Sie sind auf Seite 1von 14

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2017.2750922, IEEE
Transactions on Vehicular Technology
1

Cluster-based Cooperative MIMO-OFDMA Cellular


Networks: Scheduling and Resource Allocation
Felip Riera-Palou, Senior Member, IEEE, and Guillem Femenias, Senior Member, IEEE

Abstract—In this work, a cross-layer framework is proposed cooperation among neighboring base stations (BS) has long
for scheduling and resource allocation in the context of coopera- shown its potential to particularly improve the performance of
tive wideband wireless cellular systems supporting heterogeneous cell edge users [5], although it is not totally clear what form
traffic. To this end, the network is partitioned into clusters, each
comprising a small number of base stations (BS) that share data of collaboration is best. Very recently, the work by Hosseini
and/or control information and thus are able to apply cluster- et al. [6] has shed some light on this very specific topic by
wide cooperative multiple-antenna techniques such as network analytically and numerically comparing the capacities of N-
MIMO (N-MIMO) or large-scale MIMO (LS-MIMO). In contrast MIMO and LS-MIMO under equivalent conditions. Remar-
to previous works whose main goal was the study of purely kably, this work showed that the desired signal strength in
physical layer metrics, the framework proposed here serves to
evaluate the network performance in terms of operator-relevant an LS-MIMO system first-order stochastically dominates its
metrics such as the average user throughput, the average delay or N-MIMO counterpart while the interference-plus-noise term
different fairness indices. The cross-layer design introduced here was shown to be statistically equivalent. Consequently, the
is general enough to encompass various forms of power allocation signal-to-noise-plus-interference ratio (SINR) was found to be
or scheduling policies while being able to incorporate per-cluster larger in LS-MIMO than in N-MIMO, thus suggesting that
or per-base power constraints. Remarkably, different greedy
algorithms are introduced for both N-MIMO and LS-MIMO that LS-MIMO should be the preferred technique.
effectively exploit the multiuser and frequency diversity of the The radio management mechanisms by which the physical
system while remaining computationally feasible. A particularly layer (PHY) resources are distributed among the active users
interesting application of the proposed framework, explored in in the cell have been largely researched over the last decade
detail in this paper, is a thorough and realistic comparison of in the context of multicarrier architectures (see for example
the performance of N-MIMO and LS-MIMO that reveals the
strengths and weaknesses of each strategy in terms of different [7]–[11]). These studies have indeed served to highlight the
performance metrics. importance of an adequate user scheduling policy to fully
exploit the underlying PHY resources. Resource allocation in
the particular context of CoMP-aided OFDMA networks has
I. I NTRODUCTION been addressed in recent years (see [12] and references therein
for an extensive literature review and state-of-the-art). In [13],
Advanced multiple-input multiple-output (MIMO) techni-
Li et al. address the problem of radio resource allocation
ques are bound to keep playing a major role in the next gene-
in the downlink of a clustered network MIMO system. The
ration (so-called, 5G) of mobile networks as they have done
proposed resource allocation approach is divided in two-steps.
in current 4G systems, where network (or cooperative) MIMO
The first step focuses on inter-cluster interference mitigation
(N-MIMO) and multiuser-MIMO (MU-MIMO) have shown to
and emphasis is placed by the authors on the design and
lead to major performance improvements in comparison to the
optimization of novel cooperative fractional frequency reuse
classical MIMO techniques [1] already in use in 3G standards.
schemes, a topic that is out of the scope of our paper. The
It is widely anticipated that 5G will bring along yet a new
second step concentrates on intra-cluster resource allocation,
form of MIMO processing, namely, large-scale MIMO (LS-
the main topic of our paper, and the authors propose a sub-
MIMO or massive MIMO) [2], which is set to play a central
optimal utility-based joint scheduling and power allocation
role towards the achievement of the ambitious targets set for
algorithm subject to per-sector power constraints. The main
5G systems, alongside other techniques like densification (i.e.,
drawback of this proposal for intra-cluster resource allocation
femtocells) and mm-wave communications [3]. Remarkably,
is that only BSs with single-antenna per sector are considered
among the key requirements set by the various bodies involved
thus, on one hand, limiting the degrees of freedom that
in 5G standardization, minimum rates to be achieved for over-
can be exploited when using multiple antennas per sector
all cell as well as cell edge rates require improvements ranging
and, on the other hand, precluding the comparison of these
between 2 and 3 orders of magnitude over current 4G networks
schemes with those based on LS-MIMO-based strategies.
figures [4]. To this end, large-scale and/or cooperative MIMO
Moreover, the algorithms proposed by Li et al. in [13] adopt
techniques are expected to significantly contribute. Noticeably,
simplifying sub-optimal per-user uniform power allocation
Copyright (c) 2017 IEEE. Personal use of this material is permitted. schemes. Part of these limitations are overcome by Choi et
However, permission to use this material for any other purposes must be al. in [14], where both multiple transmit antennas per sector
obtained from the IEEE by sending a request to pubs-permissions@ieee.org. and multiple antennas per user are considered thus allowing
F. Riera-Palou and G. Femenias are with the Mobile Communications
Group, University of the Balearic Islands (UIB), Mallorca 07122, Illes Balears the implementation of pre- and post-processing MU-MIMO
(Spain). {guillem.femenias,felip.riera}@uib.es strategies to cancel inter-user interference and decompose the

0018-9545 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2017.2750922, IEEE
Transactions on Vehicular Technology
2

resulting equivalent single-user MIMO channels into parallel that improves the objective metric under consideration in
non-interfering spatial layers. However, the implementation of comparison to the previous generation. Genetic algorithms
the proposed optimization algorithms, which only focus on (GA) have been applied to the MIMO-based user scheduling
N-MIMO, may be compromised due to their high computa- problem in [22] in the context of linearly-precoded systems
tional complexity. The design of utility-based radio resource whereas [23] applied GA-based scheduling when using dirty
allocation algorithms for the downlink of a clustered network paper coding (DPC). Lately, [24] has proposed user selection
MIMO is also the main topic addressed by Huang et al. methods that combine the greedy and GA-based approaches
in [15]. The authors propose two different utility functions in the context of MU-MIMO, demonstrating the benefits that
for best effort (BE) and voice-over-IP (VoIP) traffic and can be achieved through this combination. This approach has
focus on the maximization of the sum utility experienced by been recently generalized in [25] by using a more general
the users in each cellular cluster. However, this work relies metaheuristic-based framework comprising the use of GA- and
on an unrealistic sub-optimal binary power control (i.e., in particle swarm-based optimization and applied in the context
any scheduling interval, the BS either does not transmit or of CoMP-based heterogeneous networks. The last category
transmits with full power). The joint problem of scheduling, of suboptimal scheduling techniques rely on classic convex
precoder design and power allocation in a multicell setup optimization techniques. Most optimization problems in the
has been treated in [16], whereby the precoding on each context of MIMO-OFDMA systems are inherently non-convex
cell ensures that the interference generated on the other cells due to integer-valued variables in the objective function or
in the cluster is minimized (i.e., coordinated scheduling and the problem constraints, thus requiring of a relaxation step to
beamforming). However, note that under this scheme, users actually make the problem convex and thus amenable to be
are actually assigned to a particular BS rather than a to tackled using classic optimization methods [26]. The work in
BS cluster thus somewhat restricting the scheduling process. [27] is an illustrative example of the application of convex
Moreover, the proposed design does not take into account relaxation to the scheduling problem in a MIMO-OFDMA
neither queueing aspects nor distinct QoS requirements arising context. To the best of authors’ knowledge, no previous work
from traffic heterogeneity. Very recently, authors in [17] have has exhaustively compared these three distinct categories of
considered a cooperative setup very much related to the one algorithms to the problem at hand. In fact, the appropriateness
tackled here and provide performance upper/lower bounds of one approach over the others is likely to depend on specific
for the sum rate while analyzing the effect the sharing of system constraints (power restrictions, CSIT accuracy), and
information has on the network performance but again in the moreover, the many variants within each family of algorithms
context of traffic class-unaware scenarios. makes such a comparison rather cumbersome. Therefore, in
Any resource allocation strategy is strongly affected by this work, a choice has been made to adapt a greedy-type
the specific form of the scheduling process, that is, the algorithm to the target problem notwithstanding the fact that
selection of the particular users to be served from a larger a GA- or convex relaxation-based scheduling algorithm could
potential user pool on each radio resource (time slot and/or be used instead. As mentioned towards the end of this paper,
frequency block). Indeed, as it will be seen throughout the this is indeed a very promising avenue for further research,
paper, channel-aware and/or queue-aware user selection plays although no big changes are expected regarding the relative
a fundamental role in the optimization of various network performance of N-MIMO and LS-MIMO when are both using
metrics. Scheduling has long been a very active research topic a different scheduling technique. Concluding this literature
basically due to the impossibility of implementing optimal user review, we note that a preliminary version of this paper has
selection policies since these usually involve an exhaustive been recently presented in [28], which nevertheless did not
search over a large-dimensional space that quickly becomes consider neither power allocation in detail nor multi-class
infeasible to conduct in practical designs [18]. Consequently, (heterogeneous) traffic.
different families of suboptimal scheduling strategies have This paper proposes a cross-layer framework, integrating
been proposed, each with plenty of variants. As pointed out the physical and data link control (DLC) layers, suitable
in [18], three main categories of scheduling can be found, for different clustered-based MIMO cooperative strategies in
namely, greedy, metaheuristic and convex optimization-based a downlink wideband setting. More specifically, our main
techniques. Greedy techniques construct the group of selected contributions can be summarized as:
users by adding, one at a time, the best single user that
maximizes a prescribed utility function. This procedure can 1) Exploiting the availability of both PHY layer channel
be implemented by either computing for each candidate user state information (CSI) and DLC layer queue state
the corresponding precoding vectors and resulting utility [19] information (QSI), a novel cross-layer design suitable
or by relying on simple metrics that can be computed on for cooperative cluster-based MIMO downlink scenarios
the basis of the CSI of each user and spare the precoder is proposed supporting different scheduling policies,
computation. Metrics based on the Frobenius norm of each various levels of BS cooperation, different forms of
user channel matrix [20] or the cross-correlation among the BS power constraints and, potentially, different rate and
distinct user channel matrices [21] are often used to this end. power allocation strategies. As far as we know, and
Metaheurestic-based approaches are a general class of methods unlike previous proposals that tend to solely focus on
whereby a set of initially generated feasible solutions are PHY metrics, this proposal is the first one tackling
combined and mutated to form a new generation of solutions this optimization problem in the context of SRA for

0018-9545 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2017.2750922, IEEE
Transactions on Vehicular Technology
3

cooperative MIMO-OFDMA-based systems that allows


to actually evaluate how the different forms of MIMO
processing affect various user-relevant QoS metrics.
2) For each form of cooperative MIMO processing and
in order to overcome the computationally prohibitive
optimal user selection based on an exhaustive search
approach, two novel low-complexity multiuser selection
algorithms are proposed. They are both based on greedy
principles and are shown to clearly outperform random Rnc
Rc
user selection schemes such as those considered in
[6]. Remarkably, and for the case of N-MIMO, the
proposed user selection algorithm is able to incorporate
specific operational power restrictions such as per-base
power constraint (PBPC) or per-cluster power constraint
(PCPC).
3) The proposed framework, jointly with the use of power-
ful user selection schemes, allows a meaningful compa-
rison of the two cooperative MIMO processing schemes
under consideration. In particular, the design problem
is formulated as a weighted sum rate maximization pro-
blem where the weights can serve to implement different
scheduling policies that take into account user fairness Figure 1: Cluster-based cellular layout with B = 3. The
and QoS constraints as in [10], [29]. It is worth pointing dark grey-shaded area represents the cooperation zone. Bold
out that, for the sake of analytical tractability, the com- lines identify the cooperating cells with the rest forming the
parative analysis in [6] focused on RR scheduling and, interfering clusters.
moreover, per-BS power constraints were relaxed and
shadowing effects were not accounted for. The approach
operator D (x) represents a diagonal or block-diagonal matrix
we take in this work is to cast the two schemes, LS-
with the components of vector/matrix x at its main diagonal,
MIMO and N-MIMO, into the cross-layer framework
kAkF will
√ denote the Frobenius norm of matrix A and
and numerically evaluate different performance metrics
kxk = xH x will represent the Euclidean norm of vector
while avoiding all these simplifications.
x. Superscripts (·)T and (·)H are used to denote the transpose
We note that the current view of 5G PHY becomes so-
and the conjugate transpose (Hermitian) of a matrix.
mewhat blurred when it comes to the choice of waveform
to be used: whereas some proponents advocate to continue
with the use of orthogonal frequency division multiple access II. C OOPERATIVE CLUSTER MODEL
(OFDMA) just as in 4G [4], others propose the use of We consider the downlink of a multicell tri-sectorized
alternative schemes such as filterbank multicarrier (FBMC) layout such as the one depicted in Fig. 1 where sectors from
[30] or generalized frequency division multiplexing (GFDM) neighboring BSs pointing in the same direction form a cluster
[31]. This work conservatively assumes the use of OFDMA where cooperation is allowed. That is, BSs involved in a cluster
as a means to tackle the wideband nature of the channel. have the ability to share user’s CSI and/or user’s data by
The rest of the paper is organized as follows. Section II means of a backhaul connection. The area covered by the
introduces the considered cooperative cluster model while cluster is assumed to be populated by uniformly distributed
briefly reviewing the different cooperative MIMO processing users which, in line with recent works that advocate to shift
techniques under consideration. In Section III, PHY and most of the processing to the infrastructure side, are assumed
DLC layer-abstraction models are derived, which provide the to be single-antenna. Note that this focus on single-antenna
foundation to the clustered cross-layer design proposed in this users is backed by the fact pointed out in [32] whereby
paper and fully described in Section IV. Section V provides increasing the number of antennas at the user side further
extensive simulation results highlighting the versatility of the complicates the CSI acquisition in networks where the BSs
derived model and allowing a thorough comparison between are equipped with large-antenna arrays, and furthermore, the
N-MIMO and LS-MIMO to be established in terms of rele- benefit becomes rather marginal when the scheduled users at
vant performance metrics such as average user throughput or a given instant are chosen from a large user pool. Without
average packet delay. Finally, the conclusion in Section VI loss of generality, it is assumed that the network is made
serves to recap the main findings of this work. of L clusters each formed by B sectors from adjacent cells,
This introduction ends with a notational remark. In this with each sector using NT transmit antennas1 . As shown in
paper, vectors and matrices are denoted by lower- and up-
1 We note that although examples and simulations in this work are presented
percase bold letters, respectively. The K-dimensional identity
for the specific case of B = 3, a general notation is kept throughout the
matrix is represented by I K and a (column) vector of length manuscript to emphasize the fact that, indeed, clusters can be formed by an
K with all entries equal to c ∈ C is denoted by cK . The arbitrary number of BSs.

0018-9545 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2017.2750922, IEEE
Transactions on Vehicular Technology
4

Fig. 1, let us denote by Rc the cell radius and let us define considered in this work2 [39]: Best Effort (BE), Non-Real-Time
a non-cooperation radius, Rnc ≤ Rc , that serves to separate (nRT), and Real-Time (RT). Traffic flows arriving from higher
the cooperative area from the non-cooperative one. Similar to layers are buffered into the first-in first-out (FIFO) user queues
the approach in [33], users located in the cooperative zone are at the DLC layer. At the beginning of each scheduling time
(potentially) served using the B BSs in the cluster whereas interval, based on the available joint CSI/QSI, the scheduler
users located in the non-cooperative area are only serviced and resource allocation (SRA) unit selects some bits in the
by their closest BS. Since this paper is exclusively concerned queues for transmission, which are then forwarded to the
with the cooperation aspects, only the cooperative area/users OFDM transmitter, where they are adaptively modulated and
will be considered from this point onwards. Note that the channel encoded and are allocated power and subbands based
decoupling of cooperative and non-cooperative areas can easily on the particular cooperative scheme in use.
be implemented in practice by using some form of fractional As pointed out in [6], [21], the number of degrees of
frequency reuse that assigns orthogonal resources to each area freedom per user in cluster l and sector s over subband c
[34]–[36]. It is assumed that PT is the transmit power available c
when there are Uls active users, defined as the diversity order
at each cooperative sector BS (i.e., BPT is the total power per per scheduled user on that subband/sector, is given by ζ =
c c
cooperative cluster). B(NT −Uls )+1 for N-MIMO and ζ = NT −BUls +1 for LS-
MIMO. Therefore, an LS-MIMO setup with NT = BNTNM LS
Transmission between the cluster BSs and active mobile antennas per sector is comparable to a N-MIMO system with
station (MS) in the cooperative area is organized in time- only NTNM antennas per sector [6], although these must be
frequency resource allocation units, commonly known as re- linked to the rest of the cluster through a backhaul connection
source blocks (RB). Each RB is formed by a slot in the time wide-enough so as to allow the interchange of user data. With a
axis and a subband in the frequency axis: slight abuse of notation, we will use NT to denote the number
of antennas at each BS sector whenever the specific MIMO
• In the time axis, each RB occupies a time-slot of a fixed processing does not play a role.
duration Ts , assumed to be less than the channel cohe-
rence time. Thus, the channel fading can be considered A. Channel model
constant over the whole slot and it only varies from slot
to slot. Each of these slots consists of a fixed number The propagation channel between the mth cooperating BS
No of OFDM symbols of duration To + TCP = Ts /No , and MS k is characterized by a power delay profile [40],
where To denotes the duration of payload data, and TCP common to all pairs of transmit and receive antennas, that
is the cyclic prefix duration. Slot duration and scheduling can be expressed as
time interval will be used interchangeably throughout the Lp −1
Gk,m X 2
paper. Sk,m (τ ) = σl δ(τ − τl ), (1)
• Slotted transmissions take place over a bandwidth BW , Σ
l=0
which is divided into Nf orthogonal subcarriers, out of with Lp denoting the number of independent propagation
which Nd are used to transmit data and Np are used paths, σl2 and τl being, respectively, the power and delay of
to transmit pilots and to set guard frequency bands. PLp −1 2
the lth propagation path, Σ = l=0 σl , and
The Nd data subcarriers are split into Nb orthogonal
subbands, each consisting of Nsc adjacent subcarriers and −ς
Gk,m = $dk,mk,m 10χk,m /10 Σ (2)
with a bandwidth Bb = BW Nd /(Nf Nsc ) small enough
to assume that all subcarriers in a subband experience representing the average channel propagation gain between BS
frequency flat fading. Frequency subbands in a given slot m and MS k, where $ is a constant representing the path
are indexed by the set Nb = {1, . . . , Nb }. loss at fixed reference distance (typically chosen to be 1 m)
and assumed equal for all users, dk,m is the distance between
the mth BS and MS k, ςk,m is the path-loss exponent, and
In order to shift most of the processing burden to the BS χk,m is a random variable modeling the shadow fading, which
side, user terminals are assumed to have NR = 1 receive is assumed to follow a zero-mean Gaussian distribution with
antennas. We note, however, that the proposed framework can variance σχ2 . Hence, assuming that the channel coherence time
easily be extended to support MS with multiple antennas and is greater than the time slot duration Ts , the channel impulse
multi-stream transmissions to each user (see [29], [37]). Each response between transmit antenna nT of cooperating BS m
cooperative sector is assumed to have either NTNM or NTLS and the receive antenna of MS k, over the whole scheduling
transmit antennas, depending on whether the network operates period t, can be written as
using N-MIMO or LS-MIMO, respectively. Irrespective of
Lp −1
the cooperation scheme (N-MIMO or LS-MIMO), at each X
scheduling interval and on any subband c, the cluster BSs must hnk,m
T
(t; τ ) = hnk,m,l
T
(t)δ(τ − τl ), (3)
jointly determine the U c scheduled users on subband c out of l=0

the UTOT users in the cluster. The framework to be developed 2 Using LTE terminology, real-time services would be associated to guaran-
over the next sections is able to cater for heterogeneous teed bit rate (GBR) evolved packet system (EPS) bearers, and the non GBR
traffic requirements. To this end, three service classes will be bearers would be suitable for best-effort and non-real-time services [38].

0018-9545 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2017.2750922, IEEE
Transactions on Vehicular Technology
5

−ς
with E{|hnk,m,l
R ,nT
(t)|2 } = $dk,mk,m 10χk,m /10 σl2 , where the ex- from the interfering clusters. The NTNM B × Ulc matrix WUl c [c]
l
pectation is taken with respect to the fast fading given that any represents the precoding matrix jointly conducted by all BSs
large-scale fading component remains constant as they only conforming cluster l to its scheduled users on subband c. When
vary with macroscopic changes in the users’ positions, which using a column-wise normalized ZFBF, the precoding matrix
take place on a much coarser time scale. The corresponding is easily found to be
frequency response, when evaluated over subband c (with  l
zU c ,1 [c] zlU c ,U c [c]

center frequency fc ), can be safely approximated by WUl lc [c] = kzl l [c]k . . . kzl l l [c]k , (6)
U c ,1 U c ,U c
l l l
Lp −1
X where
hnk,m
T
(c, t) = hnk,m,l
T
(t)e−j2πfc τl . (4) h i
l=0 ZlUlc [c] = zlUlc ,1 [c] . . . zlUlc ,Ulc [c]
The NT × 1 column vector collecting the channel frequency  −1 (7)
response linking the transmit array in BS m in cluster j to =HlUlc [c] (HlUlc [c])H HlUlc [c] .
the ith user in cluster l over subband c during slot t will
Matrix PlU c = D(plU c ) is a power loading diagonal matrix
be denoted by hilmj [c, t]. For notational simplicity, and since l l

processing takes place on a slot-by-slot basis, the time index with the column vector plU c representing the distribution of
l

t will be dropped from this point onwards, except where the available power among the scheduled users in the cluster.
otherwise noted. Note that the use of N-MIMO requires of each BS sector to
have knowledge of all cluster users’ CSI (in all B sectors) as
well as to have all the data to be served to the scheduled users
B. Cooperative MIMO processing review available at all B BSs.
As in [6], the use of zero-forcing beamforming (ZFBF) b) LS-MIMO case: In this setup each cluster performs
is assumed for both types of MIMO processing given its the scheduling on a sector-basis but, keeping in mind that
attractive tradeoff between complexity and performance [21]. the number of users allocated to the different sectors are not
It is worth pointing out at this stage that any other linear independent as in any given sector, a sufficient number of
precoding technique can in principle be used. For exam- spatial degrees of freedom must be employed to cancel all
ple, if more antennas could be afforded at the MS, block- users scheduled on the same subband in the other sectors
diagonalization (BD) [41] could be used to decompose the forming the cluster. Nonetheless, it should hold that the total
c c LS
multiuser-multistream channel into its constituent orthogonal number of users h in cluster l, Ul1 + · ·i· + UlB ≤ NT . Denote
modes. Depending on the particular form of MIMO coopera- by Hlm [c] = hTlmj1 [c] · · · hTlmjUj [c] the NTLS × Uj channel
tion, different reception equations arise: matrix corresponding to the transmit array from sector m in
a) Network MIMO case: Denote by Ulc the set of cluster l to all users in cluster j. Furthermore let Ulm c
represent
scheduled users on cluster l and subband c with cardinality c
the set of Ulm users scheduled on the BS sector m of cluster l
|Ulc | = Ulc = Ul1 c c
+ · · · + UlB and fulfilling Ulc ≤ BNTNM . on subband c. For later convenience the set of users scheduled
Focusing now on the central cluster (cluster 0), the received in cluster l is defined by Ulc = {Ul1 c c
, · · · , UlB }. Similar to
signal for the scheduled users on subband c for an arbitrary the N-MIMO case, a reception equation can now be derived
time instant can be expressed as for the users in cluster 0 and sector m on a subband c as
0 H
yUNM
c [c] = (HU c [c])
0 0
WU0 0c [c]P0U0c sU0c [c] yULS0m
c [c] = (H
0m
[c])H WU0m 0m
c [c]PU c sU c [c]
0m
0m 0m
L−1 L−1 B
X XX H
+ (HlUlc [c])H WUl lc [c]PlUlc sUlc [c] +η U0c [c], (5) + Hln [c] WUlnln ln
c [c]PU c sU c [c] +η U c [c], (8)
ln
ln 0m
l=1 l=1 n=1
| {z } | {z }
φU c ϕU c
0 0m

h
1 U0c
iT where, as in (5), the first term corresponds to the desired
where yUNM c =
0
y U c ... y c
0 U represents the received signal signal, the second term represents the interference from other
0
for each specific scheduled user in cluster 0, HlU c [c] = c [c] ∼ CN (0U c , I U c ). Finally, sU c [c] is
clusters and η U0m
j 0m 0m lm
c
h i
hl1 [c] · · · hlU c [c] is an NTNM B × Ulc matrix collecting the the Ulm × 1 vector of information symbols for the users in
c
j
Ulm .
channel responses between the BSs forming the lth cluster
The column-wise normalized ZFBF precoder design now
and the users in the jth cluster, and η U0c [c] ∼ CN (0U0c , I U0c ).
follows,
Without loss of generality, it is assumed that each NTNM B × 1  lm
iT zlm

h zU c ,1 [c] U c ,U c [c]
vector hlu [c] = hTl1uj [c] · · · hTlBuj [c] , where u ∈ Ujc . That lm
WUjnc [c] =
jn
kzlm [c]k
jn
. . . kzlm l
[c]k
, (9)
U c ,1 U c ,U c
is, the first NTNM entries correspond to the channel response jn jn l

from the first BS in cluster l, the second NTNM × 1 entries where


correspond to the second BS, and so on, until the last NTNM ×1
h i
lm lm
Zlm
c
Ujn [c] = z c
Ujn ,1 [c] . . . z c c
Ujn ,Ul [c]
entries correspond to the Bth cooperating BS. Furthermore,
(10)
 −1 
note that the first term in (5) represents the desired information
 H
= Hlm Hlm Hlm

.
while the second term (vector φU0c ) collects all the signals c
Ujn

0018-9545 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2017.2750922, IEEE
Transactions on Vehicular Technology
6

The subscript [·]U c implies the selection of the precoder time slot, assuming queues of infinite capacity, can then be
0n
columns associated to the users scheduled on sector n of expressed as
cluster j. Matrix Pln = D(pln c ) is a diagonal matrix
c
Uln
ln
Uln Qlk (t + 1) = Qlk (t) + Alk (t) − Rkl (t) No To , (14)
with the column vector pU c performing the power allocation
among the scheduled users.
ln
where
Qlk (t)
 
In this scheme, each sector should have CSI knowledge Rkl (t) l
= min rk (t), , (15)
No To
of the channels between the sector’s antenna array and any
user in the cluster but, unlike N-MIMO, it does not require of with rkl (t) denoting the data rate allocated to user k in cluster
a wideband backhaul among cooperating BSs as the data to l during time slot t and it will be given by
be transmitted to users in a given sector need not be shared Nb
X
with the other BSs. Notice however that the BSs conforming a rkl (t) = Nsc ρcl,k (t), (16)
cluster still need a minimal, yet crucial, level of cooperation as c=1
each sector should be aware of the users scheduled on the other where the achieved rate for user k in cluster l over subband c
cluster sectors in order to design its own precoding matrix. is given by
1  
ρcl,k (t) = log2 1 + γUAlc ,k (t) , (17)
III. PHY AND DLC LAYER MODELS T0
A. PHY Layer Modeling with A denoting a tag that can take values N M or LS.
A cross-layer resource allocation strategy that, in order to
In order to provide a model of the physical layer processing, avoid the waste of resources, selects a transmission rate
the SINR is derived for each form of MIMO processing on
Ql (t)
the basis of (5) and (8). Under the ZFBF design constraint it rkl (t) ≤ k (18)
holds that No To
is said to fulfill the frugality constraint (FC) [9]. It was shown
Uc
h i
ΣNM 0 H 0
U0c = (HU0c [c]) WU0c [c] = D ωU1 0c . . . ωU c0 , (11) in [10] that, assuming the use of queue-aware scheduling rules,
0
negligible performance improvements can be expected upon
for the N-MIMO case, and the activation of the FC. Thus, for the sake of exposition
Uc simplicity, this constraint will not be considered in this paper
h i
ΣLS
c [c] = (H
U0m
0m
[c])H WU0m
c [c] = D ωU1 0m
c . . . ωU c0m ,
0m 0m when modeling the behaviour of the DLC layer.
(12)
for the LS-MIMO scheme. Now the instantaneous SINR IV. C LUSTERED CROSS - LAYER DESIGN
experienced by an arbitrary user k, γUNMc (N-MIMO) or γULSc ,k Irrespective of the cooperation scheme (N-MIMO or LS-
0 ,k 0m
(LS-MIMO), is given by MIMO), at each scheduling interval the cluster BSs must
 2 jointly determine, for each subband c, the U c targeted users out
PU0 c ,k ωUk 0c of the UT OT users in the cluster. Without loss of generality,
γUNM
0
c
0 ,k
= 2 (13a)
σ + |φU0c ,k |2 and under the assumption that the available power per BS,
 2 denoted by PT , is equal for all BSs in the cluster, the weighted
PU0mc ,k ωUk 0m
c sum-rate maximization problem for cluster 0 can now be
γULS0m
0m
c ,k = , (13b)
σ 2 + |ϕU0mc ,k |2 formally expressed as:
• LS-MIMO:
with PU0 c ,k in (13a) (PU0mc in (13b)) denoting the kth Nb XB
0 0m ,k X X
diagonal element of P0U0c (correspondingly P0U0m c ), and φU c ,k
0 c
max
0
wk ρc0,k
U0 ,PU c ,∀c∈Nb
in (13a) (ϕU0m c ,k in (13b)) being the kth element in φ c as c
c=1 m=1 k∈U0m
U0 0

defined in (5) (ϕU0m c as defined in (8)). Nb


X
A key background result:. The aggregate interference power s.t. kWU0m
0m
0m 2
c [c]PU c kF ≤ PT for m = 1, . . . , B.
0m
experienced by an arbitrary user k in both systems, φ2U c ,k (N- c=1
0
MIMO-based system) and ϕ2U c ,k (LS-MIMO-based system), (19)
0m
are equally distributed (see [6] for the proof). A consequence • N-MIMO:
of this result, used from this point onwards, is that lumping the Nb X
X
interference power term into the noise variance (i.e., Gaussian max wk ρc0,k
U0c ,PU c ,∀c∈Nb
0
hypothesis) does not affect any comparative results between 0 c=1 k∈U0c
the two systems. Per-BS power constraint:
Nb
X 2
s.t. B a WU0 0c [c]P0U0c ≤ PT , a = 1, . . . , B (20)

B. DLC Layer Modeling F
c=1
At the beginning of a given time slot t, MS k in cluster l Per-cluster power constraint:
is assumed to have Qlk (t) bits in the queue. Independently of Nb
the form of cooperative PHY processing, if there are Alu bits X
s.t. kWU0 0c [c]P0U0c k2F ≤ BPT ,
arriving during time slot t, the queue length at the end of this c=1

0018-9545 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2017.2750922, IEEE
Transactions on Vehicular Technology
7

where wk represents the weight associated to user k in the respectively, and consequently, the optimization can now be
examined user grouping that can be set in accordance to a conducted on a subband-by-subband basis. Previous studies
prescribed scheduling policy and matrix B a , defined as have shown that this frequency-domain uniform power al-
 location does not entail a dramatic performance loss while
B a = D [ 0| .{z
. . 0} |1 .{z
. . 1} |0 .{z
. . 0} ] , (21)
it greatly simplifies the optimization as it avoids having to
(a−1)NTNM NTNM (B−a)NTNM resort to dual-domain convex optimization techniques (see for
is in charge of selecting the precoder matrix rows belonging to example [10]). For the specific case of N-MIMO with PBPC,
each BS in order to check the fulfilment of the per-BS power even assuming that the power is uniformly allocated on a
constraints in the N-MIMO case. As usual, it is assumed that per-subband basis, optimally computing the per-user power
a given subband c is exclusively allocated to a selected user allocation so as to fulfill the PBPC requires of rather complex
group Ulc (for N-MIMO) or Ulm c
(for LS-MIMO). Formally, iterative optimization procedures (see [42]). Fortunately, the
the RB allocation constraints can be captured by constraining simple power allocation proposed in [43] has been shown to
the power allocation vectors as pc ∈ P c where, yield near optimal results in the context of multiuser MIMO-
n 0
o OFDMA networks [44] and will be the method used in this
|U |
P c , pU ∈ R+ : plUlc 6= 0 → plU 0 c = 0 ∀ Ulc 6= Ul c , work. The derivation of the precoding matrix for a collection
l
(22) of users on an arbitrary subband proceeds then as follows for
with R+ denoting the set of all non-negative real numbers. both cooperative schemes:
Hence, Tis the overall
 power allocation is defined as the vector LS-MIMO: Given an arbitrary user set on sector m of
T T
p = p1 , · · · , pNb it holds that cell 0, and focusing on subband c, construct the ZFBF matrix
WU0m c [c] and normalize its columns so as to ensure they
p ∈ P = P 1 × · · · × P Nb ⊂ RS+ , (23) 0m
3
have unit power h. Form a (diagonal) i power allocation matrix
0m
where × denotes the Cartesian product (or product set). PU0mc [c] = D PU0m
0m
0m
c ,1 · · · PU c ,U c
0m 0m
whose kth diagonal
Regardless of the cooperation approach, the solution of entry will depend on whether adaptive or uniform power
the optimization problems in (19) and (20) explicitly involve allocation (APA, UPA) is conducted across the selected users,
determining the per-subband optimum user group to be served 
PT
cooperatively and the corresponding power allocation but, 
 Nb U0mc for UPA
0m +
implicitly, also require of the computation of the suitable PU0m
c ,k = (26)

N wk Nsc σ 2
precoding filters. Note that the power allocation step involves  µsc
 0
c
− (ω k ) 2 for APA,
U0m c U0m
distributing the available power at each BS in the frequency
and user dimensions and, in the specific case of (20) care needs where in the case of APA the closed-form expression is
to be taken so as to ensure that the combination of the cluster obtained by using waterfilling [45] with µ0U0m c denoting the
precoding filter and power allocation does not violate any per- +
water level required to satisfy the power constraint and [x] =
BS constraint. Unsurprisingly, the optimal solution of (19) or max(0, x). Note that for the case of APA, the waterfilling
(20) entails a phenomenal computational complexity and thus, operation is applied on the diagonal entries of ΣLS c [c] that
U0m
some simplifications are enforced. In particular, the suboptimal effectively represent the combined effects of the channel and
strategy proposed in this work relies on two approximations. beamformer.
Firstly, the optimal power allocation across subbands and N-MIMO: Let U0c , with |U0c | ≤ BNT , denote an arbitrary
users is relaxed by assuming the available power is uniformly collection of users scheduled on subband c and chosen from
distributed among the subbands (yet remains adaptive across across the B sectors. The precoder design begins by con-
users) and, secondly, the exhaustive user selection search is structing the N-MIMO ZFBF matrix WU0 0c [c], which should
substituted by a greedy technique. Both procedures are covered be subsequently complemented by a power allocation matrix
in detail over the next two subsections. P0U0c [c] that enforces a prescribed power constraint.
When relying on a PCPC design, the power allocation
A. Power allocation procedure does not need to be aware of the specific BS/sector
Under the assumption that total power PT is evenly split in the cluster each user is connected to. In this case po-
over the Nb subbands, the constraints in (19) and (20) are wer allocation resembles the one developed for LS-MIMO:
now of the form a (diagonal) power allocation matrix is defined P0U0c [c] =
h i
PT D PU00c ,1 , · · · , PU00c ,U0c whose kth diagonal entry will de-
kWU0m 0m 2
c [c]PU c kF ≤ ∀c, for m = 1, . . . , B, (24)
0m 0m Nb pend on whether adaptive or uniform power allocation (APA,
and, UPA) is conducted across the selected users,

BPT
per-BS power constraint: 
 Nb U0c for UPA
2 0 +
PT PU0c ,k = (27)

B a WU0 0c [c]P0U0c ≤ ∀c, a = 1, . . . , B Nsc wk Nsc σ 2


µ0 c
− (ω k )2 for APA,
F Nb 
U0 c
U0
(25)
per-cluster power constraint:
BPT 3 With a slight abuse of notation, the column-normalized version of the
kWU0 0c [c]P0U0c k2F ≤ ∀c, ZFBF matrix will still be denoted by the same variable WU0m [c].
c
Nb 0m

0018-9545 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2017.2750922, IEEE
Transactions on Vehicular Technology
8

where in the case of APA the resulting expression is straight- charge of power loading the symbols to be transmitted to
forwardly derived using waterfilling with µ0U0c representing the the set of selected users in the bth sector/BS and denoted by
c c
water level required to satisfy the power constraint defined by U0b . The specific power allocation factor for user k ∈ U0b is
BPT . Analogously to the LS-MIMO setup, waterfilling is ap- computed as
plied on the combined channel+beamformer gains represented  PU0bc [c]
by the diagonal entries of ΣNM U0c .

 0
U0bc for UPA
0b
If PBPC is being pursued, a more elaborated power-limiting c ,k =
PU0b 
2
+ (31)
diagonal matrix design is required. In particular, the ZFBF pre-  Nsc0 wk − Nksc σ 2
 for APA,
µ c (ω
U
) Uc
0b 0b
coder matrix will need to be suitably modified so as to ensure
that the PBPC design is enforced. To this end, let us define where in the case of APA and µ0U c is the water level required
0b

the PBPC-scaled precoding matrix W̃U0 0c [c] = WU0 0c [c]Ω0U0c [c] to satisfy the power constraint given by PU0b0c [c] when applying
where the U0c × U0c matrix Ω0U0c [c] is in charge of suitably the waterfilling algorithm.
scaling the power on each BS and whose design is presented
next. Adapting the work in [43], define the power scaling B. User selection and scheduling
B
parameters {αl }l=1 , then Ω0U0c [c] = D ([α1 Iλ · · · αB Iλ ]) with The selection of users to be served over an arbitrary time
c
λ = bU0 /Bc. Define now the B × B matrix slot t is controlled by two interrelated mechanisms. On one
hand by the policy implemented to update the user weights in
2 2
 
h 0 i[1] h i[1]
(19) and (20), and on the other hand by the search method used
0

 WU0c [c] ··· WU0c [c] B 

 1 to perform the maximization of the prescribed utility function.

F F 
0

QU0c [c] = 
 .
.. . .. .
..

, Regarding the user weights, these are typically chosen so as

 h
 i[B] 2

h i[B] 2 
to achieve a compromise between system throughput and user
W0 c [c] ··· 0 
WU0c [c] B fairness. Weights following prescribed scheduling rules can

U0
1 F F
(28) be found in [10] and references therein; important examples
h i [b] comprise:
where each WU0 0c [c] is an NT × U0c /B submatrix of • Proportional fair (PF) rule is based on a channel-aware
j
WU0 0c [c] containing the entries corresponding to the precoding traffic-agnostic scheduling rule aiming at maximizing the
coefficients of BS b to the jth group of users. Assuming logarithmic-sum-throughput of the system with
now, without loss of generality, that all BSs have the same
wk = 1/rk , ∀ k, (32)
power available PT , define the power constraint vector p =
(PT /Nb )1B . Having all these definitions at hand, the computa- where rk is the average effective data rate actually
tion of the diagonal elements of the power-constraining matrix allocated to user k, which is calculated using a moving
Ω0U0c [c] reduces to solving the linear system of equations average over a sliding window typically implemented
 −1 using exponential weighting. In particular, an estimate
2 T
α = α12 · · · αB = Q0U0c [c]
 
p. (29) of the average rate at scheduling instant t for user k in
cluster l is determined as
If any of the entries of α is negative, it reveals that the solution
found is infeasible, in which case, a conservative approach is rlk (t) = (1 − β)rlk (t − 1) + βrkl (t),
taken by defining Ω0U0c [c] = ᾱIλB , where where β is a forgetting factor used to weight the influence
v ! of the older samples. Without loss of generality, in this
2
u h i[b] i[b] 
u
0
h
0 work a value of β = 0.01 has been used, which roughly
ᾱ = tmax WU c [c] · · · WU c [c]
. (30)
b 0 1 0 B F
corresponds to an averaging window of length 100.
• Modified largest weighted delay first (MLWDF) is based
Remarkably, this simple procedure employs all the available on a channel- and queue-aware scheduling rule. The
power on the B BSs whenever possible while resorting to MLWDF scheduler aims at choosing the best combination
the conservative strategy of keeping the maximum power of queueing delay and potential transmission rate by using
transmission on one BS and scaling down the power in the weights,
other B−1 BSs. Numerical results shown in Section V confirm wk = φk WHOL,k /rk , ∀ k, (33)
the effectiveness of this approach.
Note that the power consumed by the precoder design where φm are arbitrary positive constants that can be
W̃U0 0c [c] at BS b on subband c is given by PU0b0c [c] = used to set different priority levels between traffic flows,
h i 2 h i and WHOL,k denotes the head-of-line (HOL) delay expe-
W̃U0 0c [c] where W̃U0 0c [c] collects the NT rows in rienced at the BS buffer of MS k. In order to support

b F b
W̃U0 0c [c] corresponding to the precoding matrix applied at users with absolute delay requirement Ďk and maximum
the bth BS. Similar to the LS-case, a (block diagonal) form outage delay probability requirement ξˇk , if this can be
power allocation matrix can now be formed as P0U0m c [c] =
done at all with any scheduling rule, the authors of [46]
h
01 0B
i
0b
propose to properly set the values of φk as
D PU0m c · · · PU0m
c ,U c with each term PU0m c ,U c =
h 0mi 0b
log(ξˇk )
D PU0bc ,1 · · · PU0bc ,U c being itself a diagonal matrix in φk = − , (34)
0b 0b 0b Ďk

0018-9545 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2017.2750922, IEEE
Transactions on Vehicular Technology
9

providing in this way QoS differentiation among flows Table I: Per-subband weighted maximum sum rate suboptimal
and approximately “balancing” different users’ probabi- user selection algorithm for LS-MIMO. This algorithm should
lities of QoS constraints violation. be executed on every subband 1 ≤ c ≤ Nb .
• Exponential (EXP) rule is also based on a channel- and
Inputs: Cluster 0 channel responses on subband c (hk0m0 [c], k ∈ [1, Ul ],
queue-aware throughput optimal scheduling rule that con- m ∈ [1, · · · , B]), and user weights (wk ).
siders the waiting time in the queues, the instantaneous Initialization: Potential users: UPc OT = UT OT , Selected users:
c = · · · = Uc = ∅
U01
potential transmission rates and the maximum tolerable 0B

delay requirements. Denoting by Nk the number of users 1) Select (weighted) strongest user k in any of the sectors:
k∗ = max wk log2 1 + khk0m0 0 [c]k22 ,

in the system, the weights in this case can be shown to c
k∈UP OT
be defined by
! where m0 denotes here the sector to which user k belongs. (36)
c ∗
φk φk WHOL,k − φW U0m 0 = {k }
wk = exp p UPc OT = UPc OT \ k∗ .
rk 1 + φW
ρU0c = wk∗ log2 1 + khk∗ 0m0 0 [c]k22

for all k, with ρT M P = 0
2) While U0c < NTLS and ρU0c ≥ ρT M P
Nk
1 X a) ρT M P = ρU0c
φW = φu WHOL,k . (35) b) For every k ∈ UPc OT
Nk
k=1 i) Define candidate set U0c0 = U0c , k .

c
ii) Partition user set in sectors U00 = {U00 1 , · · · , U00 B }
• Round robin (RR) totally neglects any channel/queue
iii) For each user sector partition U0c0 m .
knowledge and it can easily be implemented by setting 0
A) Select channel rows to form H0 m [c]-
wk = ∞, ∀ k , and randomly performing the user 00 m [c].
B) Construct ZFBF precoder WU c
selection. 0
0 m [c] to unit power.
00 m
C) Normalize columns of WU c
• Maximum sum rate (MSR), as RR, totally neglects any 00 m

queue knowledge and it can easily be implemented by D) Compute channel-precoder equivalent gains ΣLSU c0 [c].
0 m
setting wk = 1, ∀ k , but now user selection is conducted E) Perform multiuser power allocation using (26).
LS
F) Compute SINRs for all users in sector m, γU c .
by maximizing the utility function described in (19) and 00 m
iv) Evaluate utility ρU c0 for current candidate selection:
(20), that will exclusively depend on the CSI of each user. 0  
ρU c0 = B LS
P P
Irrespective of the user scheduling policy, and recalling the m=1 k∈U c wk log2 1 + γU c ,k
0 00 m 00 m
infeasibility of conducting an exhaustive search, two greedy- v) If ρU c0 > ρU0c then ρU0c = ρU c0 , k∗ = k.
0 0
type algorithms are proposed, one for LS-MIMO and another c) If ρU0c > ρT M P then U0c = {U0c , k∗ }, UPc OT = UPc OT \ k∗ .
one for N-MIMO, to determine appropriate (albeit suboptimal) Outputs: Selected user set U0c and corresponding power allocation P0U c .
0
user groups U0c in (19) and (20). In both cases the greedy
algorithm proceeds by adding to the already scheduled user
pool the single user that provides the highest increase in
sectors whereby some sectors may schedule more users than
(weighted) sum rate, considered here to be the utility function
others as long as the constraint on the total number of users is
to be maximized and denoted by ρU0c . This procedure is
satisfied (U c ≤ NTLS ). The iterative procedure finalizes either
repeated until the maximum number of users is reached or
when the maximum number of users in the cluster (NTLS ) has
the addition of a new user actually causes a decrease in the
been reached or no extra users can be found able to increase
utility function. Owing to the uniform power allocation in
the utility function ρU0c . For clarity, all the required steps to
the frequency domain, this greedy search can be conducted
perform the greedy user selection are compiled in Table I.
independently on each subband. Following subsections des-
In fact, note that the execution of this algorithm for all the
cribe the specifics of each greedy procedure. Nevertheless, it is
subbands c ∈ Nb constitutes a suboptimal solution to the
important to stress at this point, that should an alternative user
cross-layer design problem stated in (19).
selection approach wished to be used such as one based on a
metaheuristics or convex-relaxation, the cross-layer framework
proposed in this paper and culminated by equations (26) and D. N-MIMO user selection
(31), would be barely affected provided that the user selection
procedure returns an appropriate U0c (note that per-BS power The greedy user scheduling for N-MIMO mimics that of
constraints would likely need to be incorporated to the search LS-MIMO with the particularity that now per-base power
procedure). constraints may need to be explicitly enforced in the precoder
design. This is shown in Table II in Step 2.b.iv) where the
diagonal matrix D00 [c] is computed in accordance to the
C. LS-MIMO user selection procedure described in [43]. Note that if matrix D00 [c] = I U0c0 ,
In the particular case of LS-MIMO, the greedy algorithm the per-base constraint is relaxed to a per-cluster constraint.
proposed here proceeds on a subband-by-subband basis by Finally, note that once the allocation of all subbands has
selecting at each iteration step the best user irrespective of its been conducted, it is then possible to estimate the global rate
serving sector. Therefore, and unlike the RR scheme evaluated assigned to every user on a given time slot and it is at this
in [6], this strategy allows for the uneven load of the different point that user weights wk can be updated accordingly.

0018-9545 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2017.2750922, IEEE
Transactions on Vehicular Technology
10

Table II: Per-subband weighted maximum sum rate suboptimal ×10 6


12
user selection algorithm for N-MIMO. This algorithm should
be executed on every subband 1 ≤ c ≤ Nb .

Average served user throughput (b/s)


10
PCPC
Inputs: Cluster 0 channel responses: hk0m0 [c], k ∈ 1, Ulc ,
 
m ∈ [1, · · · , B] 8
User weights: wk .
Initialization: Available users: UPc OT = UT OT , Selected users: PBPC
c = · · · = Uc = ∅
U01 0B 6
1) Select (weighted) strongest user k in any of the sectors:
N-MIMO: APA with PCPC
k∗ = max wk log2 1 + khk0m0 0 [c]k22 ,

c 4 N-MIMO: UPA with PCPC
k∈UP OT
N-MIMO: APA with PBPC
where m0 denotes here the sector to which user k belongs. (37) N-MIMO: UPA with PBPC
c ∗
U0m 0 = {k } 2 LS-MIMO: APA
∗ LS-MIMO: UPA
UPc OT = UPc OT \k .
0
1 + khk∗ 0m0 0 [c]k22

ρU0c = wk∗ log2 0 0.5 1 1.5 2 2.5
ρT M P = 0 Average arrival user traffic (b/s) ×10 7
2) While |U0c | < BNTNM and ρU0c ≥ ρT M P
a) ρT M P = ρU0c Figure 2: Per-user throughput vs incoming traffic for RR
b) For every k ∈ UPc OT scheduling. UT OT =36 user.
i) Define candidate set U00 = U0c , k .

0
ii) Form channel matrix H0 [c].
iii) Construct ZFBF precoder WU 00 [c].
c
00 the pathloss adheres to the COST231 Type E propagation
iv) Enforce the power constraint
• if PCPC: Perform multiuser power allocation using (27).
model. Unless otherwise noted, the total number of users in the
• if PBPC: Perform multiuser power allocation using (31). cluster being served cooperatively is fixed to be UT OT = 36,
v) Compute SINRs for all users in the cluster, γU NM
c UT OT /B = 12 users are uniformly scattered within each
00
vi) Evaluate utility ρU c0 of current candidate selection: sector of the cooperation area and excluding areas where
0  
P
ρU c0 = k∈U c wk log2 1 + γU NM cooperation is not allowed (i.e., within the non-cooperation
c ,k
0 00 00 radius Rnc , see Fig. 1). In the case of N-MIMO, each sector
vii) If ρU c0 > ρU0c then ρU0c = ρU c0 , k∗ = k.
0 0 is equipped with NTNM = 5 antennas, whereas in the case
c) If ρU0c > ρT M P then U0c = {U0c , k∗ }, UPc OT = UPc OT \ k∗ .
of LS-MIMO NTLS = BNTNM = 15 antennas per sector are
Outputs: Selected user set U0c and power allocation P0U c .
0 used. Note that, in both cases, N-MIMO and LS-MIMO, the
cluster can simultaneously schedule up to 15 users from the
cooperative area.
E. Discrete rate allocation
The design introduced in (19)-(20) aims at maximising A. Single-class traffic scenario
a continuous rate metric (i.e., capacity) whereas practical
As a first scenario, a situation where all users in the system
systems require the use of discrete transmission modes whose
belong to the same traffic class is considered. For comparative
instantaneous rate is defined by a specific choice of modulation
purposes, we choose as a baseline setup an RR scheduler
format and coding rate. Adaptation to the channel condition is
that selects a fixed number of users over the cooperative area
then conducted using adaptive modulation and coding (AMC)
U c = 15 which, in both cases (N-MIMO and LS-MIMO),
whereby the transmission mode that maximizes the instanta- c
is achieved by scheduling Ulm = 5 users from each cluster
neous throughput subject to the fulfilment of a packet error
on a particular subband c. Note that, parallelizing the work
constraint is chosen. It is worth pointing out that the capacity-
in [6], in the case of LS-MIMO, each sector has NTLS = 15
based design proposed in this paper can easily be transformed c
DoFs which are exploited to transmit data to Ulm users in its
to a one using discrete rate allocation by suitably performing LS c
coverage area while cancelling NT −Ulm interferers from the
an instantaneous SINR-to-transmission mode conversion such
other two sectors. In contrast, the N-MIMO scheme utilizes
as the one shown in [10], [29].
the overall 15 DoFs provided by the BNTNM antennas to jointly
transmit data to the same number of users U c = 15. In fact,
V. N UMERICAL RESULTS as mentioned in [6], the spatial DoFs per-scheduled user is
c
Numerical results are shown here for a cluster formed by given by ζ = NTLS − BUlm + 1 = 1 for LS-MIMO and
NM c
B = 3 sectors with cell radius Rc = 1000 m operating ζ = BNT − BUlm + 1 = 1 for N-MIMO, thus making
over a bandwidth of BW = 10 MHz by means of Nb = this comparison a sensible one.
50 RBs and using a transmit power at each sector BS of Figure 2 compares the system performance between LS-
PT = 13 dBW. The non-cooperation radius has been set MIMO and N-MIMO when using RR and different power
to Rnc = 500 m. Shadowing has been modeled as a zero- allocation strategies. Performance is assessed by examining
mean log-normal random variable with a variance equal to the served per-user traffic vs the incoming per-user traffic.
6 dB and the power delay profile used to generate the fast Irrespective of the cooperative technique in use, for low traffic
fading conforms to the ETU profile from 3GPP [47] whereas arrivals, all curves begin as straight lines with unit slope

0018-9545 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2017.2750922, IEEE
Transactions on Vehicular Technology
11

×10 7
3 1

18 users
Average served user throughput (b/s)

2.5 0.95

PF scheduling

2 0.9

JFI throughput
1.5 0.85
36 users

1 0.8

RR scheduling
N-MIMO - PCPC
LS-MIMO
0.5 0.75 N-MIMO - PBPC
N-MIMO with PCPC
LS-MIMO
NMIMO with PBPC
0 0.7
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Average arrival user traffic (b/s) ×10 7 Average arrival user traffic (b/s) ×10 7

Figure 3: Per-user throughput vs incoming traffic for PF Figure 4: Throughput Jain fairness index vs incoming traffic
scheduling. UT OT =36 user. Uniform power allocation. for RR and PF scheduling. UT OT =36 user. Uniform power
allocation.

indicating that all incoming traffic is served. As traffic arrivals Noticeably, the clear advantage shown by LS-MIMO in the
increase, a saturation point is reached revealing at which case of RR scheduling is drastically reduced when considering
arrival rate traffic is not fully served. The first important point the use of a channel-aware scheduler such as PF. In particular,
to mention, and the one that serves to confirm the results in N-MIMO, when subject to a PCPC, virtually attains the
[6], is the clear superiority of LS-MIMO in comparison to N- performance of LS-MIMO whereas when a PBPC is enforced
MIMO, which on average results in a 50% gain in throughput the degradation in throughput remains around a 10%.
(in the saturation region). Nevertheless, the introduced cross- Complementing throughput results, Fig. 4 depicts the
layer framework allows us to derive further insights. Firstly, it throughput Jain’s fairness index (JFI) [48] for both RR and
can be observed that the spatial power allocation brings along PF. Recalling that a JFI close to one indicates that all users in
an almost negligible gain in the case of LS-MIMO and a rather the system attain a similar throughput, note how for low arrival
modest one for N-MIMO. In contrast, the fact that N-MIMO rates, where all users, irrespective of their channel conditions,
is subject to PBPC is seen to have a greater impact on the are able to comfortably dispatch their incoming traffic, the
throughput performance in comparison to the PCPC case. Note JFI remains very close to one regardless of the cooperative
that the PCPC performance can be taken as an upper-bound technique or power constraint. In contrast, as the arrival rate
of the performance attained when using an optimal design increases, it can be clearly appreciated how fairness in the
with PBPC such as the one proposed in [42]. Remarkably, the RR-based system quickly degrades as users suffering from
difference in saturation throughput between PCPC and PBPC poor channels are assigned, on average, the same amount of
is below 5%, thus suggesting that the simple PBPC design resources as users experiencing good channels, thus leading
proposed here is bound to perform rather close to optimality. to disparate throughput performance among users. This effect,
Figure 3 repeats the LS-MIMO vs N-MIMO comparison but although also noticeable in the case of the PF scheduling, is
now using the PF scheduler when having either UT OT = 36 considerably limited by the scheduling policy, which is able
or UT OT = 18 users in the cooperative area. Recall that PF to allocate more resources to those users who need them yet
scheduling aims at exploiting the existing multiuser diversity at the cost of sacrificing overall cluster throughput.
while preserving some degree of fairness among users. Also
note that by virtue of the greedy user selection algorithms
where the PF is embedded, and unlike the RR scheduler B. Multi-class traffic scenario
case, the number of simultaneously scheduled users on each Next, a multi-class scenario has been considered where the
subband does not necessarily reach the theoretical maximum UT OT = 36 users are partitioned into three equal-sized groups
U c = 15 as often the addition of a new user to the scheduling (12 users per group) with each group representing a different
set would bring along a decrease in the utility function whose traffic class, in particular and as mentioned in Section II, the
maximization is sought. For the sake of clarity, results only classes considered are RT, nRT and BE. In line with current
when using UPA are shown notwithstanding the fact that the LTE numerology [38], maximum allowable delays for each
APA counterparts provide marginally higher throughput in class are set to 50, 100 and 300 ms with a delay outage proba-
all cases. As expected, and concentrating on the results for bility of 0.01, 0.1 and 0.1, respectively. Given maximum delay
UT OT = 36, PF results in a much higher saturation throughput and delay outage probabilities, it is possible to define service
than RR (nearly double) while being able to sustain, without coverage [10] as the percentage of users who achieve their
losses, a traffic arrival rate that nearly trebles that of RR. QoS requirements in terms of maximum allowable average or

0018-9545 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2017.2750922, IEEE
Transactions on Vehicular Technology
12

×10 7 ×10 7
2.5 2.5

LS-MIMO - RT traffic
LS-MIMO - RT traffic
LS-MIMO - nRT traffic
LS-MIMO - nRT traffic

Average served user throughput (b/s)


Average served user throughput (b/s)

LS-MIMO - BE traffic 2 LS-MIMO - BE traffic


2
N-MIMO - RT traffic N-MIMO - RT traffic
N-MIMO - nRT traffic N-MIMO - nRT traffic
N-MIMO - BE traffic
N-MIMO - BE traffic
1.5 1.5

1 1

0.5 0.5

0 0
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Average arrival user traffic (b/s) ×10 7 Average arrival user traffic (b/s) ×10 7

Figure 5: Per-user throughput vs incoming traffic for greedy Figure 6: Per-user throughput vs incoming traffic for greedy
EXP scheduling. UT OT =36 user (12 RT+ 12 nRT + 12 BE). MLWDF scheduling. UT OT =36 user (12 RT+ 12 nRT + 12
Uniform power allocation. Per-base power constraint. BE). Uniform power allocation. Per-base power constraint.

class: whereas for BE users, and to a lesser extent for nRT


absolute delay. Having established in the previous subsection
users, delay quickly becomes unbounded effectively implying
that APA brings along little benefit to the overall system
that the BE class has ceased to be served, for RT users the
performance, the results shown next have all been generated
delay rises at a rather more moderate pace as these users
using UPA. Also, and since in most practical deployments of
capture all of the system resources. LS-MIMO undoubtedly
N-MIMO PBPC will typically be required, this has been the
outperforms N-MIMO, a fact clearly observed at an incoming
power constraint enforced. For clarity of presentation, only
traffic rate of 15 Mb/s, where the average user delay for all
channel- and queue-aware scheduling policies are considered
traffic classes is already exceeding their maxima in the case of
in this section, namely, EXP and MLWDF, noting that in the
N-MIMO while they are still kept within bounds in the case of
case of heterogeneous traffic it would not be very appropriate
LS-MIMO. To fully appreciate the merits of LS-MIMO over
to employ a scheduling strategy, such as PF, that obviates the
N-MIMO, it is enlightening to examine the service coverage
state of the queues.
performance of both schemes as shown in Fig. 8 for the EXP
Figure 5 compares the performance, in terms of served vs scheduler. In particular, it can be seen how LS-MIMO is able
arrived traffic, of N-MIMO and LS-MIMO when using the to sustain the demands of 90% of the RT users for up to 18.75
greedy EXP scheduler. Logically, irrespective of the MIMO Mbps per-user incoming rate whereas in the case of N-MIMO
technique in use, the scheduler works as intended: at low the service coverage abruptly falls at an arrival rate around
traffic arrival rates all users are served whereas as incoming 13.75 Mbps. Similar trends can also be observed for nRT and
traffic increases, first service to BE users and later to nRT BE users.
users begin to fall in an attempt to sustain the demands of Finally, Fig. 9 depicts the delay JFI performance correspon-
RT users. Eventually, as it can already be hinted in Fig. 5 for ding to the results shown in Figs. 5, 7 and 8. The overall
the highest arrival rate, even the RT users’ throughput begins (across all users) JFI delay reveals marked differences among
to saturate indicating that not all RT traffic is being served. users’ delay performance; in contrast, within each user class,
Note how differences between LS-MIMO and N-MIMO are a very high degree of fairness is achieved. This leads to the
marginal for the case of RT traffic but very remarkable for nRT conclusion that a queue- and channel-aware scheduling rule
and BE users. Demonstrating the versatility of the proposed such as the EXP scheduler rightly discriminates among the
framework, Fig. 6 presents results but now using the MLWDF different traffic classes yet attaining almost perfect fairness
scheduler. It can be observed how this scheduler provides within each class.
a different trade-off among the performance of the different
traffic classes sacrificing some of the performance of RT users
to considerably enhance that of nRT and BE users. VI. C ONCLUSION
Since the QoS is specified in terms of delay requirements, This work has proposed a PHY/DLC cross-layer design for
Fig. 7 examines the delay performance of the different user scheduling and resource allocation in cluster-based cooperative
classes for the case of EXP scheduling. In agreement with the MIMO-OFDMA networks that relies on both channel and
throughput results, delay is kept within the tolerance limits of queue state information. The framework is versatile enough so
each class roughly up to the arrival rate at which traffic begins as to take into account different forms of MIMO processing,
to stop being served (as can be seen in Fig. 5). Notice the most notably, N-MIMO and LS-MIMO, and different power
difference in delay performance in accordance with the user constraints and power allocation techniques. The scheduling

0018-9545 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2017.2750922, IEEE
Transactions on Vehicular Technology
13

1000 1

900
LS-MIMO - RT traffic
LS-MIMO - nRT traffic 0.9
800 LS-MIMO - BE traffic
N-MIMO - RT traffic LS-MIMO - RT traffic
Average user delay (ms)

700 N-MIMO - nRT traffic LS-MIMO - nRT traffic


N-MIMO - BE traffic 0.8 LS-MIMO - BE traffic
600 LS-MIMO - Overall

JFI delay
N-MIMO - RT traffic
500 0.7 N-MIMO - nRT traffic
N-MIMO - BE traffic
N-MIMO - Overall
400
0.6
300

200
0.5
100

0 0.4
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Average arrival user traffic (b/s) ×10 7 Average arrival user traffic (b/s) ×10 7

Figure 7: Average user delay vs arrival rate for greedy EXP Figure 9: Delay Jain fairness index vs arrival rate for greedy
scheduling. UT OT =36 user (12 RT+ 12 nRT + 12 BE). EXP scheduling. UT OT =36 user (12 RT+ 12 nRT + 12 BE).
Uniform power allocation. Per-base power constraint. Uniform power allocation.

100
in the case of real-time traffic.
90 Further work will progress in three directions. Firstly,
80 the framework will be extended to also include the non-
cooperative region. To this end, the effects of any frequency
70
Service coverage (%)

reuse policy will need to be taken into account in the cross-


60 layer design. Secondly, the effects of increasing the number
50
of BS antennas to what can be considered truly massive will
also be explored. Finally, the application of alternative forms
40
LS-MIMO - RT traffic
of user selection using metaheuristic techniques, on their own
30 LS-MIMO - nRT traffic or in combination with greedy approaches, will be explored,
LS-MIMO - BE traffic
20 N-MIMO - RT traffic
alongside with different precoding strategies.
N-MIMO - nRT traffic
10 N-MIMO - BE traffic
ACKNOWLEDGEMENTS
0
0.5 1 1.5 2 2.5 This work has been supported by MINECO (Spanish Go-
Average arrival user traffic (b/s) ×10 7 vernment) and FEDER under project TEC2014-59255-C3-2-
R.
Figure 8: Service coverage vs arrival rate for greedy EXP
scheduling. UT OT =36 user (12 RT+ 12 nRT + 12 BE).
Uniform power allocation. Per-base power constraint. R EFERENCES
[1] A. J. Paulraj, D. A. Gore, R. U. Nabar, and H. Bolcskei, “An overview of
MIMO communications-a key to gigabit wireless,” Proc. IEEE, vol. 92,
no. 2, pp. 198–218, 2004.
process is a major component of the design. In particular, new [2] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta,
user selection techniques have been proposed that, based on O. Edfors, and F. Tufvesson, “Scaling up MIMO: Opportunities and
challenges with very large arrays,” IEEE Signal Processing Magazine,
greedy principles and combined with different schedulers, lead vol. 30, no. 1, pp. 40–60, Jan 2013.
to a variety of tradeoffs between network performance and [3] J. F. Monserrat, G. Mange, V. Braun, H. Tullberg, G. Zimmermann, and
user fairness. The framework has been used to comparatively Ö. Bulakci, “METIS research advances towards the 5G mobile and wire-
less system definition,” EURASIP Journal on Wireless Communications
analyze in the context of ZF beamforming, the merits of N- and Networking, vol. 2015, no. 1, pp. 1–16, 2015.
MIMO and LS-MIMO showing that the huge performance [4] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K.
advantage of LS-MIMO over N-MIMO under a simple RR Soong, and J. C. Zhang, “What will 5G be?” IEEE Journal on Selected
Areas in Communications, vol. 32, no. 6, pp. 1065–1082, June 2014.
scheduling rule, vanishes considerably when applying more [5] D. Gesbert, S. Hanly, H. Huang, S. S. Shitz, O. Simeone, and W. Yu,
sophisticated schedulers. This makes the decision to use one “Multi-cell MIMO cooperative networks: A new look at interference,”
MIMO form or the other critically influenced by the impair- IEEE Journal on Selected Areas in Communications, vol. 28, no. 9, pp.
1380–1408, December 2010.
ments/costs of deploying a large number of antennas at a [6] K. Hosseini, W. Yu, and R. S. Adve, “Large-scale MIMO versus network
single site or the laying of a high-capacity backhaul among MIMO for multicell interference mitigation,” IEEE Journal of Selected
cooperating BSs. Nevertheless, if both options are feasible, Topics in Signal Processing, vol. 8, no. 5, pp. 930–941, Oct 2014.
[7] C. Zhong, L. Yang, and M. You, “Dynamic resource allocation for
LS-MIMO is confirmed as the preferred technique given its downlink multiuser MIMO-OFDMA/SDMA systems,” Circuits, Systems
significant advantage in terms of lower delay, a critical metric and Signal Processing, vol. 29, no. 6, pp. 1061–1074, 2010.

0018-9545 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2017.2750922, IEEE
Transactions on Vehicular Technology
14

[8] S. Chieochan and E. Hossain, “Adaptive radio resource allocation in [30] M. Schellmann, Z. Zhao, H. Lin, P. Siohan, N. Rajatheva, V. Luecken,
OFDMA systems: a survey of the state-of-the-art approaches,” Wireless and A. Ishaque, “FBMC-based air interface for 5G mobile: Challen-
Communications and Mobile Computing, Published online in Wiley ges and proposed solutions,” in 2014 9th International Conference
InterScience (DOI: 10.1002/wcm.696), 2008. on Cognitive Radio Oriented Wireless Networks and Communications
[9] G. Song, Y. Li, and L. Cimini, “Joint channel- and queue-aware (CROWNCOM), June 2014, pp. 102–107.
scheduling for multiuser diversity in wireless OFDMA networks,” IEEE [31] N. Michailow, M. Matthe, I. S. Gaspar, A. N. Caldevilla, L. L.
Tran. Commun., vol. 57, no. 7, pp. 2109–2121, 2009. Mendes, A. Festag, and G. Fettweis, “Generalized frequency division
[10] G. Femenias, B. Dañobeitia, and F. Riera-Palou, “Unified approach multiplexing for 5th generation cellular networks,” IEEE Transactions
to cross-layer scheduling and resource allocation in OFDMA wireless on Communications, vol. 62, no. 9, pp. 3045–3061, Sept 2014.
networks,” EURASIP Journal on Wireless Communications and Networ- [32] X. Li, E. Bjornson, S. Zhou, and J. Wang, “Massive MIMO with multi-
king, vol. 2012, no. 1, pp. 1–19, 2012. antenna users: When are additional user antennas beneficial?” in 2016
[11] G. Femenias, F. Riera-Palou, and J. S. Thompson, “Robust scheduling 23rd International Conference on Telecommunications (ICT), May 2016,
and resource allocation in the downlink of spatially correlated MIMO- pp. 1–6.
OFDMA wireless systems with imperfect CSIT,” IEEE Transactions on [33] I. D. Garcia, N. Kusashima, K. Sakaguchi, K. Araki, S. Kaneko,
Vehicular Technology, vol. 65, no. 2, pp. 614–629, Feb 2016. and Y. Kishi, “Impact of base station cooperation on cell planning,”
[12] E. Bjornson and E. Jorswieck, Optimal Resource Allocation in Coordi- EURASIP Journal on Wireless Communications and Networking, vol.
nated Multi-Cell Systems. Now Publishers, 2013, vol. 9, no. 2-3. 2010, p. 5, 2010.
[34] J. Garcia-Morales, G. Femenias, and F. Riera-Palou, “Analytical per-
[13] J. Li, C. Botella, and T. Svensson, “Resource allocation for clustered
formance evaluation of OFDMA-based heterogeneous cellular networks
network MIMO OFDMA systems,” EURASIP Journal on Wireless
using FFR,” in IEEE 81st Vehicular Technology Conference (VTC Spring
Communications and Networking, vol. 2012, no. 1, pp. 1–15, 2012.
2015), May 2015, pp. 1–7.
[14] D. Choi, D. Lee, and J. H. Lee, “Resource allocation for CoMP with [35] ——, “On the analysis of channel-aware schedulers in OFDMA-based
multiuser MIMO-OFDMA,” IEEE Transactions on Vehicular Techno- networks using FFR,” in 2015 IEEE 11th International Conference
logy, vol. 60, no. 9, pp. 4626–4632, Nov 2011. on Wireless and Mobile Computing, Networking and Communications
[15] B. Huang, J. Li, and T. Svensson, “A Utility-Based Joint Resource (WiMob), Oct 2015, pp. 786–793.
Allocation Approach for Multi-Service in CoMP Networks,” Wireless [36] ——, “Analysis and optimization of FFR-aided OFDMA-based hetero-
Personal Communications, vol. 72, no. 3, pp. 1633–1648, 2013. geneous cellular networks,” IEEE Access, vol. 4, pp. 5111–5127, 2016.
[16] W. Yu, T. Kwon, and C. Shin, “Multicell coordination via joint schedu- [37] B. Danobeitia, G. Femenias, and F. Riera-Palou, “An optimization
ling, beamforming, and power spectrum adaptation,” IEEE Transactions framework for scheduling and resource allocation in multi-stream hete-
on Wireless Communications, vol. 12, no. 7, pp. 1–14, 2013. rogeneous MIMO-OFDMA wireless networks,” in IFIP/IEEE Wireless
[17] M. Li, I. B. Collings, S. V. Hanly, C. Liu, and P. Whiting, “Multicell Days (WD), Dublin (Ireland), 2012.
coordinated scheduling with multiuser zero-forcing beamforming,” IEEE [38] C. Cox, An Introduction to LTE: LTE, LTE-Advanced, SAE and 4G
Transactions on Wireless Communications, vol. 15, no. 2, pp. 827–842, Mobile Communications. Wiley, 2012.
2016. [39] X. Wang, G. Giannakis, and A. Marques, “A unified approach to
[18] E. Castañeda, A. Silva, A. Gameiro, and M. Kountouris, “An overview QoS-guaranteed scheduling for channel-adaptive wireless networks,”
on resource allocation techniques for multi-user MIMO systems,” IEEE Proceedings of the IEEE, vol. 95, no. 12, pp. 2410–2431, 2007.
Communications Surveys & Tutorials, vol. 19, no. 1, pp. 239–284, 2017. [40] J. G. Proakis, Digital Communications, 4th ed. McGraw Hill, 2001.
[19] G. Dimic and N. D. Sidiropoulos, “On downlink beamforming with [41] Q. H. Spencer, A. L. Swindlehurst, and M. Haardt, “Zero-forcing met-
greedy user selection: performance analysis and a simple new algo- hods for downlink spatial multiplexing in multiuser MIMO channels,”
rithm,” IEEE Transactions on Signal processing, vol. 53, no. 10, pp. IEEE Tran. Signal. Proc., vol. 52, no. 2, pp. 461–471, 2004.
3857–3868, 2005. [42] R. Zhang, “Cooperative multi-cell block diagonalization with per-base-
[20] Z. Shen, R. Chen, J. G. Andrews, W. H. Heath, and B. L. Evans, “Low station power constraints,” IEEE Journal on Selected Areas in Commu-
complexity user selection algorithms for multiuser MIMO systems with nications, vol. 28, no. 9, pp. 1435–1445, December 2010.
block diagonalization,” IEEE Tran. Signal Proccessing, vol. 54, no. 9, [43] H. Zhang and H. Dai, “Cochannel interference mitigation and coope-
pp. 3658–3663, 2006. rative processing in downlink multicell multiuser MIMO networks,”
[21] T. Yoo and A. Goldsmith, “On the optimality of multiantenna broadcast EURASIP J. Wirel. Commun. Netw., vol. 2004, no. 2, pp. 222–235, Dec.
scheduling using zero-forcing beamforming,” IEEE Journal on Selected 2004.
Areas in Communications, vol. 24, no. 3, March 2006. [44] J. Pastor Perez, F. Riera-Palou, and G. Femenias, “Combining fractional
[22] V. K. Lau and M. Jiang, “Performance analysis of multiuser downlink frequency reuse with coordinated multipoint transmission in MIMO-
space-time scheduling for TDD systems with imperfect CSIT,” IEEE OFDMA networks,” in IFIP Wireless Days 2013, Nov 2013, pp. 1–8.
transactions on vehicular technology, vol. 55, no. 1, pp. 296–305, 2006. [45] S. Haykin, Communication Systems, 4th ed. Wiley, 2001.
[23] R. C. Elliott and W. A. Krzymien, “Downlink scheduling via genetic [46] M. Andrews, K. Kumaran, K. Ramanan, A. Stolyar, P. Whiting, and
algorithms for multiuser single-carrier and multicarrier MIMO systems R. Vijayakumar, “Providing quality of service over a shared wireless
with dirty paper coding,” IEEE Transactions on Vehicular Technology, link,” IEEE Communications Magazine, vol. 39, no. 2, pp. 150–154,
vol. 58, no. 7, pp. 3247–3262, 2009. 2001.
[47] Third Generation Partnership Project, Technical Specification Group
[24] R. C. Elliott, S. Sigdel, and W. A. Krzymień, “Low complexity greedy,
Radio Access Network; Physical layer aspects for evolved Universal
genetic and hybrid user scheduling algorithms for multiuser MIMO
Terrestrial Radio Access (UTRA). 3GPP Std. TR 25.814 v. 7.0.0, 2006.
systems with successive zero-forcing,” Transactions on Emerging Te-
[48] F. Kelly, A. Maulloo, and D. Tan, “Rate control for communication net-
lecommunications Technologies, vol. 23, no. 7, pp. 604–617, 2012.
works: shadow prices, proportional fairness and stability,” The Journal
[25] H. Purmehdi, R. C. Elliott, and W. A. Krzymień, “Reduced-complexity of the Operational Research Society, vol. 49, no. 3, pp. 237–252, 1998.
user scheduling algorithms for coordinated heterogeneous MIMO net-
works,” IEEE Transactions on Vehicular Technology, vol. 65, no. 8, pp.
6184–6203, 2016.
[26] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge
University Press, 2004.
[27] P. W. Chan and R. S. Cheng, “Capacity maximization for zero-forcing
MIMO-OFDMA downlink systems with multiuser diversity,” IEEE
Transactions on Wireless Communications, vol. 6, no. 5, 2007.
[28] F. Riera-Palou and G. Femenias, “On scheduling and resource allocation
in cluster-based cooperative MIMO-OFDMA cellular networks,” in
Proc. IEEE International Conference on Wireless and Mobile Compu-
ting, Networking and Communications (Wimob 2016), New York (USA),
October 2016.
[29] G. Femenias and F. Riera-Palou, “Scheduling and resource allocation in
downlink multiuser MIMO-OFDMA systems,” IEEE Transactions on
Communications, vol. 64, no. 5, pp. 2019–2034, 2016.

0018-9545 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Das könnte Ihnen auch gefallen