Coordinated Multi-Point Clustering Schemes A Survey

IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO.
2, SECOND QUARTER 2017 743
Coordinated Multi-Point Clustering

Schemes: A Survey
Selcuk Bassoy, Hasan Farooq, Muhammad A. Imran, Senior Member, IEEE, and Ali Imran, Member, IEEE
Abstract—Mobile data traffic grew by 74% in 2015 and it

is expected to grow eight-fold by 2020. Future wireless net-
works will need to deploy massive number of small cells to
cope with this increasing demand. Dense deployment of small
cells will require advanced interference mitigation techniques to
improve spectral efficiency and enhance much needed capacity.
Coordinated multi-point (CoMP) is a key feature for mitigating
inter-cell interference, improve throughput and cell edge perfor-
mance. However, cooperation will need to be limited to few cells
only due to additional overhead required by CoMP due to chan-
nel state information (CSI) exchange, scheduling complexity, and
additional backhaul limitation. Hence, small CoMP clusters will Fig. 1. Proposed Capacity Enhancements for 5G.
need to be formed in the network. This paper surveys the state-
of-the-art on one of the key challenges of CoMP implementation: 74% in 2015 and it’s expected to grow 8-fold by 2020 [1].
CoMP clustering. As a starting point, we present the need for Moreover, a 1000 fold increase in mobile data traffic is
CoMP, the clustering challenge for 5G wireless networks and pro-
vide a brief essential background about CoMP and the enabling expected for 5G beyond 2020 [2]. To enable 5G to cope
network architectures. We then provide the key framework for with this tremendous increase in data growth, following three
CoMP clustering and introduce self organization as an important development areas in the emerging wireless landscape are
concept for effective CoMP clustering to maximize CoMP gains. proposed [2]–[4].
Next, we present two novel taxonomies on existing CoMP clus- 1) Network Densification - Massive Small cell deployment
tering solutions, based on self organization and aimed objective
function. Strengths and weaknesses of the available clustering 2) Increased Spectral Efficiency - CoMP, Multiple Input-
solutions in the literature are critically discussed. We then dis- Multiple Output (MIMO), Enhanced coding techniques
cuss future research areas and potential approaches for CoMP 3) Additional Spectrum
clustering. We present a future outlook on the utilization of big Figure 1 illustrates the potential capacity gains expected
data in cellular context to support proactive CoMP clustering from each of the three key capacity enhancement proposed
based on prediction modeling. Finally, we conclude this paper
with a summary of lessons learned in this field. This paper aims for 5G [2]–[4]. Biggest capacity gains are expected from net-
to be a key guide for anyone who wants to research on CoMP work densification: a massive deployment of small cells will be
clustering for future wireless networks. required [5], [6] in search for additional capacity. Dense small
Index Terms—Coordinated multi-point, CoMP clustering, 5G. cell deployment in heterogeneous cellular networks (HetNet)
will lead to a severely interference limited network depending
I. I NTRODUCTION on the available frequency spectrum. More advanced inter-cell
UTURE wireless cellular networks will be under tremen- interference mitigation techniques will need to be deployed to
F dous pressure with the increasing data demand as the user
behaviour changes with popular high bandwidth applications.
combat interference and improve spectral efficiency. Improved
spectral efficiency will lead to much needed capacity enhance-
While smart phones become very popular, high bandwidth ment as highlighted above as one of the three key development
hungry applications like video streaming, multimedia file shar- areas for 5G.
ing etc becomes more popular. Mobile data traffic grew by CoMP or Network MIMO is the emerging technology which
has been proposed to reduce inter-cell interference and hence
Manuscript received September 10, 2015; revised April 15, 2016, improve high data rate coverage and cell edge throughout
August 19, 2016, and December 16, 2016; accepted January 17, 2017. Date of for future wireless networks. CoMP has been introduced for
publication February 1, 2017; date of current version May 31, 2017. This work long term evolution advanced (LTE-A) by the third gener-
was supported by NSF under Grant 1619346. For details about this project see:
http://bsonlab.com/ASSH/. The authors would like to acknowledge the support ation partnership project (3GPP) in Release 11 [7] and it
of the University of Surrey 5GIC (http://www.surrey.ac.uk/5gic) members for is likely to be a key feature of 5G [2]. However, coordina-
this work. tion between all cells in the network is a very complex task,
S. Bassoy and M. A. Imran are with the School of Engineering, University
of Glasgow, Glasgow G12 8QQ, U.K. (e-mail: [email protected]; due to precise synchronisation requirement within coordinated
[email protected]). cells, additional pilot overhead, additional signal processing,
H. Farooq and A. Imran are with the School of Electrical and Computer complex beamforming design and scheduling among all base
Engineering, University of Oklahoma, Tulsa, OK 74135 USA (e-mail:
[email protected]; [email protected];). stations (BSs). It will require high bandwidth backhaul links
Digital Object Identifier 10.1109/COMST.2017.2662212 due to CSI and/or user data exchange between all BSs [8], [9].
1553-877X 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.
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on March 10,2023 at 09:20:56 UTC from IEEE Xplore. Restrictions apply.
744 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO. 2, SECOND QUARTER 2017
TABLE I
To reduce this overhead, smaller size cooperation clusters are L IST OF ACRONYMS
required where coordination only takes place within the clus-
ter. Optimal CoMP clustering is one of the key challenges for
CoMP implementation for future wireless networks. Selecting
the right group of BSs for cooperation for a given user profile
is key to maximise potential CoMP gains. Trade-off between
the overhead and interference cancellation benefits needs to
be taken into account for optimum cluster size design. There
are multiple objectives for CoMP clustering and the right bal-
ance between the various efficiency/overhead indicators is a
challenge. For example, maximising spectral efficiency with
CoMP clustering can degrade energy efficiency and backhaul
limitations may prevent such cluster design. Hence a compre-
hensive clustering approach should be considered to achieve
the right balance between multiple objectives of future net-
works such as energy efficiency, load balancing and spectral
efficiency. Main scope of this article is to provide an extensive
survey of CoMP clustering techniques in the literature over the
last decade. We provide a novel taxonomy on CoMP clustering
techniques, critically discuss the strengths and weaknesses of
the available solutions in the literature. The rest of the article
is structured as follows.
In Section II, we review the relevant work on CoMP clus-
tering and show our novel contribution with this survey. In
Section III, we provide an essential background about CoMP
to the reader, main types of CoMP implementation, associ-
ated challenges and the enabling network architectures are
presented. In Section IV, we introduce a key framework for
CoMP clustering challenge and present self organising net-
works (SON) as a important platform to implement effective
dynamic CoMP clustering algorithms. In Section V, a novel implementation in [9] and [11] . Deployment scenarios and
self-organisation based taxonomy on CoMP clustering in the brief clustering reviews are presented in these works, how-
literature is introduced. Various CoMP clustering approaches ever there is no study in literature that extensively surveys
are discussed and criticised based on self organisation, com- clustering challenge for CoMP. In [5], CoMP clustering is
plexity, scalability and practical use. In Section VI, a further reviewed briefly and a subset of static overlapping clusters
taxonomy is introduced based on the aimed objective func- are presented, however this work lacks a comprehensive sur-
tion of CoMP clustering. An extensive survey of existing vey on all clustering models in literature, especially missing
clustering approaches based on different objective functions the advanced clustering techniques, i.e., dynamic and/or multi-
like spectral efficiency, energy efficiency, load balancing and objective based clustering. CoMP concept and trial results
backhaul optimisation are presented and criticised in detail. In are presented in [8] with a dynamic clustering algorithm tri-
Section VII, we discuss open research areas for CoMP clus- alled in a test network, however the paper again lacks a
tering and present potential approaches. Big Data empowered review of other available clustering models. Lee et al. [9]
prediction based CoMP clustering is identified as an impor- discuss CoMP implementation challanges and various deploy-
tant open research area for much needed low latency in future ment scenarios for LTE-A, however clustering challenge is not
wireless networks. Big Data aided spatio-temporal channel exploited in the paper. Backhaul capacity and latency require-
prediction, user mobility and user profile predictions and their ment for different CoMP schemes are investigated in [12].
potential use in proactive CoMP cluster decision making is A user-centric CoMP clustering approach is studied to inves-
detailed. Furthermore, we present future research directions tigate available backhaul capacity/latency impact on CoMP
on dynamic clustering and identify the need for comprehen- clustering. Wireless cluster feasibility is presented for differ-
sive multi-objective CoMP clustering in this section. Finally ent cluster size and backhaul capacity. However the paper
in Section VIII, we conclude with summary of lessons learnt lacks on an extensive review of other available CoMP cluster-
in CoMP clustering. The list of acronyms used in this paper ing algorithms which can be employed to dynamically adapt
is listed in Table I. to available backhaul capacity. Beylerian and Ohtsuki [13]
presents a service-aware resource allocation for non-coherent
joint transmission (JT) CoMP in cloud radio access net-
II. R ELATED W ORK works (C-RAN) architecture where a static and a user-centric
A number of works have already been conducted for CoMP clustering approach is presented. Beylerian and Ohtsuki [14]
in general [5], [8], [10] and more specifically for LTE-A propose a further resource allocation solution combining
BASSOY et al.: CoMP CLUSTERING SCHEMES: SURVEY 745
non-orthogonal multiple access (NOMA) scheme with CoMP

to exploit power and space domain multiplexing and further
improve capacity. A static clustering of a fixed cluster size of 2
is employed in this work, however both studies does not intend
to cover all clustering solutions available, especially missing
the dynamic clustering algorithms which can reduce high com-
plexity on user-centric clustering solution in large clusters of
cells. Rao and Fapojuwo [15] presents a survey on energy
efficient resource management for cooperative networks how-
ever energy efficient cooperative clustering challenge is not
reviewed extensively. A comprehensive book is published
about CoMP [10], two example clustering techniques, one
for static, one for dynamic clustering is presented however it
again fails to present an extensive review for CoMP clustering.
Coalitional game theory is introduced in [16] as an important
analytical tool to form CoMP clusters. An example clustering
algorithm is also presented for user equipment (UE) clustering
in the uplink, maximising the sum-rate capacity. Nonetheless,
the book fails to provide a review of all CoMP clustering
approaches available. An extensive survey is provided on con-
trol and data plane separation architecture (CDSA) for future
networks in [17], however this survey lacks a review on CoMP
within the CDSA architecture. Mustafa et al. [18] provides a
survey on device to device (D2D) CoMP within the CDSA
architecture and discuss CoMP clustering briefly with one
dynamic clustering example. Both papers [17] and [18] lack Fig. 2. Main Downlink CoMP Types for LTE-A [7].
a wider review of all CoMP clustering solutions available in
literature. In [19], an extensive review for self organising net-
works (SON) is provided, however CoMP clustering is not
discussed in relation to SON framework. To the best of our one of the key features. A further feasibility study for CoMP in
knowledge, there is no comprehensive survey in the litera- LTE-A is undertaken by 3GPP in Release 11 [7], where phys-
ture about CoMP clustering. This paper aims to fill this gap, ical layer aspects of CoMP is studied. Simulation results from
providing an extensive survey on the existing CoMP cluster- various sources are presented in this study where it is shown
ing approaches in literature. Two novel taxonomies on CoMP that CoMP can offer a significant performance improvement
clustering based on aimed objective and self organisation are especially at the cell edge for different network deployment
presented. Strengths and weaknesses of available solutions are scenarios [7].
critically reviewed and future research directions are identified. Lee et al. [9] show that more CoMP gains are achievable for
cell edge users in scenarios where more interference is experi-
enced. Similarly, more CoMP gains are presented for HetNet
III. C O MP - E SSENTIAL BACKGROUND scenario where pico cells experience severe interference from
In this section, we provide an essential background of CoMP macro sites.
to the reader before moving to the main scope of this article, Various levels of coordination schemes are studied in [10]
i.e., CoMP clustering. but three main downlink coordination categories are identified
Network coordination deals with inter-cell interference, by 3GPP for LTE-Advanced [7] based on the required back-
reducing the interference especially at the cell edge, resulting haul capacity and scheduling complexity. An illustration of
in much needed additional capacity and increased UE through- downlink CoMP types is given in Figure 2.
put. By making use of the shared data between coordinating 1) Joint Transmission (JT): CSI/Scheduling information
transmission points (CSI/scheduling/user data etc), inter-cell and also user data is shared between the coordinated
interference can be mitigated or even exploited as meaningful TPs. This type of coordination offers better results,
signal at the receiver. Transmission points (TP) are different however it requires high backhaul bandwidth with low
antenna ports of MIMO enabled cells which may or may not latency due to user data exchange between multiple TPs.
be located at the same place. Multiple TPs can serve to single user either coherently or
CoMP is one of the key features, standardized for LTE- non-coherently, converting interference signal to useful
Advanced to uplift the network performance. 3GPP initiated a signal. Coherent transmission refers to joint precoding
study item on LTE-Advanced in March 2008 and the require- design and synchronised transmission to achieve coher-
ments for radio interface enhancements are published in [20]. ent combining. Non-coherent JT does not require joint
To satisfy these requirements, 3GPP published the physical precoding, user data is received from multiple TPs where
layer enhancements in [21] where CoMP has been identified as data is individually precoded from each cell.
2) Dynamic Point Selection (DPS): This is a special type of • Cloud Radio Access Networks (C-RAN): Another archi-
JT where user data is transmitted from one TP only and tecture envisioned for network densification is C-RAN
serving TP is changed dynamically in each subframe where baseband processing unit (BBU) is decoupled from
based on resource availability and channel conditions. remote radio unit (RRU). A pool BBU is proposed in the
Fading conditions are exploited to select the best serving cloud where there is high bandwith front-haul between
cell at each subframe. User data is available at multiple the cloud and RRUs [22]–[24]. Baseband resource shar-
TPs similar to JT. ing can be maximised and CoMP can easily be realised in
3) Coordinated Scheduling/Beamforming (CS/CB): CSI is this architecture. Cloud can be enhanced to handle CCU
shared but user data is not shared among the coop- function and make intelligent clustering decisions for the
erated TPs so user data is only available at one TP connected RRUs. A BBU+RRU based CoMP example
but scheduling and beamforming design is coordinated has been studied in [25] for LTE-A giving promising
between the TPs. Beamforming vectors are selected such spectral efficiency gains as expected. The downside of
that interfering TPs is steered towards the null space C-RAN is the requirement for high bandwidth fronthaul.
of the interfered user to minimise interference. CS/CB Larger CoMP cluster size in C-RAN can be feasible with
require lower backhaul bandwidth when compared to JT ideal fronthaul [26] due to centralised BBUs handling
due to reduced data exchange. main CoMP functions. Concept of self organising cloud
There are two main uplink CoMP transmission categories cells is proposed in [27] where SCs within the coverage
identified by 3GPP in [7]. area of a macro BS are connected to the macro BS. Macro
1) Coordinated Scheduling/Beamforming (CS/CB): User BS then handles the decision making on which SCs to
scheduling and precoding design is done by coordina- be allocated for user data service to improve blocking
tion between the TPs however user data is only received probability, energy consumption and handover probabil-
by one TP. ity. This setup can also be easily extended to enable
2) Joint Reception: User data is received by multiple TPs CoMP and enhance macro BS to handle CoMP-CCU
jointly. Similar to downlink JT, uplink joint recep- functionality.
tion offers higher gains but with the cost of increased
complexity and higher backhaul bandwidth requirement. IV. C O MP C LUSTERING AND SON
In this section, we first discuss the key challenges
in CoMP clustering design and identify the need for dynamic
A. Enabling Technologies for CoMP CoMP clustering for maximising CoMP gains by adapting
CoMP clustering to changing network and user profile con-
The requirement for network densification for future cellular
ditions. We then propose SON as the key enabler for dynamic
networks has initiated research on a number of new network
CoMP clustering and give brief introduction on SON.
architectures to optimise increased energy consumption, sig-
nalling and complex mobility management etc. These recently
emerging radio access network (RAN) architectures will also A. CoMP Clustering Challanges and SON
help to overcome the challenges for CoMP (i.e., backhaul lim- As discussed earlier, CoMP can only be realised within
itation, complex precoding, signalling etc), enabling CoMP to small cluster of cells due to its complexity which gen-
be one of the main features of future wireless networks. erally increases with the number of coordinating cells.
• Control/Data Plane Separation Architecture (CDSA): Optimum cooperating cluster selection is key for max-
Motivated by proposed dense HetNet deployment and imising the benefits of CoMP. An illustration of CoMP
energy efficiency concerns, a control and data plane sep- clustering in a typical CDSA architecture is provided
aration architecture (CDSA) is proposed for macro BSs in Figure 3.
to provide coverage layer and handle most of the control A number of challenges need to be critically evaluated for
signalling and small cell (SC) layer under the macro BS a comprehensive CoMP clustering approach to maximise the
to provide the required data services. Reader is referred benefits of CoMP:
to [17] and [18] for two recent extensive surveys for fur- • Is it efficient to deploy CoMP? The first question which
ther reading on CDSA. CDSA is one of key enablers of need to be answered is, if it’s worth deploying CoMP
CoMP implementation where macro BSs can be enhanced for individual cells in a given network setup. Would the
to function as CoMP control unit (CCU) with strong overheads for deploying CoMP be more than the gains
backhaul links to the SCs within its coverage area. CCU it provides? As illustrated in Figure 3, cells closer to
functionality on the macro cell can handle central precod- each other need to form clusters for cooperation as the
ing design, baseband processing and can make intelligent CoMP gains would be maximised when there is severe
clustering decisions centrally within the SC layer, taking inter-cell interference which can be mitigated. However,
various efficiency metrics into account, i.e., energy effi- isolated cells may need to work without coordination,
ciency, load balancing, spectral efficiency etc. With all based on the limited amount of inter-cell interference
SCs connected to the associated macro BS, there is no experienced from other cells. In addition, users close to
need for high bandwith backhaul between the small cells the cell center may not experience high inter-cell inter-
in CDSA. ference, however cell edge users will suffer from high
network objectives will need to be considered for BS

switch-off:
– Can the remaining capacity in the cooperating cluster
cope with the traffic demand for a given quality of
service (QoS)?
– Is SINR provided by the cooperating cluster without
the sleeping cell over the minimum threshold?
– Do the cells within the cooperated set have enough
backhaul bandwidth to cope with increased traffic
when a cell is switched off for energy efficiency?
• Load Balancing / RAN Capacity/ Backhaul bandwidth:
Cooperation introduce additional capacity in the network
by improving spectral efficiency [8]. Intelligent clustering
algorithms can be employed to support load balancing by
shifting traffic from highly loaded cells to its neighbour-
Fig. 3. Dynamic Multi-Objective CoMP Clustering illustration in CDSA ing clusters. Increased cluster size can also uplift capacity
Architecture.
in hotspot areas based on network topology. However,
backhaul bandwidth requirement will also increase with
increased cluster size. Hence multiple objectives need to
interference hence, it can be more efficient to deploy be considered for intelligent CoMP clustering.
CoMP for cell edge users only. Garcia et al. [28] pre- Given the challenges for CoMP clustering design as discussed
sented a dynamic clustering scheme and suggested no above, static clustering based on a fixed topology will fail
spectral efficiency gain in employing CoMP in high to give expected gains for future networks as the network
signal-to-interference-plus-noise ratio (SINR) region due topology will be dynamically changing with on/off sleep-
to additional pilot signalling required for CoMP, reducing ing cells, user deployed cells with unknown location etc.
spectral efficiency more than the expected gains. Users Moreover, spatio-temporal distribution of users and service
are allocated CoMP clusters or CoMP is not used based demands dynamically changes. To maximise CoMP gains,
on their SINR from the local serving BS. It’s shown that clustering algorithms need to be able to accurately respond
CoMP gains are maximised when received power levels to these dynamically changing network conditions and user
from coordinating cells are close to the received power profiles. Self organised CoMP clustering algorithms can be
levels of the local serving cell. Hence it can be con- developed to make optimum clustering decisions by reading
cluded that CoMP gains vary with network density and various network data and making clustering decisions based on
CoMP may not need to be deployed for some cells based the changing conditions, maximising the objectives like spec-
on their location, user profile and the amount inter-cell tral efficiency, energy efficiency, load balancing while keeping
interference. the fairness between the users.
• How many cells in the cluster? Cluster size is another Dynamic clustering can be implemented in the SON plat-
key parameter for optimal CoMP clustering. Too small form which employs autonomous closed-loop changes in the
clusters will fail to provide full achievable gains from network dynamically. Big Data available from various sources
CoMP, on the other hand, big cluster size will lead to within the cellular network can be exploited as an input for
increased overhead on CSI feedback and backhaul capac- SON platform for proactive CoMP clustering algorithms and
ity [29]. Increased cluster size will give better weighted other SON functionalities. Accurate prediction of user profiles
sum rate [30] but with the cost of additional signal pro- and mobility based on Big Data can be employed within the
cessing and increased feedback and signalling. Moreover, SON platform for much needed lower latency on CoMP clus-
increased cluster size can lead to energy inefficiency tering design. Use of Big Data for proactive CoMP clustering
in terms of achieved bits/joule [31]. As illustrated in is further discussed in Section VII-A. A brief background for
Figure 3 for an example CDSA architecture, some clus- SON is given in the next subsection.
ters will have 6 cells, others will have 5 or 4 and some
others will reduce cluster size by switching off some cells
within the cluster for energy efficiency. Hence, there is no B. Self Organising Networks (SON)
ideal fixed cluster size, instead, cluster size needs to be SON is an emerging concept in wireless cellular networks
a dynamic parameter in the clustering algorithm which to automate some of the operational tasks in closed loop to
needs to change based on channel conditions and user overcome the challenges of a complex multi-layer network.
profile. Network conditions are monitored dynamically by exploit-
• Which cells to switch off for energy efficiency? As illus- ing Big Data from various sources and intelligent algorithms
trated in Figure 3, some cells can be switched off by are employed to effectively manage the network based on
forming intelligent CoMP clusters to enhance SINR and the changing local conditions. Dynamic CoMP clustering can
make sure minimum SINR is provided while some cells also deployed within the SON platform as an enhancement to
are switched off for energy efficiency. A number of other SON modules which utilises the Big Data for making
proactive CoMP clustering decisions. SON algorithms can be

designed as a distributed or centralised function depending
on the requirements of the tasks, especially time and scal-
ability limitations. Given the increasing complexity of the
wireless cellular networks, SON will have a strong, enhanced
presence in future networks. Future networks will need to
deploy effective SON algorithms to improve capacity and
QoS and reduce capital expenditure (CAPEX) and operational
expenditure (OPEX) by reducing labour costs.
SON has been an important part of 3GPP LTE/LTE-
Advanced standardization which has started with Release 8
and enhanced further with most recent Release 12 [32]. An
extensive survey on SON has been presented in [19].
SON is mainly categorised in three folds: Fig. 4. CoMP Clustering Taxonomy based on Self Organisation.
1) Self Configuration: This group of SON modules aim
to manage new entities integrated in the network. A
considerable amount of OPEX cost is spent for new A summary of clustering taxonomy based on self organisation
site configuration during network rollout and it will is given in Figure 4.
increase with proposed massive deployment of small Static clustering method is less complex with less signalling
cells. Self configuration algorithms aim to automate new overhead but this method is not responsive to changes in
site configuration, initial automated neighbour relations the network nodes or user locations, hence the performance
and software updates [33]. gains are limited. Semi-dynamic clustering is an enhanced
2) Self Optimisation: This group of SON modules aim to version of static clustering where a number of static clusters
optimise ongoing services in the network. Self opti- are formed and employed dynamically to improve the poten-
misation algorithms will monitor network performance tial gains. Complexity increases with additional signalling but
data and derive optimisation changes in the network performance is also improved when compared to static clus-
in open and/or closed loop, aiming to reduce OPEX tering. However, this method still lacks on truly responding
costs and also improve network spectral efficiency, to the dynamic changes in the network. Dynamic clustering
energy efficiency, network capacity and overall QoS. methods are developed to respond to network and user mobil-
Dynamic CoMP clustering can be incorporated to Self ity changes, i.e., new sites, sleeping cells, load changes etc.
Optimisation module set and implement closed-loop This scheme comes with increased complexity on scheduling
dynamic clustering decisions based on network data and beamforming design but it gives the best results, reducing
already available in the SON platform. Self optimisa- inter-cluster interference by moving the clusters dynamically.
tion is an important part of LTE/LTE-Advanced stan- Dynamic clustering can be classified in three main categories
dardisation [34] and there are already commercialised within itself based on the approach. In network-centric clus-
algorithms deployed in the current LTE networks. Self tering approach, all users in the same cluster use the same
optimisation tasks can be mainly grouped in three set of cells, however in user-centric clustering, users can be
folds [19]. assigned their own clusters which comes with additional com-
a) Load balancing plexity. Hybrid approach combines both approaches which can
b) Coverage and Capacity Improvement be a good balance of complexity vs. performance.
c) Interference Control In the subsequent subsections, we present an extensive
3) Self Healing: This group of algorithms aim to detect literature review for each category and criticise available tech-
faults in network elements, analyse the fault by gathering niques based on complexity, scalability and potential spectral
relevant information, diagnose and clear the fault. For efficiency gains.
time consuming fault restoration, self healing also aims
to perform compensation actions on neighbour cells until A. Static Clustering
the faulty cell is restored. 3GPP has standardised self CoMP coordination clusters are formed in a static way,
healing for LTE/LTE-Advanced as an important feature mostly based on topology and don’t change according to
of SON platform [35]. changes in the network. This method offers a less complex
solution which can be a good candidate to deploy in the initial
V. C LUSTERING TAXONOMY BASED phase of LTE-A deployment. Static clustering within cells in
ON S ELF O RGANISATION the co-located site is the most basic and practical option which
In this section, CoMP clustering algorithms in literature are does not require data exchange between the sites, hence not
critically discussed based on self organisation. Three main reliant on fast backhaul.
clustering types are identified: The work presented in [36] propose a static clustering
1) Static Clustering scheme, where sectors looking into each other are clustered to
2) Semi-Dynamic Clustering improve SINR. Authors assume a hexagonal grid in deploy-
3) Dynamic Clustering ment which is non-realistic in real network deployments. This
is usually the downside for most static clustering solutions. A two layer static clustering, based on regular network
In [37], static intra-site and inter-site CoMP clustering is topology is proposed in [40] to extend on static clustering.
considered with orthogonal frequency reuse where antenna This approach is then extended for several layers for dynamic
bore-sights are shifted to face into each other for extra CoMP clustering. It’s proposed for users to pick one of the available
gain. Dead-spots would be created with this new topology clusters based on power. While the solution is an improved
where small cells are proposed to fill in the dead-spots. CoMP algorithm compared to static clustering, overlapping nature of
and HetNet deployment are merged in this solution to iden- the proposed algorithm adds to the scheduling complexity and
tify locations for small cell deployment, however an idealistic require increased backhaul bandwidth. A semi-dynamic clus-
hexagonal grid is assumed again, which is unrealistic. A tering scheme is introduced in [41] where static clusters are
disjoint and overlapped static clustering model is presented formed based on hexagonal grid topology and multiple shifted
in [29] where static clusters are formed to maximise mean cluster patterns are created with different sub-channels allo-
SINR or to minimise SINR outage at possible user loca- cated for each shifted cluster. A joint, centralised scheduling
tions. In the overlapped solution, one cell can be in three is developed for this clustering type. In [42], static cluster shift
clusters where system resources are splitted into each of the idea from [41] is further enhanced with “full shift” and differ-
three clusters. Presented solution is better than the clustering ent frequency bands are allocated on shifted clusters. Static
types based on regular patterns as it can apply to realistic net- clusters are formed to maximise neighbouring cells in the
work topology. However the proposed work is not scalable as same cluster for a given hexagonal network layout. Shifted
the complexity of the solution increases with the number of clusters reduce the inter-cluster interference, maximising the
possible user locations. CoMP gain, however solution is based on hexagonal grid
A number of drawbacks for CoMP clustering have been topology which is not applicable to real networks. In [43],
investigated in [38]. Authors have investigated an inter-cell a semi-dynamic clustering scheme is proposed for downlink
interference model in HetNet scenario with pico-cells to Time Division Duplex (TDD) JT-CoMP scenario. Solution is
offload macro network. Time-domain resource partitioning is based on large size (nine cells) static clustering and creat-
considered between the macro BS and pico layer within the ing different static patterns of sub-clusters in each large static
macro BS’s coverage area. A static CB-CoMP method is cluster. Dynamically selecting sub-clusters achieves almost as
applied with centralised beamforming and scheduling for the good as large cluster spectral efficiency but with reduced com-
cluster of all pico-cells and its connected macro cell. CoMP plexity. Proposed method is not able to respond to dynamic
failed to improve the performance further from enhanced inter- changes within the static cluster, i.e., new/sleeping cells etc.
cell interference coordination (eICIC) due to the additional and also static nature of the big clusters will create inter-cluster
overhead required to implement CoMP, i.e., mainly the UE- interference.
RS signal introduced with CoMP in LTE-Advanced. In [39], In summary, semi-dynamic clusters are an improved ver-
time synchronisation limitation between coordinated cells is sion of static clusters with minimal overhead increase, however
investigated. Authors have shown that time synchronisation most solutions are based on idealistic hexagonal grid topology
will need to be taken into account for a network with large which is not realistic. Furthermore, majority of semi-dynamic
inter-site distance (7km studied), however there is minimal algorithms propose orthogonal frequency allocation from each
inter-symbol interference (ISI) issues for inter-site distance of cell to its assigned static clusters. Based on the utilisation of
< 1 km due to cyclic prefix (CP) length. dedicated bandwidth for each static cluster, proposed algo-
In summary, static clustering is an attractive approach with rithms can reduce the overall spectral efficiency. Moreover,
its significantly less complexity for initial CoMP deployment static nature of clusters is not able to respond fully to the
for LTE-A networks. Intra-BS CoMP is a promising solu- spatio-temporal changes in user profiles and the network ele-
tion which eliminates the need for high backhaul bandwidth ments. Dynamic clustering algorithms is discussed in the next
requirement between the BSs. On the other hand, inter-BS section which is mostly applicable to real network topol-
static clustering algorithms are mostly based on the assump- ogy and can dynamically adopt to changing user profile and
tion of hexagonal grid layout, which is not applicable to real network conditions.
networks. Furthermore, this method will fail to give the much
needed spectral efficiency gains and increased system capacity
for future 5G networks. Semi-dynamic and/or fully dynamic
solutions are required to respond to changing network/user C. Dynamic Clustering
profile conditions and maximise CoMP gains. Dynamic CoMP clustering is more complex with increased
signalling overhead but its more responsive to the changes
in the network. Inter-cluster interference can be minimised
B. Semi-Dynamic Clustering and cluster size for individual users can be optimised dynam-
Semi-dynamic clusters are more advanced than static clus- ically for an optimum balance. Dynamic CoMP clustering
ters where several layers of static clusters are designed to can be classified in three groups based on network elements
avoid inter-cluster interference. More than one static clustering considered for clustering:
patterns are formed where users are able to select the most suit- 1) Network-Centric Clustering
able cluster. This method also mostly relies on hexagonal grid 2) User-Centric Clustering
network topology which is unrealistic in practical networks. 3) Hybrid Clustering
consists of cell-3 only and there is interference from cell-4

and cell-11. Hybrid clustering employs larger network-centric
clusters, which improves user-1’s cluster to cell-3 and cell-4.
User-1’s SINR is improved in this clustering type but there is
still interference from cell-11. The most beneficial clustering
scheme is the user-centric one where user-1’s cluster consists
of all three surrounding cells, i.e., cell-3, cell-4 and cell-11.
Although user-centric clustering seem to be most beneficial
one, it comes with additional scheduling/precoding complex-
ity and increased backhaul requirement. The three types of
dynamic clustering are reviewed in detail in the subsequent
subsections.
1) Network-Centric Clustering: In network-centric cluster-
ing approach, cells are clustered in groups where all users
within the serving area of the clustered cells are served by all
cells or a sub-group of cells in the cluster. A simple illustra-
tion of network-centric clustering is given in Figure 5b. It is
less complex when compared to user-centric clustering, espe-
cially from scheduling point of view. However cluster edge
users suffer from inter-cluster interference. Dynamic network-
centric clustering can minimise this effect by moving cluster
boundary dynamically.
Two main methodologies are identified in the literature on
dynamic network-centric clustering.
a) Greedy algorithms: Greedy algorithms are widely
used for cooperation cluster formation in literature. Clusters
are formed iteratively, starting from a randomly chosen BS to
maximise the main objective, typically spectral efficiency. Best
cluster is formed for the randomly chosen BS, maximising
the CoMP gains, however the clusters formed in later stages
of the algorithm suffer from sub-optimal clusters. It is rela-
tively less complex but may not achieve as good results as the
other methods, i.e., game theoretic clusters. A greedy uplink
clustering algorithm is studied in [44] aiming to maximise
spectral efficiency. It’s shown that dynamic clustering with
cluster size of two cells outperforms static clustering with
much larger cluster size. A predefined fixed cluster size is
proposed which is not the optimal solution for some clus-
ters. A similar approach is employed in [45] but a dynamic
cluster size is proposed. Authors have designed a dynamic
clustering solution for uplink multi-user distributed antenna
system (MU-DAS), where one cell has a number of RRUs
placed in the cell’s coverage area with fast fiber connection
to their cell. BSs are merged based on highest interference
created to the other users. However, clustering takes only
scheduled users into account at any point in time, hence not
taking load into account for cluster formation. Starting the iter-
ations from the highly loaded cells can improve the system
throughput as the CoMP gains will be maximised for clus-
Fig. 5. Dynamic CoMP Clustering Taxonomy.
ters formed in early stages of the algorithm. Also clustering is
proposed to change with each scheduling interval in this solu-
An illustration of the three types of dynamic clustering is tion [45] which increases signalling due to high frequency
given in Figure 5. CoMP benefits are illustrated for two sam- cluster changes. Both proposed algorithms in [44] and [45]
ple users for an identical network with different clustering offer disjoint clusters where inter-cluster interference is still an
schemes. For example, user-1 in the figure is located at the important factor reducing spectral efficiency. An overlapping
edge of cell-3, receiving strong interference from cell-4 and dynamic clustering is proposed in [46] to improve network
cell-11. Network-centric clustering is the most limited scenario average sum rate and fairness. A greedy approach is consid-
where user-1 is located at the edge of the cluster. Its cluster ered starting from a random BS. Authors have shown better
results with cluster size of four with overlap size of two when network size. Computational complexity of such algorithms
compared to cluster size of eight with no overlap. The solu- can be reduced by limiting the candidate sites for coalition to
tion lacks scalability where large network size can lead to neighbour cells only. Utility function for forming coalitional
increased complexity. Overlapping clusters will also require game theoretic clusters play an important role for optimal clus-
more complex scheduling but overlap and cluster size param- ters. Utility function need to include a realistic model for
eters are introduced in the proposed algorithm to control this the cost of cluster formation and the relevant CoMP gains.
complexity. Dynamic cluster size can be self-imposed with accurate imple-
Greedy algorithms provide lower computational complexity mentation of a utility function. Also multi-objective clustering
however lack on sub-optimal clusters especially for clusters can be implemented by including multiple metrics into the
formed at later stages of the algorithm. Shortcomings of utility function, i.e., energy efficiency, load balancing, spectral
greedy algorithm can be improved by employing coalitional efficiency and backhaul bandwidth limitations.
game theory for cluster formation based on merge-split rule c) Other dynamic network-centric clustering algorithms:
for maximising system throughput. Game theory can also pro- A self organising dynamic clustering method is presented
vide distributed solutions with reduced signalling overhead in [52] where candidate clusters are formed from reported
as opposed to centralised greedy algorithms, however coali- list of cells from users. CCU is proposed to arrange clus-
tional game theoretic algorithm’s computational complexity is ter solution by listing the candidate clusters with minimal
higher than greedy algorithms [47]. Coalitional game theoretic cost, where the cost function takes into account the cluster
clustering is discussed in detail in the next paragraph. size, number of users and reference signal received power
b) Game theoretic clustering: There is an increasing (RSRP). This algorithm is a basic one where cost function
interest in applying coalitional game theory to design self- can be improved to maximise SINR / spectral efficiency for
organised, distributed cooperative clusters. A utility function more optimal solutions. It lacks on scalability with increasing
is introduced to formulate the cost and CoMP gain trade- complexity of handling high number of candidate clusters as
off for forming clusters. Proposed utility function can limit the network size / number of users increase. Time averaged
the cluster size dynamically based on BS locations and user measurements from users is considered where fast fading is
profiles. Coalitional game theory can provide distributed, eliminated. Weber’s algorithm [52] is further enhanced in [53]
stable, converging solutions to maximise CoMP gains. An by replacing the cost function based on received power levels
extensive tutorial on coalitional game theory for wireless to a utility function with the aim of maximising the weighted
communications applications is presented in [48]. sum rate. Unlike [52], Baracca et al. [53] proposed a fast
Li et al. [49] proposed a dynamic network-centric clus- changing cluster design, responding to fast fading channel
tering method employing a utility function to maximise the variations which will lead to increased signalling and possible
second best servers of the cell edge users in the same cluster. ping-pong cluster re-selections. To reduce signalling overhead,
Cluster size is fixed to two only which leads to sub-optimal cluster change frequency can be reduced to a wider time-frame
clustering for varying network conditions. Also network clus- and averaging algorithm can be used for user measurements
tering formation is based on exhaustive search for collusion, which can eliminate fast fading variations.
hence not scalable, i.e., complexity increases with network 2) User-Centric Clustering: Users are allocated their own
size. Moon and Cho [50] have studied a dynamic cluster cluster of cells individually in user-centric clustering approach.
formation algorithm which merges cells into clusters based Although this method can give better SINR/throughput gains,
on the improvement on spectral efficiency, with configurable it’s more complex, especially in terms of scheduling where
maximum cluster size and the minimum efficiency gain. This user clusters overlap with each other. To reduce complexity,
algorithm is semi-distributed where SINR measurements are user-centric clustering can be implemented in small groups of
based on pilot signal measurements but still need a CCU for cells rather than the whole network.
cluster decision-making. It implicitly takes the number of users Garcia et al. [54] have studied macro diversity CoMP with
into account and hence clusters are formed based on cell load. dynamic user-centric clustering, comparing random network
Although a more flexible cluster size is introduced in [50] and hexagonal network topologies. It’s shown that CoMP gives
when compared to [49], algorithm still lacks on scalability as higher capacity results and bigger cluster size are required in
the complexity increases with the number of BSs involved. random networks due to the random nature of BSs with more
Saad et al. [51] presented an application of a coalitional for- potential for inter-cell interference. Authors had no limitation
mation game for user clustering in the uplink, maximising on user-centric clustering which leads to complex scheduling
the sum-rate capacity with a cost function based on power between the BSs. To reduce complexity, user-centric clustering
requirements which is dependant on the distance between the can be limited to groups of cells for easier scheduling, less
users. Inspired by [51], Guidolin et al. [47] developed a coali- signalling overhead and data exchange.
tional game theoretic clustering method where utility function A three-tier clustering approach is presented in [55],
dictates average cluster size and targets for higher spectral wherein it has been proposed that cell center users will not
efficiency. It’s a distributed algorithm which does not need a use CoMP, users within the same site will use static clus-
central entity and reduces signalling overhead. SINR at the tering between intra-BS cells and a user-centric clustering
cell edge is also significantly improved when compared to a is proposed for intra-site cluster edge users. Fixed cluster
greedy algorithm. On the other hand, solution lacks on scal- size is assumed which can lead to unnecessary complexity
ability where the cluster formation complexity increase with or less efficient coordination, depending on user location and
SINR conditions. Similar complexity arises in works presented In summary, dynamic CoMP clustering is a promising
in [54] and [55] where no limitation is proposed on user- concept which can improve performance over static/semi-
centric approach to any group of cells which will lead to higher static alternatives. However, increased complexity and per-
complexity with a large number of BSs cooperating at the formance trade-off need to be evaluated for optimal solu-
same time. In [56], a user-centric inter-cell interference nulling tions. User-centric clustering provides a theoretical upper
is studied for downlink coordinated beamforming. Interference bound for maximum performance gain but it requires com-
nulling range is derived from received power levels to form plex precoding design, scheduling and increased backhaul
clusters for individual users. A threshold for relative power bandwidth [54], [55]. To reduce complexity, user-centric clus-
levels is used to identify BS clusters. tering solutions need to be limited to smaller network-centric
User-centric clustering approach is an ideal scenario to pro- clusters. Main approaches in network-centric clustering in
vide an upper limit however it is not realistic due to increased literature are greedy algorithms studied in [44]–[46] and
complexity. Hybrid clustering is discussed next, which limits more recently coalitional game theoretic approaches deployed
the user-centric approach to a group of BSs only to reduce in [47], [49], and [50]. The key balance between additional
complexity. complexity and the potential CoMP gains can be achieved
3) Hybrid Clustering: Hybrid clustering approach is the by hybrid solutions where user-centric clustering is deployed
combination of network and user-centric approaches where within network-centric clusters [57]–[59]. However, hybrid
users are allocated their own preferred cells but limited to solutions in current literature focuses either on dynamic user-
a bigger group of cells which can be dynamically changing centric approach with static network-centric clustering or
to adapt to changing network conditions. Hybrid cluster- dynamic network-centric clustering with no focus on user-
ing is driven from the complexity/throughput gain trade-off centric clustering. Further research is required to employ
where user-centric clustering is used for better throughout but dynamic clustering algorithms for both network-centric and
its complexity is kept at manageable levels by introducing user-centric clusters for more optimal solutions. A summary
network-centric clustering where users are limited to select of CoMP clustering approaches based on self-organisation and
cells only within the network-centric cluster. their shortcomings are provided in Table II.
Zhao et al. [57] developed a hybrid clustering method where
a pre-defined network-centric clustering is used for cell cen- VI. DYNAMIC C LUSTERING TAXONOMY
ter users and a number of pre-defined overlapping clusters are BASED ON O BJECTIVE F UNCTION
used for cell edge users to pick the best overlapping clus-
In this section, a novel CoMP clustering taxonomy is
ter to maximise SINR for the cell edge user. Inter-cluster
presented based on the main objective function. The main
interference on overlapping clusters is eliminated by orthog-
objective of CoMP is to mitigate interference from neigh-
onal frequency allocation. Presented solution lacks on self
bour cells and improve spectral efficiency in general but a
organisation as the pre-defined clusters are static, i.e., can’t
more comprehensive approach is required to include other
respond changes in the network (new sites, sleeping cells
metrics/limitations for CoMP clustering. Backhaul bandwidth
etc). Although overlapping cluster patterns improve cluster
limitations for CoMP implementation and energy efficiency
edge user performance, orthogonal frequency use prevents
concerns for future wireless networks need to be included in
the optimal use of the bandwidth. A simple downlink user-
comprehensive CoMP clustering algorithms. Moreover, with
centric clustering is studied in [58] where users coordinate
exponentially growing mobile data demand, better utilisation
with two best serving cells according to the received power
of system capacity with load balancing will be a key con-
levels under a bigger static cluster. Proposed static network-
cept which need to be taken into account for CoMP cluster
centric clusters will suffer from high inter-cluster interference
design. Based on our detailed literature survey, main objective
and also fixed user-centric cluster size can lead to unnecessary
functions studied are:
coordination, waste of resources and also possibly not being
1) Spectral Efficiency
able to cancel severe interference from third best server for
2) Backhaul Optimisation
some users. A self-organised, dynamic network-centric clus-
3) Energy Efficiency
tering can improve inter-cluster interference and also dynamic
4) Load Balancing
user-centric cluster size can be employed for better per-
A summary of CoMP clustering taxonomy based on objec-
formance. In [59], a hybrid clustering model for downlink
tive function is given in Figure 6. In the following subsections,
SU-COMP is studied. Authors proposed static network-centric
each objective function is critically discussed and extensive
clusters and cell edge users are proposed to have user-centric
literature review is presented.
clusters of fixed size of three within each network-centric
cluster. Authors also presented a good review of SU-COMP
scheduling and a SU-COMP joint scheduling algorithm is A. Spectral Efficiency
provided for the proposed clustering scheme. The presented Main objective of CoMP is to mitigate inter-cell interfer-
clustering scheme has low complexity, but further work is ence within the cooperating cluster. Interference cancellation
required to introduce dynamic network-centric clustering for leads to better SINR and improved spectral efficiency. Cluster
improved cluster design. Fixed cluster size is also another formation algorithms are designed to maximise spectral effi-
shortcoming of the algorithm which can generate sub-optimal ciency as a common objective, however other utilities such as
clusters. backhaul bandwidth optimisation, energy efficiency and load
TABLE II
S UMMARY OF C O MP C LUSTERING A PPROACHES BASED ON S ELF -O RGANISATION
spectral efficiency gain. Dynamic network-centric clustering

methods can further increase spectral efficiency by dynami-
cally changing CoMP clusters based on the spatio-temporal
changes in user profiles and network elements. A game-
theoretic, network-centric clustering approach is employed
in [47]. Papadogiannis et al. [44] used a greedy clustering
algorithm for uplink network-centric clustering to maximise
spectral efficiency. User-centric dynamic clustering approaches
are studied in [54] and [55] which provide an upper bound on
spectral efficiency gain but with increased complexity. Hybrid
solutions reduce this complexity where user-centric cluster-
Fig. 6. CoMP Clustering Taxonomy based on Objective Function. ing is limited only within a network-centric cluster [57]–[59].
Dynamic clustering solutions require more complex precoding
design and scheduling, and increased backhaul. Complexity
and additional requirements are reduced in semi-static clus-
balancing have also been studied in the literature. Trade-off
tering, and further simplified in static clustering with the
between spectral efficiency and other objectives for opti-
cost of reduced spectral efficiency gain. An extensive critical
mum clustering has been also in the interest of research
review of CoMP clustering solutions based on static/semi-
community [60].
static/dynamic approaches and the trade-off between complex-
3GPP identified three CoMP deployment scenarios for
ity and the additional spectral efficiency gains are provided
LTE-Advanced and released a feasibility study, presenting
in Section V. A summary of different approaches and their
simulation results from over 20 sources showing significant
shortcomings are presented in Table II.
spectral efficiency improvements by deploying CoMP espe-
cially at the cell edge [7]. The most basic, intra-site static
clustering is studied as scenario-1 and over 20% increase in
spectral efficiency is observed at the cell edge with MU-MIMO B. Backhaul Optimisation & Caching at RAN for JT-CoMP
JT-CoMP case [7]. Inter-site static clustering solutions are As discussed in previous sections, one of the key require-
employed in [36] and [37] which is not able to respond to ment of CoMP is high backhaul bandwidth and low latency.
the dynamic changes in the network and user/service profiles, Depending on the type of CoMP, backhaul requirement will
hence limiting the CoMP gains. Semi-dynamic clusters are vary. JT-CoMP will require more bandwidth due to user data
proposed in [40]–[43] where multiple static clustering patterns being shared between cooperated cells. Authors studied back-
are designed to mitigate inter-cluster interference. This type of haul bandwidth requirements for network MIMO in [61] and
approach is more responsive to the dynamic changes of the net- concluded that backhaul requirement for CSI and schedul-
work and user profile, however it still lacks on providing full ing information exchange is negligible when compared to
user data sharing. Hence, JT-CoMP require much larger back- exchange between the BSs is not required but CSI exchange
haul bandwidth than CS/CB CoMP. Backhaul requirement is is still employed for joint precoding. Deghel et al. [68]
also strongly dependant on user SINR and cluster size. Users presented a JT-CoMP solution in a limited backhaul capacity
with high SINR will demand higher throughput which will scenario BS caching is introduced to reduce required backhaul
increase backhaul demand. Biermann et al. [62] have stud- capacity for user data, hence increasing available backhaul
ied distributed JT-COMP feasibility in terms of high backhaul capacity for CSI sharing. Improved backhaul availability for
bandwidth and latency requirements especially in hotspot sce- CSI sharing improves the accuracy in CSI knowledge at the
narios where certain backhaul links are more loaded than central node, resulting in better precoding, hence improved
others. In the proposed algorithm, all CSI is sent from cooper- interference cancellation.
ated cells to the serving cell where it’s processed for precoding Realisation of CoMP depends on high backhaul band-
design and sent back from serving cell to the cooperating cells. width availability, hence CoMP clustering algorithms need to
Hence serving cell backhaul demand is more than other BSs take this limitation into account. Caching popular multimedia
in the cluster. Based on the backhaul load on each BS, a proves to be one of the ways to reduce backhaul bandwidth
dynamic serving BS reassignment algorithm is proposed by requirement for CoMP realisation. Cluster size and type of
using “forced handover” to distribute backhaul load evenly. cooperation are other factors that can change the backhaul
Zhao and Lei [63] have designed a user-centric clustering bandwidth requirement. Furthermore re-distribution of back-
strategy to minimise the backhaul data transfer for the JT- haul data transfer to less-loaded cells can be deployed for
CoMP scenario where user data exchange between the BSs better CoMP gains.
will be very high. An optimised number of links is pro-
posed for a given CoMP cluster based on minimum SINR
requirement of each user. Heuristic approach is used to reduce C. Energy Efficiency
the links based on channel strength and “Signal to Leakage” Energy efficiency has recently become an important topic
(SLR) ratio (i.e., taking signal power and also the inter- for wireless networks for both economical and environmental
ference caused to other users into account). Authors have reasons [69], [70]. It has been reported that information and
further improved this design in [64]. An optimisation problem communications technology (ICT) industry contributes 2% of
is formulated and approximate results are obtained by con- the global carbon footprint, and it’s expected to increase to 3%
vex relaxation. An iterative algorithm is followed to further by 2020 [71]. In mobile communications, more than 80% of
reduce the number of BSs in each user’s cluster. A con- the power is consumed in RAN, especially BSs [70]. As briefly
trol unit (CU) is proposed for the semi-distributed solution discussed in the introduction section, network densification is
where each BS is connected and share CSI with CU. Author’s one of the key tools to increase capacity for future wireless
approach of further user-centric clustering optimisation in a networks to meet ever increasing traffic demand which will
given network-centric cluster helps reducing wasted network severely increase energy consumption and OPEX costs. New
resources. However a trade-off between spectral efficiency and architectures like CDSA [17] and C-RAN [22]–[24] have been
backhaul bandwidth optimisation should be considered for envisioned to enable energy efficiency and reduce OPEX and
more optimal solutions. CAPEX costs in future wireless networks, mainly by providing
1) Caching at the RAN for JT-CoMP: There is an increas- small cell coverage only when it’s required. Enabling CoMP
ing interest in the research community to explore potential will also improve energy efficiency [15]. It’s been in the atten-
benefits of caching popular multimedia content closer to the tion of research to design CoMP clusters to maximise energy
user to reduce high backhaul requirement due to duplicate con- efficiency and to optimise the trade-off between spectral effi-
tent download. A significant amount of network data usage ciency and energy efficiency. On one hand, CoMP can reduce
is due to duplicate downloads of few popular multimedia cell/UE output power for a given QoS but there is also addi-
content from Netflix, Youtube, Facebook etc. Caching the pop- tional energy consumed for additional signal processing and
ular content at various points in the network, i.e., RAN, core backhaul requirement.
network or even the user devices can reduce the high back- CoMP clustering can be optimised for energy efficiency by
haul requirement and give opportunity for JT-CoMP deploy- increasing the number of sleeping BSs and/or their sleeping
ment, where high backhaul capacity is not available [65]. duration. In [72], BS sleeping with CoMP has been studied
Furthermore, caching closer to the user can improve overall for energy efficiency with static clustering and assuming one
energy efficiency. A recent study on an operational 4G net- cell is sleeping on each cluster during off-peak hours. A joint
work shows [66] that 73% of the data volume is cachable and sub-carrier and power allocation algorithm is proposed to min-
54% of the cachable data is revisited, so significant gains are imise the power requirements for coordination and compensate
possible with caching. for sleeping cell for a given QoS. Cao et al. [73] has com-
In [67], caching at the BS is proposed and an opportunis- pared the energy efficiency gains between CoMP and wireless
tic cooperative MIMO is employed without high backhaul relaying by maximising the number of sleeping cells. Based on
requirement. Cells within the same cluster are proposed to the traffic demand, it’s shown that, energy efficiency gains are
cache identical data aiming to be employed for multi-user almost constant in lightly loaded traffic conditions where net-
JT-CoMP. For users where requested data is available at the work is mainly coverage limited. In high traffic load, there is
cache, JT-CoMP is proposed. If the requested data is not almost no energy gains possible, whereas in “energy efficient
available at the cache, CB-CoMP is proposed where user-data region”, dynamic energy efficiency algorithms can provide
bigger energy efficiency. As BS density increases, the “energy Imran et al. [77] presented a distributed self organised load
efficient region” also increases and the region for larger CoMP balancing algorithm to reduce reference signal power for the
gains decreases. In [31], user-centric CoMP clustering for all congested cell to make neighbour cells more favourable and
cells within 3dB window is studied for cell switch-off on hence distribute the traffic onto neighbour cells. Another dis-
lightly loaded cells to improve energy efficiency. It’s shown tributed SON algorithm in [78] focuses on BS antenna tilt
that unnecessary increase in cluster size and imperfect channel optimisation to improve spectral efficiency at hotspots by find-
knowledge can lead to energy inefficiency. Up to 24% more ing the users centre of gravity and focusing the antenna beam
energy efficiency is observed in perfect CSI when compared to the hotspots. Kim et al. [79] presented a distributed load bal-
to imperfect CSI conditions. Authors proposed fast changing ancing solution from the idea of each BS periodically sharing
clusters responding to fast fading changes which increases sig- its average load with users and users utilise this information
nalling overhead and also imperfect CSI knowledge leads to alongside with signal quality to make the decision for cell
non-optimal clusters. Cluster change frequency can be limited association. A class of user association schemes for HetNet
to respond to large scale fading only which can average out is presented in [80] to achieve load balancing between macro
the fast scale changes for more reliable cluster formation. and small cell layer.
Besides BS switch-off, deployment costs can be reduced Centralised scheduling in emerging technologies like
and energy efficiency can be maximised by taking network C-RAN [22]–[24] and CDSA [17] can also be utilised for load
coordination into account at network planning stage. In [74], balancing. Centralised resource management entity (RME) is
a BS planning scheme is proposed to reduce the total number proposed for CDSA in [81] which will select the most suitable
of required BSs for a given coverage and traffic quality of SC for scheduling. Centralised RME can also aim to distribute
service (QoS) by inter-cell cooperation. A single user (SU) network load evenly between the SCs [17].
MIMO CoMP scheme with user-centric clustering method is As discussed in previous sections, CoMP will introduce
employed to choose the optimal BS locations for deployment spectral efficiency gain and increased throughput especially
from a number of candidate BS locations to maximise energy at the cell edge [8]. Additional capacity from CoMP can be
efficiency. A typical example of this work is to reduce the utilised for load balancing through dynamic CoMP cluster-
number of BSs required from three to two BSs where some ing based on cell load. Centralised scheduling function within
users can’t be served without the third BS if CoMP is not the CoMP cluster can be located possibly at the macro BS in
employed. CDSA, or at the pool BBU cloud in C-RAN. Self organised
Deployment of CoMP and realisation of future network CoMP clustering algorithms can be developed to dynamically
architectures like CDSA and C-RAN will enable energy effi- shift traffic from loaded cells to less-loaded neighbours while
ciency by increasing the number of sleeping cells. However, maintaining a certain level of QoS however there are no stud-
most studies in the literature are lacking the load conditions ies in the literature to our knowledge where CoMP clustering
in the network but concentrate on coverage requirements only. is used for load balancing.
With predicted mobile data growth, network capacity will be
under pressure and will require to be managed more intelli-
gently. BS switch-off with CoMP clustering algorithms will E. Multi-Objective Clustering
need to include data demand and available capacity in the As seen in aforementioned subsections, dynamic CoMP
network. Hence, a more comprehensive approach for dynamic clustering can aim to improve not only spectral efficiency
CoMP clustering should optimise energy efficiency and load but also other objectives like energy efficiency and load
balancing jointly. We discuss CoMP clustering in relation to balancing. Recent works on CoMP clustering have focused
load balancing in the next subsection. on multi-objective clustering where two objectives are opti-
mised, trade-off between the objectives have been investigated.
Katranaras et al. [60] have compared a number of static clus-
D. Load Balancing tering options for trade-off between throughput and energy
Load balancing has always been an important topic for efficiency in sparse, medium and dense deployment scenarios.
cellular networks due to non-even distribution of user traf- They have identified transmit power, inter-site distance and
fic, resulting in some BSs overloaded whereas other BSs not SINR service demands as the main inputs for this trade-off.
fully utilised. Network planning process takes traffic distribu- Li et al. [82] proposed a dynamic CoMP clustering algorithm
tion into account and BS locations are planned accordingly, with BS sleeping to maximise energy efficiency while main-
however unpredictable nature of user activities like traffic taining high achievable rate for all users. Candidate clusters
accidents, mass events etc still cause overloaded cells. With are formed by all possible combinations of groups of cells
ever increasing traffic, predicted 1000 fold increase beyond with predefined cluster size and each BS selects a suitable
2020 [2], load balancing becomes even more important in cluster from the candidate clusters by maximising achiev-
future cellular networks. Various load balancing schemes have able rate for its users. Developed algorithm then looks for
already been studied in [75] for traditional networks. A math- cell load and moves users from cells with low load onto
ematical framework for cell load and a simple load balancing other clusters to increase the number of sleeping cells and
algorithm is presented in [76]. Authors proposed to shift traf- hence better energy efficiency. Proposed clustering algorithm
fic from loaded cell to its unloaded neighbours by changing lacks on scalability as number of candidate clusters increase
the handover offset parameter in iterations. with network size, leading to high computational complexity.
TABLE III
S UMMARY OF C O MP C LUSTERING A PPROACHES BASED ON O BJECTIVE F UNCTION
Moreover, proposed algorithm fails to look at total system constraints are taken into account to make sure there is enough
capacity and load balancing aspects, i.e., BS load need to be coverage for BS switch-off. However, other constraints like
looked at for any congestion and reduce the number of sleep- RAN capacity, load balancing, backhaul bandwidth availabil-
ing BSs for much needed capacity in the network. Hence, ity also need to be considered in a realistic network for BS
energy efficiency and load balancing will need to be jointly switch-off decision. In this context, more research is required
optimised for an improved multi-objective CoMP clustering for multi-objective CoMP clustering algorithms with above
algorithm. mentioned constraints. A comparison of CoMP clustering
Available backhaul capacity is one of the biggest limita- algorithms based on aimed objective and their shortcomings
tions for CoMP, especially JT-CoMP. A number of research are provided in Table III.
works in literature focus on CoMP clustering where spec-
tral efficiency and backhaul requirement is jointly optimised. VII. F UTURE R ESEARCH D IRECTIONS
Mayer et al. [83] presented the implications of backhaul chan-
In this section, we present open research areas for CoMP
nel reliability on spectral efficiency of the clusters. It’s shown
clustering challenge. Potential use of Big Data in proac-
that, both JT and CB-CoMP scenarios give better spectral
tive CoMP clustering is reviewed in the next subsection.
efficiency results with strong backhaul reliability. However,
Its followed by open research areas in dynamic cluster-
spectral efficiency improvement reduces sharply when back-
ing approaches, reviewing the challenges on complexity/gain
haul reliability goes down. As discussed in Section VI-B1,
trade-off of dynamic clustering and the need for comprehen-
caching popular multimedia content at the BS is an emerging
sive CoMP clustering solutions to maximise not only spectral
research area for reducing backhaul requirement and hence
efficiency but also other system objectives like load balancing,
enabling JT-CoMP in limited backhaul scenarios [67], [68].
energy efficiency and backhaul optimisation.
Existing literature focuses on one limiting objective and
investigates the trade-off against spectral efficiency gains.
However, a more comprehensive CoMP clustering approach A. Big Data Empowered Proactive CoMP Clustering
need to take all limiting factors in the same algorithm for As mentioned in aforementioned sections, CoMP has the
intelligent clusters which jointly optimise backhaul bandwidth, capability to significantly improve spectral efficiency and cell
energy efficiency, load balancing and spectral efficiency. For edge throughput through cooperation of limited number of
example, BS switch-off is a widely studied concept in lit- BSs referred to as CoMP clusters. The state-of-the-art research
erature as part of CoMP clustering, however only the SINR on dynamic CoMP clustering in general have a reactive line
of action, i.e., CoMP clustering are designed/optimized with of the accessibility, retainability and integrity performances
respect to current network conditions. For example load bal- of the network. Several metrics including blocked call rates,
ancing targeted CoMP clustering will kick in when congestion access failure rates, setup times, success rate, and hand-over
is observed or diagnosed. However, in light of emerging failure rates project accessibility of the network. The dropped
5G future cellular networks personified with ambitious QoS call rates, completion times, packet data protocol context and
requirements of almost infinite capacity or zero latency [3], success rate together define the retainability of the network.
this approach will not be able to meet stringent performance The metrics like speech and data streaming quality, through-
requirements of 5G. This is because in classic dynamic CoMP put, packet jitter and delay give an idea about user perceived
clustering, certain time is required to observe the current con- quality of experience (QoE).
ditions, find optimum clustering with respect to the objective b) Cell level data: It refers to the measurements that are
function and current conditions and then trigger the appro- reported by a BS and all users within the coverage of that BS.
priate clustering action. The resultant intrinsic delay is not Examples of useful cell level data streams are measurements
compatible with 5G targeted QoE levels. Therefore for 5G, reporting uplink noise floor in terms of reference interference
CoMP clustering paradigm requires proactive or predictive power, channel based power information, physical resource
approach such that spatio-temporal future network state in block usage per cell, no. of active users per cell and minimiza-
terms of channel variation, mobility behaviour and capacity tion of driving test (MDT) measurements. MDT reports consist
requirements can be predicted beforehand. This is possible of the RSRP and reference signal received quality (RSRQ) val-
through exploitation of the cognition of context of applica- ues of the serving and neighbouring cells reported by the users
tion as well as state of the network by inferring network-level to their serving BS [84], [85].
intelligence from the massive amount of control, signalling, c) Core network level data: The core network data
and contextual data known as Big Data as proposed in [4]. includes signalling information, historical alarm logs, equip-
Key elements and sources of Big Data for mobile networks ment configuration lists and service and resource utilization
have been identified in [4]. By leveraging a dexterous combi- accounting records (Call Data Records - CDRs and Extended
nation of advanced techniques of machine learning, statistics Data Records XDRs) as well as aggregate statistics of network
and optimization, Big Data can be tapped to enable and performance metrics.
empower CoMP clustering algorithms to achieve true perfor- d) Miscellaneous data: It consists of the structured infor-
mance gains of CoMP. Endowed with predictive capabilities, mation already stored in the separate databases including
-thanks to Big Data- CoMP clustering algorithms can track, customer relationship management, customer complaint cen-
learn and dynamically build user mobility and demand pro- ter and spectrum utility maps. This also includes unstructured
files as well as channel characteristics models to predict future information such as social media feeds, specific application
user locations coupled with service requests and channel state usage patterns, and data from smart phone built-in sensors
information. This can lead to timely efficient CoMP clustering and applications.
as well as can help to alleviate high backhaul requirements. 2) Role of Big Data in Proactive CoMP Clustering:
Another advantage of exploiting Big Data in CoMP clustering In perspective of proactive CoMP Clustering, channel maps
is that, it can represent the global state of the network which built upon collected MDT reports and unified information
enables the global optimal CoMP clustering with respect to of handover traces and CDRs are potential Big Data con-
the defined objective functions such as energy efficiency, stituents that can be harnessed to predict future network state
spectral efficiency or load balancing as opposed to relying through machine learning algorithms and statistical techniques.
only on the local information that may lead to only locally Specifically, they can be utilized to predict future spatio-
optimal CoMP clustering solutions. As the current research temporal rate requirements along with the associated channel
on CoMP clustering lacks this proactive perspective and to state information as explained in subsequent subsections. This
the best of our knowledge, currently no existing work tar- can pave the way for enabling timely efficient prediction based
gets “Proactive CoMP Clustering” in general and “Big Data proactive user-centric dynamic CoMP clustering.
empowered Proactive CoMP Clustering” in particular, there- a) CoMP with big data aided channel prediction:
fore the goal of this section is to give future outlook of Accurate and timely channel estimation is one of the most
Big Data enabled proactive CoMP clustering. In subsections vital requirements of CoMP system. The coordinating BSs in
to follow, we briefly explain how Big Data can empower a CoMP cluster are typically assumed to be connected to a
prediction based proactive CoMP clustering algorithms in centralized CCU through backhaul links [86]–[88]. In FDD
terms of channel prediction, mobility prediction and user systems, each user within CoMP cluster needs to estimate and
profiling. predict the CSI from all BSs of the cluster and feed it back
1) Big Data in Cellular Networks: In the context of cellular to the serving BS which is then forwarded to CCU. Based
networks, Big Data refers to the massive amounts of control, on the available CSI, joint transmission, user scheduling or
signalling and contextual data that is being routinely produced coordinated beamforming schemes are designed. The quality
during day-to-day operation of cellular network. The potential of CSI has large impact on the performance of CoMP systems
constituents of Big Data in cellular networks are [4]: and clustering decisions. Restricted feedback and backhaul
a) Subscriber level data: The subscriber level data com- links induce different degrees of latencies resulting outdated
prises of key performance indicators obtained from a voice or measurements [89]. The X2 latencies observed in 3GPP-LTE
a data session initiated by the subscriber to give an indication deployed networks is of order 10 to 20 milliseconds [8], [90].
The outdated CSI leads to severe performance degradation in markov models have been commonly used for temporal-
CoMP systems even when the users exhibit low mobility [7]. spatial prediction purposes as in [97] and [98]. Received
Channel estimation at the terminals as well as compression signal strength indicator (RSSI) available in MDT reports can
and quantization of CSI are further sources of inaccuracy. also be utilized to predict future location as has been done
In time-varying wireless channels, channel prediction is a in [99] and [100]. The identified future location of the users
popular approach to provide up-to-date channel information along with the corresponding time stamps can be fed to the
and it is shown in [91] that the performance of CoMP sys- CoMP dynamic clustering algorithms (both user-centric as
tems is improved significantly even with the large backhaul well as network-centric) for computing optimal clusters.
latency when channel prediction is applied. State of the art Mobility behaviour of the users directly affects the CoMP
prediction techniques like Kalman and Weiner filtering make clustering decisions as CSI has small validity period for high
CoMP links more robust for CSI delays of few milliseconds speed users and clustering decisions needs to be performed fre-
and at moderate mobility [5]. Recently Doppler-delay-based quently leading to high computational overheads. One solution
prediction has been proposed wherein the channel for each can be that low speed or static users can be served by CoMP
link between a transmitter and a receiver antenna can be mod- cluster BSs, however, high speed mobile users continue to
elled by a number of multi-paths with their individual complex be served by single BS. By utilizing RSSI and the cell sizes
amplitude, delay and Doppler frequency. These parameters can information embedded in Big Data and predicted future user
be estimated for each path based upon the recent channel his- locations, CoMP clustering algorithms can be executed before-
tory embedded in Big Data and the future condition of the hand leading to significant reduction in latency and bandwidth
channel can be predicted by inserting the estimated parame- requirements.
ters into the channel model. Both Doppler delay and Kalman c) CoMP with big data aided user profiling: Call Data
prediction lead to significant improvement in Mean Squared Records (CDRs) are one of the key elements of the Big
Error (MSE) for the CSI that leads to better performance. The Data that can be harnessed from a cellular network. CDRs
powerful Big Data aided CSI prediction can be an important reflect mobile users behaviour and give out clues on how the
enabler for CoMP clustering decisions. users utilize the network resources. CDRs contain information
Big Data can also play crucial role in proactive selection about the voice calls and data usage pattern and are important
of BSs for cluster formation. One of the vital sources of Big markers of temporal-spatial capacity requirements across the
Data in mobile communications are MDT reports consisting deployed network [101]–[103]. CDRs can be utilized to pro-
of RSRP and other channel quality related metrics reported by file the network usage behaviour of the mobile users which
the users to their serving BS [84], [85]. The averaged RSRP in turn can be utilized for user-centric or behaviour-centric
values of the BSs, as reported by the UEs, can be compared CoMP clustering. By applying machine learning and statistical
to a threshold to determine which of these BSs should cooper- tools on CDRs, we can determine the capacity requirements of
ate. Based on current MDT reports, future channel conditions the users at different time periods and can utilize this profile
can be predicted through conventional time-series forecast- information to cluster the CoMP enabled BSs to satisfy the
ing methods. In case of sparse MDT reports due to small expected QoS requirements of the users.
number of users like in small cells, light-weight Grey mod- Social media feeds are another element within Big Data
elling techniques [92] that are useful for short term forecasting that give helpful insights about the interaction of the users and
can be utilized as done in [93]. The grey model can pre- expected temporal-spatial demand of network resources across
dict the next RSRP value from data points obtained in the the network. Among many online social networks, Twitter is
database. Therefore, instead of waiting for actual MDT reports, one of the popular ways users share information and experi-
the predicted RSRP measurements can be fed to the chan- ence socially on the Web. Twitter data can be mined through
nel estimation and subsequently to the cluster optimization application program interface (Twitter APIs) wherein each
algorithm that proactively adapts the cell clustering in CoMP timestamped tweet contains number of useful information like
perspective. location, number of re-tweets, number of favourites, message
b) CoMP with big data aided mobility prediction: itself and hashtags. Twitter data can be utilized to estimate
Big Data aided mobility prediction can play important role in traffic demand as number of tweets is highly correlated with
proactive CoMP clustering decisions. Mobility prediction uti- the number of people in confined places [104]. It can also
lizes persons mobility history, i.e., a series of locations and be utilized to assess networks QoE from subscriber’s perspec-
corresponding dwell times to predict this persons next location, tive [105]. The social media feeds together with the CDRs can
as well as his/her dwell time in that location [94]–[100]. In be taped to accurately model the user behaviour and can be
this way, CoMP clustering algorithms can plan in advance the utilized to optimize user-centric CoMP clustering algorithms.
clustering decisions thereby meeting the strict latency require- 3) Tapping Big Data for CoMP Clustering: Multifaceted
ments of 5G networks. Big Data as identified in [4] also and multifarious Big Data can help to enable and optimize
contains handover reports which contain Cell IDs and corre- the proactive CoMP clustering algorithms. Big Data consists
sponding timestamps whenever user is handed over to new of big pool of training datasets that is of significant advan-
cell. Several techniques such as mobility pattern matching tage for prediction techniques based on advanced supervised
using mobility database, periodicity and multi-class classifica- machine learning algorithms like deep learning methods [106].
tion and bio-inspired approaches as presented in [94]–[96] can However, Big Data comes with its own set of challenges like
be used to predict user mobility behaviour. Markov and hidden how to efficiently tap the potential of this Big Data in real
or hotspots wherein optimal large cooperative clusters, per-

forming joint transmission (JT), are formed around hotspots
and the coverage of BSs are transferred to hotspots by dynam-
ically changing the antennas beam angles. With the proposed
big data framework in place, this process can be done proac-
tively, instead of reactively, thereby further improving the QoE
CoMP can offer.
Another work [119] has proposed a cross-tier coopera-
tion in non-uniform HetNets wherein cell edge hotspot users
are served by CoMP BSs. The location of clustered users
or hotspots present in the network have also been utilized
Fig. 7. Big Data Empowered Proactive CoMP Clustering Framework. in [78] and [120]–[122] wherein network configuration param-
eters (antenna parameters) are optimized w.r.t the identified
hotspots. The underlying phenomenon is inherently the same
time that is hindered by four inherent characteristics of Big as in case of dynamic CoMP Clustering since network parame-
Data, i.e., Volume, Variety, Velocity and Veracity [4]. Big Data ters optimization is done based on hotspot location. The afore-
management tools under umbrella of Hadoop ecosystem are mentioned algorithms initiate reactively assuming hotspots
potential enablers to deal with the acute dynamicity of the Big have already formed into the network and their location is cent
Data. The main components of Big data processing platform percent known accurately. However, with Big Data Predictive
consist of [107]: Analytics, formation of hotspots can be predicted beforehand
1) Trasmission Module consisting of Flume [108] and as explained above and dynamic CoMP clustering can be per-
Kafka [109] that uploads network data in real time with formed well in time to minimise QoE degradation time. This
stable transmission to the cloud platform. is where Big Data comes into the picture.
2) Storage Module consisting of Distributed File System Very recently, some works have emerged that leverage Big
(HDFS) [110] and HBase [111] with high fault tolerance Data driven predictive analytics in mobile networks for pre-
capability. dicting hotspot formation using CDRs. The work in [107] has
3) Processing Module comprising of MapReduce [112] for performed Big Data collection, storage, and pre-processing of
parallel distributed processing, Spark [113] for cyclic CDRs and has proposed:
data flow and in-memory computing and Storm [114] i The rules for extracting location data, and constructing
for enabling real-time analysis. people trajectories
4) Management Module to monitor the whole plat- ii The methods for solving data noise (i.e., cell tower
form with Flume collecting the monitored data and oscillations)
Zookeeper [115] to modify configuration parameters of iii The algorithms for discovering common mobility pat-
each machine and equipment. terns in densely populated area
The processed statistics from Big Data can then be fed iv Identifying hotspots.
to advanced machine learning methods to model network Similarly, in [116], [117], and [123]–[125], Big Data technolo-
and user behaviour and predict future spatio-temporal net- gies and analytical algorithms have been used for predicting
work states. By knowing probable future user locations, their hotspot formation or forecasting pedestrian destinations with
expected rate requirements and estimated channel state infor- satisfactory accuracy.
mation, CoMP clustering algorithms can proactively adapt In a nutshell, Big Data driven predictive analytics predict-
themselves on the fly to cope with acute dynamics of cellu- ing the future spatio-temporal state of the network accurately
lar networks resulting in seamless quality of perception. This and using this knowledge for dynamic CoMP clustering well
framework is depicted in Figure 7. in time is the future of the CoMP clustering that can truly
4) Relevant Work on Big Data Driven Proactive CoMP unleash the real potential of CoMP and can be instrumental
Clustering: Although no existing work target proactive CoMP in improving user experience in future 5G cellular networks.
clustering leveraging Big Data explicitly yet, there exist cer- Presence of more data (Big Data) results in better and accurate
tain works, wherein dynamic CoMP clustering is performed models as it allows the data to tell for itself,İ instead of relying
targeted at hotspots, assuming hotspot location are already on assumptions and weak correlations since a weak assump-
somehow known by the network. The Big Data processing tion coupled with complex algorithms is far less efficient than
framework presented above cannot only identify the future using more data with simpler algorithms. This fact has been
hotspots but it can also predict future load, e.g., using data captured by many studies, e.g., [126] and [127] wherein results
of mobility traces and past CDR records [107], [116], [117]. suggests for a given problem, adding more examples to the
Once a hotspot is characterized, the appropriate CoMP algo- training set monotonically increases the accuracy of the model.
rithm can be leveraged to cope with high capacity demands However with aid of Big Data, as the ability to generate bet-
for hotspots. Examples of work which can leverage this idea ter predictions continues to improve, it is noteworthy that the
include study in [118]. Rezagah et al. [118] have proposed a accuracy of these predictions is only as good as the accuracy
novel cell structuring and clustering algorithm to dynamically of the underlying data (garbage-in, garbage-out). Leveraging
transfer network resources from sparse cells to crowded cells Big Data of poor quality for proactive CoMP clustering might
produce erroneous predictions, counter-productive clustering RAN caching is employed in [67] and [68] to reduce high
decisions and poor performance than that achievable with backhaul bandwidth dependency for cooperation. However,
conventional reactive dynamic CoMP clustering. backhaul bandwidth limitation is studied in isolation, not in
relation to other objective functions like spectral efficiency and
load balancing. An open research area is to develop backhaul-
B. Dynamic CoMP Clustering Challenges aware CoMP clustering algorithms which aim to maximise
1) Complexity/Gain Trade-off for Dynamic CoMP spectral efficiency and user satisfaction in relation to backhaul
Clustering: Dynamic clustering has more potential for better limitations and load balancing.
performance gains due to its ability to respond to network Another open research area is to utilise Big Data for RAN
and user/service profile changes. Inter-cluster interference can caching to compensate for the high backhaul requirement. Big
be mitigated with dynamically changing cluster boundaries. Data aided proactive caching can play significant role in JT
Both user-centric, network-centric and hybrid algorithms mode of CoMP wherein user data is shared among cooper-
have been studied in the literature. CoMP clustering research ating BSs. Such proactive caching can relax high backhaul
on user-centric approaches lacks scalability and suffer high requirements of JT CoMP.
scheduling/precoding design complexity. Network-centric c) Energy efficiency: CoMP deployment and intelligent
approaches mainly fail to provide full CoMP gains when clustering solutions can improve energy efficiency especially
compared to user-centric cluster design. Hybrid clustering with increasing the number of sleeping BSs [31], [72], [73].
provide a balance between complexity and CoMP gain BS sleeping has been employed in most works to improve
trade-off. However, existing works fail to provide fully energy efficiency, however only SINR constraints are taken
dynamic hybrid solutions. The challenge with fully dynamic into account for BS sleeping to make sure there is coverage
solution is the increased complexity especially with increased available for all users. As discussed in Section VI-E, other
scheduling and precoding design and additional overheads. constraints like system capacity and backhaul bandwidth will
More rigorous research is required on novel hybrid solutions need to be taken into account for BS sleeping. For example, a
where dynamic user-centric clustering is employed within a more realistic approach should consider load balancing while
dynamic network-centric clustering algorithm and the gains making decision for BS switch-off with the aim of maximising
of such algorithms against the complexity and additional energy efficiency. Sleeping cells may need to be switched on
overhead costs. to handle additional load in the network, however it comes
2) Multi-Objective CoMP Clustering: CoMP is envisioned with the additional signalling cost overhead and degradation
for mitigating inter-cell interference and hence increasing on energy efficiency.
spectral efficiency. Hence the primary aim for CoMP clus- Furthermore, extensive research is required on compre-
tering is to maximise spectral efficiency, however other limi- hensive multi-objective clustering algorithms to include all
tations like load balancing, backhaul bandwidth availability, limitations/objectives, i.e., spectral efficiency, energy effi-
system capacity and energy efficiency are also taken into ciency, backhaul bandwidth and load conditions into algorithm
account for improved clustering solutions. Existing literature design and analyse the trade-off between multiple objective
focuses on maximising spectral efficiency along side with one gains and associated costs. Analytical tools like coalitional
more objective, mostly focusing on backhaul bandwidth and game theory can be utilised for merging multiple objectives
energy efficiency constraints. However, a more comprehensive into a single payoff function for exhaustive multi-objective
approach is required to take all constraints into account for a CoMP clustering design. Trade-off between different objec-
realistic CoMP clustering solution. We outline the potential tives and optimum balance between these metrics is an area
research directions in multi-objective clustering as below. worth exploring further. Moreover, Big Data aided predictive
a) Load balancing: As discussed in Section VI-D, load models need to be explored further for novel proactive multi-
balancing is an increasingly important concept for mobile net- objective proactive CoMP clustering design to support much
works due to the exponential increase in data demand [1]. faster response rates required for future networks.
CoMP is likely to be deployed in interference limited networks
where there is high data demand. An interesting research area
is to develop CoMP clustering algorithms to support load bal- VIII. C ONCLUSION
ancing while spectral efficiency is maximised. A load-aware, This article provides an extensive survey on the CoMP clus-
user-centric CoMP clustering approach is presented in our tering methods for future cellular networks. We first give the
previous work [128], however further research is required to motivation for CoMP for future wireless networks and briefly
analyse fully dynamic CoMP clustering techniques and the provide an outline of CoMP implementation challenges and
trade-off between load balancing gains and potential losses on the need for CoMP clustering. We then provide a section to
spectral efficiency. give brief tutorial about different types of cooperation, associ-
b) Backhaul optimisation: A number of research are con- ated challenges and propose network architectures like CDSA
ducted for CoMP clustering which takes backhaul bandwidth and C-RAN which will enable CoMP implementation. The
limitation into account. The main contributors for high back- core of the article provides an extensive survey on CoMP clus-
haul bandwidth requirement such as cluster size [63], [64] and tering techniques available in the literature and introduce two
type of cooperation (i.e., CB or JT) [67], [68] are dynamically novel taxonomies for CoMP clustering algorithms based on
changed to adapt to limited backhaul bandwidth availability. self-organisation and aimed objective function.
Firstly, we provide a CoMP clustering taxonomy based on [14] A. Beylerian and T. Ohtsuki, “Coordinated non-orthogonal mul-
self organisation, and critically discuss static, semi-static and tiple access (CO-NOMA),” in Proc. IEEE Glob. Commun.
Conf. (GLOBECOM), Washington, DC, USA, Dec. 2016.
dynamic CoMP clustering works in literature. Dynamic clus- [15] J. B. Rao and A. O. Fapojuwo, “A survey of energy efficient
tering algorithms are further divided based on their approach, resource management techniques for multicell cellular networks,” IEEE
Commun. Surveys Tuts., vol. 16, no. 1, pp. 154–180, 1st Quart., 2014.
network-centric and user-centric approaches, their benefits and [16] E. Hossain, D. I. Kim, and V. K. Bhargava, Cooperative Cellular
shortcomings are highlighted. Wireless Networks. Cambridge, U.K.: Cambridge Univ. Press, 2011.
Secondly, we present a novel CoMP clustering taxonomy [17] A. Mohamed, O. Onireti, M. A. Imran, A. Imran, and R. Tafazolli,
“Control-data separation architecture for cellular radio access networks:
based on the objective function. CoMP clustering algorithms A survey and outlook,” IEEE Commun. Surveys Tuts., vol. 18, no. 1,
aiming for spectral efficiency, energy efficiency, backhaul opti- pp. 446–465, 1st Quart., 2015.
[18] H. A. U. Mustafa, M. A. Imran, M. Z. Shakir, A. Imran, and
misation and load balancing are extensively discussed. More R. Tafazolli, “Separation framework: An enabler for cooperative and
focus is given on comprehensive multi-objective clustering, D2D communication for future 5G networks,” IEEE Commun. Surveys
Tuts., vol. 18, no. 1, pp. 419–445, 1st Quart., 2016.
available works in literature are presented, shortcomings are [19] O. G. Aliu, A. Imran, M. A. Imran, and B. Evans, “A survey of self
identified in detail. organisation in future cellular networks,” IEEE Commun. Surveys Tuts.,
We then outline open research areas for CoMP cluster- vol. 15, no. 1, pp. 336–361, 1st Quart., 2013.
[20] 3GPP, “Requirements for further advancements for evolved univer-
ing and propose potential approaches for solutions. Proactive sal terrestrial radio access (E-UTRA) (LTE-advanced),” 3rd Gener.
CoMP clustering is envisioned to accommodate much faster Partnership Project (3GPP), TR 36.913 V9.0.0, Dec. 2009.
response rates required for 5G. We highlight the potential use [21] 3GPP, “Evolved universal terrestrial radio access (E-UTRA); fur-
ther advancements for E-UTRA physical layer aspects,” 3rd Gener.
of Big Data to empower prediction based CoMP clustering Partnership Project (3GPP), Tech. Rep. 36.814, Mar. 2010.
algorithms. Big Data in cellular networks context is identified, [22] C. Mobile, “C-RAN: The road towards green RAN,” White Paper,
2011.
and use of Big Data for channel prediction, mobility predic- [23] K. Sundaresan, M. Y. Arslan, S. Singh, S. Rangarajan, and
tion and user profiling prediction is discussed. We identify Big S. V. Krishnamurthy, “FluidNet: A flexible cloud-based radio access
Data aided prediction models to form a future outlook in pre- network for small cells,” in Proc. 19th Annu. Int. Conf. Mobile Comput.
Netw., Miami, FL, USA, 2013, pp. 99–110.
diction based CoMP clustering. We then discuss further future [24] I. Chih-Lin et al., “Toward green and soft: A 5G perspective,” IEEE
research areas on dynamic CoMP clustering complexity/gain Commun. Mag., vol. 52, no. 2, pp. 66–73, Feb. 2014.
[25] Q. Wang et al., “Application of BBU+RRU based CoMP system
trade off and multi-objective CoMP clustering algorithms to to LTE-advanced,” in Proc. IEEE Int. Conf. Commun. Workshops,
optimise load balancing, backhaul limitation, energy efficiency Dresden, Germany, 2009, pp. 1–5.
and spectral efficiency simultaneously. [26] A. Davydov, G. Morozov, I. Bolotin, and A. Papathanassiou,
“Evaluation of joint transmission CoMP in C-RAN based LTE-A
HetNets with large coordination areas,” in Proc. IEEE Globecom
R EFERENCES Workshops (GC Wkshps), Atlanta, GA, USA, 2013, pp. 801–806.
[27] T. Alsedairy, Y. Qi, A. Imran, M. A. Imran, and B. Evans, “Self organ-
[1] “Cisco visual networking index: Global mobile data traffic fore- ising cloud cells: A resource efficient network densification strategy,”
cast update, 2015–2020,” White Paper, Cisco, San Jose, CA, USA, Trans. Emerg. Telecommun. Technol., vol. 26, no. 8, pp. 1096–1107,
Feb. 2016. 2015.
[2] Q. C. Li, H. Niu, A. T. Papathanassiou, and G. Wu, “5G network [28] I. D. Garcia et al., “Impact of base station cooperation on cell plan-
capacity: Key elements and technologies,” IEEE Veh. Technol. Mag., ning,” EURASIP J. Wireless Commun. Netw., vol. 2010, no. 1, p. 5,
vol. 9, no. 1, pp. 71–78, Mar. 2014. 2010.
[3] “Looking ahead to 5G,” White Paper, Nokia Netw., Espoo, Finland, [29] P. Marsch and G. Fettweis, “Static clustering for cooperative
Jul. 2014. multi-point (CoMP) in mobile communications,” in Proc. IEEE Int.
[4] A. Imran, A. Zoha, and A. Abu-Dayya, “Challenges in 5G: How to Conf. Commun., Kyoto, Japan, 2011, pp. 1–6.
empower SON with big data for enabling 5G,” IEEE Netw., vol. 28, [30] A. S. Rodriguez and A. G. Armada, “Analysis of the cluster size
no. 6, pp. 27–33, Nov./Dec. 2014. in coordinated multipoint transmission,” in Proc. IEEE 19th Signal
[5] V. Jungnickel et al., “The role of small cells, coordinated multipoint, Process. Commun. Appl. Conf., 2011, pp. 1197–1200.
and massive MIMO in 5G,” IEEE Commun. Mag., vol. 52, no. 5, [31] G. Cili, H. Yanikomeroglu, and F. R. Yu, “Cell switch off tech-
pp. 44–51, May 2014. nique combined with coordinated multi-point (CoMP) transmission for
[6] D. López-Pérez, M. Ding, H. Claussen, and A. H. Jafari, “Towards energy efficiency in beyond-LTE cellular networks,” in Proc. IEEE Int.
1 Gbps/UE in cellular systems: Understanding ultra-dense small Conf. Commun. (ICC), Ottawa, ON, Canada, 2012, pp. 5931–5935.
cell deployments,” IEEE Commun. Surveys Tuts., vol. 17, no. 4, [32] 3GPP, “Evolved universal terrestrial radio access network (E-UTRAN);
pp. 2078–2101, 4th Quart., 2015. self-configuring and self-optimizing network (SON) use cases and solu-
[7] 3GPP, “Coordinated multi-point operation for LTE physical layer tions,” 3rd Gener. Partnership Project (3GPP), TR 36.902 R9 v9.3.1,
aspects,” 3rd Gener. Partnership Project (3GPP), TR 36.819 R11 Mar. 2011.
v11.2.0, Sep. 2013.
[33] 3GPP, “Telecommunication management; self-configuration of net-
[8] R. Irmer et al., “Coordinated multipoint: Concepts, performance, and
work elements; concepts and requirements,” 3rd Gener. Partnership
field trial results,” IEEE Commun. Mag., vol. 49, no. 2, pp. 102–111,
Project (3GPP), TS 32.501 R12 v12.1.0, Dec. 2013.
Feb. 2011.
[9] D. Lee et al., “Coordinated multipoint transmission and reception [34] 3GPP, “Telecommunication management; self-organizing networks
in LTE-advanced: Deployment scenarios and operational challenges,” (SON) policy network resource model (NRM) integration reference
IEEE Commun. Mag., vol. 50, no. 2, pp. 148–155, Feb. 2012. point (IRP); requirements,” 3rd Gener. Partnership Project (3GPP), TS
[10] P. Marsch and G. P. Fettweis, Coordinated Multi-Point in Mobile 32.521 R11 v11.1.0, Dec. 2012.
Communications: From Theory to Practice, 1st ed. Cambridge, U.K.: [35] 3GPP, “Telecommunication management; self-organizing networks
Cambridge Univ. Press, 2011. (SON); self-healing concepts and requirements,” 3rd Gener. Partnership
[11] M. Sawahashi, Y. Kishiyama, A. Morimoto, D. Nishikawa, and Project (3GPP), TS 32.541 R12 v12.0.0, Oct. 2014.
M. Tanno, “Coordinated multipoint transmission/reception techniques [36] S. S. Ali and N. Saxena, “A novel static clustering approach for CoMP,”
for LTE-advanced [Coordinated and Distributed MIMO],” IEEE in Proc. IEEE 7th Int. Conf. Comput. Converg. Technol. (ICCCT),
Wireless Commun., vol. 17, no. 3, pp. 26–34, Jun. 2010. Seoul, South Korea, 2012, pp. 757–762.
[12] T. Biermann, L. Scalia, C. Choi, W. Kellerer, and H. Karl, “How back- [37] H. Shimodaira et al., “Diamond cellular network—Optimal com-
haul networks influence the feasibility of coordinated multipoint in bination of small power basestations and CoMP cellular net-
cellular networks [Accepted From Open Call],” IEEE Commun. Mag., works,” in Proc. IEEE 24th Int. Symp. Pers. Indoor Mobile Radio
vol. 51, no. 8, pp. 168–176, Aug. 2013. Commun. (PIMRC Workshops), London, U.K., 2013, pp. 163–167.
[13] A. Beylerian and T. Ohtsuki, “Multi-point fairness in resource alloca- [38] A. Barbieri et al., “Coordinated downlink multi-point communica-
tion for C-RAN downlink CoMP transmission,” EURASIP J. Wireless tions in heterogeneous cellular networks,” in Proc. Inf. Theory Appl.
Commun. Netw., vol. 2016, no. 1, p. 12, 2016. Workshop (ITA), San Diego, CA, USA, 2012, pp. 7–16.
[39] A. M. Hamza and J. W. Mark, “A clustering scheme based on [61] D. Samardzija and H. Huang, “Determining backhaul bandwidth
timing requirements in coordinated base-stations cooperative commu- requirements for network MIMO,” in Proc. 17th Eur. Signal Process.
nications,” in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), Conf. (EUSIPCO), Glasgow, U.K., 2009, pp. 1494–1498.
Shanghai, China, 2013, pp. 3764–3769. [62] T. Biermann, L. Scalia, C. Choi, H. Karl, and W. Kellerer, “Improving
[40] F. Huang, Y. Wang, J. Geng, M. Wu, and D. Yang, “Clustering approach CoMP cluster feasibility by dynamic serving base station reassign-
in coordinated multi-point transmission/reception system,” in Proc. ment,” in Proc. IEEE 22nd Int. Symp. Pers. Indoor Mobile Radio
IEEE 72nd Veh. Technol. Conf. Fall (VTC Fall), Ottawa, ON, Canada, Commun. (PIMRC), Toronto, ON, Canada, 2011, pp. 1325–1330.
2010, pp. 1–5. [63] J. Zhao and Z. Lei, “Clustering methods for base station cooperation,”
[41] S. A. Ramprashad, G. Caire, and H. C. Papadopoulos, “A joint schedul- in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), Paris, France,
ing and cell clustering scheme for MU-MIMO downlink with limited 2012, pp. 946–951.
coordination,” in Proc. IEEE Int. Conf. Commun. (ICC), Cape Town, [64] J. Zhao, T. Q. S. Quek, and Z. Lei, “Semi-distributed clustering
South Africa, 2010, pp. 1–6. method for CoMP with limited backhaul data transfer,” in Proc.
[42] V. Pichapati and P. Gupta, “Practical considerations in cluster design IEEE Wireless Commun. Netw. Conf. (WCNC), Shanghai, China, 2013,
for co-ordinated multipoint (CoMP) systems,” in Proc. IEEE Int. Conf. pp. 3776–3781.
Commun. (ICC), Budapest, Hungary, 2013, pp. 5860–5865. [65] X. Wang, M. Chen, T. Taleb, A. Ksentini, and V. Leung, “Cache in
[43] H. Sun, X. Zhang, and W. Fang, “Dynamic cell clustering design the air: Exploiting content caching and delivery techniques for 5G
for realistic coordinated multipoint downlink transmission,” in Proc. systems,” IEEE Commun. Mag., vol. 52, no. 2, pp. 131–139, Feb. 2014.
IEEE 22nd Int. Symp. Pers. Indoor Mobile Radio Commun. (PIMRC), [66] B. A. Ramanan et al., “Cacheability analysis of HTTP traffic in an oper-
Toronto, ON, Canada, 2011, pp. 1331–1335. ational LTE network,” in Proc. Wireless Telecommun. Symp. (WTS),
Phoenix, AZ, USA, 2013, pp. 1–8.
[44] A. Papadogiannis, D. Gesbert, and E. Hardouin, “A dynamic clustering
approach in wireless networks with multi-cell cooperative process- [67] A. Liu and V. K. N. Lau, “Exploiting base station caching in MIMO
ing,” in Proc. IEEE Int. Conf. Commun. (ICC), Beijing, China, 2008, cellular networks: Opportunistic cooperation for video streaming,”
pp. 4033–4037. IEEE Trans. Signal Process., vol. 63, no. 1, pp. 57–69, Jan. 2015.
[68] M. Deghel, E. Bastug, M. Assaad, and M. Debbah, “On the bene-
[45] J. Liu and D. Wang, “An improved dynamic clustering algorithm for fits of edge caching for MIMO interference alignment,” in Proc. 16th
multi-user distributed antenna system,” in Proc. Int. Conf. Wireless Int. Workshop Signal Process. Adv. Wireless Commun. (SPAWC),
Commun. Signal Process. (WCSP), Nanjing, China, 2009, pp. 1–5. Stockholm, Sweden, 2015, pp. 655–659.
[46] S. Feng, W. Feng, H. Mao, and J. Lu, “Overlapped clustering for CoMP [69] G. Auer et al., “Enablers for energy efficient wireless networks,” in
transmissions in massively dense wireless networks,” in Proc. IEEE Int. Proc. 72nd Veh. Technol. Conf. VTC, Ottawa, ON, Canada, 2010,
Conf. Commun. Syst. (ICCS), 2014, pp. 308–312. pp. 1–5.
[47] F. Guidolin, L. Badia, and M. Zorzi, “A distributed clustering algorithm [70] V. Mancuso and S. Alouf, “Reducing costs and pollution in cellular
for coordinated multipoint in LTE networks,” IEEE Wireless Commun. networks,” IEEE Commun. Mag., vol. 49, no. 8, pp. 63–71, Aug. 2011.
Lett., vol. 3, no. 5, pp. 517–520, Oct. 2014. [71] G. Rittenhouse, S. Goyal, D. T. Neilson, and S. Samuel, “Sustainable
[48] W. Saad, Z. Han, M. Debbah, A. Hjorungnes, and T. Basar, “Coalitional telecommunications,” in Proc. Tech. Symp. ITU Telecom World,
game theory for communication networks,” IEEE Signal Process. Mag., Geneva, Switzerland, 2011, pp. 19–23.
vol. 26, no. 5, pp. 77–97, Sep. 2009. [72] S. Han, C. Yang, G. Wang, and M. Lei, “On the energy efficiency
[49] H. Li, H. Tian, C. Qin, and Y. Pei, “A novel distributed cluster com- of base station sleeping with multicell cooperative transmission,” in
bination method for CoMP in LTE-A system,” in Proc. IEEE 15th Int. Proc. 22nd Int. Symp. Pers. Indoor Mobile Radio Commun. (PIMRC),
Symp. Wireless Pers. Multimedia Commun. (WPMC), Taipei, Taiwan, Toronto, ON, Canada, 2011, pp. 1536–1540.
2012, pp. 614–618. [73] D. Cao, S. Zhou, C. Zhang, and Z. Niu, “Energy saving performance
[50] J.-M. Moon and D.-H. Cho, “Formation of cooperative cluster for coor- comparison of coordinated multi-point transmission and wireless relay-
dinated transmission in multi-cell wireless networks,” in Proc. IEEE ing,” in Proc. Glob. Telecommun. Conf. (GLOBECOM), Miami, FL,
Consum. Commun. Netw. Conf. (CCNC), Las Vegas, NV, USA, 2013, USA, 2010, pp. 1–5.
pp. 528–533. [74] Z. Niu, S. Zhou, Y. Hua, Q. Zhang, and D. Cao, “Energy-aware network
[51] W. Saad, Z. Han, M. Debbah, and A. Hjorungnes, “A distributed planning for wireless cellular system with inter-cell cooperation,” IEEE
coalition formation framework for fair user cooperation in wireless net- Trans. Wireless Commun., vol. 11, no. 4, pp. 1412–1423, Apr. 2012.
works,” IEEE Trans. Wireless Commun., vol. 8, no. 9, pp. 4580–4593, [75] A. Sharma, A. Roy, S. Ghosal, R. Chaki, and U. Bhattacharya, “Load
Sep. 2009. balancing in cellular network: A review,” in Proc. 3rd Int. Conf.
[52] R. Weber, A. Garavaglia, M. Schulist, S. Brueck, and A. Dekorsy, Comput. Commun. Netw. Technol. (ICCCNT), Coimbatore, India, 2012,
“Self-organizing adaptive clustering for cooperative multipoint trans- pp. 1–5.
mission,” in Proc. IEEE 73rd Veh. Technol. Conf. (VTC Spring), [76] I. Viering, M. Dottling, and A. Lobinger, “A mathematical per-
Budapest, Hungary, 2011, pp. 1–5. spective of self-optimizing wireless networks,” in Proc. Int. Conf.
[53] P. Baracca, F. Boccardi, and V. Braun, “A dynamic joint clustering Commun. (ICC), Dresden, Germany, 2009, pp. 1–6.
scheduling algorithm for downlink CoMP systems with limited CSI,” [77] A. Imran, E. Yaacoub, M. A. Imran, and R. Tafazolli, “Distributed load
in Proc. Int. Symp. Wireless Commun. Syst. (ISWCS), Paris, France, balancing through self organisation of cell size in cellular systems,” in
2012, pp. 830–834. Proc. 23rd Int. Symp. Pers. Indoor Mobile Radio Commun. (PIMRC),
[54] V. Garcia, Y. Zhou, and J. Shi, “Coordinated multipoint transmission Sydney, NSW, Australia, 2012, pp. 1114–1119.
in dense cellular networks with user-centric adaptive clustering,” IEEE [78] A. Imran, M. A. Imran, A. ul Quddus, and R. Tafazolli, “Distributed
Trans. Wireless Commun., vol. 13, no. 8, pp. 4297–4308, Aug. 2014. spectral efficiency optimization at hotspots through self organisation of
BS tilts,” in Proc. GLOBECOM Workshops, Houston, TX, USA, 2011,
[55] C. Zeng, P. Ren, C. Zhang, and G. Lv, “A hybrid clustering approach pp. 570–574.
in coordinated multi-point transmission system,” in Proc. 7th Int. ICST [79] H. Kim, G. De Veciana, X. Yang, and M. Venkatachalam, “Alpha-
Conf. Commun. Netw. China (CHINACOM), Kunming, China, 2012, optimal user association and cell load balancing in wireless networks,”
pp. 627–631. in Proc. INFOCOM, San Diego, CA, USA, 2010, pp. 1–5.
[56] C. Li, J. Zhang, M. Haenggi, and K. B. Letaief, “User-centric intercell [80] Q. Ye et al., “User association for load balancing in heterogeneous
interference nulling for downlink small cell networks,” IEEE Trans. cellular networks,” IEEE Trans. Wireless Commun., vol. 12, no. 6,
Commun., vol. 63, no. 4, pp. 1419–1431, Apr. 2015. pp. 2706–2716, Jun. 2013.
[57] J. Zhao, T. Zhang, Z. Zeng, Q. Gao, and S. Sun, “An overlapped [81] I. Gádor. (Jul. 2012). “EARTH Project Deliverable D3.3: Final
clustering scheme of coordinated multi-point transmission for LTE-A Report on Green Network Technologies.” [Online]. Available:
systems,” in Proc. IEEE 14th Int. Conf. Commun. Technol. (ICCT), http://bscw.ict-earth.eu/pub/bscw.cgi/d70472/EARTHWP3D3.3.pdf
Chengdu, China, 2012, pp. 479–484. [82] Y. Li, Y. Ma, Y. Wang, and W. Zhao, “Base station sleeping with
[58] W. Xu, J. Huang, F. Yang, and H. Zhang, “Dynamic cell cluster- dynamical clustering strategy of CoMP in LTE-advanced,” in Proc.
ing for CoMP in LTE-A and its calibrated system level performance IEEE Int. Conf. Green Comput. Commun., Beijing, China, Aug. 2013,
evaluation,” in Proc. IEEE 5th Int. Symp. Microw. Antenna Propag. pp. 157–162.
EMC Technol. Wireless Commun. (MAPE), Chengdu, China, 2013, [83] Z. Mayer, J. Li, A. Papadogiannis, and T. Svensson, “On the impact of
pp. 71–77. backhaul channel reliability on cooperative wireless networks,” in Proc.
[59] M. Liu, B. Li, and Z. Tong, “Cell clustering and resource allocation Int. Conf. Commun. (ICC), Budapest, Hungary, 2013, pp. 5284–5289.
scheme for CoMP SU-MIMO,” Wireless Pers. Commun., vol. 69, no. 4, [84] S. Hämäläinen, H. Sanneck, and C. Sartori, LTE Self-Organising
pp. 1399–1411, Apr. 2013. Networks (SON): Network Management Automation for Operational
[60] E. Katranaras, M. A. Imran, and M. Dianati, “Energy-aware clus- Efficiency. Hoboken, NJ, USA: Wiley, 2012.
tering for multi-cell joint transmission in LTE networks,” in Proc. [85] 3GPP, “Universal Terrestrial Radio Access (UTRA) and Evolved
IEEE Int. Conf. Commun. Workshops (ICC), Budapest, Hungary, 2013, Universal Terrestrial Radio Access (E-UTRA); Radio Measurement
pp. 419–424. Collection for Minimization of Drive Tests (MDT); Overall
Description; Stage 2,” 3rd Gener. Partnership Project (3GPP), TS [112] J. Dean and S. Ghemawat, “MapReduce: Simplified data processing
37.320 R12 v12.2.0, Sep. 2014. on large clusters,” Commun. ACM, vol. 51, no. 1, pp. 107–113, 2008.
[86] J. Zhang, R. Chen, J. G. Andrews, A. Ghosh, and R. W. Heath, Jr., [113] Apache Storm. Accessed on Feb. 8, 2017. [Online]. Available:
“Networked MIMO with clustered linear precoding,” IEEE Trans. https://spark.apache.org
Wireless Commun., vol. 8, no. 4, pp. 1910–1921, Apr. 2009. [114] Apache Storm. Accessed on Feb. 8, 2017. [Online]. Available:
[87] A. Papadogiannis and G. C. Alexandropoulos, “The value of dynamic https://storm.apache.org
clustering of base stations for future wireless networks,” in Proc. Int. [115] P. Hunt, M. Konar, F. P. Junqueira, and B. Reed, “ZooKeeper: Wait-
Conf. Fuzzy Syst. (FUZZ), Barcelona, Spain, 2010, pp. 1–6. free coordination for Internet-scale systems,” in Proc. USENIX Annu.
[88] J. Li, C. Botella, and T. Svensson, “Resource allocation for clustered Tech. Conf., vol. 8. Boston, MA, USA, 2010, p. 11.
network MIMO OFDMA systems,” EURASIP J. Wireless Commun. [116] Q. Lv, Y. Qiao, N. Ansari, J. Liu, and J. Yang, “Big data driven hid-
Netw., vol. 2012, no. 1, pp. 1–19, 2012. den Markov model based individual mobility prediction at points of
[89] 3GPP, “CoMP performance evaluation under low- interest,” IEEE Trans. Veh. Technol., to be published.
capacity/high-latency backhaul,” 3rd Gener. Partnership [117] A. Nika et al., “Understanding and predicting data hotspots in cellular
Project (3GPP), TR R1-112339, Aug. 2011. [Online]. networks,” Mobile Netw. Appl., vol. 21, no. 3, pp. 402–413, 2016.
Available: http://www.3gpp.org/ftp/tsg_ran/WG1_RL1/ [118] R. E. Rezagah et al., “Large scale cooperation in cellular net-
TSGR1_66/Docs/R1-112339.zip works with non-uniform user distribution,” in Proc. Globecom
[90] 3GPP, “Evolved universal terrestrial radio access network (E- Workshops (GC Wkshps), Atlanta, GA, USA, 2013, pp. 43–48.
UTRAN); X2 general aspects and principles,” 3rd Gener. Partnership [119] K. Yang, J. Zhang, X. Zhang, and W. Wang, “Edge aware cross-
Project (3GPP), TS 36.420 R12 v12.1.0, Dec. 2014. [Online]. tier base station cooperation in heterogeneous wireless networks with
Available: http://www.3gpp.org/ftp/Specs/html-info/36420.htm non-uniformly-distributed nodes,” IET Commun., vol. 10, no. 15,
[91] L. Su, C. Yang, and S. Han, “The value of channel prediction in CoMP pp. 1896–1903, Oct. 2016.
systems with large backhaul latency,” IEEE Trans. Commun., vol. 61, [120] M. S. Parwez, H. Farooq, A. Imran, and H. Refai, “Spectral efficiency
no. 11, pp. 4577–4590, Nov. 2013. self-optimization through dynamic user clustering and beam steering,”
[92] E. Kayacan, B. Ulutas, and O. Kaynak, “Grey system theory-based in Proc. IEEE Glob. Commun. Conf. (GLOBECOM), San Diego, CA,
models in time series prediction,” Expert Syst. Appl., vol. 37, no. 2, USA, 2015, pp. 1–7.
pp. 1784–1789, 2010. [121] N. Bitar, A. Imran, and H. Refai, “A user centric self-optimizing grid-
[93] O. Onireti et al., “A cell outage management framework for dense based approach for antenna steering based on call detail records,” in
heterogeneous networks,” IEEE Trans. Veh. Technol., vol. 65, no. 4, Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), Doha, Qatar,
pp. 2097–2113, Apr. 2016. 2016, pp. 1–6.
[122] T. V. Santana, A. Galindo-Serrano, B. Sayrac, and S. M. López,
[94] H. Ge, X. Wen, W. Zheng, Z. Lu, and B. Wang, “A history-based
“Dynamic network configuration: Hotspot identification for virtual
handover prediction for LTE systems,” in Proc. Int. Symp. Comput.
small cells,” in Proc. Int. Symp. Wireless Commun. Syst. (ISWCS),
Netw. Multimedia Technol. (CNMT), Wuhan, China, 2009, pp. 1–4.
Poznań, Poland, 2016, pp. 49–53.
[95] J. Wang and B. Prabhala, “Periodicity based next place prediction,” in [123] A. Alahi, V. Ramanathan, and L. Fei-Fei, “Socially-aware large-
Proc. Mobile Data Challenge Nokia Workshop 10th Int. Conf. Pervasive scale crowd forecasting,” in Proc. IEEE Conf. Comput. Vis. Pattern
Comput., Urbana, IL, USA, 2012. Recognit., Columbus, OH, USA, 2014, pp. 2211–2218.
[96] M. Daoui, A. M’zoughi, M. Lalam, M. Belkadi, and R. Aoudjit, [124] R. O. Sinnott and W. Chen, “Estimating crowd sizes through social
“Mobility prediction based on an ant system,” Comput. commun., media,” in Proc. IEEE Int. Conf. Pervasive Comput. Commun.
vol. 31, no. 14, pp. 3090–3097, 2008. Workshops (PerCom Workshops), Sydney, NSW, Australia, 2016,
[97] A. Ulvan, M. Ulvan, and R. Bestak, “The enhancement of handover pp. 1–6.
strategy by mobility prediction in broadband wireless access,” in Proc. [125] H.-W. Chang, Y.-C. Tai, and J. Y.-J. Hsu, “Context-aware taxi demand
Netw. Electron. Comm. Res. Conf. (NAEC), Dallas, TX, USA, 2009, hotspots prediction,” Int. J. Bus. Intell. Data Min., vol. 5, no. 1,
pp. 266–276. pp. 3–18, 2009.
[98] S. Gambs, M.-O. Killijian, M. N. del, and P. Cortez, “Next place pre- [126] M. Banko and E. Brill, “Scaling to very very large corpora for natural
diction using mobility Markov chains,” in Proc. 1st Workshop Meas. language disambiguation,” in Proc. 39th Annu. Meeting Assoc. Comput.
Privacy Mobility, Bern, Switzerland, 2012, pp. 1–6. Linguist., Toulouse, France, 2001, pp. 26–33.
[99] A. S. Sadiq, K. A. Bakar, K. Z. Ghafoor, and A. J. Gonzalez, “Mobility [127] R. Bar-David and M. Last, “Context-aware location prediction,” in
and signal strength-aware handover decision in mobile IPv6 based Revised Selected Papers From the 5th International Workshop on
wireless LAN,” in Proc. Int. MultiConf. Eng. Comput. Sci., vol. 1. Big Data Analytics in the Social and Ubiquitous Context, vol. 9546.
Hong Kong, 2011. New York, NY, USA: Springer-Verlag, 2016, pp. 165–185.
[100] T.-H. Kim and J.-W. Kim, “Handover optimization with user mobility [128] S. Bassoy, M. J. Jaber, M. A. Imran, and P. Xiao, “Load aware self-
prediction for femtocell-based wireless networks,” Int. J. Eng. Technol., organising user-centric dynamic CoMP clustering for 5G networks,”
vol. 5, no. 2, pp. 1829–1837, 2013. IEEE Access, vol. 4, pp. 2895–2906, 2016.
[101] V. Frias-Martinez, C. Soguero, and E. Frias-Martinez, “Estimation of
urban commuting patterns using cellphone network data,” in Proc.
ACM SIGKDD Int. Workshop Urban Comput., Beijing, China, 2012,
pp. 9–16.
[102] M. Ficek and L. Kencl, “Inter-call mobility model: A spatio-temporal
refinement of call data records using a Gaussian mixture model,” in
Proc. IEEE INFOCOM, Orlando, FL, USA, Mar. 2012, pp. 469–477.
[103] F. Calabrese, M. Colonna, P. Lovisolo, D. Parata, and C. Ratti, “Real-
time urban monitoring using cell phones: A case study in Rome,” IEEE
Trans. Intell. Transp. Syst., vol. 12, no. 1, pp. 141–151, Mar. 2011.
[104] F. Botta, H. S. Moat, and T. Preis, “Quantifying crowd size with mobile
phone and Twitter data,” Roy. Soc. Open Sci., vol. 2, no. 5, 2015,
Art. no. 150162.
[105] T. Qiu et al., “Listen to me if you can: Tracking user experience of Selcuk Bassoy received the B.E degree in electri-
mobile network on social media,” in Proc. 10th ACM SIGCOMM Conf. cal and electronics engineering from Middle East
Internet Meas., Melbourne, VIC, Australia, 2010, pp. 288–293. Technical University, Ankara, Turkey, in 1999. He
[106] M. M. Najafabadi et al., “Deep learning applications and challenges is currently pursuing the Ph.D. degree with the
in big data analytics,” J. Big Data, vol. 2, no. 1, pp. 1–21, 2015. University of Glasgow, researching on 5G radio
[107] Y. Qiao, Y. Cheng, J. Yang, J. Liu, and N. Kato, “A mobility analytical access network coordination and clustering tech-
framework for big mobile data in densely populated area,” IEEE Trans. niques. Since 1999, he has been working in the wire-
Veh. Technol., to be published.
less telecommunications industry, providing consul-
[108] Apache Storm. Accessed on Feb. 8, 2017. [Online]. Available:
https://flume.apache.org tancy on radio access network design, planning, and
[109] Apache Storm. Accessed on Feb. 8, 2017. [Online]. Available: optimization for global telecommunications com-
https://kafka.apache.org panies like Ericsson, T-Mobile, Telefonica, and
[110] K. Shvachko, H. Kuang, S. Radia, and R. Chansler, “The hadoop Vodafone. He is currently with Telefonica U.K. Ltd., as a Principal Engineer,
distributed file system,” in Proc. IEEE 26th Symp. Mass Stor. Syst. specialized in self-organizing networks and other smart automation tools
Technol. (MSST), 2010, pp. 1–10. for GSM/UMTS/LTE radio access networks. His research interests include
[111] L. George, HBase: The Definitive Guide. Sebastopol, CA, USA: wireless communications, network MIMO, self-optimization, coordinated
O’Reilly Media, 2011. multi-point, game theory, radio resource management, and load balancing.
Hasan Farooq received the B.Sc. degree in electri- Ali Imran (M’15) received the B.Sc. degree
cal engineering from the University of Engineering in electrical engineering from the University of
and Technology, Lahore, Pakistan, in 2009, and Engineering and Technology, Lahore, Pakistan,
the M.Sc. degree in information technology from in 2005, the M.Sc. (with Distinction) degree in
Universiti Teknologi Petronas, Malaysia, in 2014, mobile and satellite communications, and the Ph.D.
wherein his research focused on developing ad-hoc degree from the University of Surrey, Guildford,
routing protocols for smart grids. He is currently pur- U.K., in 2007 and 2011, respectively. He is an
suing the Ph.D. degree in electrical and computer Assistant Professor of Telecommunications with
engineering with the University of Oklahoma, USA. the University of Oklahoma, USA, where he is
His research area is big data empowered proac- the Founding Director of Big Data and Artificial
tive self-organizing cellular networks focusing on Intelligence Enabled Self Organizing Research
intelligent proactive self-optimization and self-healing in HetNets utilizing Center. He has been leading several multinational projects on Self-Organizing
dexterous combination of machine learning tools, classical optimization tech- Cellular Networks such as QSON, for which he has secured research grants of
niques, stochastic analysis, and data analytics. He has been involved in over $2 million in last four years, as a Lead Principal Investigator. He is cur-
multinational QSON project on self-organizing cellular networks and is currently leading two NSF funded projects on 5G SON amounting to over $750K.
rently contributing to two NSF funded projects on 5G SON. He was a recipient He has authored over 60 peer-reviewed articles and presented a number of
of Internet Society First Time Fellowship Award toward Internet Engineering tutorials at international forums, such as the IEEE International Conference
Task Force 86th Meeting, USA, 2013. on Communications, the IEEE Wireless Communications and Networking
Conference, the European Wireless Conference, and the International
Conference on Cognitive Radio Oriented Wireless Networks, on his topics of
interest. His research interests include self-organizing networks, radio resource
management, and big data analytics. He is an Associate Fellow of the Higher
Muhammad A. Imran (M’03–SM’12) received Education Academy, U.K. and a member of the Advisory Board to the Special
the M.Sc. (Distinction) and Ph.D. degrees from Technical Community on Big Data of the IEEE Computer Society.
Imperial College London, U.K., in 2002 and 2007,
respectively. He is a Professor of Communication
Systems, University of Glasgow, the Vice Dean of
Glasgow College UESTC, and the Program Director
of Electrical and Electronics with Communications.
He is an Adjunct Associate Professor with the
University of Oklahoma, USA, and a Visiting
Professor with the 5G Innovation Centre, University
of Surrey, U.K., where he has worked previously
from 2007 to 2016. He has led a number of multimillion-funded international
research projects encompassing the areas of energy efficiency, fundamental
performance limits, sensor networks, and self-organizing cellular networks.
In addition to significant funding from EPSRC, RCUK, Qatar NRF, and EU
FP7/H2020, he has received direct industrial funding from leading industries
in communications. He also led the new physical layer work area for 5G
Innovation Centre at Surrey. He has a global collaborative research network
spanning both academia and key industrial players in the field of wireless
communications. He has supervised over 25 successful Ph.D. graduates and
published over 300 peer-reviewed research papers. He was a recipient of
the Award of excellence in recognition of his academic achievements con-
ferred by the President of Pakistan, the IEEE Comsoc’s Fred Ellersick Award
2014, the Sentinel of Science Award 2016, the FEPS Learning and Teaching
Award 2014, twice nominated for Tony Jean’s Inspirational Teaching Award,
shortlisted finalist for the Wharton-QS Stars Awards 2014, the Reimagine
Education Awards 2016 for innovative teaching, and the VC’s Learning and
Teaching Award in University of Surrey. He is a senior fellow of Higher
Education Academy, U.K.

Coordinated Multi-Point Clustering Schemes A Survey

Uploaded by

Copyright:

Available Formats

Coordinated Multi-Point Clustering Schemes A Survey

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Coordinated Multi-Point Clustering Schemes A Survey

Uploaded by

Copyright:

Available Formats

IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 19, NO.

2, SECOND QUARTER 2017 743

Coordinated Multi-Point Clustering

Abstract—Mobile data traffic grew by 74% in 2015 and it

non-orthogonal multiple access (NOMA) scheme with CoMP

network objectives will need to be considered for BS

proactive CoMP clustering decisions. SON algorithms can be

consists of cell-3 only and there is interference from cell-4

spectral efficiency gain. Dynamic network-centric clustering

or hotspots wherein optimal large cooperative clusters, per-

You might also like