Cost-Aware Fronthaul Rate Allocaton to Maxmze Beneft of Mult-User Recepton n C-RAN Dora Bovz, Chung Shue Chen, Sheng Yang To cte ths verson: Dora Bovz, Chung Shue Chen, Sheng Yang. Cost-Aware Fronthaul Rate Allocaton to Maxmze Beneft of Mult-User Recepton n C-RAN. IEEE Wreless Communcatons and Networkng Conference WCNC), Mar 217, San Francsco, Unted States. 217, <1.119/wcnc.217.7925531 >. <hal-141773v2> HAL Id: hal-141773 https://hal.nra.fr/hal-141773v2 Submtted on 2 Jan 217 HAL s a mult-dscplnary open access archve for the depost and dssemnaton of scentfc research documents, whether they are publshed or not. The documents may come from teachng and research nsttutons n France or abroad, or from publc or prvate research centers. L archve ouverte plurdscplnare HAL, est destnée au dépôt et à la dffuson de documents scentfques de nveau recherche, publés ou non, émanant des établssements d ensegnement et de recherche franças ou étrangers, des laboratores publcs ou prvés.
Cost-Aware Fronthaul Rate Allocaton to Maxmze Beneft of Mult-User Recepton n C-RAN Dora Bovz, Chung Shue Chen, Sheng Yang Noka Bell Labs 7 Route de Vllejust, 9162 Nozay, France CentraleSupélec, Laboratore des Sgnaux et Systèmes 3 rue Jolot-Cure, 9119 Gf-sur-Yvette, France E-mals: dora.bovz@noka-bell-labs.com, chung shue.chen@noka-bell-labs.com, sheng.yang@centralesupelec.fr Abstract Amng to throughput enhancement n future moble networks, dense deployment of wreless access ponts APs) s ntended. Spectral effcency on the uplnk can be mproved va jont processng of users located n areas covered by several APs. Thanks to centralzed processng of such users enabled by Cloud Rado Access Network C-RAN) archtecture, low-latency mult-cell cooperaton becomes possble. However, the prce to pay to beneft from ts advantages s the cost of transferrng data from dstrbuted access ponts to the data center through lmted-capacty fronthaul lnks. A moble network operator usng C-RAN would have the objectve to maxmze data transmsson rates wth lowest possble fronthaul rate. We consder n ths work the optmal tradeoff between the amount of fronthaul allocated and the sum rate of each user group n a multple access scheme, where the wreless resource s reused among the user groups. By performng low-complexty jont recepton of sgnals quantzed accordng to attrbuted fronthaul rate, we can maxmze the beneft of uplnk transmsson n C-RAN despte constraned fronthaul lnks. Wth optmal fronthaul allocaton, the net beneft of the uplnk transmsson mproves by 1% n a usual confguraton. I. INTRODUCTION The dea of offloadng base-band functons from several cell stes to a centralzed data center arses the great opportunty of jont sgnal processng, but also the practcal challenge of data transfer between remote Access Ponts APs) and the Central Offce CO). Furthermore, n 5G archtecture, centralzaton s dentfed as a key dsrupton [1], snce t enables user-centrc processng whch can greatly facltate mult-cell cooperaton. Fndng an effcent and affordable data transfer soluton over lmted capacty fronthaul FH) lnks between APs and the CO s crucal to realze Cloud Rado Access Network C-RAN) deployments and fulfll 5G requrements [2]. Moble network operators have to offer better Qualtyof-Servce QoS) at lower cost, thus the effcent use of wreless spectrum, APs, and computatonal resources become more mportant then ever. In ths paper, we wll propose an optmzed transmsson scheme for C-RAN to ncrease uplnk UL) transmsson effcency from the User Equpments UEs) to the end of physcal layer PHY) processng. To enhance spectral effcency, low latency mult-cell cooperaton s realzed through partally non-orthogonal uplnk schedulng. We am to beneft from Multple Input Multple Output MIMO) enabled APs and centralzaton of sgnal processng whle keepng user coordnaton and recever complexty low. To deal wth lmted capacty fronthaul lnks between the APs and the CO, t s mportant to optmze the fronthaul rate allocaton n order to adapt quantzaton level of transferred sgnals accordng to the qualty of the wreless transmsson. Constraned Non-Orthogonal Multple-Access NOMA) s proposed n [3] for uplnks where users are organzed nto several subsets and wthn each of them, the users transmt smultaneously over the same wreless channel, whereas orthogonal resources are allocated to each group n order to elmnate the nterference between the groups. On the recever sde, one can use Mnmum Mean Square Error MMSE) detecton to separate sgnals of users transmttng together. Thanks to the lmted number of users n each group, the recever complexty can be relatvely low compared to that needed wth fully nonorthogonal transmsson of all users. Maxmzng the sum rate of UL mult-user transmsson through enhanced user assocaton s descrbed n [4], where the authors consder channel estmaton error and compare dfferent methods for sum rate mprovement. Jont fronthaul and power allocaton n fully orthogonal mult-user model s studed n [5]. The authors provde practcal quantzaton scheme and show that the optmzaton of fronthaul allocaton mproves sgnfcantly the overall throughput compared to unform dstrbuton of avalable fronthaul. The mprovement s obtaned by the novel scheme whch takes nto account dfferent lnk gans of receved sgnals that are forwarded on the same fronthaul lnk. Impact of sgnal compresson n C-RAN archtecture to satsfy fronthaul constrant s wdely studed, see for example [6]. In our prevous work [7], we have characterzed the optmal fronthaul rate for a sngle MIMO channel. In ths paper, we extend the optmzaton to a system of several user groups that transmt ndependently and share the fronhaul lnks. We show that optmally dstrbutng avalable fronthaul rate between mult-user transmssons allows to mprove the overall effcency of the transmsson. In an typcal scenaro, end-to-end beneft ncreases by 1% thanks to the proposed tradeoff between sum rate and fronthaul usage. The rest of the paper s organzed as follows. We descrbe the system model and the mult-user transmsson scheme n
N cell-edge ) users K/2 antennas/rrh RRH1 RRH2 Fronthaul lnks Cell PHY Cell PHY Jont MMSE detecton User PHY User PHY Central Offce Fg. 1: System model wth several user groups example M = 2 RRHs), where JD stands for jont detecton and BBU stands for base-band unt. Secton II. Then, n Secton III we develop the metrc lnkng system performance to the fronthaul rate. The optmzaton problem of fronthaul allocaton s formulated n Secton IV. Secton V provdes numercal evaluaton of the proposed fronthaul allocaton scheme and observatons. Fnally, Secton VI contans the concluson. II. SYSTEM MODEL The system that we study s depcted n Fgure 1. We consder a C-RAN system wth M RRHs located at the cellstes wth a antennas each, thus the total number of antennas s K = a M. All of the RRHs are connected to the same CO where the major part of sgnal processng s realzed for the cells. We assume that all the N users are located n the cell-edge regon and are able to communcate wth all the M RRHs. To avod nterference between them, they are beneftng from jont detecton JD) at the recever. In the system that we study, user-phy functons of all the N users are placed n the CO n order to enable mult-cell jont processng on the uplnk [8]. The users can be assocated to each other randomly or n a determnstc way to create the user groups to whch we allocate a wreless resource. Orthogonal resource blocks RBs) are allocated to dfferent groups, but users who transmt usng the same RB wll nterfere wth each other. The RRHs, after recevng the combnaton of uplnk sgnals transmtted by every group, wll forward the quantzed verson of ths sgnal to the CO through lmted-capacty fronthaul lnks. The total capacty of each fronthaul lnk s shared among all of the user groups. We assume a block fadng channel wth a coherence tme that s long enough to consder allocatng fronthaul capacty to the groups usng real-tme channel realzatons. In terms of user moblty, e.g., for pedestran users movng at 5 km/h, the transmsson characterstcs change approxmately every 72 BBU1 BBU2 ms, whch s a tme wndow large enough to update network parameters. We use the followng notatonal conventons: for random varables, upper case letters wth bold and non-talc fonts, e.g., V, for vectors, and bold and sans serf fonts, e.g., M, for matrces. Determnstc quanttes are denoted wth talc letters, e.g., a scalar x, lowercase bold for a vector v, and uppercase bold letters for a matrx M. Logarthms are n base 2 and superscrpt.) H denotes the conjugate transpose of a vector or a matrx. We have N users unformly dstrbuted n the regon covered by every RRH. The channel of each user n {1,..., N} towards all antennas at both RRHs s denoted by h n whch s a K dmensonal array followng the Gaussan dstrbuton N, R n ) wth R n C K K. The number of user groups s denoted L, there are s l users n the group l {1,..., L}. The mult-user channel of a group towards the whole set of antennas s the K s l dmensonal matrx H l. We assume that the channel s perfectly known at the recever. Though, channel estmators that are generally used result n an estmaton error, ts mpact on the fronthaul allocaton s not sgnfcant. The Gaussan channel nose vector s denoted n l N, σzi 2 K ). Power of the nput sgnal s normalzed, so that nose covarance scales wth σz 2 = 1 SNR. The receved sgnal for group l by the whole set of antennas s the K-dmensonal vector y l and t s gven by the superposton of the sgnals sent by all of the s l users n the group denoted by Π l = {π1, l..., πs l l }. y l = h π l x π l + n l 1) =1 The fronthaul capacty of the lnk between the m-th RRH and the CO allocated to group l s denoted by c l) m wth l {1,..., L} and m {1,..., M}. We can wrte ths total fronthaul rate avalable usng fronthaul rates c lk dedcated to the transmsson of the sgnal from group l receved by antenna k located at RRH m as c l) m = m a k=m 1) a+1 c lk. 2) We denote by c l) = {c l1,...c lk } the set of capacty values attrbuted to group l and c m the total fronthaul capacty avalable between RRH m and the CO. III. SUM RATE OF MULTI-USER TRANSMISSION WITH LIMITED FRONTHAUL Once user groups are created and scheduled, users transmt followng the schedulng decsons and RRHs receve the sgnals of every group - beng the superposton of the sgnals transmtted by all the UEs n the group - and forward frequency-doman I/Q symbols to the CO. We perform receve Fast Fourer Transform n the RRH to decorrelate the subcarrers and enable to select the ones that need to be forwarded
snce they are requred for the JD). Ths decorrelaton mproves also the throughput by reducng quantzaton nose n the forwarded sgnal. The channel gan depends on subcarrer frequency, so the fronthaul allocaton can change wth a dfferent schedulng decson or a modfcaton of user groups. In the followng computatons, we consder only one subcarrer per group, but the extenson to several subcarrers s straghtforward. A. Quantzaton nose We defne n the followng the equvalent quantzaton nose wth lmted capacty fronthaul for a group wth s l users transmttng towards K receve antennas. The achevable rate wth a fronthaul capacty c lk can be defned usng the mutual nformaton between the receved sgnal and the forwarded one, gven the dstorton between them. r lk mn IY lk ; Ŷlk H l ). 3) pŷlk y lk :D σd 2 lk We can derve the followng relaton between the receved sgnal power and the varance of the dstorton nose where σ 2 d lk σ 2 y lk H l 2 c lk 4) σy 2 lk H l = h kπ l 2 ) + σz. 2 5) =1 We use a scalng factor α lk n order to adapt the power of forwarded sgnal to the fronthaul capacty used,.e., α lk = σ2 y lk H l σ 2 d lk σ 2 y lk H l k {1,..., K}. 6) Scalng factors for each antenna form the matrx A l = α lk ). The dstorton s then characterzed by the dag k={1,...,k} followng upper bound: σd 2 lk α lk B. Achevable sum rate = σ2 d σ2 lk y lk H l σy 2 lk H l σd 2 σ2 y lk H l 2 clk lk 1 2 c, lk k {1,..., K}, l {1,..., L}. We can compute the achevable sum rate for a gven jont transmsson group usng the mutual nformaton between the sgnal sent by all users n the group and the one receved n the CO: =1 7) r IX l ; Ŷ l H l ) = hx l ) hx l Ŷ l, H l ) 8) We compute both entropy terms n order to fnd the lower bound of the sum rate. The frst term descrbes the quantty of nformaton sent by the users, and thus depends on the transmsson power: hx l ) = logdet2πee[x l X H l ])) = log2πe)) 9) The second term comes from the loss of nformaton between the UEs and the CO and can be computed usng the lnear MMSE covarance C e. hx l Ŷ l, H l ) logdet2πec e )). 1) For ths, we use the defnton of receved sgnal by the CO: sl ) ŷy l = A l h π l x + n l =1 11) where n = z + d s the equvalent nose contanng Gaussan channel nose and the quantzaton nose. The covarance matrx of ths equvalent nose s C N = σzi 2 K + dag k={1,...,k} σ 2 d lk α lk ). We can compute the lnear MMSE covarance based on the channel matrx C e = I sl H H l H l H H l + C N ) 1 H l. 12) Usng the nverson lemma on the lower bound of the mutual nformaton we get the followng expressng of the achevable sum rate wth a gvel channel matrx H l =1 ) r log det I sl + H H l C 1 N H l. 13) From 7) and 13), we get the achevable sum rate of a group of N users transmttng towards K receve antennas, each of them usng c j, j {1,..K} bts of fronthaul. =1 ) r log det I sl + H H l V 1 s l H l wth the equvalent nose covarance: 14) ) σ 2 V sl = σzi 2 ylk H K + dag l 2 c lk k={1,...,k} 1 2 c. 15) lk IV. FRONTHAUL ALLOCATION OPTIMIZATION After the CO receves forwarded sgnals from the RRHs, we can compute the sum rate 14) of every user group usng channel realzatons n order to evaluate the performance of the mult-user recepton n C-RAN under the constrant of lmted fronthaul capacty. Usng the sum rate expresson 14), our am s to maxmze the beneft of the transmsson gven a constrant on total rate avalable at each fronthaul lnk. We provde here a metrc enablng the allocaton of fronthaul capacty to user groups, whch should allow to maxmze the effcency of the transmsson. When the optmal fronthaul allocaton for each fronthaul lnk s computed, quantzaton of forwarded sgnal s adapted at the RRHs, n order to mprove the performance of the complete transmsson.
A. Objectve functon: net beneft of the transmsson We use the upper bound of the sum rate 14) to formulate the objectve functon allowng to maxmze the end-to-end beneft of the uplnk transmsson of N users formng L groups towards the M RRHs wth a antennas each. The parameters of ths functon are the followng: The Gaussan channel nose varance σz. 2 The average receved sgnal power from group l at antenna k gven the channel of the group l: σy 2 lk H l. The fronthaul capacty c lk wth k {1,..., K} used to forward to the CO dgtal base-band I/Q symbols of group l receved by antenna k. We denote c l) = c l1,...c lk ) T the set of capacty values for group l. The cost λ k of the fronthaul capacty used for the transmsson of the sgnal receved on antenna k. In ths work we consder only lnear fronthaul cost correspondng to explotaton costs such as transmsson energy or cost rentng the needed transmsson rate from the owner of the fronthaul nfrastructure. Note that any non-negatve convex cost functon can be used followng the deployment scenaro consdered. The followng functon characterzes the net beneft of the transmsson of group l towards the whole set of receve antennas when the fronthaul capacty allocated to the group s c l) m a m = c lk for the fronthaul lnk between RRH k=m 1) a+1 m and the CO. Gven the parameters σz, 2 σy 2 lk H l, λ k k {1,..., K}, ) K fh l, c l) ) = log det I sl + H H l V 1 s l H l λ k c lk 16) k=1 where V sl s defned n 15). The frst term of the functon f.) n 16) gves the nstantaneous sum rate durng the coherence tme block where the channel matrx H l holds and the second term s the total cost of the fronthaul transmsson for group l over the whole set of fronthaul lnks connectng the RRHs to the CO. B. Lmted fronthaul lnk rate Frst, we am to fnd the optmal capacty allocaton {c 1),..., c L) } that maxmzes the overall beneft of the uplnk transmsson of the N users dstrbuted n L groups whle the fronthaul capacty used for all groups between each RRH m and the CO s not more than c m. We need to solve the optmzaton problem ncludng ths constrant when fronthaul lnks are physcally lmted to a gven rate, for example n low capacty deployments usng other transport soluton than optcal fbers or when the allocaton of rates between varous servces explotng the same transport lnks s fxed. Let us recall that c l) m = constrant can also be wrtten as m a l=1 k=m 1) a+1 m a k=m 1) a+1 c lk, so the above c lk c m m {1,..., M} 18) Proposton 1. The problem 17) s concave, thus admts a unque soluton that gves the optmal capacty allocaton scheme. Proof: The constrant n 17) s lnear, as well as the cost term n 16), thus the concavty of the frst term of f.) s suffcent to show that the problem s concave. Notce that the K second term of 16),.e., λ k c lk s lnear w.r.t. c l), so t can k=1 be consdered as concave. The functon log deta) s concave f and only f the matrx A s non-negatve defnte. The sum of two non-negatve defnte matrces s also non-negatve defnte. Snce the dentty matrx satsfes ths condton, we only need to show that the second term of the argument of the log det.) n 16) s non-negatve defnte. The equvalent SNR matrx V sl s dagonal wth postve elements whch are ts egenvalues, thus t s postve defnte. Ths property stands also for ts nverse. If a postve defnte matrx M s multpled by another matrx and ts hermtan as B H MB, the result s also postve defnte f B s full rank. Ths s true for H H l V 1 s l H l snce the columns of H are ndependent, thus rankh) = s l. Consequently, the matrx beng the argument of log det.) s postve defnte and also non-negatve defnte, thus the frst term of 16) whch mples wth the above reasons that 17) s concave. C. Unconstraned fronthaul lnks In some deployment scenaros, very hgh capacty fronthaul lnks are avalable between RRHs and the CO. These can be consdered n practce as unlmted lnks. For example, when optcal fbers wth Dense Wavelength Dvson Multplexng are used, they can satsfy bandwdth requrements of moble fronthaul, but nstallaton and consequently usage cost may be hgh. Unconstraned optmzaton of the beneft of uplnk transmsson would allow to allocate fronthaul optmally for a gven fronthaul cost. Wthout constrant on the maxmal fronthaul rate avalable at each lnk, to solve the optmzaton problem we need to maxmze the same concave functon as n 17). Intervals of possble fronthaul capacty values are specfed accordng to the transmsson scheme, see below 19). However, they do not constran the problem n practce. Fnd {c 1),..., c L) } = subject to l=1 argmax {c 1),...,c L) } fh l, c l) ) l=1 c l) m c m m {1,..., M}. 17) Fnd {c 1),..., c L) } = wth < c lk < σ 2 y lk H l argmax {c 1),...,c L) } fh l, c l) ) l=1 k {1,..., K}, l {1,..., L}. 19)
Net beneft of UL transmsson bts/cu) 8 6 4 2 Net beneft of uplnk transmsson vs fronthaul lnk capacty Optmal alloc. λ=.1 Unform alloc. λ=.1 Optmal alloc. λ=.2 Unform alloc. λ=.2 Net beneft of the transmsson bts/cu) 15 5 Net beneft of uplnk transmsson vs fronthaul prce Unform allocaton FH lmt=24 Optmzed allocaton FH lmt=24 Unform allocaton FH lmt=4 Optmzed allocaton FH lmt=4 Optmzed allocaton unconstraned FH 2 15 2 25 3 35 4 Avalable fronthaul rate per lnk bts/cu) Fg. 2: Net beneft of uplnk transmsson wth constraned fronthaul 5.1.2.3.4.5.6.7 Fronthaul prce factor Fg. 3: Net beneft of uplnk transmsson wth varous fronthaul constrants vs. fronthaul prce factor V. NUMERICAL EVALUATION OF FRONTHAUL ALLOCATION OPTIMIZATION A. System parameters We have evaluated the results of the above fronthaul allocaton optmzaton see equatons 17) and 19)) wth N = 4 users located at the edge of 2 cells wth 1 RRH each. Note that the optmzaton problem can be solved effcently usng standard convex programmng [9]. We have K = 8 antennas equally dstrbuted between the RRHs. Channel realzatons are modeled usng ndependent one-rng scatterer model for each user [1]. Users are transmttng jontly on the uplnk by groups of 4 users each, each user group havng ts own RB. B. Effcent transmsson wth constraned fronthaul In a scenaro where we am to fnd the optmal fronthaul allocaton scheme n presence of lmted fronthaul lnks, one can be nterested to evaluate f the avalable amount of fronthaul capacty allows an effcent transmsson. Furthermore, whle desgnng fronthaul nfrastructures, the evaluaton of requred capacty to get optmal beneft from the transmsson wth expected prce values λ k ) can contrbute to correct the dmensonng of the lnks to be deployed. In Fgure 2, we compare the proposed fronthaul allocaton scheme that optmzes the net beneft of the transmsson aganst unform fronthaul allocaton for dfferent values of avalable fronthaul capacty. In case of unform fronthaul allocaton, avalable fronthaul capacty s equally dstrbuted between all groups and all antennas. When avalable fronthaul s low, both unform and optmzed allocaton result n smlar effcency, snce the constrant does not allow to acheve hgher sum rate. One can notce that optmzed fronthaul allocaton allows to acheve hgher transmsson beneft, snce n case of suffcent fronthaul, group sum rates can be mproved by allocatng more fronthaul to the receved sgnals wth hgher powers. In other words fronthaul allocaton s adapted to the varatons of channel gans for dfferent users and antennas. Even n an deal scenaro wth λ k =.1 and a relatvely hgh avalable fronthaul rate, e.g., 3 bts/channel use, the proposed allocaton scheme ncreases net beneft of the transmsson by approxmately 1%. Furthermore, by optmzng the fronthaul allocaton, n the operaton regme where the sum rate s lmted by the transmsson power, allocated fronthaul does not contnue to ncrease so that the cost remans moderate. In unform allocaton cost contnues to ncrease whle the sum rate s lmted, whch result n the drop of the beneft that we can observe n Fgure 2. Obvously, f the prce of the fronthaul usage s hgher, the transmsson s less effcent for both allocaton strateges, snce the maxmal sum rate does not ncrease whle the cost does. C. At what prce s t stll useful to transmt? When fronthaul capacty has a fxed lmtaton ncludng practcally nfnte capacty avalable), the net beneft of the transmsson can turn to negatve values when unform allocaton scheme s used and cost s too hgh. By optmzng capacty allocaton, we can avod havng hgher fronthaul cost than the utlty of the transmsson,.e., the sum rate. However, reducng allocated fronthaul rate when the prce factor s hgh would result n a decrease of the sum rate. Dependng on the transmsson requrements, settng a lower bound on the sum rate that allows to acheve the requred qualty-of-servce can be used to select whether t s useful to transmt wth a gven prce or not. We can see n Fgure 3 that the more expensve the fronthaul s, the more we mprove transmsson beneft by fronthaul optmzaton snce we can avod allocatng very hgh cost fronthaul rate. Consequently, optmzed fronthaul allocaton can reduce the sum rate of the transmsson, as we can observe n Fgure 4. If the sum rate s very low because of low FH rate that should be allocated, t can be more reasonable to not to transmt. One can also note that the optmal sum rate and net beneft values are almost equal for dfferent fronthaul constrants. A gap can be observed only for low prcng and low avalable fronthaul rate see the curve FH lmt = 24), snce we cannot acheve hgh beneft despte the optmal fronthaul allocaton.
Total sum rate bts/cu) 15 5 Transmsson sum rate wth optmzed fronthaul allocaton FH lmt=24 FH lmt=4 No constrant.1.2.3.4.5.6.7 Fronthaul prce factor Fg. 4: sum rate of uplnk transmsson wth varous fronthaul constrants vs. fronthaul prce factor Fronthaul rate used per lnk bts/cu) 25 2 15 5 Fronthaul rate allocated per lnk vs prcng factor FH lmt=24 FH lmt=4 no constrant.1.2.3.4.5.6.7 Fronthaul prce factor Fg. 5: Fronthaul rate allocated per lnk aganst prce factor D. Impact of the prce on FH lnk usage We compare the amounts of allocated fronthaul rate per lnk wth dfferent constrants ncludng nfnte) and varyng fronthaul prce n order to evaluate the dfference between constraned and unconstraned optmzaton strateges. We can observe n Fgure 5 that f avalable fronthaul rate s hgh enough see the curve FH lmt = 4) to enable to acheve maxmal sum rate, the amount of optmally allocated fronthaul s almost the same wth and wthout the constrant on avalable fronthaul for any value of the prce factor. However, wth less fronthaul rate avalable, obvously less of t s allocated per lnk even wth a low prce. For hgher prce values, we get smlar results regardless of the amount of avalable FH rate. The reason s that under low fronthaul prce, the sum rate term domnates the optmzaton, so that t s hghly motvated to ncrease the sum rate n the optmzaton, however t wll be lmted by the total amount of avalable fronthaul. For hgher prce factor, fronthaul cost must be kept low n order to maxmze the beneft of the transmsson, so the allocated fronthaul wll become smlar when the prce ncreases wth and wthout constrant. VI. CONCLUSION AND FUTURE DIRECTIONS A man lmtaton of C-RAN archtecture s the capacty lmted and expensve fronthaul lnks. In ths paper, we have analyzed a low complexty mult-user mult-cell transmsson scheme where fronthaul rate allocaton s performed to maxmze uplnk throughput. We provde a novel system model where constraned NOMA s used for several cells and mult-antenna RRHs are connected to a CO where UL JD s performed. To optmze the gan wth respect to the lmted fronthaul capacty and also prce, we allocate the FH rate based on a performance metrc explotng uplnk channel gans. We evaluate the net beneft of the transmsson whch allows a tradeoff between uplnk sum rate ncludng practcal parameters and the cost of fronthaul usage. A future work s to explore possble further mprovement of the system performance by assocatng n each group users that can acheve hgher sum rate together,.e., user groupng optmzaton. Ths would be challengng due to the ncreased system complexty and also the fact that only lmted channel state nformaton s avalable before the rado transmsson schedulng. However, we can expect further performance gans by addng the optmzaton of user assocaton. REFERENCES [1] F. Boccard, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovsk, Fve dsruptve technology drectons for 5G, IEEE Communcatons Magazne, vol. 52, no. 2, pp. 74 8, February 214. [2] Noka, 5G use cases and requrements, 215. [3] M. Al-Imar, P. Xao, M. A. Imran, and R. Tafazoll, Uplnk nonorthogonal multple access for 5G wreless networks, n 11th Internatonal Symposum on Wreless Communcatons Systems ISWCS), Aug 214, pp. 781 785. [4] Y. Ru, H. Hu, H. Y, and H. H. Chen, Robust user parng algorthm under channel estmaton errors for uplnk vrtual multple-nput multple-output systems, IET Communcatons, vol. 6, no. 3, pp. 318 323, February 212. [5] L. Lu, S. B, and R. Zhang, Jont power control and fronthaul rate allocaton for throughput maxmzaton n OFDMA-Based Cloud Rado Access Network, IEEE Transactons on Communcatons, vol. 63, no. 11, pp. 497 411, Nov 215. [6] J.-K. Kang, O. Smeone, J. Kang, and S. Shama, Jont sgnal and channel state nformaton compresson for uplnk network MIMO systems, n Global Conference on Sgnal and Informaton Processng GlobalSIP). IEEE, 213, pp. 875 878. [7] D. Bovz and S. Yang, Optmal fronthaul capacty allocaton and tranng for jont uplnk recever n C-RAN, n European Wreless, May 216. [8] D. Bovz and Y. El Mghazl, Fronthaul for 5G: low bt-rate desgn enablng jont transmsson and recepton, n IEEE Global Telecommuncatons Conference Globecom), 5G RAN Desgn Workshop, December 216. [9] S. Boyd and L. Vandenberghe, Convex Optmzaton. New York, NY, USA: Cambrdge Unversty Press, 24. [1] D.-S. Shu, G. J. Foschn, M. J. Gans, and J. M. Kahn, Fadng correlaton and ts effect on the capacty of multelement antenna systems, IEEE Transactons on communcatons, vol. 48, no. 3, pp. 52 513, 2.