Joint Scheduling and Resource Allocation with Fairness ... Scheduling and... · Joint Scheduling...
Transcript of Joint Scheduling and Resource Allocation with Fairness ... Scheduling and... · Joint Scheduling...
Joint Scheduling and Resource Allocation with Fairness Based on the
Signal-to-Leakage-plus-Noise Ratio in the Downlink of CoMP Systems1
Rana A. Abdelaal2, Khaled Elsayed and Mahmoud H. Ismail
[email protected], [email protected], [email protected]
Department of Electronics and Communications Engineering, Cairo University
Giza, Egypt, 12613
Phone: +202 3567 8839, Fax: +202 3572 3486
Abstract
Recent research has shown that Coordinated Multi Point (CoMP) transmission can provide significant
gains in terms of the overall cell capacity and cell-edge user throughput [1]. The main purpose of this
paper is to enhance the overall cell throughput, the cell-edge user’s throughput, and the fairness among
user equipment terminals (UEs) in LTE-Advanced (LTE-A) systems using CoMP. Towards that end, we
propose two novel Resource Allocation (RA) strategies based on the Signal-to-Leakage-plus-Noise-Ratio
(SLNR) for the downlink of CoMP transmission in LTE-A systems. The proposed RA strategies select
the UEs that can efficiently share the same resource block (RB) without degrading the overall throughput
by using the SLNR metric. Moreover, a fairness algorithm is proposed to achieve certain level of fairness
among the UEs and to improve the cell-edge UEs throughput. In addition, we compare the proposed
strategies to the RA based on the more common Signal-to-Interference-plus-Noise-Ratio (SINR) strategy.
The SLNR-based RA is shown to provide significant gains in throughput reaching up to 80% in the
overall system and is shown to have even less complexity than the typical SINR-based RA. Moreover, by
evaluating the proposed strategies in terms of the average cell throughput, cell-edge user throughput, and
fairness among UEs, simulations show that the proposed strategies present superior performance
compared to the more common SINR strategy. With such advantages as enhanced throughput and lower
complexity, the proposed schemes are suitable for application in practical cellular systems.
Keywords: Coordination; coordinated multi point transmission; cellular networks; LTE; interference
mitigation; Resource Allocation; fairness.
1. Introduction
Recent advances in wireless communications such as the use of multiple-input-multiple-output (MIMO)
systems and orthogonal frequency division multiplexing (OFDM) have managed to reduce the detrimental
effect of fading in wireless communication systems. Consequently, the capacity of modern wireless
cellular networks is now mainly restricted by inter-cell interference (ICI). Typically, in cellular systems, a
geographical region is divided into cells, which handle ICI through the use of pre-defined frequency reuse
patterns [2]. Although these frequency reuse schemes result in an inefficient use of the available
1 This work is part of the 4G++ research project supported by the National Telecom Regulatory Authority (NTRA)
of Egypt.
2 Rana A. Abdelaal is the corresponding author.
bandwidth, they guarantee that near cell-edge users belonging to adjacent cells do not share similar
frequencies and thus, have limited ICI. Recently, as bandwidth progressively becomes a more scarce
resource, future cellular networks shift gradually closer to the maximal frequency reuse of unity [3].
Consequently, efficient resource allocation (RA) will play an essential role in future networks in order to
guarantee that the interference level does not excessively increase.
One of the key enabling technologies that is proposed to be used in the next generation cellular system
known as LTE-Advanced (LTE-A) in order to overcome the problem of excessive ICI due to the unity
frequency reuse is coordinated multi-point (CoMP) transmission. CoMP targets increasing the average
cell and cell-edge user throughput, through limiting the ICI. There have been several previous studies in
the literature, which show that CoMP could provide significant gains in terms of achievable throughput
resulting in meeting the requirements of IMT-Advanced for next generation wireless systems [4]. These
works will be briefly covered in the following subsection.
1.1. Prior Work on CoMP
The basic idea of CoMP is to eliminate ICI through cooperation between a number of base stations (BSs)
or enhanced Node-Bs (eNBs), across several cells or sectors, under the command of a central entity that
turns interference into a useful signal. As an alternative CoMP architecture, cooperation can be achieved
inside a specific cell, where the cell can consist of several remote radio equipment (RREs), which can be
connected to a central BS or eNB (like in Fig. 1, for example). The main objectives of CoMP are to
mitigate the interference; provide high spectral efficiency over the entire cell area; and increase the
overall throughput, especially the cell-edge throughput [5]. This is made possible through the exchange of
data, control information as well as channel state information (CSI). This exchange of information occurs
over the interface connecting the central BS and the RREs (or the multiple BSs), which can be
implemented through the use of optical fibers or via dedicated radio thus making high-speed transfer of
signals possible. Full coordinated transmission is achieved among the RREs (or the BSs) through unified
radio resource management (RRM) at the central BS/eNB (or the central entity) [4]. Although CoMP
naturally increases the system complexity, it provides significant capacity and coverage benefits, making
it worth considering for constructing high capacity cellular systems [6].
The coordination in CoMP can be simple as in the techniques that focus on interference avoidance or
more complex as in the case where the same data is transmitted from multiple transmission sites. Based
on that, two approaches for CoMP are often considered. The first approach is coordinated scheduling
(CS) and the second approach is joint processing (JP), respectively. The main difference between both
approaches is the level of coordination between the transmission points. It has been previously shown that
optimizing RA for both CoMP approaches can provide high gains in terms of total throughput and
capacity enhancement.
eNB
RRE RRE
RRERRE
RRE RRE
Figure 1: Example of the proposed cell model
In CS CoMP approach, the transmission to a single scheduled UE is performed by a unique transmission
point (each UE receives the data from its serving cell only). However, the scheduling, including any
transmission weights, is dynamically coordinated between several transmission points in order to control
and/or reduce the unnecessary interference between different transmissions. In principle, the best serving
set of UEs will be selected so that the transmitter beams are constructed to reduce the interference on
other UEs, while increasing the served UE’s signal strength. Hence, the cell-edge user throughput can be
improved due to the increase in the received Signal-to-Interference-plus-Noise-Ratio (SINR). CS
approach has been studied extensively in [7-8], and it has been shown that it can achieve relatively good
performance in terms of throughput and spectral efficiency.
On the other hand, in JP CoMP approach, the data is simultaneously transmitted from multiple
transmission points to each UE. In other words, multiple transmission points are allowed to transmit the
same data to a single UE simultaneously over the same resource block (RB). In the context of LTE-A, an
RB is a collection of subcarriers that are handled together as an allocation unit available for data
transmission. When an RB is allocated to a certain UE, this means a definite frequency channel is
assigned to this UE [8]. Thus, in JP, the multi-point transmitters will behave like a single transmitter with
antennas that are geographically separated. This scheme has the potential for higher performance,
compared to coordination only in the scheduling, but comes at the expense of more stringent requirement
on the backhaul communication between the coordinating transmission points in order to share the data to
be transmitted. A JP scheme has been proposed in [8] and it has been shown that this scheme can bring
significant gains to both the average cell throughput and the cell-edge user throughput. The authors of [9]
focus on the availability of the CSI that allow BSs to coordinate. They show that although CoMP might
require a relatively moderate amount of backhaul communication, it can be quite powerful in terms of
capacity enhancement.
1.2. Review of Resource Allocation Schemes for CoMP
Lots of works in the literature have proposed various RA algorithms in conjunction with either CS-CoMP
or JP-CoMP. For example, in [10], an RA model based on the graph coloring problem has been studied
for the downlink of a JP system with carrier aggregation. It has been shown that their RA model can
efficiently be used in the downlink of JP system. This is because it can mitigate interference, which
maximizes the performance of both the network and individual UEs. Likewise, the authors in [11]
consider an RA algorithm for CS system. They focus on the overall network capacity as a measure of
system performance. It has been shown that optimizing the RA problem promises significant gains in
network capacity and system throughput. In [12], the authors discuss the RA problem in case of
coordination across BSs equipped with multiple antennas. An iterative based RA algorithm is presented.
It has been shown that the proposed algorithm is efficient and can provide appreciable performance
improvement. In [13], an RA technique was proposed for the downlink of a CS system. Although the
simulation results presented show high gains in terms of throughput, the discussed RA technique has two
main drawbacks. First, it calculates the SINR for all UEs over all the available RBs in order to allocate
the RB, which provides the highest SINR value to each UE. This allocation process is not practical as it
requires extensive computations and signaling of channel states and leads to excessive delays. The second
drawback is that the proposed allocation method does not maintain fairness among UEs. For example, if a
certain UE has a good channel (relatively good SINR), there is nothing in the algorithm proposed in [13]
that prevents allocating most of the available RBs to such a UE. In that case, the other UEs may not get
the chance to be served.
The authors in [14] proposed a JP scheduling algorithm where the BSs are grouped into a static cluster. It
has been assumed that each UE is served by all the BSs within the same cluster. The cell model proposed
is a centralized model, where there is a central scheduler available in each cluster that has the knowledge
of the CSI of all UEs being served by the same cluster. In addition, it is responsible for managing all the
resources of the cluster. Once the central scheduler has allocated the resources to the scheduled UEs, the
information of the RA is distributed to the rest of the BSs. The model proposed in [14] has two main
drawbacks. First, the scheduling information is assumed to be sent from the central scheduler to the rest
of the BSs over the X2 interface. Since large amount of data needs to be sent, intolerable delays could
occur and large-bandwidth connections are needed. The second drawback is that only the central
scheduler is responsible for allocating resources, so scheduling is done by only one BS, not jointly. This
scheduling is not necessarily efficient for the rest of BSs belonging to the same cluster.
Also, in [15], an implementation of a JP system has been tested using real-world channels and it has been
shown that JP can indeed improve the total throughput in real-world implementations. In addition, in [16],
a RA algorithm that allows RB sharing among UEs is proposed for the CoMP scheme. The algorithm
selects a group of UEs to share each available RB. The UEs with high SINR values over a certain RB are
grouped and served over this RB. The main disadvantage of this algorithm is that the group of the UEs
sharing the same RB will interfere on each other thus limiting the achieved throughput.
In [17], the authors present RA schemes to enhance the performance of CoMP systems. Their proposed
schemes provide a trade-off between system throughput and energy saving. Also an algorithm to enhance
fairness among UEs by reducing the UE starvation that may result due to allocating resources to the best
UEs is also presented. Simulation results are presented to confirm the performance in terms of system
throughput and fairness to UEs. Also in [18], RA algorithm for JP CoMP is proposed with the aim of
scheduling UEs to maximize the system throughput. The proposed algorithm consists of proportional fair
scheduling and precoding algorithm that allows each UE to be scheduled either in a single-cell or multi-
cell mode. The logarithmic relationship between rate and SINR has been used in the problem formulation.
Thus, the proposed algorithm aims at maximizing the throughput through maximizing the SINR. The
optimal solution can be obtained using exhaustive search over all possible scheduling combinations.
However, as the number of UEs grows, this incurs significant search complexity so they have proposed a
heuristic solution since the optimal solution is difficult to obtain.
1.3. Contributions
Typically in the literature, and as can be seen from the literature survey provided above, CS and JP RA
schemes are based on the use of the SINR as the performance metric that needs to be maximized in order
to increase the overall cell throughput. However, the problem of maximizing the SINR for each UE is
known to be challenging and complicated due to its coupled nature and no closed-form solution is
available yet. Instead, our work at hand is based on the maximization of the Signal-to-Leakage-plus-
Noise-Ratio (SLNR) metric, which was first introduced within the context of MIMO systems in [19].
Leakage refers to the interference caused by the signal intended for a desired UE on the remaining UEs in
the cell. In [19], the SLNR metric has been used for designing precoders with the aim of comparing with
zero forcing precoding in terms of complexity and the conditions on the number of transmit and receive
antennas. However, in this work, we use the SLNR metric with different perspective and different
objectives as will be shown later. It was shown in [19] that the use of SLNR as an objective metric
outperforms the conventional SINR. Furthermore, the leakage-based criterion leads to a decoupled
optimization problem and admits an analytical closed-form solution for the precoders design [20].
One important note is due here, although the relationship between the SLNR and the throughput is not
straightforward, it is very intuitive to conclude that by increasing the SLNR metric, the throughput will be
enhanced. As will be shown in the sequel, maximizing the SLNR is even significantly less complex than
maximizing the SINR metric of all UEs in a specific cell and leads to even better results in terms of the
achieved average cell throughput. Motivated by that, we propose two novel RA strategies for the
downlink of the CoMP system in LTE-A system, which select the UEs that can efficiently share the same
RB without degrading the overall throughput by using the SLNR metric. Thus, we do not need to
calculate all the SINR values for all UEs over all RBs as in [9-13]. Moreover, an algorithm is proposed to
achieve some degree of fairness among the UEs and to improve the cell-edge UEs throughput, a problem
that has not been tackled in [13]. The proposed model assumes coordination between RREs, thus high
speed links, such as optical fibers or dedicated radio links, can be used to connect RREs to the BS.
It is important to note that all of the prior works (with the exception of [11]) assume centralized
algorithms. Thus, even though they can achieve relatively good performance, they are too complicated to
be implemented in practice. In contrast, in this work, we assume a distributed RA algorithm. Thus, the
RA problem is solved jointly by all RREs belonging to the cell, instead of assigning resources only in a
central unit. Also, our proposed RA algorithm is more reliable since it does not depend on a single unit
only. Moreover, all of the RA algorithms in the prior work excluding [16] assume that only one UE is
allowed to use each RB. However, our proposed RA algorithm does allow a group of UEs to utilize the
same RB. Thus, it is expected that the proposed algorithm will provide better data rate. In addition, none
of the prior work discussed in the previous subsection considered fairness among UEs except for [17].
However, in this paper we take fairness among UEs into consideration as will be discussed in details in
Section 4.
It is worth mentioning, that the main concept behind the work done in [18], is the possibility of achieving
higher sum rate in some cases using single transmission point without coordination as compared to JP
CoMP. This basically happens under the assumption that there is no RB sharing. In that case, two
scenarios can happen for each RB: the cell can allocate the RB to one of its UE on a single cell mode, or
the cell can help a UE belonging to a neighbor cell by performing JP CoMP. Based on the channel that
each UE experience, a dynamic switch between non-CoMP (setup A) and CoMP (setup B) may possibly
lead to higher overall rate. However, in our proposed solution, RB sharing is allowed. Thus, a more
sophisticated setup can be achieved: setup C, where each cell can allocate the RB to more than one UE by
using JP CoMP. As will be discussed later, the group of UEs selected to share the RB, are the UEs that
interfere the least on each other. Thus, this new setup will definitely lead to higher throughout than setups
A and B. Essentially, dynamic switching between non-CoMP and CoMP does not apply to our work.
Please refer to Fig. 2 for explanation of setups A, B, and C.
In order to demonstrate the efficiency of the proposed algorithms, we present a comprehensive
comparison between four RA strategies; both proposed CS and JP schemes based on SLNR and both CS
and JP schemes based on SINR that has been addressed in [16]. By evaluating the proposed RA strategies
and fairness algorithm with different realistic propagation scenarios proposed in [21], we show significant
enhancements in terms of the cell throughput, the cell-edge user throughput, and fairness among UEs.
The remainder of this paper is organized as follows. Section 2 details the system model under
investigation. Section 3 provides the proposed resource allocation strategies. In Section 4, we discuss the
fairness algorithm and apply it to the proposed RA strategies. Section 5 discusses the simulation results,
both in terms of throughput enhancement and in terms of fairness. Finally, we draw the main conclusions
with suggestions for future research in Section 6.
2. System Model and Assumptions
In this section, we present certain important aspects of the CoMP LTE-A system design and the model
adopted to evaluate the performance of the proposed strategies is presented. We consider a cellular system
where each cell consists of one eNB, M RREs under its control, and serves K single-antenna UEs. An
example of such cell is shown in Fig. 1. There exists N RBs in the system and each of them may be
assigned to one or more UEs. We consider a 20 MHz bandwidth, consisting of 100 RBs, each RB consists
of 12 subcarriers with 15 kHz spacing, which are typical values for an LTE-A system. Channel coherence
bandwidth is assumed to be larger than the bandwidth of the RB leading to flat fading over each RB.
Uniform power allocation among RBs is considered and the overall transmit power available for each
RRE (P) is uniformly divided among the N RBs. So, the power allocated to each RB is simply Pn = P/N
(assuming all RBs are utilized by the RRE). Each RRE is assumed to have a single antenna and the UEs
are assumed to be uniformly distributed over the cell coverage area.
Figure 2: Setups A, B, and C
The classical solution for cooperating in CoMP is, as mentioned before, based on the SINR. The received
SINR at the nth RB of the kth UE is given by:
| |
∑ | |
(1)
where , is complex channel vector of the links between the kth UE and
all M RREs of the CoMP cell (as shown in Fig. 3), is the Additive White Gaussian Noise (AWGN)
power at the kth UE. is an weighting (precoding) vector that shapes the data transmitted from
the M RREs to the kth UE. The choice of the weighting vector depends on the used scheme (CS or JP) as
will be discussed in more details in Section 3.1.
Figure 3: System Model
In (1), a single RB is considered and its index (n) is dropped for simplicity of notation. As shown in (1),
the unknown variables that need to be optimized via coordination are the weighting vectors for all UEs
( in the numerator and the summation over in the denominator). Consequently, in order to select
that maximizes
for each UE, all the weighting vectors for the rest of the UEs must be taken into
account in the optimization problem. The optimal solution for such a problem is indeed NP-hard [22]. In
contrast, in this paper, the SLNR is considered as the performance metric. The SLNR (βk) at the kth UE
over the nth RB can be expressed as:
β
| |
∑ | |
(2)
As shown in (2), the numerator represents the signal intended for the kth UE and the first term in the
denominator represents the leakage caused by the signal intended for the kth UE on the remaining UEs.
The leakage concept is illustrated in Fig. 4. It can be noted that the summation in the denominator is over
instead of as in (1), which greatly reduces the computational complexity as will be shown later in
the sequel. It is thus considered more suitable for RA in practical cellular networks. The choice of the
weighting vectors { } as well as the scheduling of the UEs as proposed in this paper will be targeting
the maximization of the SLNR:
Maximize β
subject to ‖ ‖ .
(3)
The solution to this problem will be shown in the next section for both CS and JP schemes. Finally, the
achieved throughout for the kth user is given by:
∑
(4)
where is the rate achieved by the kth user over the nth RB and = zero if the kth user is not
allocated the nth RB. The overall cell throughput is thus obtained as:
∑ ∑
(5)
The achievable rate by the kth UE is directly proportional to its SINR by Shannon formula:
(6)
It is important to note that during the RA algorithm, the rate is computed according to (6). However,
while evaluating the system performance, the rate is computed differently since it depends on the
modulation and coding scheme (MCS) used. The MCS is selected according to the channel quality
indicator (CQI) that is available at the eNB through feedback. According to the LTE standard, a CQI table
of 15 entries is used, in which each CQI index refers to a specific channel coding rate and a modulation
order (QPSK, 16 QAM, and 64 QAM).
Figure 4: The leakage concept.
3. The Proposed Resource Allocation Strategies
In this section, the proposed RA strategy is discussed. The general aim of allocating resources in a CoMP
system is to maximize the throughput by reducing the interference effect, so our proposed strategy aims at
maximizing the throughput per RB, which directly leads to maximizing the overall throughput achieved
by the entire bandwidth. For each RB, a set of UEs will be selected to have their data
transmitted over the same RB. Scheduling decisions can be taken to dynamically determine which UEs
can simultaneously use the same RB.
3.1. Selection of the Weighting Vectors
In case of the CS scheme, the weighting vector determines which RRE should serve a specific UE. It is
assumed that all RREs have perfect knowledge about all channel parameters. Since, in CS, each UE is
served by only one RRE, then all the elements in are zeros except only one element will be equal to
unity, which corresponds to the serving RRE. The index of the serving RRE can be easily obtained by
solving the optimization problem in (3) through a simple exhaustive search procedure. On the other hand,
in the case of the JP scheme, the same data packet is sent to a specific UE from all RREs and thus is
not easily obtained as in the case of CS. This optimization problem has been solved in [19] in the context
of MIMO systems and the solution was found to be:
((| |
)
) , (7)
where is the identity matrix, is the eigenvector corresponding to the maximum eigenvalue
of the matrix computed in (7), and is a matrix given by:
[ ] , (8)
It is worth noting that since we assume equal power allocation, there is no need to include the power term
in designing the weighting vectors. Thus, the weighting vectors are normalized. It is also worth
mentioning that the scheduled UEs channel qualities is taken into consideration in the selection of the
weighting/precoding vectors through maximizing the SLNR. As shown in (2), the numerator of the SLNR
contains , which reflects the channel qualities of each UE. Selecting the maximum SLNR for each UE
will guarantee the selection of the best RRE(s) to serve each UE in the CS and JP CoMP schemes,
respectively. Secondly, the intended signal to each scheduled UE is accompanied by a
leakage/interference on other UEs. So, by considering the minimum leakage values from all the scheduled
UEs, we actually consider their channel qualities indirectly.
3.2. The Grouping Algorithm
The task of the grouping algorithm is to select the UEs that can efficiently share the same RB without
degrading the overall throughput. In other words, the UEs that leak the least on each other should be
assigned the same RB. In that way, the overall throughput will be enhanced and the available bandwidth
will be efficiently utilized. In both of the CoMP schemes proposed earlier, the grouping algorithm is the
same. The following steps will be repeated for every RB available. The set of UEs S is initialized to the
empty set and in both schemes will be populated via the following algorithm:
Algorithm 1: Grouping Algorithm based on Leakage
Initialization: (β )
{ } while stopping condition is not satisfied do
( )
End
The first step in the algorithm is the initialization; the UE with maximum SLNR will be chosen and set to
be the first item in the set : The UE with maximum SLNR is the one that has a relatively strong direct
channel and also has relatively low interference on other UEs. Hence, it is the UE that will leak the least
towards other UEs when sharing the RB with them. The initial element in will thus be chosen. Second,
the leakage value vector from all the elements in the set in the direction of the rest of UEs will be
computed. Leakage refers to the interference caused by the signals intended for the UEs belonging to on
the remaining UEs. The interference on UE due to the leakage from the UEs belonging to the set
can be calculated as follows:
∑ | |
(9)
where is calculated for all UEs that do not belong to the set , i. e., The UE with the
minimum amount of leakage will then be added to the set . Thus, the UE that will be affected the least
will share the RB with the UEs belonging to the set. Finally, the algorithm will continue grouping UEs till
a certain condition is satisfied. The stopping criteria are described in details in the next subsection.
3.3. The Stopping Criterion
Selecting the stopping criteria is a very important issue as it directly affects the overall achieved
throughput. We propose and investigate a number of stopping criteria as follows:
1. Marginal Utility: In this method a utility function is defined as the log function of the achieved
throughput [23], i. e.,
( ) (10)
and the following condition is checked so as to allow sharing the RB (to allow grouping):
,
(11)
where is the utility function of the UE that is currently being tested whether or not to be
added to the set, is the sum of the utility functions of the UEs that have been previously in
the group (excluding the newest UE) computed after adding the new UE and taking its
interference effect on the attainable throughputs, and is the sum of the utility functions of
the UEs in the set computed before adding the new UE. If (11) is satisfied then this UE (newest
UE) will be added, otherwise the grouping will be terminated. This approach is efficient and
provides high overall throughput gains but it does not take into consideration the case when the
throughput drops by adding a particular UE while it has not yet reached the maximum throughput
(global maximum to be reached) and so the algorithm might terminate early before reaching the
highest achievable throughput.
2. Marginal Utility with look-ahead: In this method, the same marginal utility condition in (11) is
used. However, if the condition is not fulfilled, we choose to go one level deeper into the test. In
particular, we construct a temporary set that includes the original set plus the ith UE, which is the
UE that may or may not be added to the set. Then the (i+1)th UE will be selected and the
marginal utility condition in (11) will be again checked for the temporary set with the (i+1)th UE.
If it is still not satisfied, the algorithm will terminate with the original set (not the temporary).
However, if the condition is fulfilled then the algorithm will terminate with a new set, which is
the union of the temporary set and the (i+1)th UE. This algorithm is described in more details in
Fig. 5. In general, this algorithm can include more than two stages of look-ahead; however, the
algorithm with two stages has presented high gains in terms of throughput as will be shown in
Section 5.
This stopping condition is efficient as it does not terminate once the throughput decreases as in
[16], neither does it terminate once the utility function-based condition is unfulfilled as the
marginal utility stopping condition does. Although this stopping condition is suboptimal, it is
practical and leads to high throughput gains as will be shown in Section 5.
3.4. Computational Complexity Comparison
As shown earlier, in the proposed RA strategies, the weighting vector is selected in order to maximize
(2). Maximizing the SLNR metric ( for the kth UE indeed requires less number of computations
compared to maximizing the SINR for the same UE. This is because maximizing the SLNR for each UE
is an independent process. In other words, maximizing the SLNR for a certain UE requires checking all
possible links only for this UE, and no need to check other UEs links. This is because SLNR measures the
amount of signal power indented for this UE versus the amount of leakage on other UEs due to that link.
Temp set = UE set + UE i
UE i+1
UE set UE i
o/p set = UE set + UE i
Not Satisfied
Marginal Utility condition
Satisfied
Marginal Utility condition
o/p set = Temp set + UE i+1Satisfied
o/p set = UE set
Not Satisfied
Figure 5: Marginal utility with look-ahead flow chart
In contrast, maximizing the SINR metric is much more complex process; as the SINR for each UE cannot
be optimized independently. This is because the interference at each UE is dependent on the other UEs
links. Thus, to optimize the SINR for a certain UE, an algorithm should try linking this UE with all
possible links. Also, for each possibility it should try linking other UEs with all possible links.
Consequently, the proposed strategies computational complexity is greatly reduced by considering the
SLNR as the main metric. For example, the complexity order of maximizing the SINR metric for each UE
assuming K UEs and M RREs and the CS strategy is as follows:
Number of computations in SINR-based RA
,
, otherwise.
(12)
This is because in order to maximize the SINR for a specific UE, two stages are needed. The first is to try
linking this specific UE to all RREs in the cell. The second is that while this specific UE is linked with
any RRE, all possible links between the rest of UEs and the rest of RREs should be checked. The number
of computations for the first stage is M in both cases mentioned in (12). However, the number of
computations for the second stage depends on both M and K. When , the second stage actually will
need the same number of computations for selecting RREs from the available RREs to
serve the UEs existing in the cell. Moreover, the order of selection should be taken into account.
Consequently, the number of computations for the second stage will be -permutations of
in case . When , the second stage will need the same number of computations for selecting
UEs from the available UEs to be served by the available RREs existing in the cell.
Moreover, the order of selection should be taken into account. Consequently, the number of computations
for the second stage will equal -permutations of ) in case . By multiplying the
number of computations of both stages and using the basic definition for permutations, (12) can be
obtained.
On the contrary, the complexity of maximizing the SLNR metric for each UE considering the same model
as above is simply of order M. In order to overcome the high computational complexity of maximizing
the SINR at each UE, some papers select the weighting vectors that are corresponding to the maximum
channel gain, such as in [16]. However, selecting the weighting vectors in that way does not take into
consideration the interference channels. In contrast, our proposed model maximizes the SLNR metric for
each UE, which checks the interference channels as well as the direct channel.
4. The Resource Allocation Algorithm with Fairness
The objective of the proposed fairness algorithm is to provide a certain level of fairness among UEs.
Since our proposed resource allocation strategies based on SLNR assure that each RB achieves relatively
high throughput, the overall throughput is optimized. In this section, our algorithm will be updated to
assure that the RBs are efficiently allocated to the UEs with fairness taken into account. Simulation
results have shown an improvement in the fairness among users as will be shown in the sequel.
Initially: Construct the fairness vector F that counts the number of RBs that has been assigned to
each UE in the cell, viz., [ ], where is the number of RBs allocated to the kth UE. In
addition, the weighting vectors are selected so as to maximize the SLNR (β as before. The
following algorithm will be executed:
Algorithm 2: Grouping Algorithm based on leakage and fairness
For n = 1 : N
(β )
= + 1
Normalize F and
While stopping condition is not satisfied
( )
F( ) = F( ) + 1
Normalize F and
End While
End For
The first step in the algorithm is adding the UE with the maximum β
to the set as discussed earlier.
Following that, the fairness vector F will be updated, such that
(13)
where is assumed to be the UE with the maximum β . The leakage vector from in the direction of the
rest of UEs will then be computed as discussed earlier in Section 3.1 and both the leakage and fairness
vectors will be normalized such that
and
(14)
where [ ], [ ] is the maximum element in the fairness vector
and is the maximum element in the leakage vector. The UE that is corresponding to the minimum
element in the vector [ ] will then be added to the set . By that step, the UE
that has been allocated relatively less number of RBs compared to other UEs and is affected with
relatively low leakage from the set is the one selected to join the set . This will guarantee some
fairness among UEs as it prevents allocating most of the available RBs to UEs with good channels.
Consequently, cell-edge users will get the chance to be served, unlike the RA algorithm in [13]. The
algorithm will then keep adding UEs till the stopping criterion is satisfied. The stopping criterion could be
any one of those discussed in details in Section 3.2. Since the fairness vector controls allocating resources
to UEs, the algorithm above should be repeated for all the available RBs. It worth noting that, the fairness
vector should not change until allocating all available RBs, because it is the only indicator on the number
of RBs allocated to each UE. By applying the fairness algorithm, both CS and JP schemes have provided
better fairness to the UEs as will be evident from the results shown in Section 5.
5. Simulation Results and Analysis
The RA strategies described in Sections 3 and 4 are now evaluated in this section through simulations. In
order to capture the impact of long term propagation effects on the system performance, several snapshots
are simulated and the results are averaged. To evaluate the performance of our proposed strategies, we
simulate a cellular system with the parameters shown in Table 1. In our simulations, we consider two
typical propagation scenarios; the typical urban macro-cell model and the bad urban macro-cell model,
which is same as the typical urban macro-cell scenario plus long delays. Both models are based on the
WINNER channel model [21]. It is worth noting that the number of UEs sharing the same RB | | is
restricted by the number of RREs in the cell in case of the CS scheme. That is due to the fact that, each
RRE cannot serve more than one UE over the same RB. However, there is no such limitation in case of
JP.
Table 1: Main Simulation Parameters
Parameter Value
Number of RREs per cell 6
Carrier Center Frequency (GHz) 2
Subcarrier spacing (KHz) 15
Number of RBs (N) 100
Number of subcarriers per RB 12
System bandwidth (MHz) 20
Propagation Scenarios Typical urban macro-cell and Bad urban macro-cell [21]
Number of antennas per UE One
Number of antennas per RRE One
Power distribution among RBs Uniform
UEs distribution among cell area Uniform
Scheduling algorithms CS, JP
Used modulation schemes QPSK, 16-QAM, 64-QAM
Number of UEs per set | | Less than or equal to 6 in case of CS and unlimited in case of JP
5.1. RA Strategies Without Applying The Fairness Algorithm
In order to assess the effectiveness of the proposed strategies, we compare them to the algorithms that
have been adopted for CoMP in [16]. This algorithm chooses the weighting vectors corresponding to the
highest channel gains and allocates the available RB to the UE with the highest SINR and continues
allocating the same RB to other UEs as long as the cell throughput increases. Otherwise, it finishes and
goes to the next RB.
Figs. 6 and 7 show the CS schemes with typical urban macro-cell model and bad urban macro-cell model,
respectively. The shown curves in each figure correspond to the proposed CS strategies based on SLNR
with look-ahead and with marginal utility as stopping criteria. The fourth curve corresponds to CS based
on the SINR criterion as outlined above. Both figures present the average throughput per RB, in Kbps, of
the simulated CS strategies as a function of the number of UEs per cell. As shown in both figures, the CS
strategy based on SLNR with marginal utility and look-ahead criterion exhibits an average throughput
gain of 61% as compared to that of the CS based on SINR strategy. Finally, the CS strategy based on
SLNR with marginal utility criterion presents an average throughput gain of 52% relative to the CS based
on SINR strategy. This confirms our intuition that the proposed RA strategies based on the SLNR metric
improve the overall throughput. These results show high throughput improvement, which is due to the use
of the SLNR-based criterion [19].
It is also worth noting that; throughput can be more enhanced as there are still gaps in the curves between
the rate we have reached and the available capacity in case of no interference considered. The rate
achieved in case of neglecting interference is calculated by allocating the same RB to the same UEs
selected using our proposed RA strategies.
Figs. 8 and 9 display the same results above but for the JP schemes. Again, the difference between both
figures is only in the propagation scenario used. From these figures, it is evident that, the JP strategy
based on SLNR with marginal utility and look-ahead criterion presents an average throughput gain of
24% as compared to the JP strategy based on SINR. Finally, the JP strategy based on SLNR with
marginal utility presents an average throughput gain of 7% relative to the JP based on SINR strategy. As
shown, the proposed RA strategies based on SLNR provide higher gains than that of the RA strategy
based on SINR. This is basically due to the accurate choice of the stopping criteria and also for the reason
that the weighting vectors are determined more accurately so as to maximize the SLNR metric for all
UEs.
As can be shown, the CS gains relative to the conventional scheme based on SINR are higher than that of
the JP gains. This is due to the system capacity limitations. It can be shown from both Figs. 6 and 7 that
the CS strategy based on SLNR approaches an average of 74.6% of the throughput achieved assuming no
interference. Also as shown in Figs. 8 and 9, the JP strategy based on SLNR approaches an average of
77.6% of the throughput achieved assuming no interference.
Figure 6: Throughput of the CS strategies in typical urban macro-cell scenario
20 22 24 26 28 30 32 34 36 38 4030
40
50
60
70
80
90
100
110
Number of users per cell
Th
rou
gh
pu
t p
er
RB
[K
bp
s]
CS based on SLNR and look ahead
CS based on SLNR and marginal utility
CS based on SINR
Throughput with no Interference
Figure 7: Throughput of the CS schemes in bad urban macro-cell scenario
Figure 8: Throughput of the JP strategies in typical urban macro-cell scenario
20 22 24 26 28 30 32 34 36 38 40
20
30
40
50
60
70
80
Number of users per cell
Th
rou
gh
pu
t p
er
RB
[K
bp
s]
CS based on SLNR and look ahead
CS based on SLNR and marginal utility
CS based on SINR
Throughput with no Interference
20 22 24 26 28 30 32 34 36 38 40
70
80
90
100
110
120
130
140
150
Number of users per cell
Th
rou
gh
pu
t p
er
RB
[K
bp
s]
JP based on SLNR and look ahead
JP based on SLNR and marginal utility
JP based on SINR
Throughput with no Interference
Figure 9: Throughput of the JP schemes in bad urban macro-cell scenario
It is interesting to observe that in the case of CS, the number of UEs sharing the same RB is restricted by
the number of RREs in the cell, as each RRE can serve only one UE over the same RB. However, in case
of JP, the number of UEs sharing the same RB is not related to the number of RREs, as all RREs are
serving each UE. Consequently, the JP RA scheme leads to higher throughput gains as compared to the
CS, which is evident from the results shown.
5.2. RA Strategies with the Fairness Algorithm
To evaluate the performance of the proposed fairness algorithm discussed in Section 4.1, we consider
three RA strategies. First, we consider our proposed CS based on the SLNR with the look-ahead stopping
criterion. Second, we consider our proposed JP based on SLNR with look-ahead stopping criterion. And
finally, we consider the conventional CS based on SINR for comparison purposes.
Since the main aim of the fairness algorithm is to maintain fairness among UEs specially the cell-edge
UEs, we opt to investigate the UEs throughput cumulative distribution function (CDF) before and after
applying the fairness algorithm. Figs. 10 and 11 display the user throughput CDF for different RA
strategies with/without applying the fairness algorithm. Each CDF plot is the result of 1000 different
channel realizations and for each channel realization, we assume 10 UEs per cell. So the data used for
drawing the CDF curve are 10,000 UEs throughput values. As shown in both figures, the proposed JP
scheme has the best CDF, and the proposed CS scheme has better CDF than the conventional CS based
on SINR.
From the CDF plots, we estimate the nominal cell-edge throughput which we define as the throughput
value which is less than the throughput achieved by 90% of the UEs belonging to the same cell. For
example, consider Fig. 10, which shows that the nominal cell-edge throughput for the proposed CS is 0.5
Mbps and the highest throughput achieved is around 1.2 Mbps. After applying the fairness algorithm, the
nominal cell-edge throughput is still the same 0.5 Mbps as shown in Fig. 11. However, the highest user
throughput is around 0.9 Mbps. So, we conclude from both figures that some of the UEs that are not cell-
20 22 24 26 28 30 32 34 36 38 4060
70
80
90
100
110
120
130
140
150
Number of users per cell
Th
rou
gh
pu
t p
er
RB
[K
bp
s]
JP based on SLNR and look ahead
JP based on SLNR and marginal utility
JP based on SINR
Throughput with no Interference
edge have lost some throughput, which directly means that some RBs have been saved. In order to make
sure that these RBs have been used by the cell-edge UEs, which indicates the achievability of a degree of
fairness, the entire user throughout distribution will be studied.
Considering the JP scheme, as shown in Fig. 10, before applying the fairness algorithm, the nominal cell-
edge throughput can be estimated as 0.75 Mbps and the highest throughput achieved is around 1.8 Mbps.
On the other hand, after applying the proposed fairness algorithm, the nominal cell-edge throughput is not
the same, it is 0.76 Mbps as shown in Fig. 11 whereas the highest user throughput is 1.3 Mbps. So, it is
clear that the cell-edge user throughput is enhanced. However, again the entire user throughout
distribution must be studied.
As shown in Figs. 12a and 12b, which show the probability distribution function of the throughput over
the cell for the CS scheme, most of the cell-edge users after fairness algorithm have a throughput higher
than 0.35 Mbps for the CS scheme. In contrast, most of the cell-edge users without fairness have a
throughput higher than 0.25 Mbps. This observation is valuable as it proves that the RBs saved by some
of the UEs have been allocated to cell-edge UEs, which improves the cell-edge throughput.
Figure 10: User Throughput Cumulative Distribution Curve without Fairness
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.80
0.2
0.4
0.6
0.8
1
Throughput Per User (Mbps)
CD
F
CS based on SLNR
JP based on SLNR
Conventional scheme based on SINR
Figure 11: User Throughput Cumulative Distribution Curve with Fairness
(a)
0 0.2 0.4 0.6 0.8 1 1.2 1.40
0.2
0.4
0.6
0.8
1
Throughput Per User (Mbps)
CD
F
CS based on SLNR
Conventional based on SINR
JP based on SLNR
0.2 0.4 0.6 0.8 1 1.2 1.4 1.60
100
200
300
400
500
600
Throughput Per User (Mbps)
Nu
mb
er
of
Use
rs
(b)
Figure 12: User Throughput Distribution for CS: a) without Fairness; b) with Fairness.
Figs. 13a and 13b show the throughput distribution for the JP scheme. From the figures, it can be seen the
number of UEs that used to have high throughput has been reduced after applying the fairness algorithm.
For example, more than 600 UEs out of the 10,000 UEs have been able to achieve 1 Mbps throughput
without fairness as shown in Fig. 13a. However, only less than 350 UEs have been able to achieve 1
Mbps after applying the proposed fairness as shown in Fig.13b. Obviously, the difference in the
throughput has been given to the cell-edge UEs.
It is also important to study the achievable cell-edge throughput and the average cell throughput for each
channel realization separately. Figs 14 and 15 depict the cell-edge user throughput versus the average cell
throughput before/after applying the fairness algorithm.
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
100
200
300
400
500
600
700
Throughput Per User (Mbps)
Nu
mb
er
of
Use
rs
(a)
(b)
Figure 13: User Throughput Distribution for JP: a) without Fairness; b) with Fairness
As shown in Fig. 14, before applying the fairness algorithm, we can see that most probably the
achievable average cell user throughput for the CS scheme (areas with high intensity of points) ranges
from 0.5 till around 1 Mbps and the cell-edge user throughput ranges from 0.25 till around 0.5 Mbps. In
contrast, according to Fig. 15, after fairness, the most probable achievable average cell throughput ranges
from 0.5 till around 0.9 Mbps and the cell-edge user throughput ranges from 0.35 till around 0.5 Mbps
which emphasizes the fairness enhancement concluded from the CDF curves. In order to further prove
that the proposed fairness algorithm is effective, Jain's fairness index has been computed according to
(15) before and after applying the fairness algorithm.
∑
∑
(15)
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 20
100
200
300
400
500
600
700
Throughput Per User (Mbps)
Nu
mb
er
of
Use
rs
0.7 0.8 0.9 1 1.1 1.2 1.3 1.40
50
100
150
200
250
300
350
400
450
500
Throughput Per User (Mbps)
Nu
mb
er
of
Use
rs
where is the Jain’s fairness index, is the rate achieved by the kth UE, and K is the number of UEs.
Jain’s fairness index for the proposed CS without fairness was found to be 0.6083. On the contrary, it was
found to be 0.8706 after applying the fairness algorithm [24].
As shown in Fig. 14, before applying the fairness algorithm to the proposed JP, we can see that the most
probably achievable average cell throughput (areas with high intensity of points) ranges from 0.8 till
around 1.25 Mbps and the cell-edge user throughput ranges from 0.55 till around 0.75 Mbps. In contrast,
according to Fig. 15, after fairness the most probable achievable average cell user throughput ranges from
0.8 till around 0.95 Mbps and the cell-edge user throughput ranges from 0.68 till around 0.8 Mbps, which
again emphasizes the observations obtained from the CDF curves.
Moreover, the Jain’s fairness index was found to be 0.7146 and 0.8830 before and after applying the
fairness algorithm, respectively. In order to assure that our both proposed schemes provide a higher level
of fairness compared to the conventional system based on the SINR, again Jain’s fairness index has been
calculated for the conventional system as well. It was found to be around 0.865, which is less than that of
both proposed schemes with and without fairness. From Figs. 14 and 15, it is interesting to observe that
each point can show the fairness achieved. For example, in case of JP based on SLNR when the average
user throughput is 1.2 Mbps, the cell edge user can achieve only 0.75 Mbps without applying the fairness.
However, with the fairness the cell edge user can achieve about 0.98 Mbps for the same average user
throughput. That clearly indicates the fairness achieved.
It is worth noting that, power management is also an important factor in enhancing the performance of
LTE-A CoMP systems. Thus, optimized power allocation algorithms can be performed on top of our
proposed RA algorithm with the aim of further enhancing the achievable throughput. It has been shown in
[25] that non-uniform power allocation can further improve the system performance.
6. Conclusions and Work Extensions
In this paper, we proposed two novel RA strategies based on SLNR for CoMP systems. We showed that
the use of the SLNR criterion provides significant gains as compared to the more classical SINR one.
Moreover, we showed that the proposed strategies have a significantly less computational complexity.
Using simulations, different stopping criteria have been investigated and it has been shown that using the
marginal utility condition with look-ahead criterion achieves relatively high gains and is efficient. We
also showed that the proposed fairness algorithm can enhance the cell-edge throughput and assure that the
RBs are efficiently allocated to the UEs with fairness taken into account. Future work should consider the
impact of the non-uniform power allocation, where different UEs are being served using different powers.
Figure 14: Cell-edge User Throughput versus Average User Throughput without Fairness
Figure 15: Cell-edge User Throughput versus Average User Throughput with Fairness
0.2 0.4 0.6 0.8 1 1.2 1.4 1.60.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Average User Throughput (Mbps)
Cell
-ed
ge U
ser
Th
rou
gh
pu
t (M
bp
s)
CS based on SLNR
JP based on SLNR
Conventional based on SINR
0.2 0.4 0.6 0.8 1 1.2 1.4 1.60.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
Average User Throughput (Mbps)
Ce
ll-e
dg
e U
se
r T
hro
ug
hp
ut
(Mb
ps)
CS based on SLNR
JP based on SLNR
Conventional based on SINR
References
[1] D. Gesbert, S. Hanly, H. Huang, S. Shamai, O. Simone, and W. Yu, Multi-cell MIMO cooperative
networks: A new look at interference, IEEE Journal on Sel. Areas in Communications, vol. 28, no.
9 (2010) 1380-1408.
[2] 3GPP R1-050833, Interference mitigation in evolved UTRA/UTRAN,LGE Electronics.
[3]
W. Yu, T. Kwon, and C. Shin, Joint scheduling and dynamic power spectrum optimization for
wireless multicell networks, CISS (2010) 1-6.
[4] S. Parkvall et al., LTE-Advanced - Evolving LTE towards IMT-Advanced, VTC- Fall Vehicular
Technology Conf. IEEE 68th. (2008) 1-5.
[5] Q. Wang, D. Jiang, G. Liu, and Z. Yan, Coordinated Multiple Points Transmission for LTE-
Advanced Systems, 5th Int. Conf. Wireless Communications, Networking and Mobile Computing
WiCom (2009) 1-4.
[6] P. Hosein, Coordinated Radio Resource Management for the LTE Downlink: The Two-Sector
Case, IEEE Int Communications (ICC) Conf. (2010) 1-5.
[7] L. Venturino, N. Prasad, and X. Wang, Coordinated scheduling and power allocation in downlink
multicell OFDMA networks, IEEE Trans.Veh. Technol., vol. 58, no. 6, (2009) 2835–2848.
[8] M. Sawahashi, Y. Kishiyama, A. Morimoto, D. Nishikawa, and M. Tanno, Coordinated multipoint
transmission/reception techniques for LTE-Advanced, IEEE Wireless Communications, vol.17,
no.3 (2010) 26-34.
[9] S. Kiani and D. Gesbert, Optimal and distributed scheduling for multicell capacity maximization,
IEEE Trans. Wireless Commun.,vol. 7, no. 1 (2008) 288–297.
[10] H. Bian, C. Guo, and C. Feng, An ICSGC algorithm for carrier assignment in downlink
coordinated multi-point with carrier aggregation, 2nd IEEE Int Network Infrastructure and Digital
Content Conf. (2010) 934-938.
[11] D. Gesbert, S. Kiani, A. Gjendemsj, and G. Oien, Adaptation, coordination, and distributed
resource allocation in interference-limited wireless networks, Proc. IEEE, vol. 95, no. 5 (2007)
2393–2409.
[12] H. Dahrouj and W. Yu, Coordinated beamforming for the multicell multi-antenna wireless system,
IEEE Trans. Wireless Commun., vol. 9, no. 5 (2010) 1748–1759.
[13] H. Zhang, X. Xu, J. Li, and X. Tao, Subcarrier Resource Optimization for cooperated Multipoint
Transmission, International Journal of Distributed Sensor Networks (2010) 1-7.
[14] S. Brueck, L. Zhao, J. Giese, and M. A. Amin, Centralized scheduling for joint transmission
coordinated multi-point in LTE-Advanced, Int Smart Antennas (WSA) ITG Workshop ( 2010)
177-184.
[15] V. Jungnickel et al., Coordinated Multipoint Trials in the Downlink, IEEE GLOBECOM
Workshops ( 2009) 1-7.
[16] R. L. Batista, R. B. dos Santos, T. F. Maciel, W. C. Freitas, and F. R. Cavalcanti, Performance
evaluation for resource allocation algorithms in CoMP systems, IEEE 72nd Vehicular Technology
Conference (2010)1-5.
[17] R. Devarajan, S. C. Jha, U. Phuyal, and V. K. Bhargava, Energy-Aware Resource Allocation for
Cooperative Cellular Network Using Multi-Objective Optimization Approach, IEEE Transactions
on Wireless Communications, Vol. 11, no. 5 (2012) 1797- 1807.
[18] R. Chen, R. Bendlin, E. Onggosanusi, and A. Ekpenyong, Single and Multi-Cell Scheduling in
Coordinated Multi-Poin Distributed Antenna Systems, Globecom - Wireless Communications
Symposium (2012) 4653- 4657.
[19] M. Sadek, A. Tarighat, and A.H. Sayed, A Leakage-Based Precoding Scheme for Downlink
Multi-User MIMO Channels, IEEE Transactions on Wireless Communications (2007) 1711-
1721.
[20] A. Tarighat, M. Sadek, and A. H. Sayed, A multi user beamforming scheme for downlink MIMO
channels based on maximizing signal-to-leakage ratios, IEEE International Conference on
Acoustics, Speech, and Signal Processing, vol. 3 (2005)1129-1132.
[21] IST-WINNER II, D1.1.2 ,WINNER II Channel Models, (2007), http://www.ist-
winner.org/deliverables.html, last date accessed at (11/11/2011).
[22] Z.-Q. Luo and S. Zhang, Dynamic spectrum management: Complexity and duality, IEEE J. Sel.
Top. Signal Processing, vol. 2, no. 1 (2008) 57–73.
[23] M. Al-Rawi, R. Jantti, J. Torsner, and M. Saqfors, Channel-Aware Inter-Cell Interference
Coordination for the Uplink of 3G LTE Networks, in Wireless Telecommunications Symposium
(2010) 1-5.
[24] R. Jain, D. M. Chiu, and W. Hawe, A quantitative measure of fairness and discrimination for
resource allocation in shared systems, Eastern Research Lab, DEC Research Report TR-301 (1984)
1-37.
[25] R. A. Abdelaal., K. Elsayed, and M. H. Ismail, Cooperative scheduling, precoding, and optimized
power allocation for LTE-advanced CoMP systems, in IEEE Wireless Days (2012) 1-6.