1 On the Performance of Practical Ultra-Dense Networks: The Major and Minor Factors Ming Ding‡, Member, IEEE, David L´opez-P´erez†, Member, IEEE Abstract—In this paper, we conduct performance evaluation fold from 1950 to 2000, in which an astounding 2700× gain for Ultra-DenseNetworks(UDNs), andidentifywhichmodelling was achieved through network densification using smaller factors play major roles and minor roles. From our study, we cells[2].Inthefirstdecadeof2000,networkdensificationcon- draw the following conclusions. First, there are 3 factors/models tinuedto serve the 3rdGenerationPartnershipProject(3GPP) thathaveamajorimpactontheperformanceofUDNs,andthey should be considered when performing theoretical analyses: i) a 4th-generation (4G) Long Term Evolution (LTE) networks, multi-piecepathlossmodelwithline-of-sight(LoS)andnon-line- and is expected to remain as one of the main forces to drive of-sight (NLoS) transmissions; ii) a non-zero antennaheightdif- the 5th-generation (5G) networks onward [2], [3]. 7 ference between base stations (BSs) and user equipments(UEs); 1 Indeed,in thefirstdeploymentphaseof5G, theorthogonal iii)afiniteBS/UEdensity.Second,thereare4factors/modelsthat 0 deployment of ultra-dense (UD) small cell networks (SCNs), haveaminorimpactontheperformanceofUDNs,i.e.,changing 2 the results quantitatively but not qualitatively, and thus their or simply ultra-dense networks (UDNs), within the existing r incorporationintotheoreticalanalysesislessurgent:i)ageneral macrocell network at sub-6GHz frequencies, is envisaged as p multi-pathfadingmodel based on Rician fading;ii) acorrelated one of the workhorses for capacity enhancement in 5G, due A shadow fading model; iii) a BS density dependent transmission to its large spatial reuse of spectrum and its easy manage- 3 fpaocwtoerrs;/mivo)daelsdefoterrfmuitnuirsetisctuBdSy/:usi)eradBeSnvsietryt.icFailnaanlltye,ntnhaerpeatatreern5; ment. The latter one arises from its low interaction with the ii) multi-antenna and/or multi-BS joint transmissions; iii) a macrocell tier, e.g., no inter-tier interference [2]. ] proportional fair BS scheduler; iv) a non-uniform distribution Here, the orthogonal deployment means that small cells I N of BSs; v) a dynamic time division duplex (TDD) or full duplex andmacrocellsareoperatingondifferentfrequencyspectrum, (FD) network. Our conclusions can guide researchers to down- . i.e., 3GPP Small Cell Scenario #2a [3]. In contrast, another s select the assumptions in their theoretical analyses, so as to c avoid unnecessarilycomplicated results, whilestillcapturingthe wayofdeployingsmallcellsandmacrocellsistheco-channel [ fundamentals of UDNs in a meaningful way. deployment, where they are operating on the same frequency spectrum, i.e., 3GPP Small Cell Scenario #1 [3]. 7 v I. INTRODUCTION 4 B. The Performance Analysis of Ultra-Dense Networks 6 Recent market forecasts predict that the mobile data traffic 9 volume density will keep growing towards 2030 and beyond TheperformanceanalysisofUDNs,however,ischallenging 7 the so-called 1000× wireless capacity demand [1]. This in- because UDNs are fundamentally different from the current 0 crease is expected to be fuelled by the growth of mobile 4G sparse/dense networks, and thus it is difficult to iden- . 1 broadbandservices,wherehigh-qualityvideos,e.g.,ultra-high tify the essential factors that have a key impact on UDN 0 definition and 4K resolution videos, are becoming an integral performance. To elaborate on this, Table I provides a list 7 part of today’s media contents. Moreover, new emerging of key factors/models/parameters related to the performance 1 : services such as machine type communications (MTC) and analysis of SCNs, along with their assumptions adopted in v internet of things (IoT) will contribute to the increase of the 3GPP. This list is far from exhaustive, but it includes i X massive data. Thisposesan ultimate challengeto the wireless those assumptions that are essential in any SCN performance r industry, which must offer an exponentially increasing traffic evaluation campaign in the 3GPP [4], [5]. For clarity, the a in a profitable and energy efficient manner. To make things assumptions in Table I are classified into two categories, i.e., even more complex, the current economic situation around network scenario (NS) and wireless system (WS). the globe aggravates the pressure for mobile operators and More specifically, vendors to stay competitive, rendering the decision on how • TheassumptionsontheNScharacterisethedeployments to increase network capacity in a cost-effective manner even of base stations (BSs) and user equipments (UEs). more critical. • The assumptions on the WS characterise the channel models and the transmit/receive capabilities. A. The Role of Ultra-Dense Networks in 5G Considering the 12 factors listed in Table I, a straightfor- Previous practice in the wireless industry shows that the ward methodology to understand the fundamental differences wireless network capacity has increased around one million between UDNs and sparse/dense networks would be to in- vestigate the performance impact of those factors and their ‡Ming Ding is with Data61, CSIRO, Australia (e-mail: combinations one by one, thus drawing useful conclusions [email protected]). †DavidL´opez-P´ereziswithNokiaBellLabs,Ireland(email: david.lopez- on which factors should define the fundamental behaviours [email protected]). of UDNs. Since the 3GPP assumptions on those factors were 2 TableI ALISTOFFACTORSFORTHEPERFORMANCEANALYSISOFUDNS. Factor Factor Assumption in our SG analysis Assumption in Factor and its assumption in the 3GPP [4], [5] category index (Sections II and III) our simulation NS 1 Finite BS and UE densities X X Network NS 2 Deterministic BS and UE densities Random BS and UE numbers X [Focus of this paper] Scenario (Poisson-distributed) (NS) Non-uniform distribution of BSswith some Unconstrained uniform Unconstrained uniform NS 3 constraints on the minimum BS-to-BS distance distribution of BSs distribution of BSs NS 4 Downlink, uplink, dynamic TDD (dynamic link) Downlink only Downlink only WS 1 Multi-piece path loss with LoS/NLoS transmissions X X 3D BS-to-UE distance for path loss calculation WS 2 X X considering BS/UE antenna heights Wireless WS 3 Generalized Rician fading Rayleigh fading X [Focus of this paper] System WS 4 Correlated shadow fading None X [Focus of this paper] (WS) WS 5 BS density dependent BS transmission power Constant BS transmission power X [Focus of this paper] WS 6 BS vertical antenna pattern None None WS 7 Multi-antenna and/or multi-BS joint transmissions None None WS 8 PF scheduler RR scheduler RR scheduler agreed upon by major companies in the wireless industry all It is veryimportantto note thatwe have already excluded overthe world,themoreoftheseassumptionsananalysiscan two popular NS assumptions from Table I, i.e., millimetre consider, the more practical the analysis will be. wave communications and heterogeneous network scenar- The theoretical community has already started to explore ios. Our reasons are as follows, this approach, and some of those assumptions in Table I • UDNsdonotnecessarilyhavetoworkwiththemillimetre have already been considered in various works to derive the wave spectrum, as such topic stands on its own and performanceofUDNs,e.g.,throughstochasticgeometry(SG) requires a completely new analysis [15], [16]. Hence, in analyses. In more detail, in SG analyses, BS positions are this paper, we focus on the spatial reuse gain of UDNs typically modelled as a Homogeneous Poisson Point Process at sub-6GHz frequencies, and we do not consider the (HPPP)ontheplane,andclosed-formexpressionsofcoverage new features of millimetre wave communications, such probability can be found for some scenarios in single-tier as short-range coverage, very low inter-cell interference, cellular networks [6] and multi-tier cellular networks [7]. the blockage effect, a very high Doppler shift [16]. Using a simple modelling, the major conclusion in [6], [7] • AsbrieflydiscussedinSubsectionI-A,therearetwoways is that neitherthe numberof cells northe numberof cell tiers of deploying UDNs within the existing macrocell net- changesthe coverageprobabilityin interference-limitedfully- works: the orthogonaland the co-channel deployments. loaded wireless networks. – In the former case, UDNs and the macrocell net- Recently,afewnoteworthystudieshavebeencarriedoutto workscanbestudiedseparately,makingthenetwork revisitthenetworkperformanceanalysisofUDNsundermore scenario for UDNs a homogeneousone. practical assumptions[8]–[14]. These new studies include the – In the latter case, the strong inter-tier interfer- following assumptions in Table I: i) a multi-piece path loss ence [17] from the macrocell networks to UDNs model with line-of-sight (LoS) and non-line-of-sight (NLoS) should be mitigated by some time domain inter- transmissions(WS1),ii) a non-zeroantennaheightdifference cell interference coordination techniques, e.g., the between BSs and UEs (WS2), and iii) a finite BS/UE den- almost blank subframe (ABS) mechanism [2]. Dif- sity (NS1). The inclusion of these assumptions significantly ferent from a normal subframe, in an ABS, no changed the previous conclusion, indicating that the coverage controlor data signals butonlyreferencesignalsare probability performance of UDNs is neither a convex nor a transmitted in the macrocell tier, thus significantly concave function with respect to the BS density. reducing the inter-tier interference since reference signalsonlyoccupyaverylimitedportionofatime- frequency resource block. However, when the small C. Our Contributions and Paper Structure cellnetworkgoesultra-dense,themacrocellnetwork In lightofsuch drastic changeof conclusions,onemay ask needs to convert all subframes to ABSs, so as to the followingintriguingquestion:What if we furtherconsider decreaseitsinterferencetothelargenumberofsmall moreassumptionsinSGanalyses?Willitqualitativelychange cells in its vicinity. In other words, the macrocell the recently obtained conclusions in [8]–[14]? In this paper, network would basically mute itself to clear the we willaddress thisfundamentalquestionby investigating way for UDNs, and thus again making the network theperformanceimpactofthe3GPPassumptionsofWS3, scenario for UDNs a homogeneousone. WS4, WS5 and NS2 in Table I. Therefore, we come to the conclusion that there is no 3 strong motivation to study UDNs in a heterogeneous where L is in the order of several meters for the current 4G network scenario, which is thus absent in Table I. networks. For example, according to the 3GPP assumptions The rest of this paper is structured as follows: for small cell networks, L equals to 8.5m, as the BS antenna heightandtheUE antennaheightareassumedto be10mand • SectionIIdescribesthenetworkscenarioandthewireless 1.5m, respectively [19]. system model used in our SG analysis, mostly recom- Followingthe3GPPrecommendations[4],[5],weconsider mended by the 3GPP. • Section III presents our previous theoretical results in practical line-of-sight (LoS) and non-line-of-sight (NLoS) transmissions, and treat them as probabilistic events. Specifi- terms of the coverage probability, and discusses our cally, we adopt a very general path loss model, in which the main results on the performance impact of the 3GPP assumptions of WS1, WS2 and NS1 in Table I. path loss ζ(w) is segmented into N pieces, i.e., • Section IV describes in more detail the newly consid- ζ1(w), when 0≤w≤d1 ered assumptions addressed in this paper, i.e., the 3GPP assumptions of WS3, WS4, WS5 and NS2. ζ2(w), when d1 <w≤d2 • Section V discloses our simulation results on the perfor- ζ(w)=... ... , (2) mance impact of these newly considered assumptions. ζN(w), when w >dN−1 • SpaeccttioofnthVeIrpermoaviindiensgd3iGscPuPssaiosnsuomnpttihoenspoerffWormSa6n,cWeSim7-, where each piece ζn(w),n∈{1,2,...,N} is modelled as WS8, NS3 and NS4, whichare leftas ourfuturework. • The conclusions are drawn in Section VII. ζn(w)=(ζζnnLNL(w(w))==AALnNnwL−wα−Ln,αNnL, LNoLSoSPrPorbo:bP:r1Ln−(wP)rLn(w), (3) II. NETWORK SCENARIO AND WIRELESS SYSTEMMODEL where In this section, we present the network scenario and the wirelesssystemmodelconsideredinthispaper.Notethatmost • ζnL(w)andζnNL(w),n∈{1,2,...,N}arethen-thpiece path loss functions for the LoS transmission and the of our assumptions on cellular networks are in line with the NLoS transmission, respectively, recommendations by the 3GPP [4], [5]. • ALn and ANnL are the path losses at a reference distance r =1 for the LoS and the NLoS cases, respectively, A. Network Scenario • αLn and αNnL are the path loss exponentsfor the LoS and the NLoS cases, respectively. We consider a cellular network with BSs deployed on a plane according to a homogeneous Poisson point process • In practice, ALn, ANnL, αLn and αNnL are constants obtain- 2 able from field tests [4], [5]. (HPPP) Φ with a density of λBSs/km . Active UEs are also Poisson distributed in the considered network with a density • PrLn(w) is the n-th piece LoS probability function that 2 a transmitter and a receiver separated by a distance w of ρUEs/km . We only consider active UEs in the network has a LoS path, which is assumed to be a monotonically because non-active UEs do not trigger data transmission, and decreasing function with regard to w. Such assumption thus they are ignored in our analysis. Note that the total UE has been confirmed by [4], [5]. number in cellular networks should be much higher than the number of the active UEs, but at a certain time slot and on For convenience, ζL(w) and ζNL(w) are further n n a certain frequency band, the active UEs with data traffic stacked into piece-wise functions written as demandsmaynotbetoomany.A typicaldensityofthe active (cid:8) (cid:9) (cid:8) (cid:9) UEs in 5G should be around 300UEs/km2 [2]. ζ1Path(w), when 0≤w ≤d1 In practice, a BS should mute its transmission if there Path ζ2Path(w), when d1 <w ≤d2 icselnlointUerEfecreonncneecatneddetnoeritg,ywchoincshumrepdtuiocnes[1u4n]n.eScienscsearUyEisntaerre- ζ (w)=... ... , (4) randomly and uniformly distributed in the network, it can be ζNPath(w), when w >dN−1 aΦ˜ss[u1m8]e,dththeadtethnesitayctoivfewBhSicshalissodfeonlolotewdabnyHλ˜PBPSPsd/kismtr2i.buNtiootne w“NhLer”efothrethsetrLinogSvaanrdiabthlee NPLatohS tcaakseess,trheespveacltuiveeloyf. B“Le”sidaensd, that 0≤λ˜ ≤λ, and a larger ρ leads to a larger λ˜. Details on PrL(w) is also stacked into a piece-wise function as n the computation of λ˜ can be found in [14]. (cid:8) (cid:9) PrL1(w), when 0≤w ≤d1 B.WWeirdeelensosteSbysytermthMeotwdeol-dimensional(2D)distancebetween PrL(w)=P... rL2(w), w... hen d1 <w ≤d2 . (5) a BS and an a UE, and by L the absolute antenna height dif- PrLN(w), when w>dN−1 ferencebetweenaBSandaUE.Hence,thethree-dimensional Asaspecialcase,inthefollowingsubsections,weconsider (3D) distance between a BS and a UE can be expressed as a two-piece path loss function and a LoS probability function w = r2+L2, (1) defined by the 3GPP [4]. Specifically, we use the following p 4 path loss function, ALw−αL, LoS Prob: PrL(w) ζ(w)= , (6) ( ANLw−αNL, NLoS Prob: 1−PrL(w) together with the following LoS probability function, PrL(w)= 1−5exp(−R1/w), 0<w ≤d1 , (7) ( 5exp(−w/R2), w >d1 where R1 = 156 m, R2 = 30 m, and d1 = lnR110 [4]. The combination of the path loss function in (6) and the LoS probabilityfunctionin (7) can be deemed as a special case of the proposed path loss model in (2) with the following sub- stitutions: N = 2, ζL(w) = ζL(w) = ALw−αL, ζNL(w) = 1 2 1 ζ2NL(w) = ANLw−αNL, PrL1(w) = 1−5exp(−R1/w), and PrL2(w) = 5exp(−w/R2). For clarity, this model is referred Fig.1. Theoreticalperformancecomparisonofthecoverageprobabilitywhen to as the 3GPP Path Loss Model hereafter. the SINR threshold γ = 0dB. Note that all the results are obtained using the3GPPPathLossModelintroducedinSubsectionII-B.Moreover,theBS Moreover, in this paper, we also assume a practical user densityregionsforthe4Gand5Gnetworkshavebeenillustratedinthefigure, association strategy (UAS), in which each UE is connected consideringthatthemaximumBSdensityofthe4GSCNsisintheorderof to the BS with the smallest path loss (i.e., with the largest 102BSs/km2 [2],[3]. ζ(r)) to the UE [4], [5]. We also assume that each BS/UE and gi is the multi-path fading channel gain associated with is equipped with an isotropic antenna, and that the multi-path bi. Note that when all BSs are assumed to be active, the set fading between a BS and a UE is modelled as independently of all BSs Φ should be used in the expression of I [8], [9], identical distributed (i.i.d.) Rayleigh fading [8], [11]–[14]. agg [11]–[13].However,inoursystemmodelwithidlemodeatthe small cell BSs and a finite UE density (see Subsection II-A), III. DISCUSSION ON THESTATE-OF-THE-ARTRESULTS only the active BSs in Φ˜ \bo inject effective interference into ONTHE PERFORMANCE ANALYSIS OF UDNS thenetwork,whereΦ˜ denotesthesetoftheactiveBSs.Hence, Using the theory of stochastic geometry (SG) and the theBSsinidlemodearenottakenintoaccountintheanalysis presentedassumptionsinprevioussubsections,weinvestigated of I shown in (10), due to their muted transmissions. agg the coverageprobabilityperformanceofSCNs byconsidering the performance of a typical UE located at the origin o. In B. Summary of Previous Findings such studies [12]–[14], we analysed in detail the performance As a summary, to illustrate our findings in [12]–[14], we impact of the 3GPP assumptions of WS1, WS2 and NS1 in plot the SCN performance results in terms of the coverage Table I. The concept of coverage probability and a summary probability in Fig. 1. of our results are presented in the following. The results in Fig. 1 are analytical ones validated by simulations in [12]–[14]. Note that in Fig. 1, L denotes the A. The Coverage Probability absolute antenna height difference between BSs and UEs. As The coverage probability is defined as the probability that indicated in Subsection II-B, L = 8.5m in the current 3GPP the signal-to-interference-plus-noiseratio (SINR) of the typi- assumption for small cell scenarios, while L = 0m is a cal UE is above a designated threshold γ: futuristic assumption where BS antennas are installed at the UE height. Besides, the other parameters used to obtain the pcov(λ,γ)=Pr[SINR>γ], (8) results in Fig. 1 are αL =2.09, αNL =3.75, AL =10−10.38, where the SINR is computed by ANL =10−14.54, P =24 dBm, PN =−95 dBm [4]. From Fig. 1, we can draw the following observations: Pζ(r)h SINR= Iagg+PN, (9) • Performance impact of WS1: The curve with plus markersrepresentstheresultsin [12], wherewe consider where h is the channel gain, which is modelled as an ex- the 3GPP multi-piece path loss function with LoS/NLoS ponentially distributed random variable (RV) with a mean of transmissions [4] and ignore the 3GPP assumptions of one (due to our consideration on Rayleigh fading presented WS2 and NS2, i.e., setting L = 0m and deploying an before), P is the transmission power at each BS, PN is the infinitenumberofUEs.Fromthoseresults,itcanbeseen additive white Gaussian noise (AWGN) power at the typical thatwhentheBSdensityislargerthanathresholdaround UE, and Iagg is the cumulative interference given by 102BSs/km2, the coverage probability will continuously decrease as the SCN becomes denser. This is because Iagg = Pβigi, (10) UDNs imply high probabilities of LoS transmissions i:biX∈Φ˜\bo between BSs and UEs, which leads to a performance where bo is the BS serving the typical UE, bi is the i-th degradationcausedbyafastergrowthoftheinterference interfering BS, βi is the path loss from bi to the typical UE, power compared with the signal power [12]. This is due 5 to the transition of a large number of interference paths widely adopted [5]. Hence, we consider the practical multi- fromNLoS(usuallywithalargepathlossexponentαNL) path Rician fading model defined in the 3GPP [5], where the to LoS (usually with a small path loss exponent αL). K factorindBscale(theratiobetweenthepowerinthedirect • Performance impact of WS2: The curve with square path and the power in the other scattered paths) is modelled markersrepresentsthe resultsin [13], whereweconsider as K[dB]=13−0.03w, where w is defined in (1). the 3GPP assumptions of multi-piece path loss function with LoS/NLoS transmissions and L = 8.5m [4], while B. A correlated shadow fading model (WS4 in Table I) still keeping the assumption of an infinite number of InSGanalyses,theshadowfadingisusuallynotconsidered UEs. From those results, it can be seen thatthe coverage or simply modelled as independently identical distributed probability shows a concerning trajectory toward zero when the BS density is larger than 103BSs/km2. This (i.i.d.) RVs. However, in the 3GPP, a more practical corre- lated shadow fading is often used [4], [5], [20]. Hence, we is because UDNs make the antenna height difference considerthepracticalcorrelatedshadowfadingmodeldefined between BSs and UEs non-negligible, which gives rise to another performance degradation due to a cap on in 3GPP [4], where the shadow fading in dB is modelled as zero-mean Gaussian a random variable, e.g., with a standard the signal power, resulted from the bounded minimum deviation of 10dB [4]. The correlation coefficient between distance between a UE and its serving BS [13]. • Performance impact of NS1: The curves with circle the shadow fading values associated with two different BSs and trianglemarkersrepresentsthe resultsin [14], where is denoted by τ, where τ =0.5 in [4]. we consider the 3GPP assumptions of multi-piece path loss function with LoS/NLoS transmissions and a finite C. A BS density dependent BS transmission power (WS5 in 2 UE density of 300UEs/km (a typical UE density in Table I) 5G [2]).Moreover,boththe assumptionsofL=0m and In SG analyses, the BS transmission power is usually L=8.5m are investigatedin Fig. 1. As we can observe, assumedtobeaconstant.However,inthe3GPP,itisgenerally when the BS density surpasses the UE density, i.e., 2 agreed that the BS transmission power should decrease as the 300BSs/km ,thuscreatingasurplusofBSs,thecoverage SCN densifies because the per-cell coverage area shrinks [4]. probabilitywillcontinuouslyincrease.Suchperformance Hence, we embrace the practical self-organisingBS transmis- behaviourofthecoverageprobabilityincreasinginUDNs sion power framework presented in [2], in which P varies isreferredtoastheCoverageProbabilityTakeoff in[14]. withtheBSdensityλ.Specifically,thetransmitpowerofeach The intuition behind the Coverage Probability Takeoff is BS is configured such that it provides a signal-to-noise-ratio thatUDNsprovideasurplusofBSs withrespecttoUEs, (SNR) of η0 = 15 dB at the edge of the average coverage which providea performanceimprovement, thanksto the area for a UE with NLoS transmissions. The distance from BS diversity gain and the BS idle mode operation. In a cell-edge UE to its serving BS with an average coverage more detail, as discussed in II-A, since the UE density 1 is finite in practical networks, a large number of BSs area is calculated by r0 = λπ, which is the radius of an could switch off their transmission modules in a UDN, equivalent disk-shaped coverqage area with an area size of 1. λ thus enter idle modes, if there is no active UE within Therefore,theworst-casepathlossisgivenbyANLr−αNL and 0 their coverage areas. This helps to mitigate unnecessary therequiredtransmissionpowertoenableaη0 dBSNRinthis inter-cell interference and reduce energy consumption. case can be computed as [2] η0 1010PN IV. DISCUSSION ON THEINCORPORATION OF MORE 3GPP P (λ) = . (11) ASSUMPTIONS INTOTHEMODELING ANLr0−αNL Toinvestigatewhetherthepreviousconclusionsstillholdin In Fig. 2, we plot the BS density dependent transmission more practical network scenarios, additional assumptions [4], power in dBm to illustrate this realistic power configuration [5]willalsobeconsideredinouranalysisthroughsimulations. when η0 =15 dB. Note that our modelling of P is practical, The results will be discussed in Section V. Such additional covering the cases of macrocells and picocells recommended practical assumptions are the 3GPP assumptions of WS3, in the LTE networks. More specifically, the typical BS densi- WS4, WS5 and NS2 in Table I, which are presented in the ties of LTE macrocells and picocells are respectively several 2 2 sequel. Note thatthe authorsof [10] haverecently proposeda BSs/km and around 50 BSs/km [19], respectively. As a newapproachofnetworkperformanceanalysisbasedonHPPP result, the typical P of macrocell BSs and picocells BSs are intensity matching, which facilitates the theoretical study of respectively assumed to be 46 dBm and 24 dBm in the 3GPP some of these additional 3GPP assumptions. standards [19], which match well with our modelling. A. A generalmulti-pathfading modelbased on Rician fading D. A deterministic BS/UE density (NS2 in Table I) (WS3 in Table I) In SG analyses, the BS/UE number is usually modelled In SG analyses, the multi-path fading is usually modelled as a Poisson distributed random variable (RV). However, in as Rayleigh fading for simplicity. However, in the 3GPP, a the 3GPP, a deterministic BS/UE number is commonly used more practical model based on generalised Rician fading is for a given BS/UE density [4]. Hence, we use deterministic 6 Fig. 2. The BS density dependent transmission power in dBm. (In the Fig.3. Simulated performance comparison ofthecoverage probability with 3GPP[19],46dBmand24dBmareassumedformacrocell andpicocells. more3GPPassumptions whentheSINRthresholdγ=0dB. densities λBSs/km2 and ρUEs/km2 to characterize the BS the 2G/3G systems is indeed much larger than 24 dBm, and UE deployments, respectively, instead of modelling their which is the case for Fig. 1. Nevertheless, such BS numbers as Poisson distributed RVs. densitydependenttransmissionpowerhasaminorimpact on UDNs, because BSs in UDNs usually work in a interference-limitedregion,andthusthe BS transmission V. MAIN RESULTSON THEPERFORMANCE IMPACT OF THE power in the signal power and that in the aggregated ADDITIONAL 3GPPASSUMPTIONS interferencepowercancelouteachother.Thisis obvious On top of the 3GPP assumptions discussed in Subsec- from the SINR expression in (9). tion III-B, in this section, we consider the 4 additional 3GPP assumptions described in Section IV, and study their VI. FUTURE WORK performance impacts on UDNs. More specifically, using the same parameter values for Fig. 1, we conduct simulations The performance impact of the remaining 5 assumptions to investigate the coverage probability performance of SCNs in Table I, i.e., the 3GPP assumptions of WS6, WS7, while also consideringthe 3GPP assumptionsof WS3, WS4, WS7, NS3 and NS4, are left to our future work, but briefly WS5 and NS2 in Table I. The results are plotted in Fig. 3. discussed in the following subsections. Comparing Fig. 1 with Fig. 3, we can draw the following observations: A. Discussion on The Remaining 5 Assumptions • Those 4 additional 3GPP assumptions introduced in • A 3D antenna pattern (WS6 in Table I): In SG SectionIVdonotchangethefundamentalbehaviours analyses,theverticalantennapatternateachBSisusually of UDNs shown in Fig. 1, i.e., ignored for simplicity. However, in the 3GPP perfor- – the performancedegradationdueto the transitionof mance evaluations, it is of good practice to consider 3D a large number of interfering links from NLoS to antenna patterns, where the main beam is mechanically LoS, when λ∈ 102,103 BSs/km2, and/orelectricallytilteddownwardstoimprovethesignal – the further performance degradation, due to the cap poweraswellastoreduceinter-cellinterference[4],[5]. (cid:2) (cid:3) on the signal power caused by the non-zeroantenna • Multi-antenna and/or multi-BS joint transmissions height difference between BSs and UEs, when L= (WS7inTableI):InSGanalyses,eachBS/UEisusually 8.5m and λ is larger than 103BSs/km2, and equipped with one omni-directional antenna for sim- – the performance improvement when λ is larger than plicity. However, in the 3GPP performance evaluations, ρ, thus creating a surplus of BSs and thus allowing it is usual to consider multi-antenna transmissions and for idle mode operation to mitigate unnecessary receptions[21]–[23],evenwithanenhancementofmulti- inter-cell interference. BS cooperation [24]–[26]. The consideration of multi- • The performance behaviour of sparse networks (λ ∈ antenna technologies in small cell networks expands the 10−1,100 BSs/km2) is different in Fig. 3 compared realmofUDNs,whichopensupnewavenuesofresearch with that in Fig. 1. This is mainly due to the larger BS topics for further study. (cid:2) (cid:3) transmissionpowerusedinFig.3forsparsenetworks,as • A proportional fair scheduler (WS8 in Table I): In displayed in Fig. 2, which is helpfulto remove coverage SG analyses, usually a typical UE is randomly chosen holesinthenoise-limitedregion.Basedonourknowledge for the performanceanalysis, which implies that a round ofthesuccessfuloperationoftheexisting2G/3Gsystems, Robin(RR) scheduleris employedineach BS. However, the results in Fig. 3 make more sense than those in in the 3GPP performanceevaluations,a proportionalfair Fig. 1, since the macrocell BS transmission power in (PF) scheduler is often used as an appealing scheduling 7 Thus, dynamicTDD can providea tailored configuration of DL/UL subframe resources for each cell or a cluster of cells at the expense of allowing inter-cell inter-link interference,e.g.,DLtransmissionsofacellmayinterfere with UL ones of a neighboring cell, and vice versa. The study of dynamic TDD is particularly important for UDNs because dynamic TDD is the predecessor of full duplex (FD) [40], [42] technology, which has been identified as one of the candidate for 5G. In more detail, – In an FD system, a BS can simultaneously transmit to and receive from different UEs, thus enhancing spectrum reuse, but creating both inter-cell inter- link interference and intra-cell inter-link interfer- ence, a.k.a., self-interference [42]. – The main difference between an FD system and a Fig.4. Illustration oftheidealBSdeployment inahexagonalgridnetwork. dynamic TDD one is that self-interference does not TheBSdensityisaround50BSs/km2,whichisatypical onein4G. exist in dynamic TDD [40]. The inter-link interference is expected to have a non- technique that can offer a better system throughput than minorimpactontheperformanceofUDNsinthecontext the RR scheduler, while maintaining the fairness among of a dynamic TDD or FD network, because the DL gen- UEs with diverse channel conditions [27]. Some prelim- erally overpowersthe UL, which creates high imbalance inary simulation results on the performanceof small cell among interference links in UDNs. networks considering the RR scheduler can be found in [2], [28]. B. Qualitative Results vs. Quantitative Results • A non-uniform distribution of BSs with some con- straints on the minimum BS-to-BS distance (NS3 in Finally, it is very important to point out that even if an Table I): In SG analyses, BSs are usually assumed to analysis can treat all the assumptions in Table I, a non- be uniformly deployed in the interested network area. negligible gap may still exist between the analytical results However,inthe3GPPperformanceevaluations,smallcell andtheperformanceresultsinreality,becauseofthefollowing clusters are often considered,and it is forbiddento place non-tractable factors [4]: anytwoBSstooclosetoeachother[4].Suchassumption • non-full-buffertraffic, is in line with the realistic network planning to avoid • hybrid automatic repeat request (HARQ) processes, strong inter-cell interference.It is interesting to note that • non-linear channel measurement errors, several recent studies are looking at this aspect from • quantized channel state information (CSI), difference angles [29]–[31]. In particular, a deterministic • UE misreading of control signalling, hexagonalgridnetworkmodel[6]mightbeusefulforthe • discrete modulation and coding schemes, analysis.Morespecifically,wecanconstructanidealistic • UE mobility and handover procedures, BS deployment on a perfect hexagonal lattice, and then • imperfect backhaul links, and so on; we can perform a network analysis on such BS deploy- Hence, in the context of the performance analysis for menttoextractanupper-boundoftheSINRperformance. UDNs, a high priority should be given to identifying the NotethattheBSdeploymentonahexagonallatticeleads performance trends qualitatively, rather than improving to an upper-bound performance because BSs are evenly the numerical results quantitatively. distributed in the network scenario, and thus very strong interference due to close proximity is precluded in the VII. CONCLUSION analysis [6]. Illustration of such hexagonal grid network In this paper, we have conducted a performance evaluation is provided in Fig. 4. ofUDNs,andidentifiedwhichmodellingfactorsreallymatter • A dynamic time division duplex (TDD) or full duplex intheoreticalanalyses.Fromourstudy,wehaveidentifiedthat (FD) network (NS4 in Table I): In SG analyses, most 3 factors/modelshave a major impact on the performance studies focus on the downlink (DL) network scenario as of UDNs, and they should be considered when performing in [8]–[14]. It is of great interest to see whether new theoretical analyses: conclusions can be drawn for the uplink (UL) network scenario. Different from the DL, a fractional power • a multi-piece path loss model with LoS/NLoS transmis- controlmechanismiscommonlyusedattheUEside[32], sions; [33].Moreover,anewtechnology,referredtoasdynamic • a non-zero antenna height difference between BSs and TDD, has been standardized in the 3GPP [34]–[41]. In UEs; dynamicTDD,the DL/UL subframenumberin each cell • a finite BS/UE density. or a cluster of cells can be dynamically changed on In contrast, we have found that the following 4 fac- a per-frame basis, i.e., once every 10 milliseconds [4]. tors/models have a minor impact on the performance of 8 UDNs, i.e., change the results quantitatively but not qualita- [18] S.LeeandK.Huang,“Coverageandeconomyofcellularnetworkswith tively, and thus their incorporationinto theoretical analyses is manybasestations,” IEEECommunications Letters, vol.16,no.7,pp. 1038–1040, Jul.2012. less urgent: [19] 3GPP, “TR 36.814: Further advancements for E-UTRA physical layer • ageneralmulti-pathfadingmodelbasedonRicianfading; aspects,” Mar.2010. • a correlated shadow fading model; [20] M. Ding, M. Zhang, D. López-Pérez, and H. Claussen, “Correlated shadowfadingforcellularnetworksystem-levelsimulationswithwrap- • a BS density dependent transmission power; around,” 2015 IEEE International Conference on Communications • a deterministic BS/user density. (ICC),pp.2245–2250, Jun.2015. [21] M.Ding,M.Zhang,H.Luo,andW.Chen,“Leakage-basedrobustbeam- Finally, there are 5 factors/models for future study: forming for multi-antenna broadcast system with per-antenna power • a BS vertical antenna pattern; constraintsandquantizedcdi,”IEEETransactionsonSignalProcessing, • multi-antenna and/or multi-BS joint transmissions; vol.61,no.21,pp.5181–5192,Nov.2013. [22] X. Cheng, C. X. Wang, H. Wang, X. Gao, X. H. You, D. Yuan, • a proportional fair BS scheduler; B.Ai,Q.Huo,L.Y.Song,andB.L.Jiao,“Cooperative mimochannel • a non-uniform distribution of BSs; modeling and multi-link spatial correlation properties,” IEEE Journal • a dynamic TDD or FD network. onSelectedAreasinCommunications,vol.30,no.2,pp.388–396,Feb. 2012. Our conclusions can guide researchers to down-select the [23] M. Ding, H. Luo, and W. Chen, “Polyblock algorithm-based robust assumptions in their theoretical analyses, so as to avoid beamformingfordownlinkmulti-usersystemswithper-antennapower- unnecessarily complicated results, while still capturing the constraints,” IEEE Transactions on Wireless Communications, vol. 13, no.8,pp.4560–4573, Aug.2014. fundamentals of UDNs in a meaningful way. [24] X. Tao, X. Xu, and Q. Cui, “An overview of cooperative communica- tions,”IEEECommunicationsMagazine,vol.50,no.6,pp.65–71,Jun. REFERENCES 2012. [25] M. Ding, J. Zou, Z. Yang, H. Luo, and W. Chen, “Sequential and in- [1] CISCO, “Cisco visual networking index: Global mobile data traffic crementalprecoderdesignforjointtransmissionnetworkmimosystems forecastupdate(2015-2020),” Feb.2016. withimperfect backhaul,” IEEETransactions onVehicular Technology, [2] D. López-Pérez, M. Ding, H. Claussen, and A. Jafari, “Towards 1 vol.61,no.6,pp.2490–2503,Jul.2012. Gbps/UE in cellular systems: Understanding ultra-dense small cell [26] M.DingandH.Luo,Multi-pointCooperativeCommunicationSystems: deployments,” IEEECommunications Surveys Tutorials, vol.17,no.4, TheoryandApplications. Springer, 2013. pp.2078–2101, Jun.2015. [27] J. G. Choi and S. Bahk, “Cell-throughput analysis of the proportional [3] 3GPP, “TR 36.872: Small cell enhancements for E-UTRA and E- fair scheduler in the single-cell environment,” IEEE Transactions on UTRAN-Physicallayeraspects,”Dec.2013. Vehicular Technology, vol.56,no.2,pp.766–778, March2007. [4] ——,“TR36.828:FurtherenhancementstoLTETimeDivisionDuplex [28] A. H. Jafari, D. Lopez-Perez, M. Ding, and J. Zhang, “Study on for Downlink-Uplink interference management and traffic adaptation,” scheduling techniques for ultra dense small cell networks,” Vehicular Jun.2012. TechnologyConference(VTCFall),2015IEEE82nd,pp.1–6,Sep.2015. [5] SpatialChannelModelAHG,“Subsection3.5.3,SpatialChannelModel [29] A. Al-Hourani, R. J. Evans, and K. Sithamparanathan, “Near- TextDescription V6.0,”Apr.2003. est neighbour distance distribution in hard-core point processes,” [6] J.Andrews,F.Baccelli,andR.Ganti,“Atractableapproachtocoverage arXiv:1606.03695 [cs.IT],Jun.2016. andrate incellular networks,” IEEETransactions onCommunications, [30] C.Choi,J.O.Woo,andJ.G.Andrews,“Modelingaspatiallycorrelated vol.59,no.11,pp.3122–3134, Nov.2011. cellular networkwithstrongrepulsion,” arXiv:1701.02261[cs.IT],Jan. [7] H.S.Dhillon, R.K.Ganti, F.Baccelli, andJ.G.Andrews, “Modeling 2017. andanalysisofK-tierdownlinkheterogeneouscellularnetworks,”IEEE [31] M. Ding, D. López-Pérez, G. Mao, and Z. Lin, “Microscopic analysis JournalonSelected AreasinCommunications, vol.30,no.3,pp.550– oftheuplink interference inFDMAsmallcell networks,” IEEETrans. 560,Apr.2012. onWirelessCommunications,vol.15,no.6,pp.4277–4291,Jun.2016. [8] X. Zhang and J. Andrews, “Downlink cellular network analysis with [32] T. Ding, M. Ding, G. Mao, Z. Lin, and D. López-Pérez, “Uplink multi-slope pathlossmodels,” IEEETransactions onCommunications, performanceanalysisofdensecellularnetworkswithlosandnlostrans- vol.63,no.5,pp.1881–1894, May2015. missions,” 2016 IEEE International Conference on Communications [9] T. Bai and R. Heath, “Coverage and rate analysis for millimeter-wave (ICC),pp.1–6,May2016. cellular networks,” IEEE Transactions on Wireless Communications, [33] T.Ding,M.Ding,G.Mao,Z.Lin,D.López-Pérez,andA.Y.Zomaya, vol.14,no.2,pp.1100–1114, Feb.2015. “Uplink performance analysis of dense cellular networks with los and [10] M.D. Renzo, W. Lu,and P. Guan, “The intensity matching approach: nlos transmissions,” arXiv:1609.07837 [cs.IT], vol. abs/1609.07837, Atractable stochastic geometryapproximation tosystem-level analysis Sep.2016.[Online]. Available: http://arxiv.org/abs/1609.07837 ofcellular networks,”IEEETransactions onWireless Communications, [34] M. Ding, D. López-Pérez, A. V. Vasilakos, and W. Chen, “Dynamic vol.15,no.9,pp.5963–5983, Sep.2016. TDD transmissions in homogeneous small cell networks,” 2014 IEEE [11] M.Ding,D.López-Pérez,G.Mao,P.Wang,andZ.Lin,“Willthearea International Conference on Communications Workshops (ICC), pp. spectral efficiency monotonically grow assmallcells godense?” IEEE 616–621, Jun.2014. GLOBECOM2015,pp.1–7,Dec.2015. [35] M.Ding,D.López-Pérez,R.Xue,A.V.Vasilakos,andW.Chen,“Small [12] M.Ding,P.Wang,D.López-Pérez, G.Mao,andZ.Lin,“Performance celldynamictddtransmissionsinheterogeneous networks,”2014IEEE impact of LoS and NLoS transmissions in dense cellular networks,” International Conference on Communications (ICC), pp. 4881–4887, IEEE Transactions on Wireless Communications, vol. 15, no. 3, pp. Jun.2014. 2365–2380,Mar.2016. [36] M.Ding,D.López-Pérez, A.V.Vasilakos, andW.Chen, “Analysis on [13] M.DingandD.LopezPerez,“PleaseLowerSmallCellAntennaHeights the SINR performance of dynamic TDD in homogeneous small cell in5G,”arXiv:1611.01869 [cs.IT], toappear in IEEEGlobecom 2016, networks,” 2014IEEEGlobal Communications Conference, pp. 1552– Nov.2016. 1558,Dec.2014. [14] M.Ding,D.LopezPerez,G.Mao,andZ.Lin,“Studyontheidlemode [37] H.Sun,M.Wildemeersch,M.Sheng,andT.Q.S.Quek,“D2denhanced capability with LoS and NLoS transmissions,” IEEE Globecom 2016, heterogeneous cellularnetworkswithdynamictdd,”IEEETransactions pp.1–6,Dec.2016. onWirelessCommunications,vol.14,no.8,pp.4204–4218,Aug.2015. [15] M. D. Renzo, “Stochastic geometry modeling and analysis of multi- [38] B. Yu, L. Yang, H. Ishii, and S. Mukherjee, “Dynamic TDD support tiermillimeter wavecellular networks,”IEEETransactions onWireless inmacrocell-assistedsmallcellarchitecture,”IEEEJournalonSelected Communications, vol.14,no.9,pp.5038–5057, Sep.2015. AreasinCommunications, vol.33,no.6,pp.1201–1213, Jun.2015. [16] S. Rangan, T. S. Rappaport, and E. Erkip, “Millimeter-wave cellular [39] A.K.Gupta,M.N.Kulkarni,E.Visotsky,F.W.Vook,A.Ghosh,J.G. wirelessnetworks:Potentialsandchallenges,”ProceedingsoftheIEEE, Andrews, and R. W. Heath, “Rate analysis and feasibility of dynamic vol.102,no.3,pp.366–385,Mar.2014. TDDin 5G cellular systems,”2016 IEEEInternational Conference on [17] J.Wang,X.Chu, M.Ding, and D.Lopez-Perez, “Onthe performance Communications (ICC),pp.1–6,May2016. ofmulti-tierheterogeneous networksunderlosandnlostransmissions,” in2016IEEEGlobecom Workshops,Dec.2016,pp.1–6. 9 [40] M.Ding,D.Lopez,R.Xue,A.Vasilakos, andW.Chen,“Ondynamic ofdynamicTDD:AMAClayerperspective,”arXiv:1611.02828[cs.NI], Time-Division-Duplex transmissions for small-cell networks,” IEEE Mar.2017. Transactions onVehicular Technology, vol.65,no.11,pp.8933–8951, [42] S. Goyal, C. Galiotto, N. Marchetti, and S. Panwar, “Throughput and Nov.2016. coverage for a mixed full and half duplex small cell network,” IEEE [41] M.Ding,D.Lopez-Perez,G.Mao,andZ.Lin,“Whatisthetruevalue InternationalConferenceonCommunications(ICC),pp.1–7,May2016.