ebook img

Graph-based Framework for Flexible Baseband Function Splitting and Placement in C-RAN PDF

0.4 MB·English
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Graph-based Framework for Flexible Baseband Function Splitting and Placement in C-RAN

Graph-based Framework for Flexible Baseband Function Splitting and Placement in C-RAN Jingchu Liu, Sheng Zhou, Jie Gong, Zhisheng Niu Shugong Xu Tsinghua National Laboratory for Information Science and Technology Intel Labs Department of Electronic Engineering, Tsinghua University Beijing 100084, China Beijing 100080, China Email: [email protected], [email protected] Email: [email protected] [email protected], [email protected] 5 1 Abstract—The baseband-up centralization architecture of ra- processing units (BBUs) for signal processing. The benefits 0 dio access networks (C-RAN) has recently been proposed to of such an architecture include reduced CAPEX and OPEX 2 support efficient cooperative communications and reduce de- [3], [4], efficient information exchange for cooperative com- n ployment and operational costs. However, the massive fron- munications [3], and increased flexibility due to the use of a thaul bandwidth required to aggregate baseband samples from J remote radio heads (RRHs) to the central office incurs huge general purpose platforms (GPPs) [5]. Despite these benefits, 0 fronthaulingcost,andexistingbasebandcompressionalgorithms a major challenge for baseband-up centralization is the huge 2 can hardly solve this issue. In this paper, we propose a graph- aggregationbandwidthrequirementforfronthaulnetwork.For basedframeworktoeffectivelyreducefronthaulingcostthrough a typical long-term evolution (LTE) cell configuration of 20 ] properly splitting and placing baseband processing functions in MHz wireless bandwidth and 8 antennas, 10 Gbps fronthaul T the network. Baseband transceiver structures are represented I with directed graphs, in which nodes correspond to baseband bandwidth is required in downlink (DL) or uplink (UL) to s. functions, and edges to the information flows between functions. transport baseband samples [6]. The demand for fronthaul c By mapping graph weighs to computational and fronthauling bandwidth would only be higher with even larger wireless [ costs,wetransformtheproblemoffindingtheoptimumlocation bandwidth and more antennas. to place some baseband functions into the problem of finding 1 To cope with this problem, a number of baseband com- the optimum clustering scheme for graph nodes. We then solve v this problem using a genetic algorithm with customized fitness pression algorithms have been proposed. Time-domain com- 3 function and mutation module. Simulation results show that pression algorithms [7] are simple and fast, but they only 0 proper splitting and placement schemes can significantly reduce provide limited compression performance (2−3×). Another 7 fronthaulingcostattheexpenseofincreasedcomputationalcost. 4 kindofalgorithmsperformcompressioninfrequencydomain, Wealsofindthatcooperativeprocessingstructuresandstringent 0 whichcanachieve20×compressionrate.However,frequency- delay requirements will increase the possibility of centralized 1. placement. domain methods requires a great amount of computation to 0 perform FFT/IFFT and thus may suffer from long delay. 5 I. INTRODUCTION Lorca and Cucala [8] propose a novel method that can 1 Recently, the number of smart devices has grown into significantlyreduceDLfronthaulbandwidth(30×)byrelocat- : v billions, and the wide collection of mobile applications is ing modulation and precoding processing functions from the i increasingly interleaved with our daily lives. As a result, central office back to remote sites. The intuition behind this X nextgenerationwirelesscommunicationsystemshavereceived method can be explained with some insights into baseband r a unprecedented expectations, aiming at 1000 times capacity, processing. In a sense, DL baseband processing functions 100 times data rate, billions of devices, and millisecond-level are designed to add artificial redundancies into communica- delay. To fulfill these goals, the solution envisioned so far tion signals in order to combat the impairments of wireless is a dense, cooperative [1], and heterogeneous [2] wireless channels1. Because redundancies are accumulated function- network. However, realizing such a network with the tradi- by-function along the processing chain, baseband-up central- tional,distributedarchitectureforradioaccessnetwork(RAN) ization actually needs to transport the most redundant signal. means high capital expenditure (CAPEX) and operational In constrast, the method in [8] no longer needs to transport expenditure (OPEX); and cooperative communications will the redundancies introduced by modulation and precoding, be highly limited due to the bandwidth bottleneck between therefore fronthaul bandwidth can be reduced. distributed base stations. Nevertheless, the above method does not consider the cost The baseband-up centralization architecture, e.g. C-RAN [3], is a promising solution to these problems. In this archi- 1For example, modulation turns constellation codewords, which can be tecture, radio signals are first digitized at the antenna sites represented with only a few bits (6 bits for 64 QAM), into complex constellation samples, which is usually digitized with tens of bits (30 bits by remote radio heads (RRHs), and then transported via forLTE).Otherbasebandfunctionslikechannelencodingandbeamforming, the so called fronthaul network to the centralized baseband alsointroducesothertypesofartificialredundancies. and the paper is concluded in section V. Data Center II. MODELFORMULATION To From To From upper upper upper upper Inthissection,wepresenttheproposedgraph-basedframe- layer layer layer layer work. For better understanding, we also present a concrete example for mapping baseband function splitting problems Code Decode into graph-clustering problems. Demode A. Baseband processing structures and directed graphs Werepresentabasebandprocessingstructurewithadirected graph G = (V,E). Each node v ∈ V stands for an atomic basebandprocessingfunctionsuchasFFTorMIMOdetection, Fronthaul Network andeachdirectedlinke∈E representthelogicalconnectivity Mod MIMO between the nodes it connects 3. Each node is assigned with MIMO RX TX a node weight according to the node complexity function γ : IFFT FFT V →R, which indicates the computational complexity of this processingnode.Andeachlinkisassignedwithalinkweight Radio Radio RaTdXio RaRdXio by the link bandwidth function ω : E →R, which represents TX RX the amount of information that has to be exchanged between the processing node it connects. Note some of the nodes are Cell Site 1 Cell Site 2 sources (no inbound links) or sinks (no outbound links). Each distinct path p ∈ P from a source to a sink represents a complete chain of baseband processing functions4. Fig.1. Flexiblesplittingandplacementofbasebandfunctions. The graph formulated in this way may contain cycles. The reason is that there may be mutual information exchange for accommodating additional baseband functions (computa- betweenbasebandfunctionsincooperativesystemssuchasco- tional cost) at remote sites. In reality, computational cost can operativemultipoint(CoMP)processingormulti-userMIMO. become a major constraint at remote sites due to power con- These cycles are important features of the whole system sumption and form factor considerations. Also, some wireless and have significant influence on the choice of splitting and protocols have stringent real-time requirements for processing placement.However,wedoassertthattherearenoself-cycles, baseband tasks2. Hence, the influence on processing delay whichmayappearduetoinappropriateabstractionofiterative shouldalsobeconsidered.Forthesereason,weproposein[9] processingfunction.Theinformationflowofiterativeprocess- to flexibly split and place baseband functions in the network ing functions should be embedded in the atomic baseband based on the cost profile and delay requirements of different processing functions to avoid self-cycles. applications. But an analytical framework for deciding the B. Function Splitting and graph clustering optimum splitting and placement scheme is still needed. In this paper, we present a graph-based framework for With this representation, we can express function splitting baseband function splitting and placement. We first translate and placement as graph clustering schemes ξ: V →Z, which basebandtransceiverstructuresintodirectedgraphssothatthe assignnodestoacollectionofclusters.Notethatinourmodel, splitting and placement problems can be formulated as graph- clusters have explicit physical meanings. Different clusters clustering problems (as illustrated in Fig. 1). We then propose correspondtodifferentphysicallocations(e.g.remotesitesand a genetic algorithm with customized fitness function and central office), and the nodes in the same cluster correspond mutationmoduleforthegraph-clusteringproblem.Simulation to baseband processing functions that are placed at the same resultsshowthattheproposedalgorithmcaneffectivelyreduce physicallocation.Thelinksbetweenclusterscorrespondtothe fronthauling cost. An anatomy of these results also reveals information flow to be transported by the fronthaul network. that cooperative structures and stringent delay constraints will The study of graph clustering is concerned with grouping result in more centralized function placement. nodes in order to optimize some cost/gain metric. The classic The rest of the paper is organized as follows. In section goal is to group nodes that are “close to” or “similar to” each II, we represent baseband processing structures using directed other [10]. We employ different goals in our formulation to graph and formulate the baseband function splitting and address the special concerns of baseband function splitting placement problem as a graph-clustering problem. In section and placement. The goals reflect the computational cost for III, we introduce the proposed customized genetic algorithm. accommodating nodes at some location and the fronthauling Simulation results are presented and discussed in section IV 3Weassumenodesandedgesareindexedusingwithintegervalues. 2In LTE, the processing of a subframe should be completed within 3 ms 4Theremaybemultiplepathsbetweenapairofsourceandsinkduetothe fortimelyhybridautomaticretransmissionrequest(HARQ). parallelprocessingofchannelsfordifferentusers. cost for transporting data between different locations using dataSourceDL.1.1 radioRX.1.1 fronthaul networks. Specifically,wedefineapairofcostmetrics.Thefirstmetric 0.054 0.054 1 iscomputationalcostc (i;ξ),whereiistheindexofacluster, c code.1.1 code.1.2 fft.1.1 and ξ is the clustering scheme under consideration. The com- 0.06 0.06 0.45 0.45 putationalcostistoreflectthecostofimplementingbaseband processing functions at a physical location and should thus mod.1.1 mod.1.2 MIMOrx.1.1 MIMOrx.1.2 be a function of the total node complexity inside the cluster. 0.45 0.45 0.45 0.45 Also, because computational cost often differs in different locations in real world5, clusters are allowed to have different MIMOtx.1.1 MIMOtx.1.2 demod.1.1 demod.1.2 costprofiles.Thesecondmetricisfronthaulingcostc (i,j;ξ), f 0.45 0.45 0.06 0.06 where i and j are the index of two clusters, and ξ is the clusteringscheme.Fronthaulingcostistoreflectthebandwidth ifft.1.1 decode.1.1 decode.1.2 required to transport information between different physical 1 0.054 0 .054 locations. As a result, it should be a function of the total edgeweightsbetweenclusters.Aswehaveexplainedfromthe radioTX.1.1 dataSinkUL.1.1 perspectiveofredundacy,computationalcostandfronthauling cost are in generals contradiction goals to optimize. Thus, Fig.2. Asimplifiedbasebandprocessingarchitecture different clustering scheme will result in different tradeoffs betweencomputationalandfronthaulingcosts.Forthisreason, TABLEI weaimtocharacterizethetradeoffbetweencomputationalcost NODEWEIGHTSWITHRESPECTTONODETYPE. and fronthauling cost. Index 1 2 3 4 5 6 Another important feature of our model is the path de- Type radioTX radioRX fft ifft MIMOtx MIMOrx lay constraint, which is imposed to guarantee the real-time Weight 0 0 1 1 0.5 0.5 processing of communication signals. We assume that each node on a path will impose an additional delay d(v,p;ξ) Index 7 8 9 10 11 12 to this path, where v is the index of the node and p is the Type mod demod code decode sourceDL sinkUL path under consideration. This delay function captures the Weight 0.1 0.1 0.1 2 0 0 processing and buffering latency of baseband tasks. Any valid clustering scheme should guarantee that the total delay of a path is smaller than a predefined threshold D(p): d(p;ξ) = for each type is listed in Table I. The weight value is selected (cid:80) d(v,p;ξ)<D(p). based on experimental results in [5]. v∈p The weights of links are also shown in Fig. 2. The mag- C. Example nitudes of link weights reflect the information flow between The baseband processing structure used in our simulation processingnodes.Forexample,eachMIMOrxnodegetsalink is shown in Fig. 2. We use two such baseband processing from the FFT node with a weight of 0.45 because we assume structures to represent two cells. Note this structure is just a theoverheadofcyclicprefixandcontrolsignalingis10%,and simplification of real-life physical layer baseband structure. the information after CP removal is equally divided between We only include some of the most important functions in the two processing-chains. Also, note that the link weight the DL/UL chains. Other functions such as resource map- is greatly increased/reduced after modulation/demodulation ping/demapping, channel estimation, and scrambling are ig- because we assume each 30 bit complex baseband sample nored. Also, the parameters such as node complexity and is transformed from/into a 4 bit constellation codework (16- linkweightareapproximationstoreal-worldvalues.Withthis QAM).IncaseofCoMP,weaddmutuallinks6betweenneigh- simplification, we are able to show the essence of flexible bouring MIMO modules in a cyclic fashion (e.g. MIMOtx.n.2 splitting and placement of baseband functions. Still, fine- - MIMOtx.(n+1).2 and MIMOrx.N.1 - MIMOrx.1.1). grained tuning is required if the proposed model is applied We assume computational and fronthauling cost functions in practice. to have exponential forms as shown in Table II and Table Processing nodes are labeled based on their types, the III. The computational cost at central office is zero because logical cells they belong to, and the sub-chaisn they reside provisioning computational resources in central offices is less in.Forexample,thenodeMIMOtx.1.2isaMIMOtransmitter expensive. The fronthauling cost within cell sites or the data whichresidesinthesecondDLprocessing-chain(twointotal) center is zero because internal information exchange does not of the first cell. Each node is assigned with a weight based need to use fronthaul network. The fronthauling cost between onitstypetoreflectitscomputationalcomplexity.Theweight cell sites is higher than the cost between a cell site and the centralofficebecausefronthaulnetworksareusuallyoptimized 5Forexample,itismoreexpensivetoaccommodatecomputationatremote sitesthanatcentralofficeduetoformfactor,electricity,andsiterentalcosts. 6TheweightofCoMPlinksareequaltothelinksbetweenMIMOandFFT. TABLEII A. Natural encoding and cluster seeding COMPUTATIONALCOSTFUNCTIONcc(i,ξ)USEDINSIMULATION. A key problem with GA is how to represent solutions Cellsite Centraloffice as combinations of genes (chromosome). This process is 2(cid:80)ξ(v)=iγ(i) 0 also called encoding. Good encoding should make it easy to produce legitimate offspring individuals through crossover and mutation. Here we directly use the clustering vector ξ TABLEIII FRONTHAULINGCOSTFUNCTIONcf(i,j,ξ)USEDINSIMULATION. for encoding. The advantage of this encoding scheme should be obvious when we present our crossover and mutation Clusters Cost functions.Noticethatwekeepsomenodesinafixedclusterto withincellsite 0 reflectthefactthatsomefunctionscanonlybeplaceatspecific withincentraloffice 0 locations7. Hereafter we refer to these nodes as “seed nodes”. betweencellsitesiandj 4(cid:80)ξ(e)=(i,j)ω(e) We name them in this way because the whole clustering betweencellsiteandcentraloffice 2(cid:80)ξ(e)=(i,j)ω(e) scheme is generated based on the initial cluster assignments of seed nodes. TABLEIV B. Linearly combined fitness function DELAYFUNCTIONd(p;ξ). Another important aspect is how to evaluate solutions with cellsite centraloffice a fitness function. This problem is complicated because we (cid:80) (γ(v)(cid:80) γ(v)) 0 have two (possibly contradicting) optimization objectives. To v∈p ξ(w)=ξ(v) achieve different tradeoffs between these costs, we linearly combine computational and fronthauling cost to form a single cost function. Also, we have to incorporate the path delay forcentralization.Also,weassumebasebandtasksinacluster constraints. Yet explicitly examining whether a solution vio- equally divides the computational resource, thus the delay of lates these constrains makes crossover and mutation difficult. a processing functions can be represented with the product of Henceweimplicitlyincorporatethepathdelayconstraintsasa the corresponding node weight and the total node weight in “penaltyfunction”,whichwillsignificantlydegradethefitness the cluster (Table IV). ofasolutionifthepathdelayconstraintisviolated.Summing up, the overall fitness function is as follows: III. GRAPH-BASEDGENETICALGORITHM (cid:88) (cid:88)(cid:88) F(ξ;α,β)=α c (i;ξ)+(1−α) c (i,j;ξ)) c f The clustering scheme can also be represented with a i i j discrete valued vector ξ ∈ ZN, where N is the total number (cid:88) (2) +β (d(p;ξ)−D(p))+, of baseband processing functions. The k-th entry of ξ is the p cluster index for the k-th node. With this representation, the cost functions are parameterized by ξ and the graph cluster- where 0 ≤ α ≤ 1 is the tradeoff coefficient, β > 1 is ing problem is transformed into a 2-objective combinatorial the penalty coefficient, and (·)+ is the non-negative clipping optimization problem: function. (cid:80) (cid:80)(cid:80) C. Dispersive crossover min c (i;ξ), c (i,j;ξ)) c f ξ i i j (1) The crossover function we choose is dispersive crossover. (cid:80) s.t. d(p;ξ)= d(v,p;ξ)<D(p). v∈p This crossover function selectes the genes of an offspring fromitsparentswithequalprobability.Withnaturalencoding, It is difficult to give a general analytical solution to such we are guaranteed that the offspring of legitimate parents is a problem. So we turn to genetic algorithm (GA) to find naturally legitimate. sub-optimal solutions. The basic building blocks of GA are selection, crossover, and mutation. A typical GA session is D. Graph-based mutation initialized with a population carrying a collection of genes. Mutation function helps the population’s chromosomes es- The algorithm then iteratively loop through the three basic cape from local minima. Based on the structure of our prob- building blocks until solution converges or some termination lem,wetailoredacustomizedmutationfunctioncalledgraph- conditions are met. From the perspective of computation, GA based mutation. We first define the connection matrix C, the canalsobeseenasembeddedparallelalgorithmswhichsearch entries of which take on values of either 1 or 0. C(i,j) = 1 forthe“good”solutionbysimultaneouslyexperimentingwith if and only if node i and node j are connected. Using C, multiple solutions. Although GA can be applied to many we can define the Allowed Mutation Set as A(i) = {ξ(j) | types of problems, its performance will become satisfactory C(i,j) = 1,j is seed}, which gives all the clusters that a only after some customization. For this reason, we designed a customized GA to solve the graph-clustering problems 7Radio transmitter or receiver have to be placed at distributed cell sites. described above. Assigningthemtootherphysicallocationsdoesnotmakesense. TABLEV ALGORITHMPARAMETERS. 10.5 10 Parameter Value/Type Populationsize 20 9.5 InSiteiSaleelieczdtaisotinon RaGdiroaTpxhR,-obRlalasidneidgo-Rrwaxhn,edeSolomusericlneeictDitaiLoli,nzSatiinoknUL Fronthauling cost 8.95 Increasing α Crossover Dispersivecrossover 8 Mutation Graph-basedmutation(Prob.=0.4) Delaypenaltyfactor 10 7.5 7 0 5 10 15 20 25 30 Computational cost node i is currently connected with. In summary, graph-based Fig.3. Tradeoffbetweentwocostsbyvaryingα,D(p)=30. mutation can be described as follows: we randomly change the value of an individual’s chromosome at position i to a 5 10 15 20 0.01 α 0.3 value selected from A(i). In this way, we can avoid “bad” 1 mutations because placing a node to a unconnected cluster 0.9 will only increase the total cost. 0.8 andIndtihsicsusssecstiimonu,IlVawt.ieoSnaIpMrpeUlsyuLltAthsT.eIWOcuNesRtuosEmeSiUztheLedTSpgaernaemtiectearlsgoorfiththme Distribution probability00000.....34567 example in II-C in our simulation. To avoid small dynamic 0.2 range of α, we rescale the computational and fronthauling 0.1 cost with respect to their maximum value. Other important 0 3 4 5 6 7 8 9 10 parameters of GA are shown in Table V. Note that the Node index initialization function is also graph-based, i.e. we intialize Fig.4. Clusteringschemesunderdifferentα,D(p)=30. nodes to the clusters that have connected seeds. A. Tradeoff between computation and fronthauling costs B. Influence of cooperative structures Next we show how the proposed algorithm can achieve The presence of cooperative processing structures has great different tradeoffs between computational and fronthauling influenceontheoutcomesofouralgorithm.Herewecompare costs by varying parameter α. Fig. 3 shows the average (over the simulation results of baseband structures with and without 10simulationruns)computationalandfronthaulingcostsusing CoMP. The clustering statistics (averaged over 10 simulation α ∈ [0.01,0.3]. The tradeoff between these two costs can runs) are shown in Fig. 5. As can be seen, more baseband be clearly observed: when α increases, computational cost is functionarecentralizedunderthepresenceofCoMPcompared reduced while fronthauling cost is increased. To understand with non-CoMP scenario. The reason is that, cooperating how this tradeoff is possible, we show the corresponding MIMO functions have large interconnection bandwidth re- clustering schemes in Fig. 4. The x-axis indicates the node quirements, but the fronthauling cost between distributed cell indexes,whilethey-axisindicatestheprobabilitythatacertain sites is high. Distributed placement of CoMP modules will type of node is distributed at cell sites8. As α increases incur high fronthauling cost and should thus be avoided. In (color becomes warmer), the computational cost of placing realnetworks,notallresourceblocksarescheduledforCoMP processingfunctionsatremotesitesalsoincreases.Asaresult, operation. In that case, we can combine the results of CoMP more nodes are centralized to the central office to save com- andnon-CoMPcasestosavefronthaulbandwidth.Specifically, putational resources at the expense of increased fronthauling we can centralize only cooperating MIMO functions, and cost. Also notice that, no matter what value α takes on, the leave other MIMO funcitons at cell sites. In this way, the decode nodes are always centralized. This is because these links between cell sites and central office will have less total nodes has very high computational complexity, the schemes bandwidth. which place them at remote sites has large delay penalty and C. Influence of delay constraints isprohibited.Thisphenomenonprovidesanintuitiveguideline to centralize computation-intensive functions. Wealsoinvestigatetheinfluenceofdelayconstraint.InFig. 6, we show the average (over 10 simulation runs) computa- tional and fronthauling cost under delay thresholds ranging 8We do not show distributions for radioTX, radioRX, sourceDL, and sinkULbecausetheyareseednodesandhavedeterministicclusterindex. from 1 to 20. As can be observed in this figure, differ- to further tune the graph and algorithm parameters according 1 non−CoMP to realistic baseband processing structures so that the results 0.9 CoMP can be more practical. 0.8 0.7 ACKNOWLEDGMENT Distribution probability0000....3456 sN6e1aa2Ttri0cho1hins1aP9lwr1SoocgraikrneadnmiscNeosoFpf.ooCun6hns1oid4nra0aet1di(o29n5i7n0o3,fpPtaChrrhoetigbnCryaarme(tNah:teSi2vF0NeC1aR2)tCieuosBnneda3ael1rrc6Bhg0ar0sGa1incr)to,NRuthpeoes-. of NSFC under grant No. 61321061, and Intel Collaborative 0.2 Research Institute for Mobile Networking and Computing. 0.1 0 REFERENCES 1 2 3 4 5 6 7 8 9 10 11 12 Node index [1] S. Zhou, Z. Niu, P. Yang, and S. Tanabe, “CHORUS: A framework for scalable collaboration in heterogeneous networks with cognitive Fig.5. ClusteringstatisticswithandwithoutCoMP,α=0.05,D(p)=30. synergy,”IEEEWirelessCommunications,vol.20,no.4,pp.133–139, August2013. [2] Z. Niu, S. Zhou, S. Zhou, X. Zhong, and J. Wang, “Energy efficiency and resource optimized hyper-cellular mobile communication system 13 architecture and its technical challenges,” Science China Information Science,vol.42,no.10,pp.1191–1203,2012,(InChinese). 12 [3] ChinaMobileResearchInstitute.(2014,June)C-RAN:Theroadtowards green RAN (version 3.0). [Online]. Available: http://labs.chinamobile. 11 com/cran/wp-content/uploads/2014/06/20140613-C-RAN-WP-3.0.pdf fronthauling cost109 [[54]] SJA/m/..auursLBlxttiiiihvpuna.l,,ouerxmSgUi./niakSgbZ,AshS/,go1.au4i2Pn,00.81J.Co.54fh9,Ga9vn0otiodnrrtgua,bbaloeZs.ebp,aNusMebiul.i,ssKthaae.tndidoJ.antSa[Op.pronoXolliluuns,,,e”]G“.Oi.nAnKvGutamhlioleaabbre,slCteAa:ot.imshMt’it1ctup4a-l:, ralidhar,P.Polakos,V.Srinivasan,andT.Woo,“CloudIQ:aframework forprocessingbasestationsinadatacenter,”inProceedingsofthe18th 8 annualinternationalconferenceonMobilecomputingandnetworking. Istanbul,Turkey:ACM,2012,pp.125–136. 7 [6] “CPRIspecificationv6.0:Interfacespecification,”2013. 0 5 10 15 20 25 30 Computational cost [7] D.Samardzija,J.Pastalan,M.MacDonald,S.Walker,andR.Valenzuela, “Compressed transport of baseband signals in radio access networks,” IEEE Transactions on Wireless Communications, vol. 11, no. 9, pp. Fig.6. Clusteringstatisticswithdifferentdelayconstraint,α=0.01,w/o 3216–3225,September2012. CoMP. [8] J. Lorca and L. Cucala, “Lossless compression technique for the fronthaul of lte/lte-advanced cloud-ran architectures,” in IEEE 14th InternationalSymposiumandWorkshopsonaWorldofWireless,Mobile ent delay threshold will result in different tradeoff. Smaller andMultimediaNetworks(WoWMoM),June2013,pp.1–9. [9] J.Liu,T.Zhao,S.Zhou,Y.Cheng,andZ.Niu,“CONCERT:Acloud- threshold values make distributed placement prone to higher basedarchitecturefornext-generationcellularsystems,”IEEEWireless delay penalty, thus the resulting clustering scheme favors Communications,2014,Accepted. centralization and have higher fronthauling cost. In contrast, [10] S. E. Schaeffer, “Graph clustering,” Computer Science Review, vol. 1, no.1,pp.27–64,2007. we can place more functions at remote sites when the delay bounds get looser. V. CONCLUSION In this paper, we present a graph-based framework for analyzingbasebandfunctionsplittingandplacementproblems in C-RAN. We re-express baseband processing structures with a graph model, and transform splitting and placement strategies into graph-clustering schemes. To solve the desired tradeoffs between computational and fronthauling costs, we present a genetic algorithm with customized fitness function and mutation module. Simulation results show that we can achieve arbitrary cost tradeoffs by varying algorithm param- eter. The investigation on CoMP and delay constraint also give important implications for function splitting in realistic systems. As a future work, we plan to apply the proposed framework to other baseband structures and investigate the cost tradeoff characteristics of these structures. Also, we plan

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.