User-centric C-RAN Architecture for Ultra-dense 5G Networks: Challenges and Methodologies
Cunhua Pan, Maged Elkashlan, Jiangzhou Wang, Jinhong Yuan, Lajos Hanzo
11 User-centric C-RAN Architecture forUltra-dense 5G Networks: Challenges andMethodologies
Cunhua Pan,
Member, IEEE , Maged Elkashlan,
Member, IEEE , Jiangzhou Wang,
Fellow,IEEE , Jinhong Yuan,
Fellow, IEEE , and Lajos Hanzo,
Fellow, IEEE
Abstract
Ultra-dense networks (UDN) constitute one of the most promising techniques of supporting the5G mobile system. By deploying more small cells in a fixed area, the average distance between usersand access points can be significantly reduced, hence a dense spatial frequency reuse can be exploited.However, severe interference is the major obstacle in UDN. Most of the contributions deal with theinterference by relying on cooperative game theory. This paper advocates the application of dense user-centric C-RAN philosophy to UDN, thanks to the recent development of cloud computing techniques.Under dense C-RAN, centralized signal processing can be invoked for supporting CoMP transmission. Wesummarize the main challenges in dense user-centric C-RANs. One of the most challenging issues is therequirement of the global CSI for the sake of cooperative transmission. We investigate this requirementby only relying on partial CSI, namely, on inter-cluster large-scale CSI. Furthermore, the estimation ofthe intra-cluster CSI is considered, including the pilot allocation and robust transmission. Finally, wehighlight several promising research directions to make the dense user-centric C-RAN become a reality,with special emphasis on the application of the ‘big data’ techniques.
I. I
NTRODUCTION
The fifth generation (5G) wireless system is anticipated to offer a substantially increased data throughputcompared to the fourth generation (4G) system. To achieve this ambitious goal, ultra dense networks(UDN) have been widely regarded as one of the most promising solutions [1]. In UDN, the average
Cunhua Pan and Maged Elkashlan are with the Queen Mary University of London, London E1 4NS, U.K. (Email: { c.pan,maged.elkashlan } @qmul.ac.uk).Jiangzhou Wang is with the School of Engineering and Digital Arts, University of Kent, Canterbury, Kent, CT2 7NZ, U.K.(e-mail: [email protected]).Jinhong Yuan is with the University of New South Wales, Sydney, NSW 2052, Australia. (e-mail:[email protected]).Lajos Hanzo is with the School of Electronics and Computer Science, University of Southampton, Southampton, SO17 1BJ,U.K. (e-mail:[email protected]). a r X i v : . [ ee ss . SP ] S e p Fronthaul LinksBBU Pool BBU BBU BBU
UE 1RRH 1RRH 2 UE 2UE 3UE 4 UE 5UE 6RRH 3 RRH 4 RRH 5RRH 6RRH 7RRH 8 RRH 9RRH 10RRH 11RRH 12 RRH 13
BBU
UE 7 UE 8 UE 9 (a) Disjoint cluster
BBU Pool BBU BBU BBU
UE 1RRH 1RRH 2 UE 2 UE 3UE 4 UE 5UE 6RRH 3 RRH 4 RRH 5RRH 6RRH 7RRH 8 RRH 9RRH 10RRH 11RRH 12 RRH 13
BBU (b) User-centric cluster
Fronthaul Links
Cluster 1Cluster 2 Cluster 3 Cluster 4
Fig. 1. Illustration of a C-RAN architecture: (a) Disjoint cluster, where the whole network is divided into several non-overlappedclusters, and the cluster-edge users still suffer from high inter-cluster interference; (b) User-centric cluster, where cluster is formedfrom the user side and the cluster-edge issues are eliminated. distance between users and small cell BSs is significantly reduced, hence the link quality is dramaticallyimproved, which further increases the network capacity.However, the drastic interference generated by the neighboring small cells is a limiting factor in UDN.The attainable network performance may even be decreased when the BS density is extremely high [2].Hence, the interference should be carefully managed in order to reap the potential benefits of UDN. Mostof the existing contributions deal with the interference by designing partially distributed algorithms basedon game-theoretical approaches. By adopting cooperative game theory, multiple small cell BSs exchangethe necessary information for their coordination through the wired backhaul (BH) links, which workswell for small-scale networks. However, for UDN, the heavy overhead of coordination and the increasingcost of deploying the wired BH links will preclude the application of this approach. Apart from theinterference issues, employing more small cell BSs will also increase the maintenance and operationalcosts. Hence, a new network architecture should be adopted to support reliable communications in UDN.Due to the recent developments in cloud computing, the ultra-dense cloud radio access network (C-RAN) concept has been regarded as a promising network architecture that can efficiently address theissues arising in UDN.The ultra-dense C-RAN architecture is shown in Fig. 1, which consists of three key components:
TABLE ID
IFFERENT T YPES OF
UDN D
EPLOYMENT
Type of UDN Functionality Interference Management Connectivity Deployment Cost
Small cells Fully functioning Distributed algorithm Wired backhaul High maintenance costDense C-RAN PHY layer functioning Centralized algorithm Wireless fronthaul Low hardware cost
1) Baseband unit (BBU) pool supported by the techniques of cloud computing, network functionvirtualization (NFV), software-defined networks (SDN), and so on.2) Low-cost, low power radio remote heads (RRHs) distributed over the coverage area.3) Wireless fronthaul links that connect the RRHs to the BBU pool.The main characteristic of dense C-RAN is that all the baseband signal processing units of conventionalsmall cell BSs have been incorporated in the BBU pool, where the computing resources can be sharedamong the BBUs. Then conventional full-functionality small cell BSs can be replaced by the low-cost,low-complexity RRHs, which are only responsible for the transmission/reception. Hence, RRHs can bedensely deployed at a low hardware cost to provide ultra-high throughput and seamless coverage for a largenumber of users in tele-traffic hot spots, such as airports and shopping malls. Thanks to the centralizedarchitecture of C-RAN, the global network information can be shared in the BBU pool and cooperativecommunication techniques can be realized with the aid of powerful cloud computing, such as large-scalenetwork coordination, global resource management, coordinated multi-point (CoMP) processing, etc.Hence, the cochannel interference that is a limiting factor in UDN can be efficiently mitigated under theultra-dense C-RAN architecture. Additionally, the operating status of RRHs and the computing resourcesof the BBU pool can be dynamically controlled in order to adapt to the capacity demand fluctuations ofthe users, which leads to significant energy and operational cost reductions. In [3], Tang et al. showed thatthe system cost, which consists of computation cost and the wireless transmission cost (including bothtransmit power and circuit power), can be reduced by roughly twelve percent through jointly optimizingthe number of active virtual machines (VMs) and operating status of RRHs and fronthaul links. Table Icontrasts the differences between conventional small cell network and the ultra-dense C-RAN.However, there are some challenges that need to be handled in dense C-RAN before its successfulpractical deployment. The contributions of this paper are summarized as follows:1) We first summarize the unique research challenges in dense C-RAN and the existing solutions foreach challenge. We place an emphasis on the challenge of heavy training overhead for estimatingall the channel state information (CSI) for CoMP transmission. Most of the existing works consider perfect intra-cluster CSI, which is impractical due to the limited amount of pilot resources.2) We provide a complete framework to deal with the imperfect intra-cluster CSI. One novel low-complexity pilot allocation algorithm is first proposed by considering the multi-user pilot inter-ference. Then, robust beamforming vector design is considered to account for the CSI estimationerror.3) Some promising research directions are highlighted, especially the cluster design by the aid of ‘BigData’ technique.The rest of this paper is organized as follows: in Section II, the new challenges for dense C-RAN areprovided; Section III provides a complete procedure to deal with the imperfect CSI, including both thepilot allocation and robust beamforming design. Simulation results are also shown in Section III. Finally,conclusions are drawn in Section IV along with some future research directions.II. R
ESEARCH C HALLENGES AND S TATE OF THE A RT S OLUTIONS
In this section, we summarize the challenges arising in dense C-RAN along with the state of the artsolutions.
High Computational Complexity : In dense C-RAN, the BBU pool usually supports a large number ofRRHs and the number of variables to optimize, such as beamforming-vectors will become excessive, evenin the context of cloud computing. The most common technique of reducing the complexity is to adoptthe cluster technique. In general, there are two types of cluster techniques, as shown in Fig. 1: Disjointclustering and user-centric clustering. In the disjoint cluster, all the RRHs in the network are partitionedinto several non-overlapped clusters, and the RRHs in each cluster employ the CoMP technique to servethe users within the coverage area of this cluster. However, the cluster-edge users still suffer from severeinter-cluster interference. By contrast, in the user-centric cluster, each user is individually served by itsnearby RRHs. The scheduled user is the center of the cluster. Different clusters may overlap with eachother, which will eliminate the potential cluster-edge effect. Hence, the user-centric cluster constitutesthe focus of this paper.
Stringent Fronthaul Capacity Requirement : In conventional C-RAN, the fronthaul links are typicallywired links, such as optical fibers. However, in ultra-dense C-RAN a large number of fronthaul links arerequired. Laying wired links requires high operational and maintenance costs. An attractive alternativeis to use wireless fronthaul links, such as millimeter wave (mmWave) Communications, which are muchmore scalable and cost-effective than wired links. However, the bandwidth of wireless links is muchlower than that of wired links, which means that the number of users supported by each wireless link ismuch lower. The fronthaul capacity constraint has been extensively studied, which can be divided into two categories: the compression strategy and the data sharing strategy. In the compression strategy, theBBU pool first computes the beamforming-vectors for each RRH. Then, the beamformed signals aregenerated at the BBU pool, which are compressed and sent to the corresponding RRHs. The fronthaulcapacity is related to how fine is the resolution of the compressed signals: higher resolutions require ahigher fronthaul capacity. Hence, the compression resolution should be optimized under the fronthaulcapacity constraints. In [4], Park et al. proposed to leverage joint compression of the signals of differentRRHs to better optimize the compression resolutions. In the second strategy, the beamforming-vectorscomputed at the BBU pool are directly sent to the corresponding RRHs. Then, the BBU pool shareseach user’s data directly with its serving cluster. The beamformed signals are generated at each RRH.In this strategy, the fronthaul capacity depends on the number of users served by each link. Hence, theuser-RRH associations should be optimized under the fronthaul capacity constraints. In [5], Dai et al. considered the network utility maximization problem for the user-centric downlink C-RAN, where theper-RRH fronthaul capacity constraints are explicitly taken into account.
Huge Training Overhead for CSI Estimation : To facilitate CoMP transmission, the global CSI shouldbe available at the BBU pool, which constitutes an excessive channel estimation overhead for dense C-RAN. Caire et al. [6] showed that the increasing amount of training overhead may outweigh the gainsprovided by CoMP. To reduce the overhead, a promising technique is to rely on partial CSI case underthe user-centric cluster, where each user only estimates the CSI of the links from the RRHs within itscluster (termed as intra-cluster CSI) and only tracks the path loss and shadowing outside its own cluster(termed as inter-cluster CSI), which are sent back to the BBU pool. These parameters are necessaryfor the CoMP design at the BBU pool. Indeed, the large-scale fading may be readily tracked since itchanges slowly compared to the instantaneous CSI. The design of beamforming vectors under this partialCSI case is a challenging task. Hence, there is a paucity of contributions [7]- [8] based on partial CSI.To elaborate, compressive CSI acquisition was proposed in [7] for determining the set of instantaneousCSIs and large scale fading gains. However, its complexity is high, hence cannot be readily implementedin dense C-RAN. Recently, we provided a design framework in [8] to design green transmission underpartial CSI case. However, the intra-cluster CSI was assumed to be perfect in [7] and [8], which isdifficult to achieve in practice. Hence, in the paper, we aim to consider the imperfect intra-cluster CSIcase, where both the channel estimation procedure for intra-cluster CSI and robust beamforming vectordesign are considered in the following two sections.
III. T
RANSMISSION S CHEME D ESIGNED FOR I MPERFECT I NTRA -C LUSTER
CSIFor TDD C-RANs, the training pilots sent from different users to the same RRH should be mutuallyorthogonal so that the RRH can distinguish the channel vectors from different users. The number of pilotsrequired scales linearly with the number of users, which becomes excessive for dense C-RANs. Hence,the number of time slots used for data transmission will be significantly reduced. To reduce the numberof pilots, they may be reused by a group of users under the condition that none of users in the groupshares the same RRH with the other users. It is widely recognized that the pilot reuse scheme will imposethe pilot contamination issue, which results in sizeable channel estimation errors. In the following, wepropose a two-stage optimization method to optimize the transmissions for dense C-RAN: In Stage I, anovel pilot reuse scheme is proposed; In Stage II, a robust beamforming-vector optimization algorithmis conceived by considering the pilot contamination incurred by Stage I.
A. Stage I: Novel Pilot Reuse Scheme
The pilot reuse issues have been extensively studied in massive multiple-input multiple-output (MIMO)systems [9]. The basic idea is to reuse the same pilot within the specific group of users having differentangles of arrival, which is not applicable in dense C-RAN due to limited number of antennas at eachRRH. Recently, Chen et al. [10] proposed a novel pilot allocation scheme for dense C-RANs by usingthe classic Dsatur graph coloring algorithm, which minimizes the number of pilots required for a givenset of users. However, for simplicity of implementation, the 4G LTE system suggests that the proportionof pilots designated for channel estimation should be fixed within the channel’s coherence time, i.e. for a 10 ms training period [11]. Given the fixed number of pilots, some users may not be allocated anypilots. Hence, we aim for providing a joint user selection and pilot allocation scheme for maximizingthe number of users admitted at a given number of available pilots, while satisfying the following twoconditions:1) The users sharing the same RRH must not reuse the same pilot.2) There is an upper-bound on how many time each pilot is reused, denoted as n max .The second constraint is imposed for guaranteeing a fair use of the available pilots. In the following, alow-complexity nearly-optimal algorithm is provided to deal with this problem.For a dense C-RAN with K users, we construct a ( K × K ) matrix B , where each element is given by b k,k (cid:48) = , if I k ∩ I k (cid:48) (cid:54) = ∅ and k (cid:54) = k (cid:48) , otherwise , (1)where I k denotes the set of RRHs that potentially serve user k . The above definitions mean that if user k and user k (cid:48) share common RRHs, the corresponding element is set to b k,k (cid:48) = 1 . Otherwise, the element UE 1 UE 2 UE 3UE 4 UE 5UE 6 UE 1 UE 2 UE 3UE 4 UE 5UE 6 (a) (b)
UE 1 UE 2 UE 3UE 6 (c) (d)
UE 1 UE 2 UE 3UE 4 UE 5UE 6
Fig. 2. (a) Construction of the undirected graph for the network in Fig. 1-(b); (b) The colored graph after applying the Dsaturalgorithm [10] with n max = 2 , the minimum number of required pilots is n ∗ = 3 ; (c) The user selection and pilot allocationresult after using the algorithm for Case I when τ = 2 and n max = 2 ; (d) The pilot reallocation result after using the bisectionsearch algorithm for Case II. In this pilot reallocation result, user 1 and use 2 are allocated with different pilots to reduce thepilot interference, and the same holds for user 5 and user 6, or UE 3 and UE 4. is zero. Based on matrix B , a unidirectional graph can be constructed for the network of Fig. 1-(b) inFig. 2-(a). Fig. 2-(b) illustrates the pilot allocation results after applying the Dsatur algorithm [10], whichshows that at least three different pilots are required.To solve the above pilot allocation problem, we first adopt the Dsatur algorithm to find the minimumnumber of pilots required. If this number is higher than the number of pilots available, some users shouldbe removed. Otherwise, all users can be admitted. In the latter case, some pilots may be reused by up to n max users, while some pilots are not allocated, which wastes the pilot resources. To resolve this problem,all available pilots should be used to reduce the pilot contamination. Let us denote the minimum numberof pilots required by the Dsatur algorithm as n ∗ . In the following, we provide a detailed algorithm todeal with each case: 1) n ∗ > τ ; 2) n ∗ < τ . When n ∗ = τ , the final solution is obtained. Case I: n ∗ > τ . In this case, some users should be removed. These users can be rescheduled fortransmission in other time slots or frequency bands, which is beyond the scope of this paper. Let usdefine ¯ U as the set of all users, and θ k ∆ = (cid:80) k (cid:48) (cid:54) = k,k (cid:48) ∈ ¯ U b k,k (cid:48) as the degree of the vertex for user k thatrepresents the total number users connected to this user in the constructed undirect graph. The userhaving the largest value of θ k should be deleted with high priority, since many users should be allocateddifferent pilots compared to this user. However, there may exist several users with the same value of θ k ,and randomly removing one of them will lead to a reduced performance. Hence, the user earmarked fordeletion should be carefully selected. Intuitively, the user suffering from the highest pilot interference should be removed. To this end, we define the metric η k,k (cid:48) to quantify the level of pilot interferencebetween any two users when reusing the same pilot: η k,k (cid:48) = log (cid:18) (cid:80) i ∈I k (cid:48) α i,k (cid:80) i ∈I k α i,k (cid:19) + log (cid:32) (cid:80) i ∈I k α i,k (cid:48) (cid:80) i ∈I k (cid:48) α i,k (cid:48) (cid:33) , (2)where α i,k represents the large-scale fading power from RRH i to user k . Obviously, a larger η k,k (cid:48) meanslarger pilot interference between these two users. Then, we quantify the level of pilot interference dueto user k by ξ k = (cid:80) k (cid:48) ∈K πk \{ k } η k,k (cid:48) as our figure of merit, where π k denotes the pilot index used byuser k and K π k represents the set of users that reuses pilot π k . The user with the largest ξ k should beremoved. Hence, we conceive the user selection and pilot allocation method formulated in Algorithm 1.By invoking it for the network in Fig. 1-(b), the final user selection result is shown in Fig. 2-(c). Algorithm 1
User Selection and Pilot Allocation algorithm for Case I Initialize matrix B , user set U , initial number of required pilots n ∗ obtained from the Dsatur algorithm; While n ∗ > τ Find k ∗ = arg max k ∈U θ k . If there are multiple users with the same θ k , remove the user withthe largest ξ k ; Remove user k ∗ from U , i.e., U = U /k ∗ , and update matrix B with current U ; Use the Dsatur algorithm to calculate n ∗ with B and U ; Case II: n ∗ < τ . In this case, we aim for reallocating all the available pilots to all users to additionallyreduce the pilot interference. Note that in Fig. 2-(b), there may be measurable pilot interference betweenuser 1 and user 2. If there are four pilots, these two users can be allocated different pilots as seen inFig. 2-(d). We can reconstruct an undirected graph by introducing a threshold η th . If η k,k (cid:48) > η th , user k and user k (cid:48) should be connected. Based on this idea, matrix B can be reconstructed as b k,k (cid:48) = , if I k ∩ I k (cid:48) (cid:54) = ∅ and k (cid:54) = k (cid:48) , , if η k,k (cid:48) > η th , I k ∩ I k (cid:48) = ∅ and k (cid:54) = k (cid:48) , , otherwise . (3)As expected, a smaller value of η th will result in more users becoming connected with each other andhence more pilots are required. In the extreme case of η th < min { η k,k (cid:48) } , all users become connected andthe number of pilots required is equal to K . On the other hand, if η th ≥ max { η k,k (cid:48) } , the reconstructedmatrix B reduces to the initial B defined in (1), where the number of pilots required is n ∗ . Since τ < K ,there must exist a η th value so that the number of pilots required is equal to τ . The bisection search algorithm can be adopted to find this threshold η th . Again, Fig. 2-(d) shows the pilot allocation resultsafter using this algorithm. B. Stage II: Robust Beamforming-vector Design
In Stage II, we aim for designing the beamforming-vectors by considering the pilot contaminationdue to the pilot reuse scheme in Stage I. Specifically, we formulate a user selection problem for thedense C-RAN where each RRH is equipped with multiple antennas. The following three constraints areconsidered:1) Each user’s data rate should be higher than its minimum requirement;2) Each fronthaul link capacity constraint is imposed;3) Each RRH has its individual power constraint.There are three challenges to solve this optimization problem:1) Since we consider the partial CSI case where only the inter-cluster large-scale fading parametersare available, the exact data rate of each user is difficult to obtain.2) The fronthaul capacity constraint has the non-smooth and non-differentiable indicator function,which is a mixed-integer non-linear programming (MINLP) problem that is NP-hard to solve [8].3) Due to the channel estimation error, each user suffers from residual self-interference. Conventionalweighted minimum mean square error (WMMSE) method of [12] that has been successfully appliedin the perfect intra-cluster CSI scenario [8] and [5], cannot be used for solving the problem here.We provide brief descriptions of the associated methods to address the above three challenges.First, Jensen’s inequality is used for finding the lower-bound of the exact data rate, which is moreamenable to the design of our algorithm. In [8], we have shown that for the specific scenario of non-overlapped cluster, the gap between the lower-bound and the exact data rate is within three percent forboth sparse and dense C-RAN scenarios. Hence, a simple correction factor may be used for practicalapplications.Second, the indicator function is replaced by a concave function f θ ( x ) = xx + θ , where θ is a smallpositive value. The transformed problem is the difference of convex (d.c.) program, which can beefficiently solved by the successive convex approximation (SCA) method [8].Finally, to deal with the final challenge, we adopt the semi-definite relaxation approach and formallyprove that semidefinite relaxation is tight with a probability of 1. L L Number of admitted UEs in Stage I
C a n d i d a t e s i z e t = 4 , P r o p o s e d - S t a g e I t = 4 , N o C a s e I I - S t a g e I t = 4 , C o n - S t a g e I t = 8 , P r o p o s e d - S t a g e I t = 8 , N o C a s e I I - S t a g e I t = 8 , C o n - S t a g e I C a n d i d a t e s i z e
Fig. 3. Number of admitted UEs in Stage I vs the candidate size L . The left subplot corresponds to the case when the numberof available pilots is τ = 4 while the right one is τ = 8 . C. Simulation Results
We now present our simulation results for evaluating the performance of the proposed algorithms. Thedense C-RAN is assumed to cover a square shaped area of 700 m ×
700 m. The numbers of RRHs andusers are set to I = 36 and K = 24 with the densities of 73 RRHs/km and 49 users/km , respectively.This is a typical 5G ultra-dense cellular networks as stated in [13]. Both the users and RRHs are uniformlyand independently distributed in this area. It is assumed that each user is potentially served by its nearest L RRHs. Each fronthaul link is assumed to only support three users, since mmWave communication isemployed as the wireless fronthaul link. The maximum transmit power for each RRH is 100 mW, andthe pilot power is 200 mW. The maximum pilot reuse time for each pliot is n max = 4 , and the raterequirement for each user is 4 bit/s/Hz. We compare our proposed algorithm to the following algorithms:1) Orthogonal pilot allocation (with legend “Ortho”): As the terminology implies, all users are allocatedorthogonal pilot sequences, hence the maximum number of users that can be admitted in Stage Iis equal to the number of available pilots τ . These τ users are randomly selected from K users.2) No reallocation operations for Case II in Stage I (with legend “NoCaseII”): This algorithm issimilar to our proposed algorithm, except when Case II would occur, no additional operations areperformed.3) Conventional pilot allocation method (with legend “Con”): This algorithm is similar to the “NoCa-seII” algorithm, except when Case I would occur, the users are randomly removed until the numberof required pilots is equal to τ .4) Perfect CSI estimation (with legend “Perfect”): This is the baseline algorithm, where the intra-cluster CSI is assumed to be perfectly known. L L Number of admitted UEs in Stage II
C a n d i d a t e s i z e t = 4 , P r o p o s e d - S t a g e I I t = 4 , N o C a s e I I - S t a g e I I t = 4 , C o n - S t a g e I I t = 4 , O r t h o - S t a g e I I t = 4 , P e r f e c t - S t a g e I I t = 8 , P r o p o s e d - S t a g e I I t = 8 , N o C a s e I I - S t a g e I I t = 8 , C o n - S t a g e I I t = 8 , O r t h o - S t a g e I I t = 8 , P e r f e c t - S t a g e I I C a n d i d a t e s i z e
Fig. 4. Number of admitted UEs in Stage II vs the candidate size L . The left subplot corresponds to the case when the numberof available pilots is τ = 4 while the right one is τ = 8 . Figures. 3 and 4 illustrate the number of users admitted vs the candidate RRH set size L in Stage Iand Stage II of Section III-A and III-B, respectively. It is seen from Fig. 3 that the number of admittedusers monotonically decreases upon increasing the candidate set size L . This is due to the fact that witha larger candidate RRH set size for each user, more users will become connected with each other, whichrequires more pilots. This figure also illustrates the superior performance of our proposed algorithmover the “Con” algorithm, highlighting the necessity of carefully considering the pilot interference, whenremoving users.It is interesting to observe from Fig. 4 that the number of users admitted by all algorithms (exceptthe “Ortho” algorithm) initially increases with the candidate set size and then decreases. The reason isthat when L starts to increase, a higher spatial degree of freedom is available to support more users.However, when L continues to increase, many users are rejected in Stage I, as seen in Fig. 3. This trendis different from the widely accepted concept that increasing the candidate set size will always lead tobetter performance. Hence, the channel estimation process should be taken into account, when designingthe cluster. This figure also shows the performance advantage of our proposed algorithms over the otheralgorithms. IV. C ONCLUSIONS AND F UTURE R ESEARCH C HALLENGES
We advocated the application of a user-centric dense C-RAN architecture for UDN due to its appealingfeatures such as the facilitation of centralized signal processing, low hardware cost, etc. However, we alsoidentified the challenge of requiring a heavy training overhead for estimating all the CSIs for cooperativetransmission. As a remedy, we adopted the partial CSI model, where only the large-scale inter-clusterCSI is available. The channel estimation required for intra-cluster CSI was also considered, where a novel pilot allocation scheme was proposed. Then, we developed a robust transmission design by consideringthe effect of channel estimation errors. Our simulation results verified the performance advantages of theproposed algorithm over the existing ones.Finally, we now highlight several promising research directions to make the user-centric dense C-RANmore amenable for practical implementations. Dynamic Cluster Formations : In this work, the cluster for each user was assumed to be fixed, i.e. eachuser is only connected to its nearest L RRHs. However, in practical systems, the cluster sizes for usersshould be adapted to the network state, such as the users’ rate requirements or traffic load. Additionally,the channel estimation stage should be taken into account as seen in Fig. 4, where the network performancemay even degrade with the cluster size. How to optimize the cluster size individually for each user byjointly considering the above elements remains an inspiring research direction.
User Mobility Management : User mobility is a very challenging issue in user-centric ultra-dense C-RANs. When the users move from one place to another, the cluster of RRHs assigned for serving this usershould be adaptively changed. Explicitly, an adaptive mobility management method should be developedso that the serving cluster can follow each user’s behavior (e.g. mobility and service demands) andprovide data transmission without the users’ involvement. Fortunately, the ‘Big Data’ technique relyingon machine learning is becoming mature, which can track the users’ mobility and then predict their futurelocations. By applying this, the users’ serving cluster can be formed beforehand that significantly reducesthe processing time and meets the targeted quality of experience (QoE) levels.ACKNOWLEDGMENTThis work was supported by the UK Engineering and Physical Sciences Research Council under GrantEP/N029666/1. R
EFERENCES [1] J. Andrews, S. Buzzi, W. Choi, S. Hanly, A. Lozano, A. Soong, and J. Zhang, “What will 5G be?”
IEEE JSAC , vol. 32,no. 6, pp. 1065–1082, June 2014.[2] M. Ding, P. Wang, D. Lpez-Prez, G. Mao, and Z. Lin, “Performance impact of LoS and NLoS transmissions in densecellular networks,”
IEEE Trans. Wireless Commun. , vol. 15, no. 3, pp. 2365–2380, Mar. 2016.[3] J. Tang, W. P. Tay, T. Q. S. Quek, and B. Liang, “System cost minimization in cloud RAN with limited fronthaul capacity,”
IEEE Trans. Wireless Commun. , vol. 16, no. 5, pp. 3371–3384, May 2017.[4] S. H. Park, O. Simeone, O. Sahin, and S. Shamai, “Joint precoding and multivariate backhaul compression for the downlinkof cloud radio access networks,”
IEEE Trans. Signal Process. , vol. 61, no. 22, pp. 5646–5658, Nov. 2013.[5] B. Dai and W. Yu, “Sparse beamforming and user-centric clustering for downlink cloud radio access network,”
IEEEAccess , vol. 2, pp. 1326–1339, Oct. 2014. [6] G. Caire, S. A. Ramprashad, and H. C. Papadopoulos, “Rethinking network MIMO: Cost of CSIT, performance analysis,and architecture comparisons,” in Information Theory and Applications Workshop (ITA), 2010 , 2010, pp. 1–10.[7] Y. Shi, J. Zhang, and K. B. Letaief, “CSI overhead reduction with stochastic beamforming for cloud radio access networks,”in , 2014, pp. 5154–5159.[8] C. Pan, H. Zhu, N. J. Gomes, and J. Wang, “Joint user selection and energy minimization for ultra-dense multi-channelC-RAN with incomplete CSI,”
IEEE JSAC , vol. 35, no. 8, pp. 1809–1824, Aug. 2017.[9] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, “A coordinated approach to channel estimation in large-scale multiple-antennasystems,”
IEEE JSAC , vol. 31, no. 2, pp. 264–273, Feb. 2013.[10] Z. Chen, X. Hou, and C. Yang, “Training resource allocation for user-centric base station cooperation networks,”
IEEETrans. Veh. Technol. , vol. 65, no. 4, pp. 2729–2735, Apr. 2016.[11] G. R1-074068, “Summary of reflector discussions on E-UTRA UL RS,” , Oct. 2007.[12] Q. Shi, M. Razaviyayn, Z.-Q. Luo, and C. He, “An iteratively weighted MMSE approach to distributed sum-utilitymaximization for a MIMO interfering broadcast channel,”
IEEE Trans. Signal Process. , vol. 59, no. 9, pp. 4331–4340,Sept. 2011.[13] X. Ge, S. Tu, G. Mao, C. X. Wang, and T. Han, “5G ultra-dense cellular networks,”