Distributed Multi- Cell Downlink Transmission based on Local CSI

Transcription

1 Distributed Multi- Cell Downlink Transmission based on Local CSI Mario Castañeda, Nikola Vučić (Huawei Technologies Düsseldorf GmbH, Munich, Germany), Antti Tölli (University of Oulu, Oulu, Finland), Eeva Lähetkangas (Nokia Networks, Oulu, Finland), Hadi Ghauch, Mikael Skoglund (KTH Royal Institute of Technology, Stockholm, Sweden), Sinda Smirani, Ahmed Saadani (Orange Labs, Issy- les- Moulineaux, France), and Danish Aziz (Alcatel Lucent Bell Labs, Stuttgart, Germany) 1. ABSTRACT Coordinated multipoint downlink transmission has been considered in cellular networks for enabling larger coverage, improved rates and interference mitigation. To harness the gains of coordinated transmission, a fast information exchange over a backhaul connecting the cooperating base stations is required. In practice the bandwidth and delay limitations of the backhaul may not be able to meet such stringent demands. Such impairments motivate the study of cooperative approaches based only on local CSI and which require no fast data exchange between the base stations. To this end, this article presents several distributed cooperative approaches for the downlink which were developed in the context of the EU project METIS [1]. 2. INTRODUCTION As a result of the advantages offered by coordinated multipoint (CoMP), different CoMP variants have been included for the downlink in LTE-Advanced such as coordinated beamforming (CB) and joint transmission (JT). With CB, each BS serves the users in its cell by considering the interference generated to the users in other cells. On the other hand, for JT all users in the cluster are served by all the cooperating BSs, which have available the data of the users. To enable such cooperation, data and/or CSI of all the users in the cluster has to be exchanged in a swift manner between the cooperating base stations (BSs). This imposes certain constraints on the capacity and delay of the backhaul, which may not be able to be met in practice. Furthermore, with the deployment of small cells and ultra dense networks in future communication systems, the backhaul quality might be further degraded in a cooperating area. In fact, the imperfect backhaul has been recognized as one of the key issues in fast multi-cell cooperation [2]. In this work, we focus on the downlink (DL) transmission of a cluster in a time division duplex (TDD) system, where the local channels (see Fig. 1) from each BS to all users can be obtained at each BS via uplink (UL) training and by exploiting reciprocity. Although the information exchange required to achieve global CSI at all BSs (or at a central unit) can be expected to be less than the amount of data of the users, for some applications, data may indeed be more delay tolerant than CSI which becomes outdated. By considering only local CSI at each BS, i.e. each BS is not aware of the channels from the other BSs to the users, the common assumption of global CSI for cooperative downlink transmission can be significantly relaxed. Hence, distributed schemes based solely on local CSI become highly relevant in scenarios with limited backhaul capabilities. Fig. 1: Local CSI for DL transmission of BS 1 In this article we present embodiments of such distributed cooperative approaches which were developed in METIS [1]. We discuss iterative and non-iterative schemes, which can be viewed as different CoMP variants. To this end, we first elaborate how forward-backward (F-B) training facilitates CB in a distributed manner by allowing the nodes (BSs and users) to iteratively optimize their transmitter/receiver based only on local CSI. We discuss the tradeoff due to the overhead resulting from the F-B iterations as well as the effect of imperfect CSI on distributed iterative schemes. We also comment on the required signaling to enable distributed coordination via F-B training in a challenging scenario like a dynamic TDD system. Regarding non-iterative approaches, we first present a scheme based on interference alignment which exploits outdated local CSI for distributed CB.

2 Furthermore, by assuming the data of the users to be available at the cooperating BSs, JT can be performed in a distributed non-iterative manner. In this context, we present a distributed precoding technique, which achieves one more degree of freedom (DoF) than current non-iterative approaches. 3. DISTRIBUTED COORDINATION VIA F-B TRAINING Interference coordination among cells is a non-trivial task for distributed approaches, since finding the optimum transmit/receive filters of the BSs/users in the cluster usually requires global CSI. Coordinated transmission can be achieved without CSI exchange between the nodes (BSs/users) by employing F-B training to iteratively optimize the filters of the nodes in a distributed manner. For this purpose, each F-B iteration consists of two phases: a DL (forward) phase and a UL (backward) phase. In the forward phase, the BSs transmit precoded pilots enabling the users to estimate their local CSI consisting of the equivalent channel, i.e. the cascade of the precoder and channel, of its desired link as well as of the interfering links (interference covariance matrix). The users proceed to optimize their receive filters based on the acquired local CSI. In the backward phase, the users transmit precoded pilots using the receivers computed in the previous forward phase as precoders, which allows the BSs to estimate the equivalent channels for the users in its cell and in the other cells, i.e. local CSI. With this local CSI, each BS optimizes the transmit filters for its users, which are used as pilot precoders in the forward phase of the next F-B iteration. Note that the number of orthogonal pilot resources in each forward/backward phase has to be at least equal to the number of data streams in the cluster to allow the users/bss to estimate the effective channel for each data stream. F-B training allows a fully distributed coordinated computation of transmit/receive filters of the BSs/users in the cluster, without CSI exchange over a backhaul. The filters can be computed based on different optimization criteria such as minimizing the interference leakage, minimizing the (weighted) sum MSE, maximizing the (weighted) sum rate [3,4] or maximizing the spatial separation between the signal subspace and the interference plus noise subspace (referred as max-sep [1, Section 8.10]). Algorithms which aim at interference reduction and at maximizing the effective channel of the desired link are shown to perform better than schemes that are only based on interference suppression. Moreover, if every step of the global optimization problem can be decoupled among the nodes, the distributed iterative scheme incurs in no loss in optimality (besides the overhead) compared to centralized approaches [3]. Although distributed iterative approaches may deliver close to optimum solutions, there is a tradeoff between the achievable throughput and the F-B training overhead, i.e., the gain introduced by the iterations comes at the expense of less resources available for the data. For instance, several previous works (e.g. [5]) require hundreds of iterations to converge, which would lead to a prohibitively large overhead. When taking the training overhead into account, the optimum number of F-B iterations might be small (see Fig. 3) and thus, the algorithms might be far from converging. Hence, we are interested in algorithms which require few F-B iterations, but yet achieve a good performance [6]. To view the gain with F-B training, we consider the dense urban network scenario evaluated in the METIS project, consisting of a 3x3 grid of square buildings surrounded by streets in an area of 120 m x 120 m with 9 micro BSs ( reduced Madrid Grid ) [1]. Further relevant parameters are detailed in [1, Section 8.10]. The transmit/receive filters of the BS/users in the backward/forward phase of each F-B iteration are optimized according to the max-sep scheme [1, Section 8.10]. We assume only 2 iterations for the F-B training, which results in a very low overhead. We assume perfect channel estimation at the BSs/users at each iteration. As a reference we consider an orthogonal access scheme (e.g. TDMA), where each user is assigned a separate transmission block and served with eigenbeamforning based on the SVD of the channel. Fig. 2 depicts the cumulative CDF of the user rate, comparing our proposed scheme based on F-B training and the benchmark for several configurations. We fix the number of antennas and streams at the users to N=4 and d=2, respectively, while varying the number of antennas M at the micro BSs. As the number of BS antennas increases, the performance of our scheme increases massively (~200% gain for M=10), while improvement of the orthogonal access scheme is quite insignificant. Although these results do not take the overhead into account, the fact that the total number of streams and number of F-B iterations is small, implies the overhead is not significant. Fig. 2. Gain due to F-B training with 2 F-B iterations

3 We now discuss further relevant aspects of distributed CB via F-B training such as signaling and imperfect CSI. We first consider a scenario where distributed coordination is a challenging task and discuss how F- B training can be employed in this context. 3.1 DYNAMIC TDD To satisfy the increasing demand of mobile traffic in future communication systems, small cells can be deployed along with macro/micro-cells. In such scenarios, the UL and DL traffic may vary significantly over time and among neighboring cells. Existing technologies do not support fully flexible UL/DL ratio and flexible link direction switching. Dynamic TDD on the other hand, allows to adaptively allocate resources according to the asymmetric UL/DL traffic and provides improved overall resource utilization. One of the main challenges in dynamic TDD systems is handling cross link interference generated by the asymmetric traffic between cells, i.e. user-to-user interference and BS-to-BS interference. Although this interference can be mitigated through a centralized coordinated resource allocation and beamforming, it requires explicit feedback of the user-to-user channels and a fast CSI exchange between BSs. Such centralized approach can be circumvented by employing F-B training to convey information to the BSs about the interference and receivers at the users [1,3,4]. An adaptive frame structure is however required, such as the one presented in [7], with a bidirectional control signal part embedded in each subframe and time separated from the data payload for either the downlink or uplink. Demodulation reference signal (DMRS) symbols are located e.g. in the first symbol of the data part and can be precoded with the data, enabling the receivers to estimate the equivalent channel of its desired and interfering links. The proposed subframe structure together with a short subframe length, such as 0.25 ms in dense deployments, enables fully flexible UL/DL switching and achieving tight latency targets for both control and data. We now evaluate the performance of distributed CB in a dynamic TDD system consisting of three cells: two operating in the DL and one in the UL, with 4 users per cell. Each BS has 4 antennas and serves only the users in its cell, which are equipped with 2 antennas. The path loss from a BS to the in-cell users is normalized to 0 db. The ratio of the transmit power to noise variance in the DL and UL is 20 db and 4 db, respectively. The simulation environment is further defined by the separation between DL users in different cells, between the UL users and DL users and between the UL BS and the DL BSs, for which we define the relative path loss as α, β and δ, respectively. For the small scale fading we assume Rayleigh fading. The transmitters/receivers of the BSs/users in the cluster are obtained via F-B training such that sum rate is maximized via an iterative equivalent weighted MSE minimization [1,3,4]. For further details of the algorithm refer to [3,4]. We assume perfect channel estimation of the desired link and interference covariance matrix at each iteration. For the UL precoded pilots we employ the MMSE receivers [3, Eq. (4)], while the DL precoded pilots are obtained from solving the transmitter MSE minimization problem [3, Eq. (15)]. The alternating algorithm is detailed in [3, Algorithm 3]. The relative signaling overhead γ for one F-B iteration is assumed to be 1% (γ=0.01), so the total overhead after n F-B iterations is ρ = n γ. The overhead ρ is taken into account in the actual sum rate with F-B training. As a reference we consider an uncoordinated beamforming design, where the intercell interference produced to users in other cells is not taken into account in the design of the BS transmitters. Fig. 3 depicts the sum rate of the three cells in the cluster as a function of the normalized overhead ρ. Different scenarios are considered through different α,β and δ settings. Compared to uncoordinated beamforming, the performance with distributed CB via F-B training improves significantly for small values of α,β and δ, which represent high interference scenarios. We can clearly observe the tradeoff due to the training overhead. The peak rate in all curves based on F-B training is obtained with 8 to 12 F-B iterations, i.e. less than 12% overhead is required for the peak rate for all scenarios. The sum rate with F-B training can still be better than the baseline scheme with 30% to 60% overhead depending on the scenario. Therefore, the employed algorithm for distributed CB based on F-B training performs very well especially when complex interference conditions are present. Fig. 3. Actual sum rate vs. normalized overhead at SNR = 20 db for different settings 3.2 IMPERFECT CSI Although many existing works on distributed schemes based on F-B training consider the training overhead,

4 they usually assume perfect estimation of the CSI at each iteration. In practice, however, only estimated CSI is available, such that the receive filters are based on estimated local CSI and hence, not entirely matched to the true channel conditions. Despite this fact, only few works consider the impact of imperfect CSI on the distributed iterative schemes [8,9]. To observe the effect of imperfect CSI, consider a cluster of three cells with one user per cell. We assume the BSs and users to be equipped with two antennas and that each BS transmits one stream to its own user. All channels are modeled as Rayleigh fading. The average sum rate with F-B training for three different algorithms: minimum leakage, max SINR and MMSE are depicted in Fig. 4 for the case of perfect CSI and imperfect CSI, for which we assume the variance of the estimation error σ e 2 =0.1. The same number of F-B iterations is assumed for each curve. For both cases, we can see the effect of the imperfect CSI becoming more pronounced at higher SNRs, such that already at SNR=10 db, the sum-rate degradation for the MMSE approach can be up to 15%. Although the maximum SINR algorithm provides the maximum rate with perfect CSI, this is no longer the case when considering imperfect CSI, where in fact the MMSE algorithm performs better as it is more robust to estimation errors. Fig. 4: Effect of Imperfect CSI for different algorithms Note that the convergence of the algorithms with imperfect CSI might be questioned, since at each iteration we experience a different estimation error. However, simulations show that the iterative algorithms converge (averaging over the estimation error) to a fixed bias that depends on the variance of the estimation error and SNR. In fact, a faster convergence of the algorithms can be achieved compared to the perfect CSI case. Moreover, in case there is a fixed power budget available for the channel estimation, there is a tradeoff between the number of iterations and the amount of pilot power per iteration. It turns out, however, that it is preferable to use less number of iterations with more power per iteration for the pilot transmission. 4. DISTRIBUTED INTERFERENCE COORDINATION Previously we have discussed how distributed CB can be performed via F-B training, enabling the BSs to iteratively obtain information about the channel and interference at the users. We now focus on noniterative schemes for distributed interference coordination, whereas before each BS serves only the users in its cell. Current solutions in this context include centralized approaches as eicic in LTE, which require a fast backhaul and are mainly based on limiting the resources in some parts of the network. In the following we discuss a distributed interference coordination approach based on interference alignment, i.e. the spatial dimensions at the transmitter are employed to align the interference at the users in a given subspace, whereas the spatial dimensions at the users are used for interference suppression. Due to the limited number of antennas we have a limited number of DoF and hence, it is not always possible to align the interference in one subspace. We thus propose to align the intracell interference with only partial intercell interference (ICI), consisting of the direction with maximum average ICI, i.e. the principal eigenvector of the ICI covariance matrix. We refer to this scheme as multi user inter cell interference alignment (MUICIA) [10]. After the users have estimated their ICI covariance matrix, the BSs obtain the partial ICI via feedback from the users. The users cannot estimate, however, the current ICI covariance matrix prior to the DL transmission. Hence, we propose each BS employs outdated ICI information from the previous transmission. The BS thus designs the beamforming vectors such that the intracell interference of the current transmission is aligned at the users with the outdated partial ICI subspace. For this, each BS requires the local channels to its users (obtained from UL training) and the outdated principal eigenvector of the ICI covariance matrix of its users (obtained via feedback). Due to the possible misalignment between the current ICI and the outdated ICI, there might be some interference leakage in the desired signal subspace. This interference leakage, however, can be minimized with a proper robust receiver design. To evaluate the performance of MUICIA, we consider a scenario with three cells each having two users, with two antennas at each BS and user. We assume the spatial channel model from [11]. The users are placed in each cell according to a given average SINR, i.e. depending on the average signal strength and ICI. At each BS we assume the channels of its users are perfectly known and the partial ICI from the previous transmission is available without delay. For

5 the interference suppression, the users employ the MMSE algorithm. Further simulation parameters can be referred from [1, Section 8.7]. We compare the performance against two baseline approaches: one using the SVD of the serving channel matrix for the transceiver design, and zero-forcing precoding (EZF) based on the principal eigenvector of the channel matrices. We plot the average cell rate of the different approaches as a function of the average SINR of the users in Fig. 5. Simulation results show that MUICIA outperforms the baseline schemes in the complete range of SINR. This is due to the fact that in this scenario, the ICI is very strong and we are able to find a dominant ICI direction for alignment as well as for suppression SVD 2 EZF 1 MUICIA Average SINR [db] Fig. 5: Performance comparison of MUICIA with other transmit precoding schemes Mean Cell Rate [b/s/hz] 5. DISTRIBUTED JOINT TRANSMISSION In the previous sections, we have discussed how coordination can be achieved in a distributed manner via iterative and non-iterative approaches. Further examples of non-iterative distributed CB schemes based on local CSI include distributed linear precoding like distributed ZF and the distributed virtual Signal-to-Interference-plus-Noise ratio approach [12]. By assuming the data of the users to be available at all BSs, these latter works can be extended such that JT with only local CSI can be performed in a distributed manner [12]. The data of all users can be available at all BSs, either by buffering of delay tolerant applications or by exploiting the envisaged possibility of data sharing/caching in future wireless networks [13]. The number of DoF of current non-iterative distributed linear approaches, however, remains the same with local or global data (e.g. equal to the number of antennas at each BS). On the other hand, one more DoF can be achieved with the destructive interference addition (DIA) scheme [14] as it is explained in the following. With DIA, the beamforming vectors at each BS are computed such that the equivalent interference channels are forced to be equal to a set of predefined values. The predefined values are chosen offline and agreed among the BSs so the interference resulting from the joint transmission adds up to zero at the users. This is in contrast to distributed ZF, where the equivalent interference channels resulting from each transmission are all forced to be equal to zero. With two antennas at each BS, for example, two equivalent interference channels can be forced to a predefined value with DIA when computing the 2-dimensional beamforming vector to a given user. This means that three users can be served under such interferencefree condition, which implies three DoF for DIA, in contrast to two DoF (the number of antennas at each BS) for the current solutions. Note that with DIA, the beamforming vector to a given user does not depend on the channel between the BS and the user. When computing the DIA beamforming vectors, it is likely that the equivalent interference channels cannot be set to the predefined values without violating the power constraint at each BS. In this case, each BS needs to scale its beamforming vectors with a per BS specific factor to fulfill its power constraint. The resulting interference, however, will not add up to zero at the users, but will lie on the same direction. We refer to this approach as DIA uncoordinated power control (UPC), which offers some gain at medium SNR but does not achieve the DoF of DIA (See Fig. 6). To guarantee the destructive interference addition, all BSs should actually scale their beamforming vectors with the same factor: the factor of the BS with the most critical power constraint. Such scaling might lead to some BSs transmitting with less than their available power. In any case, a fast exchange of this factor among the BSs is required for DIA. Fig. 6: Performance of Distributed Precoding and DIA To observe the performance of DIA, we depict in Fig. 6 the average sum rate as a function of the cell edge SNR for a cluster of three BSs each with two antennas serving three single-antenna users, which are uniformly distributed in the cluster. The inter-bs distance is 50 m and cell edge SNR refers to the

6 average SNR that a user at the same distance from all three BSs would experience, i.e. in the middle of the cluster (see Fig. 1). We consider path loss and Rayleigh fading. For comparison we consider distributed maximum ratio transmission (MRT), distributed ZF and cooperative multicell precoding (CMP) [12]. Due to the higher number of DoF, DIA outperforms all the baseline schemes at high SNR, which saturate due to multi-user interference. The baseline schemes can avoid this saturation by dropping one user and serving instead only two (two DoF). With this strategy, however, DIA eventually outperforms the baseline schemes, due to the three DoF with three users. Moreover, DIA is outperformed at low SNR, since the optimum strategy is to maximize the strength of the desired link, which DIA does not consider. Since the DIA beamforming vectors only depend on the interference channels, there is no coherent addition of the joint transmission. Although this is not relevant in an isolated cluster at high SNR, this might be an issue with out-of-cluster interference, which can be partially mitigated through interference floor shaping [15]. Finally, we point out that DIA can also be extended to the case of multi-antenna users [14], which is not possible for the current non-iterative distributed linear precoding schemes. 6. Conclusion We have presented a survey of distributed cooperative schemes for the multi-cell downlink scenario. Motivated by the fact that several deployments in future wireless communication networks might have a backhaul with limited capabilities, we consider approaches based solely on local CSI. This assumption eases the burden on the backhaul by avoiding CSI exchange between the cooperating BSs, thus addressing core aspects for small cell integration. For this purpose, we have considered distributed coordination based on F-B training to gradually refine the transmit/receive filters of the nodes in a fully distributed manner. Several relevant issues of such iterative schemes have been addressed like the training overhead, signaling and imperfect CSI. We have also shown how F-B training can be employed to manage interference in a dynamic TDD system. Besides iterative approaches, we have also discussed non-iterative schemes for distributed cooperative transmission. In this context, we have presented a scheme for interference coordination which exploits outdated local CSI. Furthermore, joint transmission based on local CSI has also been discussed where a distributed scheme was presented which outperforms current noniterative approaches at high SNR. 7. Acknowledgment Part of this work has been performed in the framework of the FP7 project ICT METIS, which is partly funded by the European Union. The authors would like to acknowledge the contributions of their colleagues in METIS, although the views expressed are those of the authors and do not necessarily represent the project. 8. References [1] Final performance results and consolidated view on the most promising multi-node/multi-antenna transmission technologies, METIS project deliverable, available at [2] Coordinated multi-point operation for LTE with non-ideal backhaul", 3GPP TR , [3] P. Komulainen, A. Tölli and M. Juntti, Effective CSI signaling and decentralized beam coordination in TDD multi-cell MIMO systems, IEEE Trans. Signal Processing, vol. 61, no. 9, pp , May [4] P. Jayasinghe, A. Tölli and M. Latva-aho, Bidirectional Signaling Strategies for Dynamic TDD Networks, Proc. of IEEE SPAWC, Jun [5] K. S. Gomadam, V. R. Cadambe and S. A. Jafar, Approaching the Capacity of Wireless Networks through Distributed Interference Alignment, Proc. of IEEE GLOBECOM, Dec [6] H. Ghauch, T. Kim, M. Bengtsson and M. Skoglund, Distributed Low-Overhead Schemes for Multi-stream MIMO Interference Channels, IEEE Trans. on Signal Processing, vol. 63, No. 7, Apr [7] E. Lähetkangas, K. Pajukoski, J. Vihriälä et al., Achieving low latency and energy consumption by 5G TDD mode optimization, Proc. of IEEE ICC, Jun [8] H. Shen and B. Li and M. Tao and X. Wang, MSE-based transceiver designs for the MIMO interference channel IEEE Trans. on Wireless Communications, vol. 9, no. 11, pp , Nov [9] S. Smirani and A. Saadani and Y. Yuan-Wu, "On the Distributed Approaches with Imperfect Channel Estimation for MIMO Interference Channel", Proc. of IEEE SPAWC, Jun [10] D. Aziz and A. Weber, "Transmit precoding based on outdated interference alignment for two users multi cell MIMO system," Proc. of IEEE ICNC, Jan

7 [11] 3GPP, Further advancements for E-UTRA physical layer aspects (Release 9), 3rd Generation Partnership Project (3GPP), TR , Mar [12] E. Björnson, R. Zakhour, D. Gesbert and B. Ottersten, Cooperative Multicell Precoding: Rate Region Characterization and Distributed Strategies with Instantaneous and Statistical CSI, IEEE Trans. on Signal Processing, vol. 58, no. 8, pp , Aug [13] N. Golrezaei, A. F. Molisch, A. G. Dimakis, and G. Caire, Femtocaching and device-to-device collaboration: A new architecture for wireless video distribution, IEEE Comm. Magazine, vol. 51, no. 4, pp , Apr [14] Y. Long, N. Vucic and M. Schubert, Distributed Precoding in Multicell Multiantenna Systems with Data Sharing, Proc. of European Wireless, May [15] EU Artist4G project, Deliverable 1.4, Interference Avoidance Techniques and System Design, Jul