Interval
Next Coherence Interval
Next Coherence Interval
Figure 2: MSC of the partially distributed method.
Figure 3: MSC of the fully distributed method.
UL = 𝑁Ci × 𝑁sc × 𝑏 IQ
𝑅IQ , (1a)
Δ𝑇Ci
DL = 𝑅UL ,
𝑅IQ sc
𝐾 × 𝑏 CSI × 𝐶𝑁BW
, fc IQ (1b)
UL =
DL = 𝑅DL = 𝐾 × 𝑅UL , 𝑅CSI , fc , (2a)
𝑅IQ , pd IQ,fd IQ (1c) 𝑠CSI × Δ𝑇OFDM
UL
UL DL UL 𝑅 CSI , fc
where 𝑅
IQ is the UL IQ rate of all methods, 𝑅IQ,fc , 𝑅CSI,pd = , (2b)
DL and 𝑅DL are the DL IQ rate of the fully cen-
𝑅IQ Δ𝑇ls
, pd IQ,fd UL = 0,
tralized, partially distribute and fully distributed, re- 𝑅CSI , fd (2c)
spectively, 𝑁 Ci is the number of OFDM symbols sent
in each coherence period, 𝑁sc is the number of sub-
where 𝐾 is the number of UEs, 𝑁 sc is the number of
carriers used in the OFDM signal, 𝑏IQ is the number
of bits used to represent each IQ sample, and Δ𝑇Ci
subcarriers, 𝑏 CSI is the number of bits used to repre-
sent each CSI coe cient, 𝐶BW is the coherence band-
is the time in seconds of the coherence interval. The
width, 𝑠 CSI is the number of OFDM symbols used to
equations in (1) show the required rate to transport
transport the CSI coe cients, Δ𝑇OFDM is the period
all subcarriers on each OFDM symbol. More specif-
of an OFDM symbol, and Δ𝑇ls is the large-scale inter-
ically, (1a) is the uplink rate for all methods, (1b) is
val duration. 𝑠𝐶𝑆𝐼 > 1 indicates that the CSI could
the downlink rate for the fully centralized method, and
be transported along with more than one OFDM sym-
(1c) is the downlink rate for the partially and fully
distributed methods. The distributed methods in (1c)
bol, and 𝑁 sc /𝐶BW indicates that one estimation can be
require multiplication of the DL rate for every AP be-
used by 𝐶 BW subcarriers simultaneously, reducing the
amount of estimations necessary. The peak FH rate
cause in total 𝐾 OFDM symbols are sent to every AP,
used by the partially distributed method is divided by
one for each user.
the large-scale interval because data is only sent to the
The peak FH rate happens at the beginning of the
CPU when the large-scale coe cient changes.
coherence interval. The rate used by the UL of the
CSI samples for the fully centralized methods is shown Finally, the peak FH rate per AP is the sum of IQ
in (2a), the partially distributed method is shown and CSI during UL:
in (2b), and the rate used by the fully distributed
method is 0 in (2c) because the AP does not send CSI
UL = 𝑅UL + 𝑅UL .
𝑅total
to the CPU. IQ CSI (3)
Table 1: FH Usage per AP. Acknowledgements
Metric Fully Partial. Fully This work was partially supported by Innovation Cen-
Central. Distrib. Distrib. ter, Ericsson Telecomunicações S.A. and CNPq.
𝑅IQ (Mbps) 134.4
𝑈𝐿
134.4 134.4
References
CSI (Mbps) 179.2
𝑅𝑈 𝐿
4.48 0
𝑅total (Mbps) 313.6
𝑈𝐿
138.88 134.4 [Nay17] E. Nayebi, et al, Precoding and Power Op-
𝑅IQ
𝐷𝐿 (Mbps) 134.4 2150.4 2150.4 timization in Cell-Free Massive MIMO Sys-
tems, IEEE Transactions on Wireless Com-
3 System Model and Results munications, vol. 16, no. 7, pp. 4445-4459
Jul. 2017.
In this work, we consider a scenario with 𝐾 = 16 UEs
and 𝑀 = 128 APs. The coherence interval is the same [Bor19] M. N. Boroujerdi2019, et al, Cell Free Mas-
as in LTE, Δ𝑇 OFMD = 1 ms with 𝑁Ci = 14 OFDM sym- sive MIMO with Limited Capacity Fron-
bols in-between each period, and the large-scale inter- thaul, Wireless Personal Communications,
val is Δ𝑇 ls = 40 ms. The coherence bandwidth available vol. 104, no. 2, pp. 633-648 2019.
is 𝐶BW = 12 subcarriers, and the total number of use- [Bjo20] E. Björnson, et al, Scalable Cell-Free Mas-
sc = 600. Each IQ and CSI sample
ful subcarriers is 𝑁
sive MIMO Systems, IEEE Transactions on
is represented with 𝑁IQ = 𝑁CSI = 16 bits, and takes
Communications, vol. 68, no. 7, pp 4247-
𝑠CSI = 1 OFDM symbol to transport the UL CSI. Us- 4261, 2020.
ing the mentioned con guration in (1), (2) and (3), we
obtain the FH throughput shown in Table 1 for each [Int19] G. Interdonato, et al, Scalability Aspects of
AP. Cell-Free Massive MIMO, ICC 2019 - 2019
The results on Table 1 has two important informa- IEEE International Conference on Commu-
tions: the fully centralized approach requires a higher nications (ICC), pp 1-6, 2019.
UL tra c on FH, but the DL tra c can be lower than
[Ngo17] H. Q. Ngo, et al, Cell-Free Massive MIMO
the others approaches to implement CF-mMIMO.
Versus Small Cells, IEEE Transactions on
As indicated in other works [Int19] and shown on Wireless Communications, vol. 16, no. 3, pp
Table 1, the fully centralized ZF approach requires a 1834-1850, Mar. 2011.
high UL tra c, in order to transport the CSI from APs
to CPU. However, the same table shows that the DL [Li2019] L. Li et al., Enabling Flexible Link Capac-
FH rate can be considerably lower than the distributed ity for eCPRI-Based Fronthaul With Load-
approaches, especially if the number of users 𝐾 is high, Adaptive Quantization Resolution, IEEE
and some works [Nay17, Pal19] showed that ZF can Access, vol. 7, pp. 102174-102185, July 2019.
outperform CB in terms of max-min user rate.
[Ngu17] L. D. Nguyen, el al, Energy E ciency in
On the other hand, if each CSI samples were repre-
Cell-Free Massive MIMO with Zero-Forcing
sented with more bits than IQ samples, the total UL
Precoding Design, IEEE Communications
tra c of the fully centralized approach could be sub-
Letters, vol. 21, no. 8, pp 1871-1874, 2017.
stantially high, resulting in a higher probability of a
constrained FH link. In this scenario, in order to guar- [Pal19] F. R. Palou, el al, Trade-os in Cell-free
antee scalability, the distributed approaches would be Massive MIMO Networks: Precoding, Power
more advantageous [Int19]. Allocation and Scheduling, 2019 14th Inter-
national Conference on Advanced Technolo-
gies, Systems and Services in Telecommuni-
4 Conclusion cations (TELSIKS), pp. 158-165, 2019
[Zha20] Y. Zhao, et al, Power Allocation in Cell-Free
CF-mMIMO provides high capacity and fairness due
Massive MIMO: A Deep Learning Method,
to the high number of APs and UE centric approach,
IEEE Access, vol. 8, pp 87185-87200, 2020.
but in practical systems, resources such as computa-
tional power and FH link capacity are limited, mak-
[Lar2019] L. M. P. Larsen, et al, A Survey of the
ing scalability an issue when the network grows larger
Functional Splits Proposed for 5G Mobile
[Int19, Zha20]. This work provides insight regarding
Crosshaul Networks, in IEEE Communica-
the FH link requirements for CF-mMIMO networks
tions Surveys & Tutorials, vol. 21, no. 1, pp.
by comparing the data rate used during UL and DL
146-172, Firstquarter 2019.
with the CB and ZF precoding methods.