Selection strategies for distributed beamforming optimization

EUROPEAN COOPERATION IN THE FIELD OF SCIENTIFIC AND TECHNICAL RESEARCH COST 2100 TD(10)11036 Ålborg, Denmark 2010/June/02-04 EURO-COST SOURCE: Institute of Communiation Networks and Computer Engineering (IKR) Universität Stuttgart Germany Seletion strategies for distributed beamforming optimization Matthias Kashub, Thomas Werthmann, Christian M. Blankenhorn and Christian M. Mueller Institute of Communiation Networks and Computer Engineering (IKR) Universität Stuttgart Pfaffenwaldring 47 D 70569 Stuttgart Germany Phone: +49 711-685-67966 Fax: +1 711-685-57966 Email: matthias.kashub@ikr.uni-stuttgart.de

Seletion strategies for distributed beamforming optimization Matthias Kashub, Thomas Werthmann, Christian M. Blankenhorn and Christian M. Mueller Abstrat Inter-ell interferene (ICI) is a major performane-limiting issue in OFDMA networks. The amount of ICI an be drastially redued if diretional antennas are used and if resoure alloations are oordinated among neighboring base stations. We present an interferene oordination algorithm based on main-lobe steering beamformer antennas. The algorithm is implemented in the base stations and works in a distributed way. In ase a mobile terminal experienes high ICI, a base station tries to find a better resoure alloation and possibly initiates a sequene of resoure alloation hanges in neighboring base stations. Eventually, a better resoure alloation with less ICI is found. In this artile, we fous on different resoure alloation strategies and evaluate their impat on oordination gain. Index Terms Cellular radio aess networks, Interferene oordination Introdution 1.1 Motivation For ellular radio aess networks, the ever-growing demand for apaity makes spetrum a sare resoure. To use the available spetrum as effiiently as possible, these networks use the same frequeny resoures in all base stations (reuse 1) and try to maximize the spetral effiieny using only these resoures. In reuse 1 senarios, the ahievable spetral effiieny is limited by the interferene from neighboring base stations, whereas noise does not play a signifiant role. Therefore, the goal of most interferene oordination mehanisms is to redue the interferene from neighboring ell setors on the transmissions in the urrent ell setor. Espeially for users at the ell edge this improves the hannel quality. We onsider the downlink of ellular OFDMA networks suh as IEEE 802.16e WiMAX or 3GPP LTE. We assume the base station to be equipped with multiple antennas, that are used in ombination with main-lobe steering beamformers. 1.2 Related Work There are several approahes to deal with inter-ell interferene [3]. While lassi stati frequeny reuse sheme as used for GSM are relatively ineffiient, more sophistiated shemes have been proposed. With frational frequeny reuse, a ell is split into different areas. Primary and seondary frequeny resoures are defined and alloated to these areas. The so-alled soft reuse shemes further redue transmission power for seondary resoures [1]. The partitioning of the resoures is usually done during network planning or in a semi-stati way on a large time sale. Higher gain is ahieved with more dynami shemes, whih adapt to traffi variations and user distribution [2]. When base stations are equipped with multi-antenna arrays, multiple signal soures an be distinguished and the spatial domain an be utilized to avoid interferene at the transmitter side [3]. The oordination of transmissions of multiple base stations with diretional antenna systems is denoted as oordinated beamforming in the 3GPP speifiations [4]. In this paper, we investigate an interferene oordination sheme using main-lobe steering antenna arrays in the downlink of an OFDMA mobile ommuniation system. Previous work has shown that large gains an be ahieved by joint optimization of the sheduling and beamformer orientations of multiple base stations [5], [6]. In ontrast to the oordination algorithm presented here, the algorithms in [5], [6] are not fully distributed. An algorithm similar to the presented one is used in [7]. There, interferene in a 60 GHz WPAN is mitigated by hanging the set of used resoures when the transmission is jammed by interferene. The authors did not deal with the situation when all available resoures are used at the same time, whereas we onsider a reuse-1 senario with full transmit power and full system load in all ells. In [8], the authors desribe an interferene oordination algorithm for LTE using odebook-based preoding matries. Eah mobile station measures the hannel to the neighbor ells and signals whih preodings the neighbor ells shall avoid. In ontrast to our algorithm their algorithm requires more measurement by the mobile stations and more signaling between mobile stations and base stations as well as between base stations. The remainder of this doument is strutured as follows. The following setion presents the oordination algorithm. Setion 3 desribes the applied evaluation methodology. The oordination algorithm is evaluated in Setion 4. Setion

5 onludes the paper. 2. Interferene oordination algorithm The overall goal of the presented algorithm is to improve the systems apaity by using beamforming and by direting the beams to the targeted reeivers. Steering the beams provides a degree of freedom, whih we exploit to minimize the overall interferene. Sine the geographial position of the mobiles varies over time, the basestation adjusts the beams ontinuously. We also assume non-greedy traffi, whih imposes additional boundary onditions. 2.1 Assumptions The presented algorithm was designed to be real world implementable. Thus, we assume, mobile stations annot measure single interfering signals but only the interferene sum. In urrent and future ellular networks, mobile stations have or will have information on the RSSI and the CQI. The interferene power and the SIR an be derived from these values. For our simulations, we onsider on-off traffi. In order to derive throughput values, there are always 8 ative mobile stations in eah setor. An extension to different numbers of users is straightforward. Sine we designed the algorithm for low measurement overhead, we do not assume frequeny seletive sheduling. The algorithm uses logial hannels, that are reated by frequeny diverse permutation of sub-arriers to average out fast fading hannel variation. 2.2 Key idea Assigning radio resoures and steering the beams to their assoiated mobile stations is one degree of freedom, that base stations an exploit to improve the interferene situation. Keeping the resoure assignment stati for several subsequent frames gives the base station the possibility to measure the interferene level that its mobile stations reeive. In ontrast, only hanging the resoure assignment an improve a suboptimal interferene situation. These are onfliting goals. Figure 1: Priniple of oordination algorithm The presented algorithm tries to find the optimal point in time for hanging resoure assignment. The approah is to keep resoure assignment over multiple frames as onstant as possible. Fig. 1 shows an example of this behavior for TTIs 1 to 3, where the resoure assignments do not hange although the interferene situation hanges. As soon as the interferene level supersedes a ertain level, the base station hanges the resoure assignment of the affeted mobile stations. In the given example, this is the ase for the purple (third) mobile station after TTI3. Algorithm 1: Interferene mitigation algorithm 1: for all TTI do 2: for all mobile station do 3: Obtain reports of previous TTI (γ and i) 4: Calulate ost funtion: (γ,i) and hange probability p := s() 5: if (random value u(0,1) < p) then 6: Choose pivot partition and perform hange 7: end if 8: Shedule to assigned resoure partition 9: end for 10: end for Algorithm 1 gives a pseudo ode representation of the algorithm. At eah TTI, a base station estimates the interferene situation by evaluating report messages from the mobile stations. In order to find an interferene-minimal resoure assignment in a deentralized fashion, base stations need to hange the resoure assignments and observe the effets of the hange. A base station's deision whether to perform a hange is based on a ost value. So far, we onsidered the following ost funtions α, where γ represents the SINR in db and i the interferene sum in dbw/hz measured for the resoures of the orresponding mobile station:

CONST : IFSUM : SIR : IF SIR : SIR ( γ, i) = 0 ( γ, i) ( γ, i) = i = γ ( γ, i) = i γ CONST IFSUM IF SIR The result of a Bernoulli Experiment determines, whether the base station will perform the resoure assignment hange. The following s-urve transforms the ost values into probability values for the Bernoulli Experiment: 0 2 x α δ 2 2δ s( x) = 2 α x + δ 1 2 2δ 1 x α δ α δ < x α α < x α + δ else The parameters α and δ define the position of the s-urve's turning point and the width of the s-urve. Simulation studies have shown that δ has only little influene on the algorithms behavior, while α diretly influenes the frequeny of resoure assignment hanges, sine it defines the ost threshold of a hange. One a hange deision is made to hange the resoures of a mobile (referred to as swapper mobile), the base stations selets a seond mobile (referred to as pivot mobile) to exhange resoures. Different strategies to selet a pivot mobile will be introdued and evaluated in the next setion. 2.4 Pivot seletion strategies Given the assumption that all ells are fully loaded and the resoures assigned to the mobiles are of equal size. Then seleting a resoure partition is equivalent to seleting another mobile station and swap with its resoure partition. This seletion an either be performed randomly or by taking into aount additional information about the interferene situations. We use the following two soures of additional information: 1) Mobile stations measure the SINR on all resoure partitions and report the results to their basestation. This information allows a basestation to predit the hannel quality a mobile would get on a resoure where it was not served before. The reporting is expensive, sine it is performed on the air interfae. 2) Neighboring basestations ommuniate their sheduling deision among eah other. For eah resoure, they indiate the geographial position of the served mobile. This position is quantized to one of six "subsetors" per setor. We assume that for eah pair of subsetors in the system, the expetation of the interferene reeived at a mobile in the first subsetor from a basestation serving a mobile in the seond subsetor is known. This information an be derived from network planning alulations. From the sheduling information of the basestations and the interferene relations between the subsetors, a basestation an predit the interferene a mobile would get after a hange in the resoure alloation. This predition an be performed for loal mobiles as well as for mobiles served by foreign basestations. The ommuniation between the basestations is performed on the bakhaul and therefore heaper than the signaling from mobiles to base stations. We investigated different strategies for the seletion of the pivot MS, whih will be desribed in the following paragraphs. We apply additional randomization to eah strategy to avoid loal hill limbing. RANDOM The simplest strategy is to hoose a pivot mobile from all mobiles exept the swapper mobile randomly with equal probability. This does not require any additional measurement, reporting or ommuniation. SWAPPER This strategy strives to optimize the interferene level of the swapper mobile. Therefore, for every resoure the hannel quality the swapper mobile would get is estimated (see Fig. 2). The resoure with the best hannel quality is then seleted with the highest probability. The mobile whih is urrently served on the hosen resoures is used as pivot mobile. The strategy ignores whether the pivot mobile gains or suffers from the swapping. SWAPEE This strategy strives to optimize the interferene level of the pivot mobile. For all mobiles exept the swapper mobile, the basestation estimates the hannel quality they would get on the resoure released by the swapper mobile (see Fig. 2). The mobile whih would get the highest gain from getting this resoure assigned is hosen as the pivot mobile. The strategy ignores whether the swapper mobile gains or suffers from the swapping. BOTH This strategy is a ombination of the strategies SWAPPER and SWAPEE. The pivot mobile is hosen suh that the expeted gain for the swapper mobile and for the pivot mobile is optimized jointly. ALTRUISTIC This strategy hanges the resoure assignment to minimize the interferene aused to mobiles in neighboring setors. To predit the interferene whih would be aused to foreign mobiles after a swapping of the resoure alloations, the seond estimation sheme of those introdued above has to be used. This strategy ignores whether swapper or pivot mobile gain or suffer from the swapping. (1) (2)

Figure 2: Relevant hannel quality values for the pivot strategies SWAPPER and SWAPEE. 3. Simulation methodology The performane evaluation has been onduted using an LTE-like system level simulator based on the IKR SimLib [9]. Our senario onsists of 19 three-setorized sites on a regular grid with 1000m inter-site distane. Eah setor is equipped with a main-lobe steering beamformer, onsisting of a linear array of 4 setor antennas. The antenna pattern is derived from measurements of a real setup as desribed in [12]. The influene of satterers is not taken into aount. For eah resoure blok, the main lobe is pointed at the reeiving mobile station. The auray of the beam is 1 and ideal traking of the mobiles is assumed. The 1000 mobile stations are uniformly distributed and move with a onstant speed of 10 km/h. They are equipped with a single isotropi antenna without gain. The pathloss model orresponds to [10]. Shadowing is lognormal with std. deviation 8dB and time-orrelated with orrelation distane 50m and inter-site orrelation fator 0.5 [11]. We do not model any small-sale fading, given that the algorithm is assumed to be used with a frequeny diverse permutation sheme. To avoid any border effets wraparound is used. The 10MHz system bandwidth is divided into 8 equally sized alloation units. The sheduling interval is 1ms and the sheduler assigns one of these alloation units to eah mobile station. The SINR value is alulated for eah alloation unit aording to the sheduling and the beamformer orientations of the other base stations. Only the downlink is regarded. Feedbak in the uplink diretion is assumed to be ideal and without delay. The traffi model in an on-off model with a mean length of the ON phase of 0.2s. During the ON phase, a mobile terminal always has data to send. The traffi model is designed suh that the number of ative mobile stations remains onstant. If the ON phase of one mobile station in a ell ends or a mobile station leaves a setor, one of the previously inative mobile stations in this ell is swithed on. 4. Performane evaluation results Figure 3: Average SINR for different pivot seletion strategies using information soure (1).

Figure 4: Average SINR for different pivot seletion strategies using information soure (2). Fig. 3 shows the average hannel quality over all mobiles in the simulation for different pivot strategies with information soure (1). The parameter alpha, whih is shown on the x axis, determines the probability to hange resoure alloations (as desribed in setion 2.3). The left border of the plot (low values of alpha) orresponds to a swap probability lose to one (random sheduling), while the right border orresponds to swap probability lose to zero (stati sheduling). In both ases, the algorithm annot oordinate, beause the resoure alloation annot onverge. The hannel quality ahieved with this parameterization an therefore serve as referene. A oordination gain an be ahieved with values of alpha in the range of -160 to -140. This orresponds to a swap probability of about 17% to 0.2%. The simplest pivot seletion strategy RANDOM does ahieve a oordination gain of about 0.6 db. This oordination gain is expeted beause the algorithm deides to hange a resoure alloation when a mobile suffers high interferene, only. Even if the pivot mobile is seleted randomly, the algorithm eventually onverges to a good situation. The oordination gain an be inreased by using more sophistiated pivot seletion strategies. The strategies SWAPPER and SWAPEE ahieve a oordination gain of about 1 db. With the strategy SWAPPER, the number of hanges required to get a mobile to the optimal resoure is dereased. However, a single resoure may be the best one for multiple mobiles. In that ase, the mobiles "ompete" for this resoure, whih leads to an osillation and drastially redues the interferene oordination gain in the neighboring setors. In ontrast, the strategy SWAPEE prohibits this ompetition for good resoures, beause it prohibits that the pivot mobile is displaed to a resoure with a low hannel quality. However, the probability that the situation is improved for the swapper mobile is not higher as with the strategy RANDOM. The strategy BOTH, whih ombines the strategies SWAPPER and SWAPEE, leads to a slightly higher oordination gain of about 1.2 db. Even when the pivot mobile is seleted suh that the hannel situation is optimized, eah swap auses unpreditable hanges for the neighboring setors. Therefore, the algorithm annot ahieve any oordination gain when the swap probability is too high. Fig. 4 shows the gain ahieved with seleted pivot strategies in ombination with information soure (2). The performane of the strategy RANDOM is the same as in Fig. 3, beause this strategy does not utilize the additional information. The strategy BOTH performs muh worse than with information soure (1). The reason for this behavior is that the estimation based on the quantized mobile positions reported by the neighbor setors has a low quality. The strategy ALTRUISTIC shows a slightly better result. In ontrast to all other strategies, this strategy strives for a globally optimal resoure alloation by making use of the knowledge of interferene relations between subsetors. It does therefore have the hane to redue the amount of those swaps, whih are done as reation to a swap in a neighboring setor. Nevertheless, it does not take into aount whether the hannel quality of the swapper mobile is improved. In general, the strategies based on a hannel measurement by the mobiles perform better. However, ompared to the interferene estimation based on ommuniation between basestations, they introdue a higher overhead on the air interfae. The strategy ALTRUISTIC leads to the best results when the measurement by the mobiles is not used. 5. Conlusion We proposed a distributed interferene oordination algorithm based on main-lobe steering beamformer antennas. The omplexity of the algorithm is low, and it an operate even without additional hannel measurements of the mobile terminals. Its basi priniple is to keep the resoure alloation of a mobile station stati, unless a predefined interferene and hannel quality threshold is exeeded. In this ase, the resoure alloation of this mobile is hanged, possibly initiating a sequene of resoure alloation hanges in neighboring base stations. Over a few frames, the system onverges to more favorable resoure alloations that ause less inter-ell interferene. In this artile, we presented results for different resoure alloation strategies. The oordination gains ahieved by the

respetive strategies have been evaluated by simulation under an on/off traffi model. It has been shown that even if the time-frequeny resoures of a mobile terminal are seleted by random, a onsiderable improvement in SINR is ahieved. More sophistiated resoure alloation strategies try to predit the impat of a resoure alloation hange on the hannel qualities of mobiles in the same and in neighboring ells. This predition is done using a history of hannel measurements reported by the mobiles in the loal ell or inter-bs signaling in ombination with a oarsegrained stati interferene relation table from network planning. ACKNOWLEDGEMENT The authors would like to thank Hardy Halbauer and his group of the Alatel-Luent Bell Labs Germany and Magnus Proebster for the many valuable disussions and omments on this work. 6. Referenes [1] S. Halpern, Reuse partitioning in ellular systems, in Vehiular Tehnology Conferene, 1983. 33rd IEEE, vol. 33, 1983, pp. 322 327. [2] A. Stolyar and H. Viswanathan, Self-organizing dynami frational frequeny reuse for best-effort traffi through distributed inter-ell oordination, in INFOCOM 2009. The 28th Conferene on Computer Communiations. IEEE, 2009, pp. 1287 1295. [3] G. Boudreau, J. Paniker, N. Guo, R. Chang, N. Wang, and S. Vrzi, Interferene oordination and anellation for 4G networks - [LTE part II: 3GPP release 8], Communiations Magazine, IEEE, vol. 47, no. 4, pp. 74 81, 2009. [4] 3GPP, TR 36.814 - further advanements for E-UTRA physial layer aspets (release 9), v1.0.0, 2009. [5] M. Neker, Interferene oordination in ellular ofdma networks, Network, IEEE, vol. 22, no. 6, pp. 12 19, 2008. [6], A graph-based sheme for distributed interferene oordination in ellular OFDMA networks, in Vehiular Tehnology Conferene, 2008. VTC Spring 2008. IEEE, 2008, pp. 713 718. [7] M. Park, P. Gopalakrishnan, and R. Roberts, Interferene mitigation tehniques in 60 GHz wireless networks, Communiations Magazine, IEEE, vol. 47, no. 12, pp. 34 40, 2009. [8] L. Liu, J. Zhang, J. Yu, and J. Lee, Inter-ell interferene oordination through limited feedbak, submitted to International Journal of Digital Multimedia Broadasting, 2009. [9] Institute of Communiation Networks and Computer Engineering, IKR Simulation Library, http://www.ikr.unistuttgart.de/content/ikrsimlib/, Stuttgart, Germany, 2010. [10] E. Damosso and L. Correia, COST 231 (Digital mobile radio towards future generation systems), Final report, 1999. [11] ETSI, TR 101 112, UMTS 30.03, V3.2.0, 1998. [12] M. Neker, Towards Frequeny Reuse 1 Cellular FDM/TDM Systems, in Proeedings of the 9th ACM/IEEE International Symposium on Modeling, Analysis and Simulation of Wireless and Mobile Systems (MSWiM 2006), Otober 2006. [13], A Novel Algorithm for Distributed Dynami Interferene Coordination in Cellular OFDMA Networks - Communiation Networks and Computer Engineering Report No. 101, Ph.D. dissertation, Universität Stuttgart, 2009.