The Effect of System Architecture on Net Spectral Efficiency for Fixed Services

Transcription

1 The Effect of System Architecture on Net Spectral Efficiency for Fixed Services John Naylon November 25, 2011 Abstract We consider the net spectral efficiencies of pointto-point and point-to-multipoint fixed service architectures for mobile broadband backhaul. The properties of mobile broadband backhaul traffic are examined with reference to a network case study. The network case study also provides measurements of the improvement in spectral utilisation possible with a point-to-multipoint architecture. Considering the projected growth in mobile data consumption, we argue that the gains possible with a point-to-multipoint architecture are increasing with trends in RAN design towards high peak rates and small serving cells. Our concluding argument is that technology-neutral, block assignment licensing that permits point-to-multipoint architectures for fixed services can improve net spectral efficiency and alleviate frequency band congestion. Figure 1: Point-to-point star topology 1 Fixed Service Architectures Figures 1 and 2 illustrate the simplest point-topoint(p-p) and point-to-multipoint(p-mp) fixed service architectures. The fundamental difference between the two systems relates to the channel accessarrangements. InaP-Psystem,agivenradiofrequency(RF)channel, orinthecaseofa frequency division duplex system, a pair of channels, is statically allocated to a pair of transceivers. Thereisnochannelarbitrationintheformofa medium access control(mac) scheme required, because the channel is only available to this single pairoftransceivers. Thistypeofsystemmaybe Figure 2: Point-to-multipoint topology 1

2 2 MOBILE BROADBAND BACKHAUL TRAFFIC CHARACTERISTICS 2 conceived of as a circuit-switched network architecture, where the circuit is the statically-allocated RF channel. In a point-to-multipoint system, by contrast, channel arbitration of some kind is a necessity. The central station broadcasts downstream traffic with some kind of additional identifying addressing field. All of the remote terminals(rts) areabletoreceiveanddecodeatleasttheaddressing fields and thus can discriminate traffic which is destined for them from traffic destined forotherrts. Aformofmediumaccesscontrol isusedontheupstreamdirectionwherebyanrt canindicatethatithasdatatotransmitandis granted access to the channel while transmission by other RTs is suppressed. The upstream and downstream directions are thus multiplexed, and the usual multiplexing methodologies may be employed; some form of dynamic time division multipleaccessismostcommon.ap-mpsystemembodies a packet-switched architecture. 2 Mobile Broadband Backhaul Traffic Characteristics Figure 3 is a representative seven day sample of downstream traffic destined for a tri-cellular, HSPA+21.6MbpsnodeB 1. Figure 3 illustrates one of the key properties of mobile broadband backhaul traffic, namely that itisbursty. Thischaracteristicisaresultofthe backhaul traffic being the composition of many independent data sources which are themselves bursty.forexample,seefigure4,whichisatrace of a smartphone s downstream data demand over theperiodofanhourduringwhichtheuserwas askedtousethehandsetintensively. Weusethe peak-to-mean ratio of such samples as a measure of thedegreeofburstinessofthesample.asample withauniformratehasapeak-to-meanratioof1.0 by definition and arbitrarily large figures are possible. Figure 5 shows the peak-to-mean ratio of the downstream backhaul traffic as measured for a 1 Wewillconsiderdownstreamtraffichereunlessotherwise stated, since data rates on the 3GPP downlink are normally higher than on the uplink, and consequently the downstream backhaul demands are higher than upstream. The analysis is, however, equally applicable to traffic in the uplink direction. group of 922 node Bs(again, tri-cellular HSPA+ 21.6Mbps). This network-wide analysis illustrates thattheburstinessobservedinfigure3isauniversalphenomenoninbackhauloverawiderangeof peak download speeds. There is no particular pattern observable in terms of an increase or decrease inpeak-to-meanratioasthepeaknodebbandwidth demand rises. However we can observe an average peak-to-mean ratio of 3.9 across this broad range of node B traffic profiles. An immediate consequence of this observed peak-to-mean ratio is that we can accurately estimate the utilisation of a point-to-point link carrying a typical node B s backhaul traffic. Clearly foragivenlinkwewillprovisionatleastthepeak backhauldataratesincewedonotwantthebackhaul to constrain the RAN. If we conservatively assume that we provision precisely the peak requirement,thenthemeanvolumeofdatacarriedby thelinkisrelatedtothepeakbythecarrieddata s peak-to-mean ratio. Thus the average utilisation of suchalinkcanbeatmostthereciprocaloftheaverage peak-to-mean ratio, or approximately 25% given that the average peak-to-mean ratio is 3.9. Considering the range of peak throughput valuesshownalongthehorizontalaxisoffigure5 also begins to illuminate a second property of mobile broadband backhaul; namely that backhaul bandwidth demand is not highly correlated between serving cells. Since these are tri-cellular node Bs, the theoretical instantaneous peak bandwidthshouldbe3 21.6=64.8Mbps,required whenthepeakdownloadspeedoccursatthesame timeinallthreecells.inpractice,thehighestpeak observed is around 27Mbps. It may appear counter-intuitive that backhaul bandwidth demand is not highly correlated given the well-known diurnal load pattern(as may be discerned in figure 3 for example). However, the timing of the actual peaks in demand, rather than the daily swells, is fundamentally random as they are driven by the individual locations and actions of a large population of independent users. When considering the cross-correlation between two sources, it is the samples which lie furthest fromthemeanwhichdominatethesum,i.e.the peaks and anti-peaks. To quantify the degree of correlation between node Bs backhaul demands, we examine pairs of nodebsthataregeographicallyclosetoonean-

3 2 MOBILE BROADBAND BACKHAUL TRAFFIC CHARACTERISTICS 3 Node B Measured Downstream Traffic (Mbps) Peak Rate: 23.2Mbps Mean Rate: 5.5Mbps Peak-to-Mean Ratio: Aug Aug Aug Aug Aug Aug 2011 Figure3:Downstreamtrafficforatri-cellularHSPA+nodeBovertheperiodofaweek. Smartphone Measured Downstream Traffic (Mbps) Peak Rate: 11.4Mbps Mean Rate: 0.1Mbps Peak-to-Mean Ratio: :20:00 14:30:00 14:40:00 14:50:00 15:00:00 15:10:00 15:20:00 Figure 4: Downstream traffic for a smartphone over the period of one hour. Peak-to-Mean Ratio of Downstream Traffic Node Bs Average Peak-to-Mean Ratio: Peak Downstream NodeB Backhaul Requirement (Mbps) Figure 5: The peak-to-mean ratio of downstream backhaul traffic plotted against the peak downstream backhaul traffic demand for a set of 922 HSPA+ tri-cellular node Bs.

4 2 MOBILE BROADBAND BACKHAUL TRAFFIC CHARACTERISTICS 4 other, and compute the product-moment correlation coefficient of each pair of backhaul traces. By consideringonlynodebswhichareclosetoone another, we actually select the pairs which are most likely to be well-correlated because they are servingsimilaruserpopulations(e.g.asetofnodebs servingthecentralbusinessdistrictofacitycanbe expectedalltobebusyduringofficehoursandall tobequietatnight).here,ourdefinitionof close isthatthenodebsarewithinthesamep-mpsector in the backhaul network. A histogram of the resulting coefficients is showninfigure6alongwithafittednormaldistribution. A coefficient of 1.0 would indicate perfect correlation between two signals; that is all peaks and anti-peaks occurring in perfect synchronicity, while a coefficient of 1.0 would indicate perfect anti-correlation; that is the peaks from the first signal being perfectly synchronised with the antipeaks from the second signal. A zero coefficient indicates no correlation at all. The measured mean coefficient of correlation amongst the close node Bsisonly0.16,indicatingveryweakcorrelation 2. Suchaverylowdegreeofcorrelationinpeak backhaul requirements is assumed in the Next Generation Mobile Networks Alliance s Guidelines for LTE Backhaul Traffic Estimation[1] to recommend that, when provisioning backhaul for N enodebs,oneshouldprovisionalowerboundof: Node B Pairs Fitted distribution, µ= Figure 6: Histogram of the pairwise Pearson s cross-correlation coefficient between the backhaul demands of 2262 pairs of geographically close HSPA+tri-cellularnodeBs.Afigureof1.0would indicate perfect correlation. max(peak, N busy time mean) This provisioning rule, shown diagrammatically in figure 7, relies on the observation empirically verified here, that peak requirements are statistically unlikely to occur simultaneously in nearby serving LTE cells. 2.1 Summary This section has illustrated that mobile broadband backhaul traffic has the two essential characteristics that allow multiple sources to be multiplexed onto a channel and to realise a statistical multiplexing gain: the sources have a non-uniform rate distribution, and the sources are uncorrelated to 2 Computingthecoefficientofcorrelationamongstallpairs in the network(which lies within a single time-zone) gives a mean of 6. Figure 7: NGMN provisioning guidelines for multiple enodeb backhaul, reproduced from[1].

5 3 COMPARATIVE EFFICIENCY CASE STUDY 5 some extent. Best practice in backhaul dimensioning takes into account these characteristics and the statistical multiplexing gain(sometimes known as trunking gain ) that arises from them. Poor average link utilisation of approximately 25% results from carrying this bursty packet data over circuitswitched P-P fixed services. 3 Comparative Efficiency Case Study The usual measure of spectral efficiency for a microwavelinkisthenumberofbitspersecondper Hertz(bps/Hz) achieved by the link. This captures the efficiency of the link s modulation and coding scheme, but does not consider the nature ofwhatisbeingsent.clearlyifthereisnodatato send, and we consequently send idle or framing patterns, the radio channel is still occupied but no useful work has been done. We therefore define net spectral efficiency as being the product of spectral efficiency in bps/hz and the mean end-to-end utilisation of the link. Were-examinedatafromthesamesetofHSPA+ node Bs previously referenced to quantify the net spectral efficiency gains possible with P-MP. To illustrate the methodology, we first consider a single P-MP sector containing eight node Bs connected via Ethernet to P-MP remote terminals(rts). Wefirstmeasurethepeakdataraterequiredfor eachnodeb.logically,sincewedonotwishto constrainthethroughputoftheraninthebackhaul section of the network, we must provision atleastthepeakrequirementforeachnodeb.for the purpose of this exercise, we will provision exactly the peak. If we therefore consider how much bandwidthintotalwillberequiredforapointto-point system to carry this traffic, the answer isthesumofthepeaksofeachdataset. Thisis showngraphicallyinfigure8,andtheactualsum is 123.2Mbps. If instead we envision a point-to-multipoint systemcarryingthetraffic,ateachinstantwesum the combined data requirements and then take the peakofthissummeddatasetasbeingthetotal bandwidth required. Naturally, the sources being only weakly correlated, some of the peaks and troughs in the different demands cancel out, thus Spectrum Channel Net Spectral Required Utilisation Efficiency (MHz) (%) (bps/hz) P-P P-MP P-MP Gain Table 1: Comparison for the example sector. givingusareductionintheoveralltotal. Thisis showninfigure9,andtheactualsumis77.9mbps. Because this data originates from a live network which is actually backhauled using a P-MP system, we can compare the theoretical, on-paper multiplexingshowninfigure9withwhatisactually achieved in practice. Differences arise because there are some overheads associated with the channel arbitration (which can lead to the actual total data rate being higher than theory) and also because there are additional multiplexing gains occurring below the sampling rate used togatherthedatapresentedhere(whichleadto theactualtotaldataratebeinglowerthantheory). Figure 10 shows that the correspondence between theory and practice is indeed very close. The actual peak downstream rate in the sector was 77.7Mbps. We may now calculate the amount of spectrum needed and the mean channel utilisation for the P- PandP-MParchitecturesinthiscase.WewillassumethattheRFchannelsizecanbechosenarbitrarily. In practice this is not usually the case since the channels will conform to a preferred channel arrangement recommendation specifying channel widths of 3.5MHz, 7MHz, 14MHz etc. In additionwewillnotnormallyhavetheaprioriknowledge of the exact traffic statistics that would be necessary for this precise dimensioning. Nevertheless, we make this simplifying assumption for boththep-pandp-mpcasesoasnottointroduce any bias. We will also assume a spectral efficiency of8bps/hzforbothtypesofsystem,whichisapproximately representative of the commercial state oftheartin2011. Theresultsareshownintable1. Tocompute the spectrum required, we simply take the peak

6 3 COMPARATIVE EFFICIENCY CASE STUDY Mbps Mar Mar Mar Mar Mar Mar Mar 2011 Figure 8: Point-to-point provisioning: Eight node B backhaul traces as transported across eight separate channels sized precisely for each trace s peak. Mean traffic: 39.7Mbps 7 6 Mbps Mar Mar Mar Mar Mar Mar Mar 2011 Figure 9: Point-to-multipoint provisioning: The same eight node B backhaul traces multiplexed onto a single channel sized for the peak of the joint trace. The traces are shown alternately black and grey.

7 3 COMPARATIVE EFFICIENCY CASE STUDY Mbps Mbps Mar Mar Mar Mar Mar Mar Mar 2011 Mean error: 1 Mbps 13 Mar Mar Mar Mar Mar Mar Mar 2011 Figure 10: The theoretical multiplexing of the eight node Bs backhaul requirements(solid grey) and the actually measured downstream sector data rate(black line). The lower figure shows the theoretical rate minus the measured rate, showing a mean variance of just 1Mbps between theory and practice.

8 4 RANTRENDS 8 (123.2Mbpsin thep-p case)anddivideby the spectral efficiency of 8bps/Hz to give a result in thep-pcaseof15.4mhz.tocomputethechannel utilisation we divide the mean bandwidth used by the amount of bandwidth provisioned. The mean bandwidth used in this example is 39.7Mbps in both the P-P and P-MP case(while statistical multiplexing gain means that the peaks of the individualdemandsdonotsum,themeansaresimply additive). Thus for example the utilisation in the P-MPcasehereis: 39.7Mbps 77.7Mbps =51.1% Number of sectors Finally our net spectral efficiency is the system spectral efficiency of 8bps/Hz multiplied in each case by the utilisation. We use the actual measured P-MP data rates and not the theoretical summation for all these calculations. We can quantify the amount of statistical multiplexing gain achieved by P-MP by dividing any ofthesemeasures. Inthiscasethereisagainof 1.6.Thatis,werequire1.6timesmorespectrumto carrythesamedatawithp-p;orweusethespectrum1.6timesmoreefficientlyinthep-mpcase; orournetspectralefficiencyis1.6timeshigherin the P-MP case. 3.1 Results The network in this casestudy consists of922 HSPA+ node Bs being backhauled via 237 P-MP sectorsat26ghz. Foreachofthesesectors,we calculate the statistical multiplexing gain as described above for the example sector. The results areshowninfigure11.thereareanumberofdegenerate cases where there is only a single remote terminal in the P-MP sector. In this case, naturally, thereisaunitygain. Themodalgainacrossthe network is 1.5. Thisisaverysignificantincreaseintheefficiencyofuseofspectrum. Bywayofcomparison, to achieve an equivalent increase by using higher-order modulation, it would be necessary for a point-to-point system to move from using 256-QAM to 4096-QAM Statistical Multiplexing Gain Figure 11: Histogram of the statistical multiplexing gain across the network. 4 RANTrends The most significant trends forecast for RAN deployment in the forthcoming years are the introductionofthe3gppltestandardandtheongoing densification of networks. Both these trends are driven by a need to increase network capacity to supply an exponentially growing demand for mobile broadband data from end users. 4.1 NetworkDensification Overall RAN network capacity increases in linear proportion to the number of serving cells. Thereforetheuseofnetworkdensificationasatechnique to supply capacity to meet an exponential demand leads to a super-linear increase in the numberofnodebsinaunitareaifotherfactors remain constant. If there are a greater number of nodebsperunitarea,thenthenumberofnodebs perp-mpsectorwillincreaseaslongasthep-mp sectors are not capacity limited. If we re-plot the statistical multiplexing gain resultsfromthecasestudyagainstthenumberof remote terminals per P-MP sector on the x-axis, as infigure12,weobservethattheamountofgain

9 5 CONCLUSION 9 Statistical Multiplexing Gain VectaStar P-MP Sectors Best fit Number of Remote Terminals per P-MP Sector Figure 12: The statistical multiplexing gain which is achieved by using a P-MP architecture versus a P- P architecture, plotted against the number of remote terminals per P-MP sector to illustrate increasing gain as the network becomes denser. increases with the number of remote terminals. This is intuitive: the more independent sources we multiplex together, the greater the likelihood that apeakinonesourcewillbe canceled byanantipeak in another. This is extremely strong empiricalevidencethatthebenefitsofp-mpintermsof net spectral efficiency gain increase with increasing network density. Thereisafurtherlikelygaininbenefitdueto network densification. If the number of node Bs increases then the average number of active UEs pernodebwilldecrease.theresultofthisisthat there is less multiplexing of the individual UEs offered loads occurring in the average node B itself. This in turn means that the peak-to-mean ratioofthebackhaulloadofferedbythatcellincreases(equivalently we can think of this as meaning that the backhaul load becomes more similar to a single UE s load figure 4). Finally, the proliferation of ever more numerous node Bs with smaller service radii(the small cell phenomenon) drives the simplification of the node B,primarilyforcostreasons. Thuswhileacurrent node B will typically be tri-cellular, and often multi-band,asmallcellnodebismostlikelytobe uni-cellular and single band. Once again the net resultofthisislessmultiplexingintheaccesspart of the network more burstiness in the traffic andhenceagreaterneedformultiplexinginthe next stage of the network. 4.2 More Sophisticated RAN Technology Thepeakspectralefficiencyfora2 2MIMOLTE system is approximately 8.6bps/Hz. Simulations give varying results for the mean spectral efficiencyachievedbyasystemwithamaturerealworld user population; for example[2] gives a figure of 1.3bps/Hz. The consensus appears to be in the range 1 1.5bps/Hz which implies that the peak-to-mean ratio of LTE backhaul traffic could beintherangeof Asnotedinsection2,themeanutilisationofP- Pfixedservicesisatbestequaltothereciprocal of the peak-to-mean of the traffic carried, so these simulation results imply a degradation in mean P- Pefficiencyfromacurrentaverageof25%tosomewhere in the range of %. ForaP-MPsystem,aswehaveseen,thesystem s net spectral efficiency is no longer directly related to the peak-to-mean ratio of an individual offered load. Therefore if we assume that backhaul demands amongst enode Bs remain weakly correlatedasintoday sran,wecanexpectthe net spectral efficiency benefit of P-MP over P-P to increase. 5 Conclusion The consumer shift from voice services to mobile broadband has changed the characteristics of

10 REFERENCES 10 the backhaul traffic in mobile networks, with major implications for backhaul topologies and spectrum utilisation. Our analysis of an HSPA+ network illustrates that backhaul traffic is both highly bursty and very weakly correlated. P-P fixed service topologies popular during the voiceeraarenotabletocarrydataofthiskindefficiently, since they must be over-provisioned to cope with its peaks, but will then generally operate at a much lower utilisation. The analysis shows that a P-MP topology increases spectral utilisation very significantly, resulting in a modal increase in net spectral efficiency of 50%. Looking to the futuredevelopmentoftheran,weforeseeanincrease in the degree of burstiness of backhaul trafficandanincreaseinsitedensity.weobservethat thebenefitsofp-mpversusp-pintermsofnet spectral efficiency gain increase with both of these trends. We therefore consider that, to alleviate congestion in fixedservicefrequencybands, it is advantageous to use P-MP systems because of the reduction in overall spectrum needed to deliver equivalent service. We suggest that further harmonisation of existing frequency bands for P-MP fixedservicewillbebeneficialasthedemandfor fixed service links for mobile broadband backhaul rises. We consider that opening additional bands to technology-neutral fixed service with block assignment would would also be beneficial, allowing co-existence with existing P-P services and smooth migration to more efficient P-MP systems. Revision 1.2 (created at November 18, 2011 by jbpn) First draft for review Revision 1.1 (created at October 24, 2011 by jbpn) Initial revision References [1] Next Generation Mobile Networks Alliance, Guidelines for LTE Backhaul Traffic Estimation, July [2] Ofcom, Predicting Areas of Spectrum Shortage, April Revision Log for speceff.tex Revision 1.4 (created at November 25, 2011 by jbpn) First release revision Revision 1.3 (created at November 22, 2011 by jbpn) Second draft for final review