Anda di halaman 1dari 5

Distributed Precoding with Local Power Negotiation for Coordinated Multi-Point Transmission

Dennis Hui
Ericsson Research 200 Holger Way, San Jose, California 95134 Email: dennis.hui@ericsson.com

AbstractCoordinated multi-point (CoMP) transmission has recently received much attention due to its promise in effectively mitigating the inter-cell interference in a cellular network. In this paper, we compare two different architectures for CoMP transmission. The rst architecture partitions all basestations in the network into disjoint clusters with signal transmissions within each cluster being coordinated independently of one another at a centralized location of each cluster. In the second architecture, the processing for coordinated transmission is distributed among overlapping clusters, each centered around one cell. For this architecture, we propose a simple method for computing linear multi-user precoders for all clusters in a distributed manner, where the precoders are chosen not only to maximize throughput to the served users but also to reduce interference with other users in overlapping clusters. A low-complexity iterative algorithm is proposed for negotiating transmit power levels among adjacent clusters to ensure the power constraints at all transmission points are satised. Through system simulations using practical channel models, we show that the distributed architecture substantially outperforms the cluster-wise centralized architecture in maximizing the user data throughputs in the network.

neighboring CoMP clusters. This has the potential of creating a large service quality disparity among users within the network. In order to reduce the fraction of users being adversely affected by uncoordinated inter-cluster interference, the number of cells included in each CoMP cluster is preferred to be large, posing challenging latency problems over the backhaul connecting distant cell sites to the JPC, which in turn impose practical limitations on the maximum size of each CoMP cluster.

Users in CoMP cluster 1

Users in CoMP cluster 2

I. I NTRODUCTION One of the major impediments to achieving higher spectral efciency in a cellular network is the mutual interference among users in adjacent cells utilizing the same radio resources. Substantial research efforts have been devoted to reducing the adverse impact of inter-cell interference. In particular, coordinated multi-point (CoMP) transmission has recently received much attention in both industry and academia, cf. [1] [2] [8], due to its promise in effectively mitigating inter-cell interference through the use of multi-user precoding across the antennas of multiple basestations (BSs). A simple architecture for CoMP transmission is to partition all cells in a cellular network into disjoint independent groups, referred to as CoMP clusters, as depicted in Fig. 1. Basestations within each of these xed CoMP clusters are connected, forming a coordination set, to a centralized, joint processing center (JPC) where transmissions to all users within that cluster are coordinated. In this cluster-wise centralized architecture, although each CoMP cluster provides an extended geographical area over which inter-cell interference within the cluster can be controlled, the users near the edge of a CoMP cluster can still experience substantially degraded service quality compared to those near the center of the CoMP cluster due to the uncoordinated interference from other
Fig. 1.

Users in CoMP cluster 3

Illustration of a Cluster-wise Centralized CoMP Architecture

II. D ISTRIBUTED F RAMEWORK In this paper, we propose a distributed architecture for CoMP transmission where CoMP clusters are allowed to overlap with each other. In this architecture, each cell is associated with a cell-specic (CS) CoMP cluster, centered around the serving cell. Users in any given cell are served by the set of coordinating BSs in its own CS CoMP cluster. For example, Fig. 2 shows two overlapping CS CoMP clusters (green and yellow) whose serving cells are adjacent to each other and are at the center of their respective CS CoMP clusters. The precoding weights of each CS CoMP cluster are computed almost independently (as explained later) by a coordinating processor, which is presumably located at the BS of its serving cell. Since the CS CoMP clusters closely mingle with each another, the interference generated by one CS CoMP cluster can severely affect many other overlapping clusters. To reduce its impact on the overall system performance, a geographical area, termed the region of interference avoidance (RoI), is associated with each CS CoMP cluster such that (some or

User in Targeted Cell j

Region of Interference Avoidance for Cell j

Fig. 2.

Set of Coordinating Basestaions for Cell j Illustration of a Distributed CoMP Architecture

all) users in the RoI are taken into account by the CS CoMP cluster when computing the precoding weights in such a way that minimizes the interference generated to these users. The RoI of each cluster is assumed to contain its serving cell. Since each BS antenna can be used to serve multiple CS CoMP clusters simultaneously, a mechanism for resolving potential conicts in the power requirements among different CS CoMP clusters is needed to avoid violating the power constraints of each BS. To this end, we propose a simple iterative distribtuted algorithm for negotiating power requirements among overlapping CoMP clusters. Convergence properties of the proposed algorithm will be discussed. Using such an algorithm, we evaluate the system performance of the distributed approach of CoMP transmission and compare it with that of the cluster-wise centralized approach. III. P RECODER C OMPUTATION In this section, we describe a simple, effective method of computing the precoding weights of all BSs in a distributed manner. Let J denote the set of indices of all cells, and their corresponding CS CoMP clusters, in a cellular network. Based on the framework described above, consider the situation where each CS CoMP cluster j J uses nb,j BSs in its coordination set Cj to transmit signal to Kj users in its serving (or center) cell while trying to limit interference to other Lj users in the RoI. Let nt (i) denote the number of available transmit antenna for each BS i Cj , and the total number of transmit antenna for this CS CoMP cluster be nt,j = iCj nt (i). Let Sj = {qj (1), qj (2), , qj (Kj )} denote the set of global user indices for the Kj users in the serving cell, and let Aj = {qj (Kj + 1), qj (Kj + 2), , qj (Kj + Lj )} denote the set of other Lj user indices in the RoI but not in Sj , where qj () denotes a mapping from the local user indices within the CS CoMP cluster j to the corresponding global user indices. For each k Sj Aj , let Hj,k be the nr,k by nt,j channel response matrix from the transmit antennas of cooperating BSs in Cj to the receiver antennas of user k, where nr,k denotes the number of receive antennas at user k.

Let Pj,k denote a nt,j -by-ns,k precoding matrix of user k to be served by the CS CoMP cluster j, where ns,k denotes the number of data streams transmitted to user k. For notational simplicity, we let Pj [Pj,qj (1) , Pj,qj (2) , , Pj,qj (Kj ) ] and Hj [HH j (1) , HH j (2) , , HH j (Kj +Lj ) ]H . We assume j,q j,q j,q that the processor of CS CoMP cluster obtains, on a regular basis, the instantaneous channel state information (CSI) Hj from all users in its serving cell and all the users to be served in the RoI either through feedback or through measurement in the reverse link in a Time-Division Duplex (TDD) system. To facilitate low-complexity, distributed computation of {Pj }jJ , we propose to divide each precoding matrix Pj into two components: a tentative precoding matrix Pj j,q (1) , Pj,q (2) , , Pj,q (K ) ] and a power back-off (or [P j j j j scaling) factor j . The tentative precoding matrix Pj represents the desired precoding matrix when no other CS CoMP cluster in J needs to share the antenna resources of any BSs in Cj . When there is a possible sharing of antenna resources among overlapping CoMP clusters, a power backoff, as specied by j , is used to scale the tentative precoding weights so that the power requirements of all antenna elements are satised. The overall precoding matrix is simply given by Pj = j Pj . Given a ns,k -by-1 information-bearing symbol vector sk (normalized such that Esk sH = 1) for each user k k, the transmitted signal from the coordinating BSs in the CS CoMP cluster j to all Kj users in the serving cell is given by Pj,k sk = j xj = Pj,k sk . (1)
kSj kSj

The main idea here is to let each CS CoMP cluster determines the tentative precoding matrix Pj independently while using a low-complexity distributed algorithm to negotiate the power back-off factor j among overlapping CS CoMP clusters. There are many methods of computing Pj as a function of Hj , such as the oft-cited zero-forcing precoding [2] [3], and the MMSE-based multi-user precoding techniques, cf. [4] [5] [6] [7]. In our simulations, we adopt the transmit Wiener lter (TxWF) precoder [7] due to its good performance-complexity tradeoff and its lack of restriction in the size of Hj . More precisely, the tentative precoding matrix Pj,qj (i) for user qj (i), for i = 1, 2, , Kj , is computed as [ ( )1] Pj,qj (i) = HH Hj HH +(j)I j j qj (i) (2)

1 Lj +Kj 2 where (j) PT nr,qj (l) n,qj (l) . mj,i l=1 i l=1 ns,qj (l) , [A]:,m : n denotes those columns of the matrix A positioned between its mth and nth column inclusively, 2 I denotes the identity matrix, n,k denotes the variance of noise-plus-interference observed at user k, PT represents the maximum total transmit power of all antennas in all BSs in Cj , and k denotes a matrix for allocating different relative powers to different data streams.

:,mj,i1 +1:mj,i

IV. P OWER N EGOTIATION Once Pj is determined at the serving cell of a CS CoMP cluster, the corresponding tentative antenna weights, as speci-

ed by Pj , are sent to each BS in the coordination set Cj . To resolve any potential conicts in power demand among different overlapping clusters, we present here a simple algorithm, which we called iterative sum-min algorithm, for computing a power back-off factor j for each CS CoMP cluster by iterative communications among neighboring serving cells, which is described as follows and illustrated in Fig. 3.
Processors Antennas Compute and report power margin at each antenna

and use j

(n)

to update {j (m)}m as (m) = j


(n)

(n)

(n+1)

j (m).

(n)

(6)

Scale weights by minimum margin

Step 6 : Increment iteration index n, and go back to Step 2 until certain stopping criteria is satised. For (n) example, j is very close to 1 for every CS CoMP cluster j J. Step 7 : The power back-off factor for the CS CoMP cluster nf 1 (l) j is computed as j = l=0 j , where nf is the nal value of the iteration index. In the following, we study the convergence properties of the iterative sum-min algorithm. Lemma 1. For each antenna m , the sequence of power margins {(n) (m)}n1 generated by the sum-min algorithm satises (n) (m) 1 and (n+1) (m) (n) (m) for all n 1. Proof: For n 1, it follows from (4) that for all m , (n) (m) = Pt (m) j (m) Pt (m) min ] j
(n1) (n)

Fig. 3. Graphical Representation of Iterative Sum-Min Algorithm (different branch colors represent different CS CoMP clusters)

Let be the global index set of all antennas in the network, and let B(m) denote the index of the BS where antenna m resides. Also let [A]k,l denote the (k, l) element of matrix A. Iterative Sum-Min Algorithm: Step 1 (Initialization): Set the iteration index n to 0. Processor (located at the serving cell) of the CS CoMP cluster j computes, for each j in the network, the (0) initial power requirement j (m) as [ ] 2 (0) j (m) = Pj rj (m),l , (3)
l

j:B(m)Cj

j:B(m)Cj

[ Pt (m)

m :B(m )Cj

(n1) (m )
(n1)

(m)

(n1) (m)

= 1, (m)

j:B(m)Cj

for each antenna indexed by m in the coordination set Cj , where rj (m) denotes the row of the matrix Pj that corresponds to the antenna m. Step 2 : Processor sends antenna power requirements (n) {j (m)}m to all respective neighboring BSs (or their associated processors) in the set Cj \{j}. Step 3 (Sum): Upon receipt of the power requirements (n) {j (m)}m from neighboring BSs for each antenna m, the processor whose serving cell possesses the antenna computes the power margin (n) (m) as (n) (m) = Pt (m) j (m)
(n)

where the second equality follows from (5) and (6), and the last equality follows from the denition of (n1) (m). Now, again from (4), we have (n) (n+1) (m) j:B(m)Cj j (m) = (n+1) (n) (m) (m) j:B(m)Cj j (n) j:B(m)Cj j (m) [ ] = (n) (m ) (n) (m) min j j:B(m)Cj
m :B(m )Cj

(4)

j:B(m)Cj

j:B(m)Cj

j (m)
(n)

(n)

j:B(m)Cj

1 j (m)

= 1,

where Pt (m) denote the maximum power constraint of antenna m. Step 4 : Each processor sends power margins {(n) (m)}m of all the antennas of its serving cell to processors of neighboring serving cells that utilize these antennas to transmit signals to their respective users. Step 5 (Min): Upon receipt of the power margins of all antennas in BSs in Cj , the processor of the jth CS CoMP cluster, for each j, computes the minimum power margins over all the antennas given by j
(n)

where the inequality follows from (n) (m) 1 for all m . Hence, the sequence {(n) (m)}n1 is monotonically nonincreasing. Lemma 2. For each CS CoMP cluster j J, the sequence (n) of minimum power margins {j }n1 generated by the (n) sum-min algorithm satises j 1 for all n 1 and (n) lim j = 1.
n

min
m:B(m)Cj

(n) (m)

(5)

Proof: j 1 for all n 1 is obvious from Lemma 1 and (5). It also follows immediately from Lemma 1 and (5) that (n) the sequence {j }n1 is a monotonically non-increasing

(n)

sequence for all j J. Therefore, the sequence {j }n1 has a limit, denoted by cj , which is no less than one. Now suppose cj > 1, then from (6), we have that for some m , (n) (1) j (m) (cj )n1 j (m) as n , which is a contradiction since Pt (m) (n) (n) j (m) j (m) = (n) Pt (m) < (m)
j:B(m)Cj

(n)

23

21 4 2 40 20 34 1 39 53 33 5 14 52 25 35 6 44 54 26 10 7 45 56 37 18 16 57 55 38 19 36 17 15

1000
42 22 3

500
41 30 11 27 8 46 31 12 50 47 32 13 51 48 29 28 9 43 24

for all n 1, where the equality above follows from (4). From these lemmas, one can conclude that after only one iteration, the resulting power back-off factors {j }jJ is guaranteed to satisfy the power constraints of all antennas in the network (since (1) (m) 1 for all m ). Moreover, the algorithm is guaranteed to converge, and additional iterations after the rst iteration yield higher values of {j }jJ so that the power of each antenna can be more fully utilized. V. N UMERICAL R ESULTS In this section, we present system simulation results of an OFDM cellular communication system employing CoMP transmission in the downlink using the cluster-wise centralized and the distributed architectures described above. The system conguration follows to large extent the 3GPP Long Term Evolution (LTE) standard [9]. For the distributed case, a total of 57 hexagonal cells or sectors served by 19 sites are simulated, as depicted in the Fig. 4. The coordination sets of those CS CoMP clusters near the edge of the layout are wrapped around, forming a total of 57 overlapping CS CoMP clusters, each centered at one cell. For the cluster-wise centralized case, each CoMP cluster is formed by grouping 9 neighboring cells (or sectors) served by 3 sites with sectorized antennas, as depicted in Fig. 5. A total of 7 CoMP clusters (i.e. 63 cells) are simulated in this layout. For both layouts, radio signals are wrapped around from one end of the layout to another end. Moreover, each cell is assumed to serve one scheduled user in each time frame and frequency subcarrier. All cells are also assumed to be synchronized in time, and the cyclic prex is assumed to be sufciently long for the propagation delays of each (CS) CoMP cluster. Each user is assumed to have 2 receive antennas and to move at a speed of 3 km per hour. Other system simulation parameters are summarized in Table I. Fig. 6 compares the performance of the cluster-wise centralized and the distributed CoMP architectures, in terms of system throughput (sum data rates of all users) versus data throughput of the 5-percentile users, which is indicative of the data throughput of users located near cell edges. The curves labeled by Cn and Dn are for the centralized and the distributed architecture, respectively, where n denotes the number of cells per (CS) CoMP cluster. For example, the curve labeled D1 represents the performance for the case when the coordination set of each CS CoMP cluster only contains the serving BS. The notation C=k means the BSs of k tiers of neighboring cells or sectors are connected to form the coordination set, while the notation I=m means m additional tiers of neighboring cells or sectors are in the region

49

500

1000 1000 500 0 500 1000

Fig. 4. Cell Layout with Sectorized antennas for Distributed CoMP Architecture (red dots denote site locations)

of interference avoidance. The multi-user TxWF precoder is used for both cluster-wise centralized and distributed cases. For comparison, the green curve represents the performance of a non-CoMP system using the xed codebook of precoding matrices as dened in LTE [9]. As shown in Fig. 6, the distributed architecture yields substantially better performance, in both the cell-edge user data rate and the total system data throughput, over the clusterwise centralized architecture with a smaller number of cells per CoMP cluster. Moreover, enlarging the region of interference avoidance beyond the cells covered by the coordination set can give signicant additional gain, at the expense of more CSI measurements and/or feedback for the network. Note that at high load, the gain in system throughput attained by the distributed architecture with a large RoI diminishes slightly, due to the limitation in the transmit dimension of the precoder and the increased number of users in RoI. Fig. 7 shows the distributions of user data rates at an averaged load of half a user per cell. As shown, compared
13 10

1000
31 28 34 17 33 29 35 36 22 19 25 23 24 20 26 27 41 21 8 5 30 14

16 15 11 18 4 1 7 6 2 9 40 37 43 42 38 44 45 39 53 50 47 54 3 52 51 48 62 49 46 59 56 63 12 61 60 57 58 55

500
32

500

1000

1500

1000

500

500

1000

1500

Fig. 5. Cell Layout with Sectorized antennas for Centralized CoMP Architecture with 9 Cells Per CoMP cluster (red dots denote site locations)

DistDL 7 Cells per CSCoMP Cluster with 4x2 TxWF Ideal CSITx 3 NonCoMP C9TxWF D1TxWF (C=0,I=1) D7TxWF (C=1,I=0) D7TxWF (C=1,I=1)

2.5 Celledge (5 pct) bitrate [bps/Hz]

algorithm was introduced for computing the power back-off factors in a distributed manner, and its convergence properties were analyzed. Through system simulations with practical channel models, we found that the distributed architecture with the proposed distributed method of computing multiuser linear precoders can lead to substantially better system performance than the cluster-wise centralized architecture. ACKNOWLEDGMENT This work was partly supported by the Sixth Framework Programme via the COOPCOM project (http://www.coopcom.eu.org). The author would like to thank Kambiz C. Zangi, George J ngren, and Magnus A. o Olsson for valuable comments and suggestions. R EFERENCES

1.5

0.5

0 0.5

1.5

2 2.5 3 3.5 System Throughput [bps/Hz/sector]

4.5

Fig. 6.

Performance Comparison of Centralized vs. Distributed CoMP

to the centralized architecture, the distributed architecture is capable of supporting substantially higher data rates for all users with a smaller number of cells per CoMP cluster. VI. C ONCLUSIONS In this paper, we studied the system performance of linear multi-user precoding under a distributed architecture for CoMP transmission in a cellular network. We propose a simple method for computing the precoding matrices in a distributed manner by separating them into a tentative part, which is computed independently by different cell-specic CoMP clusters, and a scaling part for power back-off, which is negotiable across CoMP clusters. The tentative precoding matrices are computed in such a way that maximizes the desired user data rates in the serving cell while minimizing the interference to other users in the surrounding region. A simple sum-min
Modulation Coding Link Adaptation Channel Model Scattering Environment Base Antennas Bandwidth Frequency Reuse Inter-site distance user Receiver Data Trafc Model Scheduling Tx Power per antenna Antennas Sectors per CoMP cluster QPSK, 16QAM, 64 QAM Practical Turbo Codes Ideal (i.e. based on perfect channel quality measurements) 3GPP SCM [10] Suburban-macro [9] 120-degree antenna (with no down tilt) 5 [MHz] 1/1 500[m] Ideal MMSE-SIC Full buffer Round Robin 5 [Watts] Base: 4 ; UT: 2 Cluster-wise centralized: 9 sectors (xed) Distributed: 1 or 7 sectors (cellspecic) Cluster-wise centralized: 7 Distributed: 57 (cell-specic, with wrap-around) TxWF = Transmit Wiener Filter TABLE I S YSTEM
SIMULATION PARAMETERS

[1] S. Parkvall, E. Dahlman, A. Furuskr, Y. Jading, M. Olsson, S. Wnstedt, K. Zangi, LTE-Advanced Evolving LTE towards IMT-Advanced, in Proc. VTC 2008-Fall, pp. 1-5, Sept. 2008. [2] M.K. Karakayali, G.J. Foschini and R.A. Valenzuela, Network coordination for spectrally efcient communications in cellular systems, IEEE Wireless Communications, vol. 13, pp. 56-61, August 2006. [3] M. Joham, W. Utschick, and J.A. Nossek, Linear Transmit Processing in MIMO Communications Systems, IEEE Trans. Sig. Proc., vol. 53, no. 8, pp. 2700-2712, August, 2005. [4] A. Mezghani, M. Joham, R. Hunger and W. Utschick, Transceiver Design for Multi-User MIMO Systems, Proc. IEEE Workshop on Smart Antennas, CityplaceUlm, country-regionGermany, Mar. 2006. [5] M. Joham, K. Kusume, M.H. Gzara, W. Utshick, Transmit Wiener Filter for Downlink of TDDDS-CDMA Systems, Proc. IEEE 7-th Int. Symp. On Spread-Spectrum Tech. & Appl., pp. 9-13, Prague, Czech Republic, Sept 2-5, 2002. [6] H. Karaa, R.S. Adve, and A.J. Tenenbaum, Linear Precoding for Multiuser MIMO-OFDM Systems, Proc. ICC 2007, pp. 2797-2802, 2007. [7] S. Shi and M. Schubert, MMSE Transmit Optimization for Multi-User Multi-Antenna Systems, in Proc. ICASSP 05, Mar. 2005. [8] A. Papadogiannis, placeE. Hardouin, and D. Gesbert, Decentralising Multi-Cell Cooperative Processing: a Novel Robust Framework, EURASIP J. on Wireless Comm. and Net., vol. 2009, pp. 1-10, Apr. 2009. [9] E. Dahlman, S. Parkvall, J. Skld and P. Beming, 3G Evolution HSPA and LTE for placeMobile Broadband, 2nd ed., Academic Press, 2008. [10] G. Calcev et al. A wideband spatial channel model for system-wide simulations, in IEEE Trans. on Vehicular Technology, Vol. 56, No. 2, Mar. 2007.
DistDL 7 Cells per CSCoMP Cluster with 4x2 TxWF Ideal CSITx; (load = 0.5) 1 0.9 0.8 Pr(User Data Rate <= abscissa) 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 NonCoMP C9TxWF D1TxWF (C=0,I=1) D7TxWF (C=1,I=0) D7TxWF (C=1,I=1) 1 2 3 4 5 Data Rate [bps/Hz] 6 7 8 9

Number of CoMP clusters

Precoding Schemes

Fig. 7. User Data Rate Distributions of Centralized and Distributed CoMP at Medium Load (averaged 1/2 user per cell)

Anda mungkin juga menyukai