1260 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 29, NO. 6, JUNE 2011

Transcription

1 26 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 29, NO. 6, JUNE 2 : A Practical Interference Management in Heterogeneou Wirele Acce Network Kyuho Son, Member, IEEE, Soohwan Lee, Student Member, IEEE, Yung Yi, Member, IEEE and Song Chong, Member, IEEE Abtract Due to the increaing demand of capacity in wirele cellular network, the mall cell uch a pico and femto cell are becoming more popular to enjoy a patial reue gain, and thu cell with different ize are expected to coexit in a complex manner. In uch a heterogeneou environment, the role of interference management (IM) become of more importance, but technical challenge alo increae, ince the number of cell-edge uer, uffering from evere interference from the neighboring cell, will naturally grow. In order to overcome low performance and/or high complexity of exiting tatic and other dynamic IM algorithm, we propoe a novel low-complex and fully ditributed IM cheme, called (REFerence baed Interference Management), in the downlink of heterogeneou multi-cell network. We firt formulate a general optimization problem that turn out to require intractable computation complexity for global optimality. To have a practical olution with low computational and ignaling overhead, which i crucial for low-cot mallcell olution, e.g., femto cell, in, we decompoe it into per-bs (bae tation) problem baed on the notion of reference uer and reduce feedback overhead over backhaul both temporally and patially. We evaluate through extenive imulation under variou configuration, including the cenario from a real deployment of BS. We how that, compared to the cheme without IM, can yield more than 4% throughput improvement of cell-edge uer while increaing the overall performance by 7%. Thi i equal to about 95% performance of the exiting centralized IM algorithm (MC-IIWF) that i known to be near-optimal but hard to implement in practice due to prohibitive complexity. We alo preent that a long a interference i managed well, the pectrum haring policy can outperform the bet pectrum plitting policy where the number of ubchannel i optimally divided between macro and femto cell. Index Term Interference management, heterogeneou wirele acce network, femto cell, reference uer, power control, uer cheduling, feedback reduction, ditributed algorithm; I. INTRODUCTION MANY reearcher from networking and financial ector forecat that by 24, the total mobile data traffic throughout the world will grow exponentially and reach about Manucript received 8 June 2; revied 3 December 2. Thi work wa upported by IT R&D program of MKE/KEIT [KI237, Ultra Small Cell Baed Autonomic Wirele Network]. Some part of thi work wa preented at WiOpt 2, Avignon, France. Thi work wa performed while the firt author wa with KAIST a a Ph.D. candidate. K. Son i with the Department of Electrical Engineering, Viterbi School of Engineering, Univerity of Southern California, Lo Angele, CA 989 ( kyuho.on@uc.edu). S. Lee, Y. Yi and S. Chong are with the Department of Electrical Engineering, Korea Advanced Intitute of Science and Technology (KAIST), Daejeon 35-7, Korea ( hlee@lanada.kait.ac.kr, {yiyung, ongchong}@kait.edu). Digital Object Identifier.9/JSAC //$25. c 2 IEEE 3.6 exabyte per month, 39 time increae from 29 [], [2]. Puhed by thi exploive demand mainly from bandwidthhungry multimedia and Internet-related ervice in broadband wirele cellular network, communication engineer eek to maximally exploit the pectral reource in all available dimenion. Small cell uch a pico and femto cell, eem to be one of the mot viable and economic olution [3]. Of recent ignificant interet i the femto cell deigned for uage in a home or an office and deployed by uer, utilizing uer Internet connection, e.g., cable or DSL (Digital Subcriber Line) Internet ervice a a backhaul. Macro and pico cell are controlled by mobile network operator and ue dedicated backhaul. Small cell are alo conidered a a way of incrementally increaing coverage and/or capacity inide the initial deployment of macro cell. In addition to the advantage of pectrum reue efficiency, the network operator are alo attracted by financial benefit becaue mall cell can reduce both capital (e.g., hardware) and operating (e.g., electricity, ite leae and backhaul) expenditure. In heterogeneou multi-cell network with a mixture of macro and mall cell, interference i a major obtacle that can impair the potential gain of mall cell and it pattern i highly divere [4], e.g., interference in macro-to-macro, femto-to-femto, and macro-to-femto. A the number of mall cell increae, the number of uer at cell edge uffering from low throughput due to evere interference alo grow. In particular, femto bae tation (BS) are intalled in an adhoc manner without being planned by uer, not the network operator, which alo increae the technical challenge of interference management (IM). To mitigate interference, a traditional frequency reue or more enhanced cheme uch a fractional frequency reue (FFR) [5] and it variation [6], [7] can be utilized. All thee cheme repreent tatic IM algorithm for pure macro cell network, where a pecific reue pattern i determined a priori by the network operator at offline. However, in reality, BS are not uniformly deployed over the network. In particular, femto BS are purely controlled by uer, which implie that they may often be intalled by uer and even exiting one may be turned on/off dynamically. Under thi ituation, having the tatic reue pattern i naturally inefficient. Recently, everal dynamic IM algorithm have been propoed to addre thi problem [8] [4]. They can ignificantly improve the performance over the tatic cheme, but many of them uffer from prohibitively high complexity and meage paing among neighboring BS.

2 SON et al.: : A PRACTICAL INTERFERENCE MANAGEMENT IN HETEROGENEOUS WIRELESS ACCESS NETWORKS 26 In a ingle-carrier multi-cell network etup, everal BS coordination cheme [8] [] have been propoed under the aumption of binary power control, i.e., each BS tranmit data with it given maximum power or zero. There are everal recent work in a multi-carrier multi-cell network [], [2]. Venturino et al. [] propoed everal algorithm which perform multiple iteration loop in a lot for uer cheduling and power allocation. Although they can achieve near-optimal performance, all of them are centralized algorithm which are too complex to be implemented in practice. Stolyar et al. [2] propoed algorithm that adjut BS power much more lowly than per-lot uer cheduling. Thi time-cale eparation doe implify the problem olution and reduce the complexity, but may lead to non-negligible performance lo. To tackle the cro-tier interference between macro and femto cell, there alo have been everal approache by Sprint, Ericon [3] and Chandraekhar et al. [4] that adjut the tranmit power of femto BS baed on the relative location with repect to the macro BS. The power control alo ha been handled in ad-hoc network. Chiang et al. [5] howed that in high SINR (ignal to interference plu noie ratio) regime nonconvex power control optimization problem can be tranformed into convex optimization problem through a geometric programming technique. In addition, there ha been reearch on mitigation interference in a MIMO (multiple input multiple output) etting [6]. The part of dynamic IM algorithm, in particular, the power control component, wa tudied in the context of multi-tone DSL network (ee [7] and the reference therein for a nice urvey). Indeed, the wired multi-tone DSL model with crotalk can be interpreted a a pecial cae of the wirele multi-carrier cellular model with inter-cell interference when (i) only one uer exit per cell and (ii) wirele channel are tationary and (iii) ditributed operation among BS are not crucial. In fact, the power control of our propoed olution i motivated by ASB (Autonomou Spectrum Balancing) [7], [8] in the DSL network that ue the idea of reference line. However, in the IM over multi-cell wirele network, much more challenging iue till remain for practical implementation, e.g., joint operation with multi-uer cheduling in each cell, dynamic election of reference uer over time-varying channel, and mall meage paing among neighboring cell. The fundamental challenge in the dynamic IM are that (i) BS power control problem itelf (even if cheduled uer in each BS are fixed) i formulated by a highly nonconvex optimization [], [2], [5], (ii) it i tightly coupled with multi-uer cheduling, and (iii) heavy meage paing i uually required to coordinate BS power. In thi paper, we aim at developing an IM cheme coniting of joint power allocation and uer cheduling which i practical in term of low complexity and mall meage paing, but yet the cheme achieve near-optimal performance. Low complexity and mall meage paing i particularly eential for low-cot olution uch a femto BS becaue they are typically made of cheap device for price competitivene and connected to the lowpeed reidential cable or DSL Internet connection, not to the high-peed dedicated backhaul. Another important iue in heterogeneou network i the way of haring pectrum between macro and femto cell. If network operator adopt a pectrum plitting policy where macro and femto cell orthogonally ue the reource for convenience of implementation, they can be free from the macro-to-femto interference. However, uch a plitting policy need to determine the ratio of optimal plitting that varie depending on the configuration of cell, e.g., denity and poition, reulting in pectrum inefficiency. Alternatively, a pectrum haring policy between macro and femto cell can be adopted to maximally reue the reource. However, the macro-to-femto interference may harm the ytem performance unle appropriate IM algorithm are employed. It will be intereting, epecially for the network operator, to anwer which of thee two police i better under what circumtance. Motivated by the above, we propoe a novel IM algorithm, called (REFerence baed Interference Management), whoe core feature are ummarized a follow. ) With the notion of reference uer, each BS can approximate the interference impact of all other cell with a ingle uer to which the BS generate the mot ignificant interference. Thi abtraction ubtantially implifie the problem, reulting in the power control algorithm with low computational and ignaling overhead. 2) Due to the nonconvexity of the power control problem, different initial power etting may lead to different olution. We empirically how that running the power control algorithm with the power ued at the previou lot a the initial power can have effect of removing multiple loop without much performance degradation. 3) In the original power control baed on the reference uer, it require to feedback per-uer information in a cell to the neighboring BS at every lot. We reduce uch a heavy meage paing overhead over backhaul both temporally and patially. 4) ha a nice feature of incremental deployability that partial deployment in ome pecific region, probably tarting from the region that experience mall capacity due to evere interference, ufficiently increae the capacity in thoe region, but not affecting other region. Thi i a deirable property for network operator who cannot afford to upgrade the IM module in all BS. 5) We alo demontrate that, a long a an appropriate IM uch a i adopted, the performance of the pectrum haring policy i much better than the bet performance of the pectrum plitting policy that the number of ubchannel i optimally divided between macro and femto cell. The remainder of thi paper i organized a follow. In Section II, we formally decribe our ytem model and general problem. In Section III, we propoe a reference uer baed power allocation and uer cheduling with feedback reduction idea. In Section IV, we analyze the computational and ignaling complexity. In Section V, extenive imulation under variou configuration demontrate the performance of the propoed algorithm compared to previou algorithm. Finally, we conclude the paper in Section VI.

3 262 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 29, NO. 6, JUNE 2 II. SYSTEM MODEL AND PROBLEM DEFINITION A. Network and Traffic Model We conider a wirele cellular network coniting of multiple heterogeneou BS, where N macro, N pico and N femto are the et of macro, pico and femto BS, repectively. Denote by K =.. {,...,K} and N = N macro N pico N femto = {,...,N} the et of uer and BS, repectively. BS and uer are equipped with one tranmit and one receive antenna, repectively. Each uer i aumed to be connected to a ingle BS. Denote by K n the (nonempty) et of uer aociated with the BS n, i.e., K = K K N and K n K m =ø,for n m. A full buffer traffic model with infinite data packet in the queue for each uer at it aociated BS i ued to conider bet-effort traffic. Macro and pico BS can in general exchange information very fat with each other becaue there are connection between them via high-peed wired dedicate backhaul directly or through a bae tation controller. On the other hand, femto BS can be aumed to exchange information lowly through uer Internet connection a backhaul. B. Reource and Allocation Model We conider a ytem where a ubchannel i a group of ubcarrier a the baic unit of reource allocation. Aume that there are S number of ubchannel and all BS can ue all the ubchannel for data tranmiion, i.e., univeral frequency reue. Denote by S =. {,...,S} the et of ubchannel. We focu on the downlink tranmiion in the time-lotted ytem. At each lot, each BS need to determine (i) which uer i cheduled on each ubchannel and (ii) how much power i allocated for each cheduled uer on each ubchannel. Uer cheduling contraint: In regard to (i), denote by. I (t) = [I k,n (t) : k K,n N] the uer cheduling indicator vector, i.e., I k,n (t) =when BS n chedule it aociated uer k on ubchannel at lot t, andotherwie. Furthermore, we denote the uer cheduled by BS n on ubchannel at lot t by k(n,, t). Reflectingthatatmot only one uer can be elected in each ubchannel for each BS, we hould have: I k,n (t), n N, S. () Power contraint: In regard to (ii), denote the tranmit power of BS n on ubchannel at lot t by p n (t). The vector containing tranmit power of all BS on ubchannel. i p (t) = [p (t), p N (t)] T. In parallel, the vector containing tranmit power of all ubchannel for BS n i p n (t) =[p. n (t), p n S (t)]t. Each BS i aumed to have the total power budget and pectral mak contraint: p n (t) P n,max, n N, (2) S p n n,mak (t) P, n N, S. (3) In practice, a typical tranmit power of macro BS i around 43dBm, which i 2 3dBm higher than that of mall BS. For notational implicity, the time-lot index (t) i dropped unle confuion arie. C. Link Model In thi paper, we do not conider advanced multiuer detection or interference cancellation, and hence the interference from other BS i treated a noie. We focu on the pectrum level coordination, i.e., finding multi-channel power allocation of each BS in order to improve ytem performance by mitigating the interference. For a given power vector p, the received SINR for uer k from BS n on ubchannel can be written a: γ k,n (p )= g k,n p n m n gk,m p m + σ k, (4) where p n and g k,n repreenting the nonnegative tranmit power of BS n on ubchannel and the channel gain between BS n and uer k on ubchannel during a lot, repectively; σ k i the noie power. The channel gain i time-varying and take into account the path lo, log-normal hadowing, fat fading, etc. Following the Shannon formula, the achievable data rate [in bp] for uer k on ubchannel i given by: r k,n (p )= B ( S log 2 + ) Γ γk,n (p ), (5) where B denote the ytem bandwidth; Γ denote the SINR gap to capacity which i typically a function of the deired bit error ratio (BER), the coding gain and noie margin, e.g., Γ= ln(5ber).5 in M-QAM (quadrature amplitude modulation) [9]. Note that r k,n (p ) i the potential data rate when the uer k i cheduled for ervice by BS n on ubchannel and it actual data rate become zero when another uer i cheduled, i.e., r k,n (p, I )=I k,n r k,n (p ). We aume that Γ=and drop B/S mainly for implicity, but our reult can be readily extended to other value of Γ and B/S. D. General Problem Statement Our objective i to develop a lot-by-lot joint power allocation ( ) and uer cheduling algorithm that determine p(t) t= and ( I(t) ) t=, where p(t). = (p (t), S). and I(t) = (I (t), S). The long-term achieved throughput vector R = (R k : k K), where R k = t lim t t τ = S rk,n (p (τ), I (τ)), i the olution of the following optimization problem: (Long-term P) : max U k (R k ) (6) k K ubject to R R, (7) where U k ( ) i a concave, trictly increaing, and continuouly differentiable utility function for uer k; R R K + i the et of all achievable rate vector over long-term, referred to a throughput region. Note that thi optimization problem i challenging to olve, ince we aim at deviing an intantaneou, ditributed algorithm even though the contraint et R i neither available to the uer nor the BS. More performance gain may be achieved by canceling inter-cell interference uing ignal level coordination, uch a CoMP (Coordinated Multi Point Tranmiion and Reception) addreed in the LTE-Advanced, which i beyond the cope of thi paper.

4 SON et al.: : A PRACTICAL INTERFERENCE MANAGEMENT IN HETEROGENEOUS WIRELESS ACCESS NETWORKS 263 With the help of the tochatic gradient-baed technique in [2] that elect the achievable rate vector maximizing the um of weighted rate where the weight are marginal utilitie at each lot, it uffice to olve the following lot-by-lot problem (P) which produce the long-term rate that i the optimal olution of the (Long-term P). (P ): max h(p, I) = w k r k,n (p, I ) (8) p,i k K S ubject to I k,n, n N, S, (9) p n P n,max, n N, () S p n P n,mak, n N, S, () where w k > i the derivative of it utility w k = du(r k ) dr k Rk =R k (t) that can be interpreted a the weight of uer k for the lot. For example, we can et w k a the invere of it average throughput /R k (t) to achieve proportional fairne among uer [2]. Since the ytem objective i a nonconvex function of the tranmit power and i alo tightly coupled with integer variable of the cheduling indicator, the problem (P ) i a mixed-integer nonlinear programming (MINLP). Unfortunately, it i known in [22] that even the implified problem, in which uer cheduling iue i eliminated (i.e., K n = for all n N), i computationally intractable. To find a global optimal olution, we need to fully earch the pace of the feaible power for all BS with a mall granularity along with the all poible combination of uer cheduling. Thu, even for a centralized algorithm, it may not be feaible in practical ytem to olve (P ) at each lot. III. : REFERENCE USER BASED INTERFERENCE MANAGEMENT A. Joint Power Allocation and Uer Scheduling Uer cheduling for fixed power allocation We now preent our propoed approach to olve the problem (P ). Notefirt that for any given feaible power allocation, the original problem can be decompoed into intra-cell uer cheduling problem. Lemma 3.: For any fixed feaible power allocation p, the original problem (P ) can be reduced to N S independent ubproblem for each BS n and ubchannel a follow: max w k I k,n r k,n (p ) (2) I ubject to I k,n. (3) Proof: For the given power allocation p, we can rewrite h(p, I) a follow: h(p, I) = w k n N S = [ n N S I k,n w k I k,n r k,n (p ) ] r k,n (p ). A w k and r k,n (p ) are given parameter, we only have to invetigate dependencie among I k,n. Since the contraint (9) for the given BS n and ubchannel doe not affect the other BS and ubchannel at all, the original problem can be decompoed and i equivalent to individually olving the N S ubproblem in (2) and (3) for each BS and ubchannel. Accordingly, an optimal uer cheduling algorithm under the given power p i eaily obtained by {, if k = k(n, ) = arg max w I k,n k r k,n (p ), =, otherwie. Power allocation for fixed uer cheduling (4) For any given uer cheduling I, the original problem reduce to the following power allocation problem: max p ubject to w k(n,) log 2 (+ n N S p n P n,max, S g k(n,),n p n m n gk(n,),m p m +σ k(n,) n N, p n P n,mak, n N, S. Solving the above problem require the knowledge of all interference channel gain acro cell and noie power, forcing it to operate in a centralized fahion. To overcome thi complexity and develop a ditributed cheme with mall meage paing, we introduce the concept of reference uer. LetN (n) be the et of neighboring BS 2 of BS n, and further denote by A(n, ) the et of all cheduled uer on ubchannel in N (n), i.e., A(n, ) = {k(m, ) m N(n)}. The reference uer for BS n on ubchannel i defined a the uer among A(n, ) which ha the tronget channel gain between the uer and BS n on ubchannel. We eparately denote the index of BS to which the reference uer belong and the reference uer by ref n and k(ref n,), repectively. We will elaborate on the way of chooing the reference uer hortly at the end of thi ubection. Once the reference uer i fixed, each BS trie to find it own power allocation taking into account jut one reference uer per ubchannel intead of olving the above problem conidering all (N number of) cochannel uer in the network. Thi approximation come from the intuition that jut conidering the wort-cae uer may be a good approximation of the cae when all the uer are included. The problem (P n ) 2 If there i any chance that uer in BS n will handed over to certain adjacent BS, then we conider uch a et of BS a the neighboring BS of BS n, denoted by N(n). Thi et can be determined a priori by mobile network operator at the time of deployment and/or maintained in online baed on the ignal trength between the BS. )

5 264 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 29, NO. 6, JUNE 2 to be olved by each BS n can be written a follow. ( ) (P n ): max w n p log g n 2 + pn n g n,m p m + σn ubject to S m n + S w refn log 2 (+ m ref n g refn p refn g refn,m p m + σrefn (5) p n P n,max, (6) S p n P n,mak, S. (7) Note that we replace the index of cheduled uer k(m, ) with the correponding index of BS m to keep our notation imple. 3 For any given uer cheduling and reference uer election, the correponding optimal power allocation mut atify Karuh-Kuhn-Tucker (KKT) condition [23]. In particular, let λ n denote the nonnegative Lagrange multiplier aociated with the total power budget contraint (7). Then, the optimal λ n and p n mut atify the following equalitie: p n = [ w n λ n ln 2 + t n where t n = m n gn,m p m + σ n ]P n,mak g n, w refn g refn,n γ refn l N grefn,l p l + σrefn, ) (8) λ n( p n P n,max) =, (9) where [ ] b. a = min [max [,a],b] and γ refn ( ) i the received SINR a defined earlier in (4). Note that the t n term can be interpreted a a taxation (or penalty) term. If the reference uer i cloe to the BS n (i.e., high interference channel gain g refn,n ), then the value of t n increae. Conequently, BS n lower it power level to reduce the harm to the reference uer. Since the modified problem (P n ) i till nonconvex [], [5], (8) and (9) are the firt-order neceary condition. Therefore, there might exit a duality gap to the optimal primal olution. However, encouraged by the tate-of-the-art aymptotic reult [22] that the duality gap become zero when the number of ubchannel i large, we develop an effective approximation algorithm for the problem (P n ) baedonthe condition (8)-(9). Note that a fixed point equation of p n in (8) i a monotonic function of λ n. Thu, it can be olved via a fat biection method. Starting from an initial power allocation and λ n, we calculate the power p n and taxation term t n in (8) for all ubchannel. If the um of updated power exceed P n,max,thenλ n i increaed. Otherwie, λ n i decreaed. With the updated power, we repeat thi until the equation (9) hold. If no poitive value of λ n matche the equality, then λ n i et to be zero. In the latter cae (interfering too much), the BS n doe not ue all of it available power. 3 We ue the index of BS intead of the index of cheduled uer a follow: w n w k(n,), g n gk(n,),n, g n,m g k(n,),m, γ n γ k(n,),n, σ n σk(n,) g refn,m, w refn w k(ref n,), g refn g k(refn,),m, γ refn γ k(refn,),refn and σ refn g k(refn,),refn, σ k(refn,). TABLE I GENERAL ALGORITHM DESCRIPTION : Power initialization 2: repeat (uer cheduling loop): 3: Uer cheduling 4: Neighboring environment abtraction 5: repeat (power allocation loop): 6: Update the effect to the neighboring environment 7: Power allocation 8: until p converge or max # of iteration i reached 9: until I converge or max # of iteration i reached Remark 3.2: If the taxation term t n i ignored, our power allocation algorithm i reduced to the water-filling (WF) algorithm [24], where each BS act elfihly in order to maximize it own performance. By adding thi term t n >, each BS operate in a ocial way by conidering the reference uer and lower the water-filling level, which could lead to a globally better olution. General algorithm decription: joint uer cheduling and power allocation TABLE I decribe a conceptual peudo-code of the general algorithm for our problem. At each lot, each BS tart from a proper power allocation. For the given power, each BS firt execute the uer cheduling and then abtract what i happening in the neighboring environment. For example, the concept of reference uer can be ued a one of the abtraction method. For thi given uer cheduling and neighboring environment abtraction, the BS iteratively update it power and effect to the neighboring environment until p converge or the maximum number of iteration i reached. Then each BS repeat the uer cheduling and the neighboring environment abtraction for the updated power and goe into the power allocation loop again. Thi procedure i repeated until the uer cheduling I converge or the maximum number of iteration i reached. The general algorithm not only ha a prohibitively high computational complexity due to inner and outer loop, but alo require multiple information exchange per lot between BS to reflect the updated interference level followed by the updated power. However, the multiple feedback in a ingle lot are practically impoible becaue each uer can end it own meaurement information to the BS once per lot. To overcome thi complexity, we will propoe a implified algorithm in ubection III-E, which execute uer cheduling and power allocation tep-by-tep without loop. B. Online Reference Uer Selection Method Baed on the notion of reference uer, each BS need to conider the only one uer intead of all cheduled uer in the network on each ubchannel. From BS n point of view, although it tranmit power interfere with all the cheduled uer on the correponding ubchannel in other BS, the effect will be critical epecially for the the uer who ha the tronget channel gain (or the mot influenced uer from BS n). Thi dominant victim uer can be found in the neighboring (or adjacent) BS N (n). Therefore, we propoe an online reference uer election

6 SON et al.: : A PRACTICAL INTERFERENCE MANAGEMENT IN HETEROGENEOUS WIRELESS ACCESS NETWORKS 265 BS3 BS4 BS2 BS BS5 BS7 BS6 Step. Receive Information about the cheduled uer from neighborign BS. Fig.. Step 2. Select a reference uer. Online reference uer election method. BS ` BS7 Step 3. Solve an approximate per-bs power allocation optimization problem. method, in which each BS n chooe the reference uer on each ubchannel a follow: Reference Uer Selection Rule k(ref n,) i the reference uer on ubchannel, where ref n =arg max m N (n) gk(m,),n. (2) It i worthwhile mentioning that the reference uer i independently and locally elected by each BS on each ubchannel and thu no centralized coordination i neceary. Fig. depict an example of our online reference uer election procedure. The rule in (2) provide a guideline for chooing a reference uer. There may be other method uch a (i) making one virtual uer by averaging channel of the cheduled uer from neighboring BS, and (ii) electing multiple reference uer (e.g., elect the M wort uer). A we will how later in ubection V-A, uch variation do not lead to the high performance improvement, compared to the high increae in complexity. C. Feedback Reduction To determine a reference uer at each lot, each BS n require (F) the channel gain g k(m,),n of the cheduled uer in neighboring cell m N(n). We call thee uer the candidate of reference uer. Once the reference uer i elected, to calculate a taxation term for power control, additional information about the reference uer i neceary. The following are the required information for the reference uer: (F) The weight of the reference uer w k(ref n,), (F2) The received ignal trength of the reference uer g k(refn,),refn p refn, (F3) The noie plu interference trength of the reference uer m ref n gk(refn,),m p m + σ k(refn,). The information for the candidate need to be collected by neighboring BS and be forwarded to BS n. However, femto BS ue their Internet connection a backhaul, and thu meage exchange may not be reliable and fat enough. Beide, there i no guarantee on the feedback latency. Even for macro BS with a dedicated backhaul network, the per-lot meage exchange may be a large overhead. We preent a more practical olution to reduce the backhaul feedback overhead both temporally and patially. Temporal feedback reduction: Intead of the per-lot meage exchange for all the information, (i) each candidate uer firt calculate the time-average of the information and end them to it aociated BS infrequently (ay, every T lot), (ii) and then the BS broadcat the information about all candidate uer to it neighboring BS through wired backhaul. The only thing that the BS exchange at each lot i the indexe of the cheduled uer. Each BS will maintain a table that contain thee average of candidate uer. Once each BS receive the indexe of cheduled uer from neighboring BS, then it ue the information in the table correponding to the indexe. Note that the low feedback mechanim can be applied to femto BS. Spatial feedback reduction: Firt, we reduce the amount of infrequent feedback by making macro BS end the information only for edge uer. Thi idea come from the intuition that the uer in the center of cell are not likely to be elected a the reference uer becaue the center uer do not receive too much interference. Second, we eliminate the per-lot feedback of indexe of cheduled uer for femto BS. The indexe of cheduled uer (very mall amount of information) can be eaily exchanged between neighboring macro BS at each lot through dedicated backhaul. However, thi per-lot meage exchange i not poible for femto BS becaue their backhaul do not provide any guarantee on the feedback latency. To overcome thi difficulty, we propoe an alternative olution for the femto BS that doe not require per-lot meage exchange at all. In the propoed olution, the femto BS obtain the indexe of cheduled uer by overhearing downlink control meage (e.g, DL-map in the IEEE 82.6e [25]) from neighboring macro BS. Thi may need light modification frame tructure for the femto BS in the current wirele tandard. The remaining challenge i on the revere direction, i.e., ending the indexe of cheduled uer in the femto BS to the neighboring macro BS. We pay attention to one of the main feature of femto cell, that i, the mall coverage. In other word, the ditance between uer in a femto BS are relatively much horter than the ditance from the neighboring macro BS. Thu, from neighboring macro BS perpective, it eem reaonable to aume that the uer in the femto cell are patially located at the ame point. Baed on thi patial implification, the neighboring macro BS imply can pick any uer in the femto cell and conider a the candidate uer. D. Initial Power Setting Our algorithm require an initial power value to compute the power allocation (ee line in Table I). Since our problem i a nonconvex problem, different tarting point may lead to different olution with different peed. The following three trategie for the choice of initial power are carefully invetigated in thi paper. ) Uniform rule. The power allocation alway tart from the ame point for every lot. Each BS uniformly plit it maximum tranmiion power to all ubchannel, i.e., (t) =P n,max /S. p n,init

7 266 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 29, NO. 6, JUNE 2 TABLE II COMPLEXITY COMPARISON OF VARIOUS ALGORITHMS Algorithm Computational complexity Signaling complexity (inter-bs) Uer cheduling Power allocation Per-lot feedback Periodic feedback EQ O(SK) Zero Zero Zero WF O(SK) O(S) or O Zero Zero log 2 P max ɛ MGR O(SK) O(S)+n n p vo(sk) Zero N n S P O(SK) O(S) or O log max 2 macro: ρs &femto:zero ρ K ɛ n AS MC-IIWF T (O(SK)+O(SN)) Complete information i aumed. 2) Random rule. Each BS randomly chooe the initial power level for each ubchannel between and P n,max /S, and then each BS cale it up with an appropriate weight P n,max / pn,init (t) to ue up the total tranmiion power budget, i.e., pn,init (t) =P n,max. 3) Previou rule. Each BS tart from the power ued at the previou lot, i.e., p n,init (t) =p n (t ). We will demontrate later in ubection V-A that although we avoid multiple loop in a lot for power allocation and uer cheduling, the performance of the previou rule i likely to remain unchanged while the other rule loe the performance much. Thi i becaue in ome ene the previou rule exploiting temporal correlation ha the effect of iteration for power allocation in a lot-by-lot manner. Thi reult encourage u to deign a implified algorithm in ubection III-E, which get rid of multiple loop in a lot and execute uer cheduling and power allocation equentially with the previou rule. E. : Reference Baed Interference Management We now propoe our final algorithm, called (REFerence baed Interference Management), in Table III that merge all the component developed in the above. adopt the previou rule for initial power etting (ee line ), and ue the notion of reference uer for the neighboring environment abtraction and limit the number of reference uer to one (ee line 3). While the general algorithm ha uer cheduling and power allocation loop (ee line 2 and 5 in Table I), execute uer cheduling (ee line 2) and power allocation (ee line 5) equentially without loop. Thi tep-by-tep approach can not only be done very fat in a lot, but it require the feedback from each uer jut once per lot. Through extenive imulation, uch a imple algorithm will be hown to be efficient. IV. COMPLEXITY ANALYSIS In thi ection, we analyze the computational complexity and inter-bs ignaling complexity of compared to conventional equal power allocation (EQ) and elfih waterfilling (WF), a well a MGR (Multi-ector GRadient) [2] and MC-IIWF (MultiCell Improved Iterative Water Filling) []. Table II ummarize the reult of complexity analyi. Computational complexity conit of two part: the complexity from uer cheduling and power allocation. Uer cheduling ha a linear complexity O(SK) with the number TABLE III : REFERENCE BASED INTERFERENCE MANAGEMENT : Power initialization p n (t) p n (t ) 2: Uer cheduling according to (4): k(n, ) = arg max w k r k,n (p n ). 3: Reference uer election according to (2). 4: Taxation update according to (8): 5: Power allocation via biection: [a, b] [,λ n max]. while P pn P n,max <δ, Set λ n =(a + b)/2 and update p n according to (8). if P pn >P n,max, then [a, b] [λ n,b], ele P pn <Pn,max, then [a, b] [a, λ n ]. end while of uer for each ubchannel for all algorithm except MC- IIWF. For power allocation, EQ ha zero complexity. WF can be obtained by either an exact algorithm O(S) or an iterative algorithm (i.e., a biection method that converge to a olution with a certain error tolerance ɛ) O ( P log max ) 2 ɛ [24]. The only difference between WF and i the taxation term conidering the reference uer. Thu, the complexity of power allocation for i baically the ame a that for WF. MGR in [2], one of the tate-of-the-art dynamic IM algorithm, adjut the power allocation for every n p > lot (relatively lowly) and condene the complexity for updating power to /n p. However, MGR ha additional complexity from a virtual cheduling that need to be run n v time per lot. Accordingly, the total computational complexity for power allocation i high, n p O(S) +n v O(SK). Another recently developed MC-IIWF in [] i the centralized algorithm that ha iteration loop for power allocation and uer cheduling. Let T be the number of iteration needed for iteration loop. Then, the total computational complexity i equal to T (O(SK)+O(SN)). Now let u invetigate the inter-bs ignaling complexity. EQ and WF do not require any inter-bs meage paing overhead becaue they are autonomou algorithm without conidering neighboring BS, but at the cot of performance reduction, a will be hown in Section V. MGR adjut the power allocation lowly o that it require not per-lot but periodic feedback, N n S (enitivity information for neighboring BS and all ubchannel). MC-IIWF, a centralized algorithm, aume a central control unit to have complete information. require the periodic feedback about the candidate uer for the reference uer, ρ K n AS, whereρ i the average percentage of edge uer and A =4i the number of required information (F) (F3) about the reference uer. Note that

8 SON et al.: : A PRACTICAL INTERFERENCE MANAGEMENT IN HETEROGENEOUS WIRELESS ACCESS NETWORKS 267 while macro BS require the per-lot feedback for the indexe of cheduled uer at each ubchannel, femto BS do not. In ummary, the computational complexity of i the ame a that of WF and i much lower than that of tate-ofthe-art dynamic IM algorithm uch a MGR and MC-IIWF. For ignaling complexity, although the feedback per lot-wie manner i challenging, the information need to be exchanged only between neighboring macro BS at each lot are jut the indexe of cheduled uer. We believe that uch mall amount of information can be eaily exchanged through high-peed dedicated backbone. V. PERFORMANCE EVALUATION We verify the ytem performance through extenive imulation under variou topologie and cenario. Firt, in order to verify the effectivene of by varying everal tunable parameter and comparing it with other algorithm, a two-tier macro-cell network compoed of 9 hexagonal cell i conidered in ubection V-A. Second, in order to provide more realitic imulation reult, a real 3G BS deployment topology coniting of heterogeneou environment (urban, uburban and rural area) i conidered in ubection V-C. Third, a heterogeneou network topology with mall cell inide macro cell i alo conidered in ubection V-D. In our imulation, macro and mall cell are loaded with 2 and 4 uer, repectively and they are uniformly ditributed in each cell. All uer are aumed to have a logarithmic utility function, i.e., U(R k )=logr k, but the other utility function enforcing more fairne are alo conidered in our technical report [26]. We conider a ytem having 6 ubchannel each of which conit of multiple ubcarrier. The maximum tranmit power of macro and mall BS are 43dBm and 5dBm [27], repectively. In modeling the propagation environment, an ITU PED-B path lo model log (d[m]) for macro cell and an indoor path lo model log (d[m]) for mall cell with db penetration lo due to wall are adopted. Jake Rayleigh model (with the peed of 3km/h), where the channel coefficient vary continuouly over lot, i adopted for fat fading. The channel bandwidth and the time-lot length are et to be MHz and m, repectively. i compared to conventional EQ and WF, a well a MGR [2] and MC-IIWF [] developed recently. A performance metric, the geometric average of uer throughput (GAT) and the average of edge uer throughput (AET) are ued. We ue GAT ince maximizing thi metric i equivalent to our ytem objective (um of log throughput). We conider AET a the average of the bottom 5% of uer throughput (i.e., 5th percentile throughput), which can be regarded a a repreentative performance metric of cell-edge uer. A. Effectivene of Propoed Algorithm Fig. 2(a) how the GAT performance of for the number of reference uer per ubchannel and that of EQ a a baeline. A mentioned in Remark 3.2, without a reference uer i reduced a WF. taking reference uer into conideration can obtain higher performance gain. It i noteworthy that conidering only the one reference uer per ubchannel i efficient enough becaue it can obtain more Fig. 2. GAT [Mbp] GAT [Mbp] WF EQ Number of reference uer per ubchannel (a) The number of reference uer Previou Uniform Random (X,X) (X,O) (O,O) (uer cheduling loop, power allocation loop) (b) Iteration loop and initial power Effect of everal tunable parameter. than 97% of the performance conidering all the ix neighbor. Fig. 2(b) how the effect of iteration loop and initial power: (i) adding uer cheduling loop and/or power allocation loop give additional performance gain from any initial power etting and (ii) uing power at the previou lot a an initial power outperform other two trategie. However, for the cae in which power at the previou lot i ued a an initial power, the performance gain from adding power allocation loop i marginal. Fig. 2(b) i an encouraging reult that lead u to deign the algorithm without loop and ue the previou power a an initial power. We conjecture that thi i becaue in ome ene uing the previou power rule ha the effect of iteration not in a lot but over a erie of multiple lot. To verify thi tatement, we provide additional imulation in a linear two-cell network where the ditance between BS i 2km. Each cell ha two group of uer: center and edge uer, whoe ditance from the aociated BS are within 2-4m and 7-9m, repectively. We adopt the imple network configuration to gain inight eaily a well a for eay of preentation, but all dicuion can be extended to the general cae. Fig. 3(a) how the time-erie of tranmit power on different ubchannel. Although the power do not eem to quite converge, they remain the certain level for everal dozen of lot. In Fig. 3(b), we plot the average tranmit power level on

9 268 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 29, NO. 6, JUNE 2 Average tranmit power [W] Tranmit power [W] Tranmit power [W] Slot BS BS Slot (a) Time erie of tranmit power for different ubchannel BS BS Subchannel index (b) Average tranmit power between 75 and 25 lot Fig. 3. Tranmit power level for different ubchannel in a linear two-cell network. different ubchannel during the period between 75 and 25 lot. A can be clearly een, each BS excluively utilize five ubchannel of high power and hare ix ubchannel of low power with the other. We further invetigate the relationhip between the uer group and their ubchannel on which they are cheduled. Interetingly, for the mot of time (more than 98% of lot), the center and edge uer are erved by the et of ubchannel with low and high power, repectively. In other word, it i highly likely that on each ubchannel a uer will be elected by the cheduler who ha imilar channel condition to the uer elected at the previou lot. Thu, given imilar uer cheduling over conecutive lot, the power allocation with the previou power can be interpreted a if it ha iteration over a erie of multiple lot. Thi explain why that execute uer cheduling and power allocation tep-by-tep in a lot can achieve a good olution without much performance degradation. In Fig. 4, we alo tet the effect of outdated feedback information about reference uer. We conider two type of uer with different peed: nomadic uer (or tationary uer that have fixed path lo and hadowing factor, but have time-varying fat-fading) and mobile uer (moving fat with peed of 6km/h). For nomadic uer, the GAT performance degradation i relatively mall, even though we chooe a Fig. 4. Fig. 5. GAT [Mbp] CDF Nomadic uer Mobile uer Period of feedback The effect of outdated feedback information MC-IIWF 43% EQ MGR WF MC-IIWF Throughput [Mbp] Comparion with other algorithm. long feedback period uch a 2 lot. For mobile uer, the GAT performance naturally decreae due to the error of feedback information a the feedback period increae. In practical ytem, different type of uer with different peed are expected to coexit. In uch an environment, to reduce the amount of feedback while maintaining the performance degradation marginal, it i eential to adaptively control the period of feedback for the different peed uer, e.g., T = 2 lot for nomadic uer ( 3km/h) and T =lot for mobile uer ( 6km/h) 4. B. Performance Comparion with Other Algorithm Now we compare the performance of with other four algorithm: conventional EQ, elfih WF, MGR [2] which adjut power allocation infrequently, compared to perlot bai uer cheduling, and MC-IIWF [] which i a centralized algorithm achieving the near-optimal performance. Fig. 5 how the CDF (cumulative ditribution function) of the throughput of entire uer in the network for different algorithm. Compared to EQ, WF and MGR, can improve the throughput for all uer in the network. Particularly, we can oberve higher improvement (43% improvement in 4 We ue the implified verion of random waypoint model [28], where each uer tart from an initial point, randomly chooe it detination, and move toward it at a given peed. After reaching the detination, it repeat thi proce unite the end of imulation time.

10 SON et al.: : A PRACTICAL INTERFERENCE MANAGEMENT IN HETEROGENEOUS WIRELESS ACCESS NETWORKS 269 [km] 5 Urban Suburban Rural GAT [Mbp] EQ WF.4 Fig [km] Real 3G BS deployment map..2 Urban Suburban Rural (a) GAT (geometric average of uer throughput) AET compared to EQ) for uer achieving low throughput, i.e., uer at cell edge. Thi i due to the fact that IM i mainly targeted for performance improvement of celledge uer. In addition, can achieve about 95% of the performance of near-optimal MC-IIWF in term of two repreentative throughput metric (GAT and AET) a well a the arithmetic average of uer throughput (AAT). It i omewhat urpriing that uch a imple ditributed algorithm can obtain a imilar performance to the centralized algorithm that i hard to implement due to prohibitive complexity. AET [Mbp] EQ WF C. Topology with Real BS Deployment: Urban, Suburban and Rural Environment Fig. 6 depict the map of BS layout that we ue for more realitic imulation. It i a part of real 3G network operated by one of the major mobile network operator in Korea. There are a total of 3 BS within 2 km 2 rectangular area. We aume that the number of BS per unit area i proportional to the uer denity. In other word, the average number of uer per cell i almot imilar becaue BS in an urban environment cover a mall area and BS in a rural environment a large area. Under thi aumption, we generate uer one-by-one in the rectangular area and attach them to the cloet BS until each BS will have 2 uer. We chooe thi partial map to include a challenging cenario that three environment are mixed together. We teted everal other map, and obtained imilar or even better performance of. We examine three different zone: urban (5 BS in km 2 ), uburban (5 BS in 2 6 km 2 ) and rural (8 BS in 9 9 km 2 ) area 5. Fig. 7 how GAT and AET performance under urban, uburban and rural environment. A expected, we can obtain high performance improvement in the urban and uburban environment. However, almot low or no gain i found in the rural environment, which mean that the IM doe not take much effect in a pare topology, which follow our intuition. For example, in the urban area, the performance gain of are 6% and 42% in term of GAT and AET, repectively. Another nice feature of i incremental deployment. Suppoe that we implement our algorithm only on the BS 5 In order to ee clearly how the denity of BS affect the performance gain, the ditance between BS in uburban and rural zone are increaed by.5 and 2 time, repectively. Fig. 7. Fig. 8. Normalized gain Urban Suburban Rural (b) AET (average of edge uer throughput) Throughput performance in real BS deployment topology Partial Effect of partial deployment. Full in a pecific area. While the BS inide thi area perform well a we want, the BS in the boundary of the area doe not. Thi i becaue they may not receive information about reference uer from the ome of it neighboring BS on which our algorithm i not implemented. Even in uch a cae, our algorithm will automatically reduce to WF. Thu, it perform like WF at leat and better than EQ. Compared to the full deployment cae, the partial deployment cae where only 5 BS (mainly elected from the urban area among 3 BS) are equipped with can achieve more than 85% gain a hown in Fig. 8. The reult encourage the mobile

11 27 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 29, NO. 6, JUNE 2 Ditance [m] Only one mall BS Two ymmetric mall BS Two aymmetric mall BS Macro BS3 Macro BS Macro BS Ditance [m] (a) Heterogeneou network topology 3 2 Home Home 2 Small BS A Small BS B : Symmetric cae : Aymmetric cae GAT [Mbp] EQ WF Spectrum haring 6% 38% GAT [Mbp] Spectrum haring 7% 56% EQ WF 2 Spectrum plitting 2/6 4/6 6/6 8/6 /6 2/6 4/6 Spectrum plitting ratio (b) Five mall cell per a macro cell 2 Spectrum plitting 2/6 4/6 6/6 8/6 /6 2/6 4/6 Spectrum plitting ratio (c) Ten mall cell per a macro cell Fig. 9. Performance in the heterogeneou network topology. network operator to upgrade it BS incrementally from the urgent one, e.g., denely located BS experiencing heavy interference, and thu low capacity. D. Heterogeneou Network (Macro + Small BS) Now we conider a heterogeneou network topology having everal mall cell inide macro cell a hown in Fig. 9(a). A a kind of mall BS, we conider the femto BS that are deployed and provide a high-peed indoor acce mainly to home uer. We need to reflect femto-to-femto cell interference a well a macro-to-macro and macro-to-femto cell interference. To thi end, we conider the mixture of three femto BS deployment cae: (i) only one femto BS cae (no femto-to-femto interference), (ii) two ymmetric femto BS cae (trong femto-to-femto interference): two femto cell are adjacent with each other and each femto BS i located in the center of the home, (iii) two aymmetric femto BS cae (very trong femto-to-femto interference): two femto cell are adjacent with each other and femto BS in home i located at the border between home. The lat cae can often happen becaue uer locate their own femto BS wherever they want without conidering next door neighbor. In Fig. 9(b) and 9(c), we compare the performance of the pectrum haring policy (i.e., univeral frequency reue) between macro and femto cell with that of the pectrum plitting policy where macro and femto cell orthogonally ue the reource. Note that the performance curve for the pectrum haring policy and the pectrum plitting policy are repreented by olid and dotted line, repectively. The x-axi repreent the ratio of ubchannel ued by macro cell among all 6 ubchannel. In the cae of five femto cell per a macro cell in Fig. 9(b), there exit mall cro-tier (macro-to-femto) interference. Thu, the pectrum haring policy i alway better than the pectrum plitting policy. For example, even the performance of EQ without any interference management in the pectrum haring policy i higher than or equal to the performance of in the optimal pectrum plitting (at 8/6). If the number of femto cell increae, then the portion of macro uer who will ee more and cloer femto cell increae. Conequently, it i highly probable that their performance are degraded by the evere cro-tier interference. A hown in Fig. 9(c) with ten femto cell per a macro cell, the bet performance in the optimal pectrum plitting (at 6/6) can catch up with the that of EQ and WF in the pectrum haring policy. However, if the propoed IM algorithm, e.g.,, i adopted in the pectrum haring policy, then we can mitigate the cro-tier interference, reulting in the better performance

12 SON et al.: : A PRACTICAL INTERFERENCE MANAGEMENT IN HETEROGENEOUS WIRELESS ACCESS NETWORKS 27 than any cae in the pectrum plitting policy. Note that the performance improvement of our in the pectrum haring policy compared to EQ in the pectrum haring and pectrum plitting policie are 38% and 6% in Fig. 9(b), and they become larger a the number of femto cell increae, i.e., 56% and 7% in Fig. 9(c). The pectrum plitting policy i expected to be widely ued rather than the pectrum haring policy in an early tage of femto cell deployment becaue it make the femto cell eaily coexit with macro cell without worrying about the macro-to femto interference. However, the reult in Fig. 9 enlighten u on the potential gain of the pectrum haring policy. Therefore, we believe that, if the IM algorithm become more mature in the near future, then the pectrum haring policy will be adopted in order to maximally exploit the pectral reource due to the exploive traffic demand VI. CONCLUSION Heterogeneou acce network, coniting of cell with different ize and ranging from macro to femto cell, will play a pivotal role in the next-generation broadband wirele network. They can increae the network capacity ignificantly to meet the exploive traffic demand of uer with limited capital/operating expenditure and pectrum contraint. One of the bigget challenge in uch environment i how to effectively manage interference between heterogeneou cell. To tackle thi challenge, thi paper developed, which i an efficient low-complex and fully ditributed IM in downlink heterogeneou multi-cell network. Our key idea i to ue the notion of reference uer, which can patially implify the impact of all other neighboring cell by a ingle virtual uer and reult in the power control algorithm with low computational and ignaling overhead. In order for to be implemented even on the femto BS, we alo further reduced the feedback over backhaul both temporally and patially. Through extenive imulation and complexity analyi, we demontrated that not only perform well but alo i practically implementable. We alo concluded that a long a appropriate IM algorithm uch a are adopted, the pectrum haring policy can outperform the bet pectrum plitting policy where the number of ubchannel i optimally divided between macro and femto cell. ACKNOWLEDGMENTS The author would like to thank the anonymou reviewer for their helpful comment that greatly improved the quality of thi paper. The author alo would like to thank Prof. Mung Chiang, Prof. Jianwei Huang and Dr. Pachali Tiaflaki for their helpful dicuion. REFERENCES [] Cico viual networking index: Global mobile data traffic forecat update, 29-24, Feb. 2, [Online] Available: n75/n827/whitepaperc html. [2] Data, data everywhere, Feb. 2, [Online] Available: economit.com/pecialreport/diplaystory.cfm?toryid= [3] V. Chandraekhar, J. G. Andrew, and A. Gatherer, Femtocell network: aurvey, IEEE Commun. Mag., vol. 46, no. 9, pp , Sept. 28. [4] 3G home nodeb tudy item technical report, 3rd Generation Partnerhip Project (3GPP), TR25.82, v8.2., Aug. 28. [5] R. Giuliano, C. Monti, and P. Loreti, WiMAX fractional frequency reue for rural environment, IEEE Commun. Mag., vol. 5, pp. 6 65, June 28. [6] K. Son, S. Chong, and G. de Veciana, Dynamic aociation for load balancing and interference avoidance in multi-cell network, IEEE Tran. Wirele Commun., vol. 8, no. 7, pp , July 29. [7] Soft frequency reue cheme for UTRAN LTE, 3GPP Std. R-5 57, May 25. [8] S. Da, H. Viwanathan, and G. Rittenhoue, Dynamic load balancing through coordinated cheduling in packet data ytem, in Proc. IEEE INFOCOM, San Francico, CA, Mar. 23, pp [9] K. Son, Y. Yi, and S. Chong, Adaptive multi-pattern reue in multi-cell network, in Proc. WiOpt, Seoul, Korea, June 29, pp.. [] A. Gjendemj, D. Gebert, G. E. Øien, and S. G. Kiani, Binary power control for um rate maximization over multiple interfering link, IEEE Tran. Wirele Commun., vol. 7, no. 8, pp , Aug. 28. [] L. Venturino, N. Praad, and X. Wang, Coordinated cheduling and power allocation in downlink multicell OFDMA network, IEEE Tran. Veh. Technol., vol. 58, no. 6, pp , July 29. [2] A. L. Stolyar and H. Viwanathan, Self-organizing dynamic fractional frequency reue for bet-effort traffic through ditributed inter-cell coordination, in Proc. IEEE INFOCOM, Rio de Janeiro, Brazil, Apr. 29, pp. 9. [3] Home NodeB Output Power, 3GPP TSG Working Group 4 meeting TSG-RAN WG Contribution R , June 27. [Online]. Available: ran/wg4 Radio/TSGR4 43bi/ Doc/. [4] V. Chandraekhar and J. G. Andrew, Uplink capacity and interference avoidance for two-tier femtocell network, IEEE Tran. Wirele Commun., vol. 8, no. 7, pp , July 29. [5] M. Chiang, C. W. Tan, D. Palomar, D. O Neill, and D. Julian, Power control by geometric programming, IEEE Tran. Wirele Commun., vol. 6, no. 7, pp , July 27. [6] R. Zakhour and D. Gebert, Ditributed multicell MIMO precoding uing the layered virtual SINR framework, to appear in IEEE Tran. Wirele Commun., 2. [7] P. Tiaflaki, M. Diehl, and M. Moonen, Ditributed pectrum management algorithm for multiuer dl network, IEEE Tran. Signal Proce., vol. 56, no., pp , Oct. 28. [8] R. Cendrillon, J. Huang, M. Chiang, and M. Moonen, Autonomou pectrum balancing for digital ubcriber line, IEEE Tran. Signal Proce., vol. 55, no. 8, pp , Aug. 27. [9] A. J. Goldmith and S.-G. Chua, Variable-rate variable-power mqam for fading channel, IEEE Tran. Commun., vol. 45, no., pp , Oct [2] A. L. Stolyar, On the aymptotic optimality of the gradient cheduling algorithm for multiuer throughput allocation, Operation Reearch, vol. 53, no., pp. 2 25, Jan. 25. [2] Y. Liu and E. Knightly, Opportunitic fair cheduling over multiple wirele channel, in Proc. IEEE INFOCOM, San Francico, CA, Mar. 23, pp [22] R. Cendrillon, W. Yu, M. Moonen, J. Verlinden, and T. Botoen, Optimal multiuer pectrum balancing for digital ubcriber line, IEEE Tran. Commun., vol. 54, no. 5, pp , May 26. [23] S. Boyd and L. Vandenberghe, Convex Optimization, t ed. Cambirdge Univerity Pre, 24. [24] D. P. Palomar and J. R. Fonolloa, Practical algorithm for a family of waterfilling olution, IEEE Signal Proce. Lett., vol. 53, no. 2, pp , Feb. 25. [25] Part 6: Air Interface for Fixed and Mobile Broadband Wirele Acce Sytem, IEEE Std. 82.6e-25, Feb. 26. [26] K. Son, S. Lee, Y. Yi, and S. Chong, Practical dynamic interference management in multi-carrier multi-cell wirele network: A reference uer baed approach, Available at kio/dim technical report.pdf, Technical Report, Dec. 29. [27] Apect of WINNER+ pectrum preference, WINNER+ Deliverable D3.2, May 29. [28] J. Broch, D. A. Maltz, D. B. Johnon, Y.-C. Hu, and J. Jetcheva, A performance comparion of multi-hop wirele ad hoc network routing protocol, in Proc. ACM Mobicom, Dalla, TX, Oct. 998, pp

13 272 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 29, NO. 6, JUNE 2 Kyuho Son (S 3-M ) received hi B.S., M.S. and Ph.D. degree all in the Department of Electrical Engineering from Korea Advanced Intitute of Science and Technology (KAIST), Daejeon, Korea, in 22, 24 and 2, repectively. He i currently a pot-doctoral reearch aociate in the Department of Electrical Engineering at the Univerity of Southern California, CA. Hi current reearch interet include interference management in heterogeneou cellular network, green networking and network economic. He ha been erving a a Web Chair of the 7th International Sympoium on Modeling and Optimization in Mobile, Ad Hoc, and Wirele Network (WiOpt 29) a well a IEEE SECON 2 Workhop on Green and Sutainable Communication Network (GASCoN). Soohwan Lee (S ) received hi B.S. degree in the School of Electrical Engineering and Computer Science from Kyungpook National Univerity, South Korea, in 29, and hi M.S. degree in the Department of Electrical Engineering from Korea Advanced Intitute of Science and Technology (KAIST), South Korea, in 2. He i currently a Ph.D tudent in the Department of Electrical Engineering at KAIST. Hi current reearch interet include interference management in heterogeneou cellular network, green wirele networking, and ecurity management in cellular network. Yung Yi (S 4-M 6) received hi B.S. and the M.S. in the School of Computer Science and Engineering from Seoul National Univerity, South Korea in 997 and 999, repectively, and hi Ph.D. in the Department of Electrical and Computer Engineering at the Univerity of Texa at Autin in 26. From 26 to 28, he wa a pot-doctoral reearch aociate in the Department of Electrical Engineering at Princeton Univerity. Now, he i an aitant profeor at the Department of Electrical Engineering at KAIST, South Korea. Hi current reearch interet include the deign and analyi of computer networking and wirele communication ytem, epecially congetion control, cheduling, and interference management, with application in wirele ad hoc network, broadband acce network, economic apect of communication network, and green networking ytem. He ha been erving a a TPC member at variou conference uch a ACM Mobihoc, Wicon, WiOpt, IEEE Infocom, ICC, Globecom, ACM CFI, ITC, the local arrangement chair of WiOpt 29 and CFI 2, and the networking area track chair of TENCON 2. Song Chong (S 93-M 95) received the B.S. and M.S. degree in Control and Intrumentation Engineering from Seoul National Univerity, Seoul, Korea, in 988 and 99, repectively, and the Ph.D. degree in Electrical and Computer Engineering from the Univerity of Texa at Autin in 995. Since March 2, he ha been with the Department of Electrical Engineering, Korea Advanced Intitute of Science and Technology (KAIST), Daejeon, Korea, where he i a Profeor and the Head of the Communication and Computing Group of the department. Prior to joining KAIST, he wa with the Performance Analyi Department, AT&T Bell Laboratorie, New Jerey, a a Member of Technical Staff. Hi current reearch interet include wirele network, future Internet, and human mobility characterization and it application to mobile networking. He ha publihed more than paper in international journal and conference. He i an Editor of Computer Communication journal and Journal of Communication and Network. He ha erved on the Technical Program Committee of a number of leading international conference including IEEE INFOCOM and ACM CoNEXT. He erve on the Steering Committee of WiOpt and wa the General Chair of WiOpt 9. He i currently the Chair of Wirele Working Group of the Future Internet Forum of Korea and the Vice Preident of the Information and Communication Society of Korea.