comparison between benefits and costs of offload of mobile...

39
1 TPRC 2015: 43rd Research Conference on Communications, Information and Internet Policy (Revised version, Nov. 2015. Newer and extended version under peer review in journal) Comparison between Benefits and Costs of Offload of Mobile Internet Traffic Via Vehicular Networks Alexandre K. Ligo †‡ a , Jon M. Peha b , Pedro Ferreira c , João Barros d Carnegie Mellon University, USA University of Porto, Portugal Abstract Dedicated Short Range Communications (DSRC) is an emerging technology that connects automobiles with each other and with roadside infrastructure. The U.S. Department of Transportation may soon mandate that cars be equipped with DSRC to enhance safety. This work finds that if they do, then DSRC networks could also be an important new way to provide Internet access in urban areas that is more cost- effective than expanding the capacity of cellular networks. By combining our simulation model with data collected from an actual vehicular network that is operating in Porto, Portugal, we estimate how much Internet traffic can be carried on vehicular networks that would otherwise be carried by cellular networks under a variety of conditions. We then compare the benefits of cost savings of reduced cellular infrastructure due to offload with the cost of the DSRC vehicular network, to determine whether the former exceeds the latter. Although we find that the benefits from Internet traffic alone are not enough to justify a universal mandate to deploy DSRC in all vehicles, i.e. benefit of Internet access alone is less than total costs, the majority of DSRC-related costs must be incurred anyway if safety is to be enhanced. Thus, soon after a mandate to put DSRC in new vehicles becomes effective, the benefits of Internet access through vehicular networks in densely populated areas would be significantly greater than the remaining costs, which are the costs of roadside infrastructure that can serve as a gateway between DSRC-equipped vehicles and the Internet. Moreover, benefit of Internet access would exceed DSRC infrastructure cost in regions with lower and lower population densities over time. Keywords: mobile Internet , vehicular networks, mobile data offload, Dedicated Short Range Communications, DSRC, benefit-cost analysis, social welfare a Alexandre K. Ligo, Ph.D. Student, Carnegie Mellon University and University of Porto, [email protected] b Jon M. Peha, Professor, Carnegie Mellon University, [email protected], www.ece.cmu.edu/~peha/bio.html c Pedro Ferreira, Professor, Carnegie Mellon University, [email protected], pedro-ferreira.org d João Barros, Professor, University of Porto, [email protected], web.fe.up.pt/~jbarros

Upload: others

Post on 25-Jun-2020

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

1

TPRC2015:43rdResearchConferenceonCommunications,InformationandInternetPolicy

(Revisedversion,Nov.2015.Newerandextendedversionunderpeerreviewinjournal)

ComparisonbetweenBenefitsandCostsofOffloadofMobileInternetTrafficViaVehicularNetworks

AlexandreK.Ligo†‡a,JonM.Peha†b,PedroFerreira†c,JoãoBarros‡d†CarnegieMellonUniversity,USA

‡UniversityofPorto,Portugal

AbstractDedicated Short Range Communications (DSRC) is an emerging technology thatconnects automobiles with each other and with roadside infrastructure. The U.S.DepartmentofTransportationmaysoonmandatethatcarsbeequippedwithDSRCtoenhancesafety.Thisworkfindsthatiftheydo,thenDSRCnetworkscouldalsobeanimportantnewwaytoprovideInternetaccessinurbanareasthatismorecost-effective than expanding the capacity of cellular networks. By combining oursimulation model with data collected from an actual vehicular network that isoperating inPorto,Portugal,weestimatehowmuchInternettrafficcanbecarriedonvehicularnetworksthatwouldotherwisebecarriedbycellularnetworksunderavariety of conditions. We then compare the benefits of cost savings of reducedcellularinfrastructureduetooffloadwiththecostoftheDSRCvehicularnetwork,todeterminewhethertheformerexceedsthelatter.AlthoughwefindthatthebenefitsfromInternettrafficalonearenotenoughto justifyauniversalmandatetodeployDSRCinallvehicles, i.e.benefitofInternetaccessaloneis lessthantotalcosts, themajorityofDSRC-relatedcostsmustbeincurredanywayifsafetyistobeenhanced.Thus, soon after a mandate to put DSRC in new vehicles becomes effective, thebenefitsof Internetaccess throughvehicularnetworks indenselypopulatedareaswould be significantly greater than the remaining costs, which are the costs ofroadside infrastructure that can serve as a gateway between DSRC-equippedvehiclesandtheInternet.Moreover,benefitof InternetaccesswouldexceedDSRCinfrastructurecostinregionswithlowerandlowerpopulationdensitiesovertime.

Keywords:mobileInternet,vehicularnetworks,mobiledataoffload,DedicatedShortRangeCommunications,DSRC,benefit-costanalysis,socialwelfare

aAlexandreK.Ligo,Ph.D.Student,CarnegieMellonUniversityandUniversityofPorto,[email protected],Professor,CarnegieMellonUniversity,[email protected],www.ece.cmu.edu/~peha/bio.htmlcPedroFerreira,Professor,CarnegieMellonUniversity,[email protected],pedro-ferreira.orgdJoãoBarros,Professor,UniversityofPorto,[email protected],web.fe.up.pt/~jbarros

Page 2: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

2

1 IntroductionStandardizedtechnologynowexiststhatwouldsupportvehicularnetworks,whicharemeshnetworksrunningInternetprotocolssuchasIP.Theroutersinavehicularnetworkareplacedinautomobiles,mostofwhicharemoving,andininfrastructureplacednearroadsforthispurpose.Thistechnologymaysoonbewidelydeployed,primarilyasawayofenhancingautomotivesafety.Extensiveresearchisbeingdoneonthepotentialsafetybenefitsofvehicularnetworks(Kenney2011;U.S.DepartmentofTransportation2015;Mecklenbraukeretal.2011).Thispaperinvestigatesanentirelydifferentuseofvehicularnetworks–asanewwaytoprovideInternetaccess,especiallyformobiledevices.

Thereisstrongmotivationtofindnewcost-effectiveapproachesasInternettrafficovermobilenetworkshasbeengrowingsteadily(Sandvine2014).Somepredictthatgrowthwillbesustained;Ciscoforecastsatenfoldincreaseinmobiletrafficoverthenextfiveyears(Cisco2015).Partofthegrowthisexpectedtocomefromin-vehicleInternetusage,bothbecausevehicleoccupantswillincreasinglyusesmartphonesandlaptops,andbecauseofvehicularinfotainmentplatformssuchasAppleCarPlay,AndroidAuto,andcarmaker-proprietarysystems.Expandingcapacityofcellularnetworkstomeetsuchdemandgrowthwouldbecostly,asitwouldrequiresignificantlymorespectrum,capital,orboth.However,ifpartofthetrafficcouldbeoffloaded,i.e.deviatedfromthemacrocellularnetworkstoalternativenetworks,thenthedemandgrowthmightbemetwhileaddingfewernewcellsandthereforeincurringlowercosts.VehicularnetworksareapossiblealternativeforbringingInternetaccesstodevicesinautomobiles,aswellasdevicescarriedbypedestriansorplacedinlocationsnearroads.

Thispapershowsthatundersomeimportantcircumstances,vehicularnetworkscanprovideInternetaccessatlowercoststhanwouldbeincurredintoday’scellularnetworks.ThepaperanalyzesthecostsandbenefitsofInternetaccessthroughvehicularnetworksthatuseanemergingtechnologycalledDedicatedShortRangeCommunications(DSRC).ThedevelopmentofDSRCtechnologyisprimarilymotivatedbyroadsafetyapplicationssuchastheexchangeofwarningmessagesbetweenvehiclesenrouteofcollisionratherthanInternetaccess.TheUnitedStatesDepartmentofTransportation(USDOT)isexpectedtoproposerulemakingin2016tomandateDSRCinallnewvehicles(U.S.DepartmentofTransportation2015),andtheU.S.FederalCommunicationsCommission(FCC)hasalreadyallocated75MHzofspectrumforIntelligentTransportationSystems(U.S.FederalCommunicationsCommission;Kenney2011).TheDSRCstandardsallowpartofthe75MHzallocatedintheU.S.(and50MHzthathasbeensimilarlyallocatedintheEuropeanUnion)tobeusedforapplicationsotherthansafety(UzcateguiandAcosta-Marum2009;CampoloandMolinaro2013;Kenney2011).Non-safetyapplicationsincludevehicleandroad-relatedservicessuchasnavigationandtollcollection,aswellasInternetaccess(Zeadallyetal.2010;CampoloandMolinaro2013).

Ourcost-benefitanalysiswillinformimportantdecisionsregardingwhetherresourcesshouldbeinvestedinvehicularnetworksforthepurposeofInternetaccess,ratherthanjustvehicularsafety.ThisincludesdecisionsaboutwhetherDSRC-equippedroadsideinfrastructureshouldbedeployed,whethervehiclesshouldbeequippedwithDSRCdevices,andwhetherspectrumshouldbeallocatedforintelligenttransportationsystems.Wedeterminewhetherdecisionstoincurtheseinvestmentswouldincreasesocialwelfarebycomparingtherelevantcoststobenefits,andignoringanysunkcosts.OnedecisioniswhethertoinvestinroadsideinfrastructureforInternetaccess.InscenarioswhereDSRCspectrumhasalreadybeenallocated,asisthecaseinseveralregionsworldwide(Zeadally

Page 3: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

3

etal.2010),andwherethereisalreadyamandatetodeployDSRCOnboardUnits(OBUs)forsafetypurposes,thenspectrumandOBUcostsaresunk.Inthesescenarios,thedeploymentofroadsideinfrastructureexclusivelyforInternetaccesswouldincreasesocialwelfareifandonlyifthebenefitofInternetaccessexceedsRSUcoste.Thispaperwilldeterminewhenthatisthecase,andwhichfactorsaremostinfluential.Inparticular,wefindthatdeploymentofthisinfrastructureforInternetaccessindenseurbanareasislikelytoincreasesocialwelfarefairlysoonafteramandatetoputOBUsinvehiclesbecomeseffective.OtherdecisionsincludewhethertoallocatespectrumandmandateOBUsinthefirstplace,ifthesestepsarenottakenforsafetyreasons.InsituationsinwhichbenefitofInternetaccessexceedsalltypesofDSRCcost,thensocialwelfareisincreasedbymandatingDSRCdevicesinallvehiclesandallocatingspectrumregardlessofwhethertherearesafetyorothertypesofbenefit.Thispaperwillalsodeterminewhenthisisthecase.

SomecarriersandresearchersareconsideringuseoffixedWi-Fihotspotsthatoffloadtrafficfromstationarydevicesthatareincloseproximity,orvehiculardatatrafficthatistoleranttodelays(AT&T2015;Comcast2013;Comcast2015;Balasubramanian,Mahajan,andVenkataramani2010;Eriksson,Balakrishnan,andMadden2008;K.Leeetal.2010;Balasubramanianetal.2008),andtherehasbeenresearchontheresultingeconomicimpact(Markendahl2011;J.Leeetal.2014).However,vehicularnetworksoffernewopportunitiesforInternetaccessthatarequitedifferentfromwhatispossiblewithWi-Fihotspots,andthisrequiresnewanalysis.

ThebenefitsofvehicularnetworksaredifferentfromWi-Fihotspotsbecausethetrafficcarriedisdifferent.Wi-FiisoftenagoodsolutionforuserswhoarestationaryfortheperiodwhentheyareaccessingtheInternet,butitisofteninadequateforuserswhoaccesstheInternetwhilemoving.OnereasonisthatWi-Fihotspotsrequire1-8secondsforauthentication(Bychkovskyetal.2006;Murray,Dixon,andKoziniec2007),whichmustoccurbeforeanewconnectioncanbeestablished.Thisisoflittlevaluetocarstravellingathighspeeds.Incontrast,DSRClinkscanbeestablishedinjust300milliseconds(IEEE2010a;MussabbirandYao2007).Thus,whilevehicularnetworkscouldservesomeofthesameusersasWi-Fi,e.g.apedestrianwhoisnearbothaWi-Fihotspotandabusystreet,vehicularnetworkscanbringInternetaccesstomanyusersinmovingvehiclesthatarenotservedwithWi-Fi.Inaddition,thecostsassociatedwithvehicularnetworksarequitedifferentfromthecostsoftypicalWi-Finetworks,whicharegenerallymicrocellular.AsshowninFigure1,vehicularnetworksbasedonDSRCaremeshnetworkscomprisedofvehicle-to-vehicle(V2V)andvehicle-to-infrastructure(V2I)wirelesslinks,whichmeansthatinformationcantravelalongmultihoppathsfromvehicletovehicletovehiclebeforefinallyreachingfixedinfrastructure.Inameshnetwork,thedevicesusedtoaccessthenetworkalsoserveasinfrastructurethatbringsaccesstoothers.Asaresult,arelativelysmallnumberoffixedDSRCroadsideunits(RSUs)canconnectalargenumberofvehicle-borneDSRCunitstotheInternet.ItalsohelpsthatDSRClinkscanbelongerthantypicalWi-Fihotspots,i.e.upto1000-meterdistancesiftherearenoobstructions,and250-350metersinclutteredurbanareasf.AlthoughfarfewerfixeddevicesareneededtocoveranareaswithavehicularnetworkthanwithWi-Fi,thosefixedDSRCdevicesarealsomore

eTheestimationofbenefitsandcostsassumesthattheroadsideinfrastructureinquestionisusedonlyforInternetaccess,andnotforsafetyandotheruses.ThereareotherpossibilitiestoincreasesocialwelfarewheninfrastructureusedforsafetyandInternetisshared,whichissubjectforfuturework.fAsmeasuredinthecityofPorto,Portugal.

Page 4: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

4

expensive,inpartbecausetheymustoperateoutdoorsinhostileconditions,andinpartbecausetheyarenotcurrentlymassproduced.

Figure1.DSRC-basedcommunications.V2IandV2Vlinkscanbeusedtodisseminatesafetymessages,andalsoforwardIPpacketsbetweenvehiclesandroadsideinfrastructureforInternetapplications.MultipleV2Vhopsthroughintermediatevehiclescanbeusedfortwoendpointstoconnect

OurmethodologyconsistsofcombiningageneralsimulationmodelthatwedevelopedwithdatacollectedfromanactualvehicularnetworkthatisoperatinginPorto,Portugal.ThefirststepistoestimatehowmuchvehicularInternettrafficthatwouldotherwisebecarriedbycellularproviderscaninsteadbecarriedbyaDSRC-basednetworkunderavarietyofconditions.Toachievethis,wedevelopedsoftwarethatsimulatestherateatwhichdataistransferredbetweenvehiclesandRSUs.Oursimulationemploysrealisticrepresentationsoftheelementsofavehicularnetworkthatgreatlyaffectthroughputrates,includingthelocationofvehiclesandRSUs,thesignallossbetweendevices,andtheDSRCprotocolitself.Someofthatrealismcomesfrommeasurementdatatakenfromthecity-scaletrialinPortugal.Forexample,ourmodelsofvehicletrafficpatternsarebasedinpartonlocationdatacollectedfrom900busesandtaxisbetween2012and2015.

ThenextstepistoestimatecostsandbenefitsofInternetaccessthroughvehicularnetworksundergivenconditions.Today,nearlyalltrafficfrommobiledevicesmustbecarriedoveramacrocelltothenearestcellulartower(asdiscussedabove,lesscostlyalternativessuchasWi-Fihotpotssometimesexistforstationarydevices,butusuallynotfordevicesthataremoving).Inacapacity-limitedcellularnetwork,areductionoftrafficfrommobiledevicesthatmustbecarriedinthebusyhourallowseachcelltowertoprovideadequatecapacityoveralargerarea,therebyreducingthenumberofcostlytowersthatacellularoperatorneedstocoveragivenregion.WedefinethebenefitofInternetaccessthroughvehicularnetworksinagivenscenarioasthecostsavingsfromreducingthenumberofcelltowers.ThisiscomparedtothecostsofDSRCRSUs,spectrumorOBUs.Inthisanalysis,weconsiderawiderangeofvaluesforimportantfactorssuchaspopulationdensity,DSRCpenetration,datarateperDSRC-equippedvehicle,andvariousunitcosts.

Thispaperisorganizedasfollows.Section2describestheDSRCnetworkoperatinginPortoforInternetaccess,andwhichdataisbeingusedfromitforthispaper.Section3explainsthesimulationmodel,thebenefit-costanalysisandtheirunderlyingassumptions.Section4containstheresultsthatarerelevanttoanswertheresearchquestionsproposed.Section5endsthepaperwiththeconclusions,aswellasthelimitationsandopportunitiesforfuturework.

V2I$>">">">">">">"

V2V$

)")")")")")")")")")")"

V2V$)")")")")")")")")"

RSU$

OBU*equipped$vehicles$

Internet$

Page 5: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

5

2 PortoVehicularNetworkforInternetAccessandDatasetPortoisthesecondlargestcityinPortugal,witha2011populationof237,000inanareaof41.4km2(InstitutoNacionaldeEstatistica2011).InSeptember2014,theurbanbusauthoritycompanyofPortostartedofferingfreeWi-Fiserviceforitspassengers24hoursaday,7daysaweek.Eachofits477urbanbuseshasanOBUequippedwithaWi-Fihotspotforpassengerstoconnectto.AllbushotspotsshareasingleWi-FinetworkIDwithopenaccess:uponactivatingaWi-Fienabledsmartphoneorcomputerinthebus,apassengercanuseittofindthecommonWi-Finetworkname,andafteropeningawelcomewebpage,he/shehasInternetaccesswithnopasswordrequired.

EachbusOBUhasitshotspotcoupledtoarouterthatrelaysallpassengertrafficto/fromtheInternet.Eachpacketisrelayedthroughoneoftwopossiblepaths.ThepreferredisthroughtheuseofDSRC,forwhichtherewere27RSUsgdeployedatfixedlocationsofthecitysuchastrafficlightsh:busescanconnecttoRSUseitherdirectlyorthroughmultihopconnectionsusingotherbuses.IfnoRSUiswithinrangeofasingleormultihopconnection,thentheOBUtransferdatathroughthecommercialLTEnetwork.BothOBUsandRSUsaredeployedandmaintainedbyVeniamNetworks.

AsofMarch2015,over2.7TBweretransferredbyPortobuspassengersthroughDSRCandcellular.Overthefirstquarterof2015acompoundmonthlygrowthrateof35%wasobserved.TheobservedvolumetransferredthoughDSRCvarieswithlocation,withthemajorityoftheRSUsbeingconcentratedindowntown,wheretheratiobetweenthenumberofbytestransferredthroughDSRCandthetotalnumberofbytescanreachasmuchas70%atpeakhours.

Metadataregardingthebusnetworkstateandusageiscollectedandstoredperiodically.ThatincludesbytestransferredthroughDSRCandcellular,signalstrengthbetweentheendpointsofthewirelesslinks,andGPSpositionsofthebuses.

Moreover,thereisdataabouttaximobilityinPorto.Ofthecityestimatedtotalof800taxis,GPSpositionsof435vehicleswerecollectedduringonemonthin2012andsharedforuseinthispaper.

ThedatafromthePortobusnetworkandtaxipositionrecordsthatwereusedinthispaperissummarizedinTable1.AsdescribedinSection3,realbusandtaxipositionsareusedtosimulateaDSRCnetworkwithvaryingconditionsofpopulation,OBUpenetration,andloadofInternetdata.RealmeasurementsofDSRCsignalswerecomparedwiththephysicalmodelofthesimulation.

gAsofMarch2015.hTherearemoreDSRCRSUsdeployedinthemetropolitanregionofPorto,namelyintheharborarea.Nevertheless,theseareforcommunicationswithcargotrucksandhavenotbeenusedforoffloadingofInternettraffic.See(Ameixieiraetal.2014):thesametypeofRSUandOBUequipmentareusedinboththetruckandbusnetworks.

Page 6: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

6

Table1.Datausedfortheanalysis

DatafrombusDSRCnetworkcollectedfromOctober2014toMarch2015

DataItem NumberofObservations Description

Wi-Fisessions 477buses:106sessions PerWi-Fisession:numberofbytestransferred,startandendtimes,andanidentifierofthebus

Datavolume/position/signalper15-secondperbus

477buses:240*106datapoints

Per15-secondinterval,perbus:bytestransferredoverDSRC,bytestransferredovercellular,dateandtime,GPSposition,receivedsignalstrengthfromRSU(ifV2I-connected)orpeerbus(ifV2V-connected),identifierofthebus,identifierofthepeerifV2IorV2V-connected

RSUpositions 27RSUs PerRSU:GPSpositionandheight

DatafromtaxiscollectedinMarch2012

DataItem Value Description

Positionpersecondpertaxi

435taxis:120*106datapoints

Persecond,pertaxi:time,GPSposition,andanidentifierofthevehicle

3 MethodologyTheanalysisinthispaperevaluatesthevolume,benefits,andcostsofInternetaccessthoughvehicularnetworks,underseveralscenariosrepresentingdistinctvaluesoftheparametersthatmostaffectresults.Theanalysisisperformedintwomainsteps,asillustratedinFigure2.

Figure2.Summaryofsteps,inputsandoutputsofthemethodology

ThefirststepinthemethodologyistoestimatepotentialofInternetaccessthroughavehicularnetwork.ToachievethiswedevelopedawirelessnetworksimulationmodelwhichsimulatestherateatwhichdataistransferredbetweenvehiclesandRSUsthroughsingleormultiplehops.Thatmodelmakesuseofrepresentationsoftheelementsthatmostinfluencethethroughputrates:locationsofRSUsandvehicles,signallossbetweenthem,andmultiplevehiclesandRSUsexchangingdatasimultaneouslyinthesamearea,atwhichcompetitionfortheuseofthewirelessmediumishandedbytheDSRCandInternetprotocols.Portodataisusedinthreeways:first,busandtaxiGPSpositionsfromPortoare

Page 7: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

7

usedtodeterminethepositionsofthevehiclesinthesimulation.Second,thereceivedsignalmeasuredinthebusesisusedtoverifywhetherthesimulatedsignalloss(whichinfluencestransmissionrangesandinterference)iscompatiblewithmeasuredloss,onaverage.Third,measureddataratesthroughDSRCfromPortoarecomparedwiththesimulatedrates,inordertoverifywhetherthelatterisareasonableapproximationofrealdataratesthroughDSRCundersimilarconditionsofnumberofvehiclesandRSUs,anddatarates.

OnepossiblequestionabouttheaboveiswhydataratesthroughDSRCaresimulated,whenmeasuredratesfromarealvehicularnetworkareavailable.TheansweristhatweareinterestedinlearningthepotentialofInternetaccessthroughvehicularnetworkswithcharacteristicsthatdifferfromthePortonetwork,withrespecttothequantityorconcentrationofvehiclesandRSUs,typesofvehicles,volumeofInternettrafficdemanded,availablebandwidthforInternetaccess,etc.WiththemodelwesimulateconditionsthatrepresentcitiesotherthanPorto,andfutureperiodswithhigherpenetrationofDSRCdevicesorhigherInternettraffic.

ThesecondstepistousethedataratethroughDSRCtoestimatethebenefitandcostofInternetaccess.ToaccomplishthisweconsiderthebenefitasthesavingsaccruedfromthedifferencebetweenthenumberofmacrocellulartowersthatwouldbenecessaryifthereisnoInternetaccessthroughDSRC,andthe(lower)numberoftowersnecessarywhenpartofthetotaltrafficisoffloaded.Costsofvehicularnetworksareofthreetypes:DSRCOBUs,spectrum,andRSUs.Whilethequantityofonboarddevicesandamountofspectrumareamongthedefinitionsthatcharacterizeascenarioofanalysis,theamountofinfrastructuredeployedforeachscenarioisestimatedattheoptimalquantityofRSUsthatmaximizesthedifferencebetweenbenefitofInternetaccessandinfrastructurecost.

Locationcharacteristics,i.e.whetheragivenareaisurban,suburbanorrural,influencebothsteps.Dataratesareinfluencedbysignalpropagationcharacteristics,whichdifferbetweenurbanandruralareas.Moreover,thosedataratesareonlyrelevantwherethecellularnetworksarecapacity-limited,whichalsoisaconditiontypicalforurbanareas.Ontheotherhand,thosesamedataratesareexpectedtobehigherinurbanareas,becauseofthehigherpopulationdensities.Therefore,resultsarelikelytobemoresubstantialinurbanareas,whichmakethemtheprimaryfocusofthisanalysis.

EachstepisdescribedinSections3.1and3.2,respectively,andthenumericalvalueschosenforthebasecasescenarioanditsvariationsaredescribedinSection3.3.

3.1 NetworkSimulationThesimulationmodelrepresentsawirelessnetworkofDSRCRSUsconnectedtotheInternet,andvehiclesequippedwithDSRCOBUsthatexchangeInternettrafficwiththoseRSUs.TransfersofdatapacketsaresimulatedbetweenpairsofvehiclesandbetweenvehiclesandRSUs.Atanygiventime,packetstreamsflowbetweeneachconnectedvehicleandoneRSUwhichservesasagatewaytotheInternet,eitherdirectlyorthroughmultiplehopswithothervehiclesactingasrelays.Thesedatatransfersaresimulatedatthetransport,network,linkandphysicallayersusingthens-3networksimulator(“Ns-3NetworkSimulator”2015).Aparticularsetofparametersusedinarunofthesimulationisreferredinthispaperasasimulationscenario.

Theunderlyingassumptionsofthenetworksimulationmodelaredescribedbelowinthefollowingorder.First,mobilityandnetworktopology,theuseofDSRCspectrum,the

Page 8: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

8

estimationofthroughputrates,andendpointsfortrafficflowsaredescribed.Then,thedescriptionisseparatedbycommunicationlayer,beginningwiththetransportlayerandthenproceedsonebyoneuntilthephysicallayer.

VehiclemobilityandRSUlocations.Resultsobviouslydependonthelocationsofvehiclesineverytimeinterval.ArealisticmodelofvehiclepositionsisderivedfromthelogsofvehicleGPSreadingsfromPorto.GPSreadingsarecollectedeverysecondfortaxis,soeveryfifthreadingmarksthebeginningofatimeinterval.GPSreadingsforbusesarecollectedevery15seconds,sowegetpositionsinterpolatedevery5seconds.ThepositionsofvehiclesotherthanbusesarealsoderivedfromtheGPSlogsoftaxis.Especiallyinurbanareas,mobilitypatternsofprivatevehiclesarelikelytobesimilartothoseoftaxis,althoughperhapsnotidentical.Vehiclemobilityissimulatedasaseriesof“snapshot”positionsin5-secondintervals,meaningthatrepresentationsofvehiclesarecreatedinthesimulationwithstaticpositions.Then,communicationsbetweenvehiclesandRSUsaresimulated,andthethroughputratesareestimated,representingawirelessnetworkwithnon-movingnodescommunicatingfor5seconds.Afterthesimulationruncompletesandthroughputratesarecalculatedforonetimeinterval,theprocessrepeatsforthenext5-secondinterval:thepositionsofthevehiclesarechangedtorepresentthenetworktopologyforthenext5seconds,thecommunicationssimulationandthroughputrateestimationisperformedagainforthereferredinterval,andsoon.

ResultsalsodependonthelocationsofRSUs.ThesimulationacceptsRSUdensityasaninputvariable,andthenplacesRSUswheretheyarelikelytodothemostgood.Thus,RSUsshouldbesetinplaceswithalargenumberofvehiclesatpeakhours.Morespecifically,agivennumberkofRSUsareplacedusingthek-meansclusteringheuristic(Moore2001),withpeak-hourvehiclepositionsastheinput.Thealgorithmisapopularapproachtodivideanumberofobservations(vehiclelocations,inourcase)intokregions,andfindtheoptimalcentroidforeachregion,withrespecttominimizingthedistancebetweeneachobservationandthecentroid.ThepositionsfoundforthecentroidarethenusedtoplacetherepresentationsoftheRSUsinthesimulationbeforeitisrun.Foreachsimulationscenario,thenetworkissimulatedmultipletimeswithinfrastructuredensityrangingfrom0to10RSUs/km2.

VehicleandRSUantennasareplacedinatri-dimensionalspace.XandYcoordinatesrepresentlongitudeandlatitude,respectively,andaregivenbytheGPSdata.Zcoordinatesrepresenttheheightofantennas.AllRSUantennashaveaheightof7meters,whichistheaverageheightofPortoRSUsasinMarch2015.Busantennashaveaheightof3meters(averageofsingledeckbusesinPorto),andallothervehicleshaveheightof1.5meters(whichisconsistentwithpreviousworkin(Bobanetal.2011)).

UseofDSRCspectrumforInternetaccess.75MHzofspectrumallocatedforDSRCisusedinseven10MHzchannels,ofwhichoneisreservedforcontrolandmanagementofallchannels,andtwoothersarereservedforsafetyapplications(IEEE2010b).WeassumethefourremainingchannelsareavailableforInternetaccess,andeachvehicleOBUandeachRSUisequippedwithfourradios.

Itisassumedthateachpacketstreamsflowusesonechanneli.Thechanneltobeusedateachhopoftheflowischosenastheleastusedchannelintheareasimulated.

iMorechannelsmaybeallocatedifthedatatobesentinaflowexceedsthecapacityofthechannel,butthisisnotthecasefortheresultspresentedinthispaper.

Page 9: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

9

Estimationofthroughputofthevehicularnetwork.ThethroughputrateviaDSRCforeachvehicleisdefinedasthedatathroughputachievablewhenthatvehiclereceivesdatafromaRSUitisconnectedto(eitherthroughasingleormultiplehops).WeassumethatthetrafficsentdownstreamtoanygivencarequalsthesumofthethroughputovertheDSRCnetworktothatcarandthethroughputoverthecellularnetworktothatcar.Thesameisassumedfortraffictravelingupstreamfromeachcar.Theseassumptionsareaccurateiftheamountoftrafficthatislostandtheamountoftrafficthatisunnecessarilysentonbothnetworksarebothnegligible.Thisisreasonableaslongasthecellularnetworkisalwaysavailableandhasenoughcapacitytocarryalltrafficthatcannotbecarriedoverthevehicularnetwork.

Steady-statethroughputthroughDSRCareestimatedforeach5-secondintervalbasedonthepositionsofallvehiclesatthebeginningoftheinterval.Thissimplifyingassumptionignoresthefactthatvehiclesmovecontinuouslyduringtheinterval,sothroughputwouldactuallychangegraduallyratherthanjumpevery5seconds.Thisformofanalysismaymisssomeofthefluctuationsindatarateasobservedbyamovingvehicle,butitallowsforagoodapproximationofthroughputwhenaveragedovermanytimeintervalsaslongasvehiclescanswitchoffbetweenthevehicularnetworkandaubiquitouscellularnetworkasneededsothatdataratefluctuationshavelittleeffectonthetotalamountoftrafficsentandreceived.Thisisareasonablefirst-orderestimateifthetimetoestablishV2VandV2Ihopsisnegligible,andthisswitchingtimewithDSRCisexpectedtoberoughly300milliseconds(IEEE2010b;MussabbirandYao2007).Toestimatesteady-statethroughputinagiventimeinterval,thesimulationisfirstrunforanextendedwarm-upperiodbeforestatisticsaregathered.Thewarm-uppartofthesimulationrunsfor8seconds,andafterthatstatisticsarecalculatedforthedatareceivedinonesecond.This8-secondwarm-upperiodwasobtainedbyexperimentation–allscenariossimulatedresultedinthroughputclosetothemeanafterthatperiod,andmostdosolessthan1secondafterthebeginningoftheinterval.

EachDSRC-equippedvehicleistheendpointofoneandonlyonebidirectionalflow,whileeachRSUmaybetheendpointforzero,oneormoreflows,uptothenumberofvehicles.However,anyvehiclecanalsoserveasarelayfordataofaflowthathasanothervehicleasadestination,incaseofmultihopcommunications.Protocol-specificdataincludeacknowledgmentsandretransmissionsinalllayers.However,thoseprotocolmessagesarenotconsideredinthestatistics–onlythenumberofapplication-layerdatabytesreceivedandsentbythevehicleperunitoftimeisconsideredinthethroughput.

Endpointsfortraffic.EachRSUisagatewaytotheInternetwhichagivenvehicleconnectsto.Weonlymodelthetrafficonthevehicularnetwork,i.e.betweenvehiclesandRSUs,sowetreattheRSUasifitweretheendpointofatransport-layerconnectionratherthanmerelyagateway.

Transportlayer.Ateachinterval,aTransmissionControlProtocol(TCP)connectionissimulatedbetweeneachvehicleandRSUitconnectsto.TCPisusedbecauseitisthemostcommontransportprotocolusedintheInternet(Raoetal.2011;Zambelli2009).

TheTCPMaximumSegmentSize(MSS)usedis2244bytes,whichisthemaximumsizeofthepacketthattheIEEE802.11linklayersupportswithoutfragmentation(2304bytes),minus60bytesforthelinkandIPheaders(WangandHassan2008).ThatMSSisroughlysimilartotypicalvaluesTCPconnectionstraversing802.11networks.

Networklayer.IPpacketsareroutedbetweentwoendpointsthroughthepathwiththesmallestnumberofhops.IfthereareneighboringRSUs,theonewiththehighestreceived

Page 10: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

10

signalischosenastheendpoint.Otherwise,thevehiclesearchesitsneighborsinrandomsequence,fora2-hoppathtoanyRSU.Thesearchstopsforthefirstpathfound.Ifnoneisfound,thenthesearchcontinuesfora3-hoppath.Ifnonepathisfound,thenitisassumedthevehicleisunreachablebyanyRSU.Thismethodisasimplificationbecauseroutingalgorithmsinvehicularnetworksareanongoingresearchtopic(Meireles2015;LiandWang2007;Meireles,Steenkiste,andBarros2012;Wisitpongphanetal.2007).

Linklayer.Themediaaccesscontrol(MAC)sublayerintheDSRClinklayeristheonespecifiedintheIEEE802.11pamendment(IEEE2010a)oftheIEEE802.11standards.AllpacketstransferredinallhopshavethesameprioritywithrespecttotheIEEE802.11puserprioritylevels.

Physicallayer.Thereceiversensitivitythresholdis-94dBm.Ahopisusedbetweentwonodesonlyifsignalstrengthatthereceiverexceeds15dBabovethesensitivitythreshold.ThisisthecriteriadeterminedempiricallyinthebusnetworkofPortoastheminimumqualityforthepairsofnodestotransferdata.Whenthehopisused,packetsarereceivedatanerrorratethatalsodependsonthesignal-to-interference-plus-noiseratio(SINR),asdescribedin(LacageandHenderson2006)and(“Ns-3NetworkSimulator”2015).

Thetransmittedpoweris14.6dBm,obtainedfrommeasurementsattheequipmentoutput,whichisalsoconsistentwith(Cardoteetal.2012)and(Bai,Stancil,andKrishnan2010),andthegainsofthetransmissionantennasare16dBiand5dBifortheRSUsandvehicles,respectively,whichareconsistentwiththesettingsoftheequipmentinthePortobusnetwork.

Thereceivedsignaliscalculatedaccordingtothepropagationlossmodelfrom(Meiniläetal.2009)(urbanmicrocellB1variant).ItwasthepreferredmodelbecauseitisvalidfortheDSRCband(5.9GHz),anditexplicitlymodelstwoothercharacteristicsthatarerelevantinvehicularnetworks:whetherthosenodesareinline-of-sight(LOS)ornon-LOS(NLOS)(Meiniläetal.2009;Zhaoetal.2006),andtheantennaheightsofvehiclesandRSUs(Mecklenbraukeretal.2011;Meiniläetal.2009).ForLOS,thelossLisgiven(indB)as

𝐿 = 𝑃𝐿$%& + 𝑁~(0, 𝜎)

where𝑃𝐿$%& = 𝐿/ + 10𝑛𝑙𝑜𝑔6/(𝑑)isthepathlosscalculatedasareferenceloss𝐿/andafunctionofthedistanced(meters)andthepathlossexponentnrepresentingthedegreeofattenuation.NisaGaussianrandomvariablewithzeromeanandrepresentslarge-scalefadingeffectssuchasshadowingoftheLOSpathbyobstacles.ForLOSthevaluesare

𝑛 =2.27𝑓𝑜𝑟𝑑 < 𝑑>?4𝑓𝑜𝑟𝑑 ≥ 𝑑>?

𝐿/ =41 + 20𝑙𝑜𝑔6/

𝑓5 ∗ 10D

𝑓𝑜𝑟𝑑 < 𝑑>?

9.45 − 17.3𝑙𝑜𝑔6/ ℎ6 − 1 − 17.3𝑙𝑜𝑔6/ ℎI − 1 + 2.7𝑙𝑜𝑔6/𝑓

5 ∗ 10D𝑓𝑜𝑟𝑑 ≥ 𝑑>?

𝜎 = 3

wherefistheDSRCfrequencyinHz,h1andh2aretheheightsofthevehiclesand/orRSU,and𝑑>? = 4 ℎ6 − 1 ℎI − 1 𝑓/𝑐 (cisthespeedoflightinm/s).

ForNLOS,

𝐿 = 𝑃𝐿$%& + 20 − 12.5𝑛 + 10𝑛𝑙𝑜𝑔6/ 𝑑 + 3𝑙𝑜𝑔6/𝑓

5 ∗ 10D+ 𝑁~(0, 𝜎)

Page 11: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

11

𝑛 = 𝑚𝑎𝑥(2.8 − 0.0024𝑑, 1.84)

𝜎 = 4

Othermodels(Zhaoetal.2006)providesimilarpathlossandshadowingparametersnotsubstantiallydifferentfromthoseshownabove.

EachintervaleachlinkisassumedtobeinLOSorNLOSaccordingwithprobabilityProb(LOS)estimatedas(Calcevetal.2007)

𝑃𝑟𝑜𝑏(𝐿𝑂𝑆) =𝑑 − 300300

𝑓𝑜𝑟𝑑 < 300

0𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

(Asplundetal.2006)and(Meiniläetal.2009)proposeexpressionswhichresultssimilarLOSprobability.

Inadditiontopathlossandshadowing,somemodelsincludezero-meanrandomvariablestorepresentfast-fadingeffectssuchasmultipathpropagationandDopplerspread(Mecklenbraukeretal.2011).Inoursimulationmodel,theestimatedpathlossandshadowingcomponentsareassumedtobeconstantovereach5-secondinterval,andtheeffectoffast-fadingisassumedasnegligible,asweestimateaveragelossesacrossmanylinksratherthanpredictthelossofaparticularlink.

ThedifferencebetweenthemediansimulatedlossandthemedianlossmeasuredinPortobusesislessthan5dBformostdistancesshorterthan200meters,whichshowstheassumedmodelisareasonableapproximationfortheobservedloss.Forexample,atadistanceof100mbetweenaRSUandabus,themedianmeasuredlossis92dBwhilethesimulatedlossis95dB.Morethan95%ofthehopsobservedinthePortonetworkareshorterthan200meters.

3.2 Benefit-CostAnalysisThesecondstepofthemethodologyistouseDSRCthroughputtoestimatebenefitsandcostsofInternetaccessatpeak-hours.Ourdefinitionsofcostsandbenefitareindependentofwhoincursthosecostsandwhoderivesthosebenefits.Thisallowsustoquantifytheimpactofdeployinganewkindofinfrastructureontotalsocialwelfarewithoutmakinganyassumptionsaboutthingslikewhopaysforbuildingandoperatingtheroadsideinfrastructure,whethertheoperatorofroadsideinfrastructurechargesfortheservice,whopaysfortheservice,orhowmuch.Goodanswerstothequestionscanbefoundifandonlyifanewsystemwouldincreaseoverallsocialwelfare.

WedefinethebenefitofInternetaccessthroughvehicularnetworksasthenetpresentvalueofcostsavings,whichwederiveunderthefollowingassumptions.Allmacrocellularcarriersintheregionbeinganalyzedareassumedtobecapacity-limitedinsteadofcoverage-limited.Inacoverage-limitedsystem,acarrierdeploystheminimumnumberoftowerstomeetcoveragerequirements,andtherewillstillbemorecapacitythanneededeveninthepeakhour.InternetaccessthroughDSRCisnotvaluableinaregionthatalreadyhasexcessunusedcapacity.Incontrast,inacapacity-limitedsystem,acarrierdeploysenoughtowerstomeetcapacityrequirements,whichmeansthesystemisexpectedtooperateatfullcapacityduringpeakhours.Therefore,Internetusageinvehiclesasanewsourceofmobiletrafficshouldbemeteitherviacapacityexpansionofthemacrocellularnetworks,orviaoffload.Toservemoreusersorhigherrateperuser,acapacity-constrainedcarrierthatisalreadyusingcurrenttechnologythroughoutthespectrumavailabletoit

Page 12: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

12

mustdeploynewtowers,resultinginasmallerareapercelltodelivermorecapacityperarea.

Besidesdeploymentofnewtowers,therearetwootherwaystoincreasemacrocellularcapacitythatmaypreservetheexistingmacrocellulartopology.Oneistheacquisitionofmorespectrum,whichincreasesthecapacitypertower.Theotherwayischangingtheefficiencyofthetechnologyemployedpertower,suchasmigratingfrom3Gto4Gequipment,oraddingequipmenttoincreasethenumberofsectorspertower.Sincenetworkdesignerswillgenerallychoosetheapproachforexpandingcapacitythatismostcost-effectiveatthetime,themarginalcostofincreasingcapacityislikelytobesimilarforallavailableapproaches(TanandPeha,2015).Weassumeforthisanalysisthatthedeploymentofnewtowersisthepreferredmethodtoincreasemacrocellularcapacity.Carriersdodeploytowerswhentheyneedcapacity,inpartbecausespectrumisdifficultandcostlytoobtain,andcarriersthatneedmorecapacityinaregionareoftenalreadyusingcurrenttechnologythere(Clarke2014)andhaveoftendeployedthemaximumnumberofsectorsallowedbythattechnology.

ItisassumedthatineveryintervaldeviceswillsendasmuchtrafficaspossibleovertheDSRCnetwork.TheamountoftrafficcarriedthroughDSRCequalsthereductionintheamountoftrafficcarriedthroughcellular,meaningthatdevicesswitchbetweentheDSRCandmacrocellularnetworkwithnegligibledisruption,withnodatabeinglostortransmittedinduplicitythroughbothnetworks.

WedefinethebenefitofInternetaccessthroughvehicularnetworksinagivenscenarioasthenetpresentvalueofcostsavingsfromreducingthenumberofcelltowersthatwouldotherwisebeneededtocarrythatpeak-hourtrafficifitwasnotcarriedthroughDSRC,assumingcellularcarriersarelimitedbycapacity.

TheNPVofthebenefitofInternetaccessperkm2is

𝑁𝑃𝑉𝐵 = 𝜌[\]^_`ab^c[ ∗ 𝐶`ab^c

where𝜌[\]^_`ab^c[isthetotalnumberofmacrocelltowers“saved”perunitofareaduetoInternetaccessthroughvehicularnetworksand𝐶`ab^c istheaverageNPVpermacrocelltower.

Whencalculatingtherelationshipbetweencostandcapacity,weassumethatifthereissufficientcapacitydownstreamthenthereisalsosufficientcapacityupstream,andthatcarriersareusingFrequencyDivisionDuplexing(FDD)sospectrumcanbelabeledaseitherupstreamordownstream.Thisisreasonablebecausedownstreamtrafficrateshavebeengrowingfasterthanupstreamrates(Sandvine2014),andmosttier-1carrierscurrentlyuseFDD(Engebretson2012).Inacellularnetwork,themaximumdownstreamcapacity𝑏𝑝𝑠_𝑚𝑎𝑥ghiinbitspersecondperunitofareaisgivenby

𝑏𝑝𝑠_𝑚𝑎𝑥ghi = 𝑠[^j`ac ∗𝑏𝑤𝐹𝑅

∗ 𝜌`ab^c[ ∗ 𝑁[^j`ac[

where𝑠[^j`ac istheaveragedownstreamspectralefficiencyinbitspersecondperhertzpersector,bwisthetotalbandwidthpermacrocellularcarrierusedfordownstreamtransmission,𝐹𝑅 ≥ 1isthefrequencyreusefactor,𝜌`ab^c[isthenumberoftowersperkm2

and𝑁[^j`ac[isthenumberofsectorspermacrocellulartower.

Inordertoserveallfluctuationsofdemand,themaximumcapacityshouldequalorexceedthedataratedemandatpeakhours.Therefore,if𝑏𝑝𝑠1ghi isthepeak-hour,downstream

Page 13: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

13

dataratedemandperunitofareafrommacrocellswhennoInternetaccessthroughDSRCtakesplace,thenumberoftowersnecessary𝜌1perunitofareais

𝑏𝑝𝑠1ghi = 𝑠[^j`ac ∗𝑏𝑤𝐹𝑅

∗ 𝜌1 ∗ 𝑁[^j`ac[

Let𝑏𝑝𝑠2ghi alsobethedataratedemandfrommacrocells,butwhenpartofthetrafficiscarriedthroughDSRC,tobeservedby𝜌2towers.Thedifferencebetween𝑏𝑝𝑠1ghi and𝑏𝑝𝑠2ghi isthetrafficoffloadedperunitofarea:

𝑏𝑝𝑠1ghi − 𝑏𝑝𝑠2ghi = 𝑏𝑝𝑠𝑂𝑓𝑓 = 𝑠[^j`ac ∗𝑏𝑤𝐹𝑅

∗ 𝜌[\]^_`ab^c[ ∗ 𝑁[^j`ac[

thenthenumberofsavedmacrocelltowersis

𝜌[\]^_`ab^c[ =𝑏𝑝𝑠𝑂𝑓𝑓 ∗ 𝐹𝑅

𝑠[^j`ac ∗ 𝑏𝑤 ∗ 𝑁[^j`ac[

wherebpsOffisthedownstreamDSRCthroughput.

ThetotalcostofInternetaccessthroughDSRCperkm2𝑁𝑃𝑉𝐶 consistsofthreetypesofcosts:

𝑁𝑃𝑉𝐶 = 𝑁𝑃𝑉𝐶m&n + 𝑁𝑃𝑉𝐶%>n + 𝑁𝑃𝑉𝐶&o^j`cph

where𝑁𝑃𝑉𝐶m&n, 𝑁𝑃𝑉𝐶%>n and𝑁𝑃𝑉𝐶&o^j`cpharetheNPVperkm2ofthecostsofRSUs,OBUsandDSRCspectrum,respectively,andaregivenas

𝑁𝑃𝑉𝐶m&n = 𝜌m&n ∗ 𝐶m&n

𝑁𝑃𝑉𝐶%>n = 𝜌%>n ∗ 𝐶%>n

𝑁𝑃𝑉𝐶&o^j`cph = 𝜌&o^j`cph ∗ 𝐶&o^j`cph

where𝜌m&n isthenumberofRSUsforInternetaccessdeployedperunitofarea,whichisassumedtobeindependentandnotsharedwithRSUsdeployedforsafetyorpurposesotherthanInternetaccess,𝜌%>n isthenumberofOBUsdeployedperunitofarea,𝜌&o^j`cphistheamountofDSRCspectruminMHztimesthepopulationdensity,and𝐶m&n, 𝐶%>n, 𝐶&o^j`cpharetheNPVperRSU,OBU,andMHzofspectrumperperson(alsoknownasthecostperMHz-pop),respectively.

Thecomparisonbetweenthebenefitandcostsdefinedabovedependsonthedecisiontobemade,i.e.someofthecoststhatarerelevantforonedecisionmaybeirrelevantforanother.Forexample,inthecontextofasafetymandate,DSRCspectrumisallocatedandOBUsarepurchasedforsafetyreasons.Inthiscase,spectrumandOBUcostsaresunkwithrespecttoInternetaccess,andadecisiontodeployRSUinfrastructureincreasessocialwelfareifandonlyifbenefitofInternetaccessthroughDSRCexceedsRSUcosts.

However,ifthereisnosafetybenefitderivedfromthemandate,thenspectrumandOBUcostsarenotsunk,andsocialwelfarewillincreaseonlyifbenefitofInternetaccessexceedsallDSRCcosts:RSUs,OBUsandspectrum.

Inthisanalysis,weassumeparametersthataffecttheNPVofcostandbenefitarestatic,andwillusenumericalvaluesthatarereasonablefordecision-makersthatarelookingseveralyearsintothefuture.Inreality,someoftheseparametersarechangingovertime,althoughinwaysthataresometimeshardtopredictmorethanafewyearsintothefuture.Ingeneral,weexpectthatDSRCthroughputwillincreaseovertime,becauseboththenumberofDSRC-equippedvehiclesontheroadandthedatarateperDSRC-equippedvehicleare

Page 14: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

14

likelytoincreaseovertime.Asthroughputincreases,thebenefitsandthecostsofInternetaccessthroughDSRCbothincrease,e.g.moreRSUsaredeployedtocarrymoretraffic.AslongasnetbenefittendstoincreasewithDSRCthroughput,andthroughputincreasesovertime,thenaresultthatbenefitsexceedcostsinagivenscenariogenerallymeansthatdeploymentinthecurrentplanninghorizonwillincreasesocialwelfare.Ontheotherhand,ifnetbenefitincreaseswithDSRCthroughputandcostsarefoundtoexceedbenefitsinthecurrentplanninghorizon,thatdoesnottelluswhetherornotcostswillstillexceedbenefitsinthefuture.

3.3 BaseCaseScenarioThisSectiondescribesthebasecasenumericalvaluesfortheassumptionsusedintheestimatesofthethroughputofInternetaccessviaDSRC,benefitandDSRCcosts.Someofthosevaluesaresubjecttouncertainty,changewithlocation(suchasthepopulationdensity),and/orareexpectedtoevolveintime(suchaspenetrationandtrafficpervehicle).Therefore,wewillalsoconsiderscenarioswhereoneassumptionvalueisvariedatatimefromitsbasecasevalueinordertoobservetheeffectofeachassumptionontheresults.

Themonetaryvaluesthatfollowareinconstant2014dollarsj.BenefitandcostNPVsarecalculatedatarealdiscountrateof7%overahorizonof10years.ThediscountrateisconsistentwiththeraterecommendedbytheU.S.OfficeofManagementandBudgetforbenefit-costanalysisoffederalprograms(OfficeofManagementandBudget1992).Otheranalysesusesimilarrates(HallahanandPeha2009;Markendahl2011;MarkendahlandMäkitalo2010;Hardingetal.2014).The10-yearhorizonislongenoughtoevaluatethelifetimecostsofthemainelementsofthemodel.Forexample,RSUlifetimewasestimatedtobe10yearsinanalysisfortheU.S.Dept.ofTransportation(Wrightetal.2014).AnOBUlifetimeof10yearsisconsistentwithestimatesfromtheU.S.DOTfortheaveragelifetimeofcarsintheU.S.(Santosetal.2011).Althoughsomecostssuchasmacrocellulartowersareincurredforalongerhorizon,becauseofthe7%discountrate,theirNPVisprimarilydeterminedinthefirst10years.

Thebasecasevaluesare:

Populationdensity.Wemakethesimplifyingassumptionthatpopulationdensityisconstantthroughouttheregionbeinganalyzed.Forthebasecasethepopulationdensityischosenas5000people/km2,whichisrepresentativeofPorto(5,600)(InstitutoNacionaldeEstatistica2011)whereourmeasurementsweretaken,aswellaslargecitiessuchasBoston(5000people/km2),Chicago(4,600),Miami(4,300)(UnitedStatesCensusBureau2015),London(5,000)(UKOfficeforNationalStatistics2012),andTokyo(5,900)(TokyoMetropolitanGovernment2014).Populationdensitycanbemuchgreater,e.g.Paris(21000people/km2)(INSEE2013)oritcanbenegligible.

Numberofvehiclesontheroadatpeakhourspercapita.AssumedasinTable2,whichiscalculatedastheproductofvehiclesownedpercapita,fractionoftimevehicleisinuse,andratioofpeak-hourusagetoaverageusage:

jWhenvaluesarebasedonoldersources,theyareadjustedusingtheU.S.ConsumerPriceIndex.

Page 15: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

15

(i) ThenumberofvehiclesownedpercapitaintheU.S.varieswithpopulationdensityandwasestimatedusingdatafrom(UnitedStatesCensusBureau2015),andisshowninTable2k.

(ii) TheU.S.NationalHouseholdTravelSurvey(NHTS)reportsthatavehicleisusedfor57minutesperday,onaverage(Santosetal.2011).

(iii) AlsofromtheNHTS,weestimatedtheratiobetweenthenumberofvehiclestravellingatpeakhoursandthenumberofvehicletravellingatalltimesoftheday,as1.94(calculatedfromFigure12of(Santosetal.2011)).

Table2.Numberofvehiclesontheroadatpeakhourspercapita,asafunctionofpopulationdensity

Peopleperkm2

Numberofvehiclesownedpercapita

Numberofvehiclesontheroadatpeakhourspercapita

10 1 0.077200 0.8 0.0611000 0.65 0.0502000 0.6 0.0463000 0.55 0.0425000 0.44 0.03412000 0.22 0.017

Numberofbuses/Numberofvehiclesontheroad.Inthebasecase,weassume1.4%ofthevehiclesontheroadarebuses.BasedonNHTSdata,about10%oftotalpassenger-kmperyearintheU.S.aretravelledinpublictransit(calculatedfromTable7of(Santosetal.2011)).Assuming1.5astheaveragecaroccupancy(NHTS,Table16)and11astheaveragebusoccupancyintheU.S.(Puchalsky2005),wecalculatedtheaverageratiobetweenthenumberofbusesandthetotalnumberofvehiclesontheroadas10%*1.5/11≈1.4%.Itisassumedthatthisratioappliesforpeakhours,andmostpublictransitkmaretraveledinbuses.

DSRCPenetration.Thebase-casevalueofpenetrationofDSRCOBUsinvehiclesis25%.Thisisreasonableforadecision-makerlooking5to10yearsaheadinthecontextofagovernmentmandatetodeployDSRCinvehicles.TheU.S.DepartmentofTransportationforecaststhatallnewscarswouldbesoldwithOBUswithin3yearsafterasafetymandateiseffective(Hardingetal.2014).TheaveragelifetimeofanewcarintheU.S.is11years(Wile2014),soaslongasabout9%ofallvehiclesarereplacedeachyear,itisreasonabletoexpectpenetrationwillreach25%afterafewyearsfollowingthemandate–indeed,(Hardingetal.2014)estimateis5to6years.

DatatrafficperDSRC-equippedvehicleontheroad.Forthebasecase,weassumethatinany5-secondintervalduringthepeakhour,50%oftheDSRC-equippedvehiclesontheroadareendpointsfordatabeingcontinuallyat800kbps(totaldownstreamandupstream).Theremainingvehiclesarenotendpointsfortraffic,althoughtheymayrelaypacketsforothervehiclesinmultihopconnections.ThustheaveragetrafficperDSRC-equippedvehicleontheroadis400kbps.ThisisconsistentwiththeDeutscheTelekom kWeusedthedatafromtheAmericanCommunitySurvey2013,atthecountylevel,availableatcensus.gov/programs-surveys/acs/.TheU.S.averageisroughly0.9.

Page 16: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

16

predictionthatvehiculartrafficwillreach5GB/monthinthe“nextyears”(DeutscheTelekom2013),iftheaveragevehicleisontheroad57minperdayasdiscussedabove,andaveragedatarateisthesamewheneverthevehicleisinuse.Inreality,dataratesvaryfromvehicletovehicleatanygiventime,butsinceRSUsaretypicallyinrangeofdozensofDSRC-equippedvehiclesatalltimesduringpeakhour,thissimplifyingassumptionshouldhavelimitedeffectonaggregatethroughput.

Shareofdownstreamtraffic.Weassumethatwhileavehicleistransferringdata,90%ofthedataflowsinthedownstreamdirection(RSUtovehicle).InPortoDSRCnetwork92%ofasessionvolumeisdownstream,onaveragel,and(Sandvine2014)reportsasimilarratioforthemonthlyusagepermobiledeviceintheU.S.

Unitcostofmacrocellulartower.Thebase-caseassumptionforNPVofcostpermacrocelltowerover10yearsis$750,000.Wherecarriersareleasingspaceonexistingcelltowers,thiscostincludesleasingfees.Wherecarriersbuildtheirowntowers,adecadeofleasingfeesisreplacedbyCAPEX.A10-yearNPVof$750,000isroughlyconsistentwithsomepreviousestimatesthatvarybetween$650,000(FederalCommunicationsCommission2010a),$800,000(HallahanandPeha2011),and$900,000(Newman2008),in2014dollars.

Macrocellularspectrumefficiency.Weassumedthedownstreamaverageefficiencyofamacrocellas1.4bps/Hz/sectorforthebasecasevalue,whichisanacceptedvalueforLTE-FDDrel.8,assessedbythe3GPP(Sesia,Toufik,andBaker2011).Somedeviceswillbemorespectrallyefficient,suchasthoseusingLTE-Awhichisexpectedtohaveanefficiencyof2.4bps/Hz/sectorormore(Sesia,Toufik,andBaker2011),whileusageoflessefficientdevicesalsocontinues(withefficienciesbelow1bps/Hz/sector,asestimatedin(Clarke2014)).

Sectorspermacrocell.Weassumedeachmacrocellisdividedin3sectors,whichisconsistentwithatypicalmacrocellconfiguration(Sheikh2014).

Macrocellularbandwidth.Weassumedthatanynewtowerdeployedinacapacity-limitedregionwouldbeconstrainedbythebandwidthavailablefordownlink,andwouldoperateoveradownlinkbandwidthof70MHzpersectorinthebasecase.Atier-1providerisestimatedtoholdroughly30MHzofdownlinkspectrumforLTE,onaverage(Goldstein2015),andspectruminuseforLTEisestimatedasabouthalfoftotalspectrumformobilebroadbandm.Substantialamountsofnewspectrumareexpectedtobeallocatedbyregulators(FederalCommunicationsCommission2010b),butitseffectiveusemaytakeseveralyearsforreallocation,auctioningandactualdeployment.Thuswechose70MHzasabasevalueofdownlinkspectrumforthenextfewyears.

Macrocellularfrequencyreusefactor.Weassumedafrequencyreusefactorinmacrocellsof1.Thisisconsistentwithatypicalmacrocellularnetworkconfigurationwithcurrenttechnology(WannstromandMallinson2014).

UnitcostofDSRCRSU.WechosetheaverageNPVover10yearsofaDSRCRSUas$14,000.ThisisbasedonU.S.DOTestimates(averageannualcostbetween$2,000-3,000(Wrightetal.2014),includingreplacementcostsevery5to10years).However,inSection4wewillconsidervariationsofmorethan50%fromthebasecasevalue,asconditionsaboutinfrastructureavailabilitymayvary.Forexample,theCityofPortodeployedRSUsfora

lAsinMarch2015.mSeee.g.(Clarke2014),Table3,or(SprintNextelCorporation2011).

Page 17: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

17

Capexofbetween$1,200-4,000,byplacingRSUsinexistingstructures(trafficpoles,buildings,etc.)alreadyownedbythecityandalreadyequippedwithenergyandbackhaulaccess.ThecostperRSUcouldbealsobelowerifRSUsdeployedforInternetaccessaresharedforsafetyorvice-versa,thoughsharingdependsonmanyissues,includingwhethertheoptimalplacementofRSUsforInternetaccessmatchestheplacementforsafetycommunicationsandwhetherdevicesforInternetaccessanddevicesthataresafety-criticalareplacedundersharedcontrol.ForthebasecasevalueofthecostperRSU,nosharingisassumed.

Ontheotherhand,costscanbesignificantlyhigherifnewpoles,energyandcommunicationsinfrastructurehastobebuiltentirely(Wrightetal.2014).

UnitCostofDSRCOBU.Forthecaseofthevehiculardevice,ourbase-caseassumptionforNPVofthecostofaDSRCOBUis$350.ThisisbasedonU.S.NHTSAestimates(Hardingetal.2014)consideringfourradiointerfacesandantennaspervehicle.However,massproductionofOBUscoulddriveunitcostsdown,especiallyinthecontextofamandate.Becauseofthis,wevarytheNPVperOBUfrom$350downto$50inSection4.

UnitcostofDSRCspectrum.ForthecostofDSRCspectrum,wechoseavalueof$0.10perMHzperpopulation(MHz-pop).Thisvalueisuncertain,asthecostofspectrumdependsonfrequency(Keransetal.2011;TanandPeha2015;Alotaibi,Peha,andSirbu2015;Peha2013),andthemarketvalueabove5GHzisnotwell-established.

4 ResultsThisSectionpresentsthesimulatedDSRCthroughput,benefitandcostresultsforthebasecasescenario,andhowthoseresultsvaryifbasecasevalueschangeeitherbecauseofvariationsacrosscitiesorregions(suchaspopulationdensity),anticipatedchangesovertime(suchaspenetrationortrafficpervehicles),oruncertaintyaboutthebasecasenumericalassumptions.Section4.1showsthroughput,benefitandcostsforthebasecasescenario,whereinSections4.2to4.7wediscussthevariationsinvaluesofpopulationdensity,penetration,trafficpervehicle,unitcostsofDSRC,andunitcostandbandwidthofmacrocellulartowers.

4.1 BaseCaseScenario–Benefit,CostsandVolumeofInternetAccessThroughDSRC

Figure3showsthroughputasafunctionofRSUsperkm2underbasecaseassumptions.Thethroughputincreaseswithhigherquantitiesofinfrastructure,asthenumberofvehiclesthatcanconnecttoaRSUincreases.However,themarginalgainsinoffloadratedecreaseastheRSUdensityexceeds2perkm2.ThismattersbecausewhileincreasingRSUdensityincreasesDSRCthroughputandthereforebenefit,italsoincreasescost.ThiscanbeseeninFigure4,whichshowsbothbenefitandcostsasafunctionofRSUdensityunderthesameassumptions.Figure4showsthatforthebasecasevalues,themaximumdifferencebetweenbenefitsandcostsoccursatinfrastructuredeploymentof1RSUperkm2.Atthisoptimalquantityofinfrastructure,benefitsexceedthecostofRSUsby50%.IfthespectrumhasalreadybeenallocatedandtheOBUsarealreadybeingpurchased,asislikelytooccurifDSRCisdeemedtobeimportantforsafetyapplications,thenthosearesunkcosts.

Page 18: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

18

Consequently,thebenefitsofdeployingRSUsexceedthecosts,anddoingsowillincreasesocialwelfare.However,Figure4alsoshowsthatbenefitofInternetaccessisconsiderablylessthanthecostofOBUs,muchlessthecombinedcostofRSUs,OBUs,andspectrum.Thus,thevalueofdeploymentofvehicularnetworksforInternetaccessalone,i.e.withoutconsiderationoftheimprovementsinhighwaysafety,arenotsufficienttojustifythedeploymentofOBUsandtheallocationofspectruminthebasecasescenario.

Infrastructure Density (RSUs/km2)

Figure3.Averagetrafficofferedandoffloadrateatapeakhour,forthebasecasescenario

Infrastructure Density (RSUs/km2)

Figure4.Benefitandcostforvaryinginfrastructuredensity,forthebasecasescenario

0.5 1 2 3 4 6 8

Peak

-hou

r Tra

ffic

(Mbp

s/km

2 )

0

4

8

12

16

20

Offloaded

Offered: 400 kbps per DSRC-equipped vehicle

0.5 1 2 3 4

Bene

fit a

nd C

ost N

PV (U

SD/k

m2 )

0K

50K

100K

150K

200K

Benefit

Cost: RSU+OBU+Spectrum

Cost: RSU+OBU

Cost: RSU

Page 19: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

19

Thenumberofsimulatedobservationsishighenoughsuchthatstatisticalsignificanceinthisgraph(andthegraphsthatfollow)issufficienttosupportconclusions.Forthebasecasescenario,foreachvalueofRSUdensity,1000vehiclesweresimulatedina20km2regionovera50-secondperiod.Wecalculateaveragethroughputovertimeforeachofthevehiclessimulated.Ifwemakethesimplifyingassumptionthatthethroughputsofthese1000vehiclesaremutuallyindependent,thenthemeanthroughputacrossallvehiclesis170kbps,whichisabout40%ofofferedload,andtheconfidenceintervaliswithin7%ofthemean.Thesamplestandarddeviationis180kbps.ThesamplestandarddeviationisaslargeasthesamplemeanbecausethedistributionofDSRCthroughputisclosetobimodal;withbasecaseassumptions,offloadiszeroin30%oftheobservations,wherevehiclescannotconnecttoaRSUeitherthroughsingleormultihoppaths,andDSRCthroughputequalstheofferedtrafficin30%oftheobservations.

4.2 PopulationDensitySection4.1showedthatdeployingRSUscanincreasesocialwelfareinthebaselinecase,whichcorrespondstoadenselypopulatedcitysuchasPorto,Chicago,orLondon.However,thatmaynotbethecaseeverywherebecauseboththecostsandthebenefitsofInternetaccessthroughvehicularnetworksarelikelytodependgreatlyonpopulationdensity.Thehigherthepopulationdensityinanarea,itisexpectedthatthenumberofvehiclesownedbythatpopulation,andthenumberofvehiclesontheroadatpeakhours,willbothincrease.Thereforeitisexpectedthatmorein-vehicleOBUswillbeused,andmoreRSUswillbedeployedforthosevehiclestoconnectto.Ononehand,thismakesDSRCcostsofOBUandRSUsincreasewithpopulationdensity.Ontheotherhand,throughputperunitofarea,andhencethebenefit,arealsoexpectedtoincrease.Thus,thisSectionexaminestheeffectofpopulationdensityonbenefitandcost.

Figure5andFigure6showthroughput,benefitandcostsasafunctionofpopulationdensity.Trafficpervehicle,penetration,unitcostsandspectrumparametersareheldconstantatbasecasevaluesforallpopulationdensities.BenefitandthecostofRSUsinFigure6dependonthequantityofRSUsforeachpopulationdensity,whichischosenasfollows.ForthevaluesofpopulationdensityinwhichtheNPVofbenefitofInternetaccessexceedstheNPVofcostofRSUs,thenumberofRSUschosenisthequantitythatmaximizesthedifferencebetweentheNPVofbenefitandtheNPVofRSUcost(thethroughputforeachscenarioisactuallysimulatedforadiscretesetofRSUquantitiesn.AlinearfitisperformedsuchthatfractionalRSUquantitiesarealsopossible).ForthepopulationdensityvaluesinwhichtheNPVofbenefitislowerthantheNPVofRSUcostforanyquantityofRSUs,theoptimalquantityisobviouslyzero,leadingtozerobenefitofInternetaccessandzeroRSUcostforthosepoints.However,forthesepointsweinsteadcalculatethequantityofRSUsasalinearextrapolationfromthepopulationdensityrangewhichtheNPVofbenefitisgreaterthantheNPVofRSUcost.ThisshowshowfarfromRSUcoststhebenefitofInternetaccesswouldbeinthose“negative”regions,thoughitdoesnotshowtheoptimalcost(becauseoptimalRSUcostandbenefitwouldbothbezerointhisregion).

Figure5showsthatofferedtrafficincreasesrapidlyasafunctionofpopulationdensity,whichisexpectedconsideringconstanttrafficpervehiclebutanincreasingquantityof

nEachscenarioissimulated12times,onetimeforeachofthefollowingvaluesofRSUdensity:0.25,0.5,0.75,1,1.25,1.5,2,3,4,6,8,and10RSUs/km2.Thenumberofpointswaschosenforsimulationtimereasons,sincethens-3modeliscomputationallyexpensive.

Page 20: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

20

vehiclesforhigherpopulationdensities.DSRCthroughputalsoincreaseswithpopulationdensity,thoughatalowerpacethanofferedtraffic.ThishappensbecausetheincreasedInternettrafficoriginatedfromagreaternumberofvehiclesperkm2causesmorecompetitionfortheuseofthewirelessmedium,andlimitsoffload.

Figure6showsthatbenefitincreasesfasterthanRSUcost.ThereasonisthatthroughputgrowsroughlyproportionallytopopulationbuttheoptimalnumberofRSUsrisesatalowerpace:for12000people/km2,thedifferencebetweenbenefitandRSUcostisfourtimeshigherthanfor5000people/km2areas.ThedensityatwhichbenefitofInternetaccessisequalorgreaterthanRSUcostdependsonpenetration,trafficandunitcosts;forthebasecasevaluesoftheassumptions,the“threshold”valueis4000people/km2.IfdecisionsaboutwhethertodeployRSUsaremadeonacity-widebasis,thismeanscitieswithpopulationdensitiesatleastasgreatasChicagoorPortoowouldbenefitfromRSUdeployment,atleastincaseswherethereisalreadyspectrumallocatedandamandateofDSRCOBUsforsafetypurposes.However,RSUscouldbedeployedwithinanareamuchsmallerthanacity,andmanycitieswithmoremodestpopulationdensityhavesomeneighborhoodswithpopulationdensityover4000peopleperkm2.

Population Density (people/km2)

Figure5.Averagetrafficofferedandoffloadrateatapeakhourforvaryingpopulationdensitiesandotherparametersfixedatbasecasevalues,optimalRSUquantityateachpoint(i.e.atRSUquantitythatmaximizestheNPVofbenefitminustheNPVofcostforeachpopulationdensity:1to2RSUs/km2)

oSee(GoverningInstitute2015)forpopulationdensitiesofthoseandothercities.

5K 10K 15K 20K

Peak

-hou

r Tra

ffic

(Mbp

s/km

2 )

0

12

24

36

48

60

Offloaded

Offered

Page 21: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

21

Population Density (people/km2)

Figure6.Benefitandcostforvaryingpopulationdensities(andotherparametersatbasecasevalues),andoptimalRSUquantityateachpoint

Figure6showsthatbenefitgrowsfasterthanRSUcostforawiderangeofpopulationdensities.However,thetrendisdifferentforOBUcost,whichgrowsmuchfasterthanbenefit.ThereasoncanbeseenfromFigure7,whichshowsvehicleownershipandvehicleusageasafunctionofpopulationdensity,usingbasecaseassumptionsofvehicleownershippercapita,timeontheroadpervehicleandpeakhourratio.Ownershipreferstothetotalnumberofvehiclesavailableperunitofarea.Intheeventofamandate,ownershipdetermineshowmanyvehicleswillhaveOBUsinstalled,andthetotalOBUcost.Ontheotherhand,vehicleusageisthenumberofvehiclesontheroadatpeakhours.AvehicleequippedwithDSRCwillonlyhavetrafficcarriedwhileontheroad,andonlythepeak-hourthroughputisrelevantforbenefit.Thus,Figure7helpsexplainwhyOBUcostsaresignificantlyhigherthanbenefitofInternetaccess,underamandatescenario.Overlocationswithincreasingpopulationdensities,andunderuniformDSRCpenetration,vehicleownershiprisesfasterthanvehicleusage,makingOBUcostsrisefasterthanbenefits,atleastforbasecasevaluesoftheotherparameters.Thismaynotbetrueforallassumptions.Forexample,ifOBUscostless,thentotalOBUcostwouldgrowmoreslowlywithrespecttopopulationdensity,buttheOBUcostswouldneedtobelowerthatbaselinebyoneorderofmagnitudeforOBUcostsnottoincreasefasterthanbenefits.

5K 10K 15K

Bene

fit a

nd C

ost N

PV (U

SD/k

m2 )

0K

40K

80K

120K

160K

200K

Benefit

Cost: RSU+OBU+Spectrum

Cost: RSU+OBU

Cost: RSU

Page 22: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

22

Population Density (people/km2)

Figure7.Averagevehicleownershipandusage(i.e.vehiclesontheroadatpeakhours)forpopulationdensityranges

Ontheotherhand,Figure8showswhathappensiftheratiobetweenthequantityofDSRC-equippedvehiclesinuseandthetotalquantityofvehiclesownedisdifferentthaninthebasecaseassumption.Forthisgraph,thepopulationdensityisheldinthebasecasevalue(5000people/km2),aswellaspenetration,trafficpervehicle,unitcosts,spectrumcharacteristicsandnumberofvehiclesowned.Whatisvariedisthenumberofvehiclesontheroadatpeakhourpercapita,meaningtheratiobetweenthatandthenumberofvehiclesownedchanges.Theratiovalueof0.08correspondstothebasecase,andislikelytovaryamongcitieswithcomparablepopulationdensities,inpartduetofactorsliketheavailabilityofpublictransportation(EuropeanCommission2012).AsFigure8shows,thenetbenefitofdeployingRSUswillbegreaterinacitywherealargerfractionofcarsareontheroadinpeakhours.

Ifvehiclesareequippedvoluntarilyratherthanbecauseofamandate,thenFigure8isrelevantforadifferentreason.Ifadoptionisvoluntary,ownersofvehiclesthatareofteninusearemorelikelytoadopt,andthiswouldalsohavetheeffectofincreasingtheratioofDSRC-equippedvehiclesontheroadatpeakhourtototalcarsthatisshowninFigure8.Thus,ifmanyoftheDSRCequippedcarsaredrivenextensively,asiscertainlythecasefortheDSRC-equippedbusesandtaxisinPorto,thenthiswillalsoincreasethenetbenefitofdeployingRSUs.

102 103 104

Vehi

cle

Den

sity

(veh

/km

2 )

0

500

1000

1500

2000

2500

3000

peak-hour usage

ownership

Page 23: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

23

Number of Vehicles Using DSRC at Peak-hour /Number of Vehicles Owned

Figure8.BenefitandcostforvaryingratiosbetweenthequantityofDSRC-equippedvehiclesinuseandthetotalquantityofvehiclesowned(andotherparametersatbasecasevalues),andoptimalRSUquantityateachpoint(1to2RSUs/km2)

4.3 PenetrationLikepopulationdensity,OBUpenetrationislikelytoaffectbenefitandcosts,althoughunlikepopulation,penetrationmayincreaserapidlyovertime.Withhigherpenetration,itisexpectedthatboththenumberofDSRC-equippedvehiclesandthenumberofDSRC-equippedvehiclesontheroadatpeakhourswillincrease.ThereforeitisexpectedthatmoreRSUsforthosevehiclestoconnecttowillbenecessary.Ononehand,thismakesDSRCcostsofOBUandRSUsincreasewithpenetration.Ontheotherhand,DSRCthroughputperunitofarea,andhencethebenefit,arealsoexpectedtoincrease.ThisSectionexaminestheeffectofpenetrationonbenefitandcost.

Figure9andFigure10showthroughput,benefitandcostsasafunctionofOBUpenetration,assumingthepopulationdensity,quantityofvehicles,trafficpervehicle,unitcostsandspectrumparametersareheldconstantatthebasecasevaluesforallvaluesofpenetrationconsidered.BenefitandthecostofRSUsdependonthequantityofRSUsforeachpenetration,whichischoseninthesamewayasinthepreviousSection.

Figure9showsthetotalamountoftrafficofferedincreasesrapidlyasafunctionofpenetration,whichisexpectedconsideringanincreasingquantityofvehiclesforhigherpenetrations.TheDSRCthroughputisalsohigher.Ifpenetrationincreasesovertimeasexpected(especiallyifthereisamandate),thenDSRCthroughputwillincreaseovertime.Sincebenefitisdefinedasafunctionofthroughput,RSUsareexpectedtobedeployedonlyinareaswherethepotentialratesarehighenoughforbenefittoexceedRSUcost,aslongasspectrumandOBUcostsaresunk.Therefore,thegrowthofDSRCthroughputovertimewouldeventuallycausethepotentialbenefitofInternetaccesstoexceedthecostofRSUsinregionswherethisisnotinitiallythecase.

0.05 0.1 0.15 0.2 0.25 0.3

Bene

fit a

nd C

ost N

PV (U

SD/k

m2 )

50K

100K

150K

200K

Benefit

Cost: RSU+OBU+Spectrum

Cost: RSU+OBU

Cost: RSU

Page 24: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

24

DSRC Penetration (@ pop.dens.=5000/km2)

Figure9.Averagetrafficoffloaded(andoffered)atapeakhour:varyingDSRCpenetration(andotherparametersatbasecasevalues),optimalRSUquantityateachpoint(0.8to1.6RSUs/km2)

Figure10showsbenefitandcostsasafunctionofpenetration,withallparametersexceptpenetrationandpopulationdensityinthebasecasevalues.Thetophorizontalaxisshowspenetrationforalowerpopulationdensity(2000people/km2),whilethebottomhorizontalaxisshowspenetrationforthebasecasepopulationdensity(5000).Figure10showsthatasOBUpenetrationincreases,benefitincreasesfasterthanRSUcost.Thus,itisamatterofwaitinguntilpenetrationishighenoughthatbenefitofInternetaccesswouldexceedRSUcost.Ifpenetrationishighenough,itwillremainhighenough.Ontheotherhand,incitieswhereRSUdeploymentdoesnotresultinbenefitexceedingRSUcostwithinthecurrentplanninghorizon,thismaychangeafterafewyearsaspenetrationincreases.Moreover,benefitwillexceedRSUcostsoonerforcitieswithhigherpopulationdensity.UnderthenumericalassumptionsofFigure10,benefitofInternetaccessexceedsRSUcostswhenpenetrationis0.19orgreaterinacitywithpopulationdensityof5000,andwhenpenetrationis0.37orgreaterinacitywithpopulationdensityof2000peopleperkm2.

AlthoughbenefitofInternetaccessincreasesfasterthanRSUcostsaspenetrationincreases,OBUcostincreasesmuchfasterthanbenefit.ThereasonisthatpenetrationaffectsOBUcostandofferedtrafficlinearly,buttheformerincreasesmuchfasterthanthelatter,atleastforbasecasevalues.Moreover,benefitdependsonDSRCthroughput,whichincreaseslowerthanofferedtraffic.EvenifallofferedtrafficwerecarriedthroughDSRC,benefitwouldbenomorethantwiceascurrentlyestimated,andOBUcostwouldstillincreasefasterwithpenetration.

Thisresultmeansthatforthebasecaseassumptions,OBUcostfarexceedsbenefitofInternetaccessregardlessthepenetration.Andaspenetrationisexpectedtoincreaseovertime,thenthedifferencebetweenOBUcostandbenefitisalsolikelytoincrease.Inthissituation,iftherewereamandatewithnobenefitsotherthanInternetaccess,whichcouldonlybetrueifDSRChadnosafetybenefitswhatsoever,thensocialwelfarewoulddecrease.

0 0.1 0.2 0.3 0.4 0.5

Peak

-hou

r Tra

ffic

(Mbp

s/km

2 )

0

8

16

24

32

Offloaded

Offered

Page 25: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

25

DSRC Penetration (@ pop.dens.=2000/km2)

DSRC Penetration (@ pop.dens.=5000/km2)

Figure10.BenefitandcostforvaryingvaluesofDSRCpenetration(andotherparametersatbasecasevalues),andoptimalRSUquantityateachpoint.Eachxaxisreferstoadifferentpopulationdensity

4.4 CostperOnboardUnit(OBU)InordertoinvestigatewhetherbenefitofInternetaccessthroughDSRCwouldeverexceedallcosts,includingtheOBUcostthatdominatedinthebasecase,thisSectionexaminestheeffectoftheOBUunitcostontotalbenefitandcost.

Figure11showsbenefitandcostsasafunctionofOBUunitcost,forthebasecasevaluesofpopulationdensity,thequantityofvehicles,penetration,traffic,RSUandmacrocellularunitcosts,andspectrumparameters.ThequantityofRSUsischosentomaximizethedifferencebetweenbenefitandRSUcost.IfamandatewastobejustifiedbyInternetaccessonly,thenbenefitofInternetaccessaloneshouldexceedallDSRCcosts.Figure11showsthattotalOBUcostwouldexceedRSUandspectrumcostscombinedundertheseassumptions,andthatthesumofRSUandOBUcostswouldexceedbenefitofInternetaccessevenifthecostperOBUfallsbymorethan80%from$350to$50.

ItisonlypossibleforthecostperOBUtodecreaseovertherangeshowninFigure11ifDSRCismass-deployedatascalecomparabletoWi-Fi.InthephysicallevelDSRCisspecifiedbytheIEEE802.11pstandard(IEEE2010a),whichismostlyanadaptationoftheWi-Fi802.11astandardforthe5.9GHzband.Wi-Firadioswithantennascurrentlycostnomorethanafewtensofdollars,andperhapsDSRCOBUcostscoulddropifitismassproduced.Butevenifthishappens,Figure11showsthatbenefitstilldoesnotexceedtotalOBUcost.

However,ifthereisamandateinwhichspectrumisalreadyallocatedandOBUsarepurchased,thenspectrumandOBUcostsaresunk.Inthisscenario,sincebenefitofInternetaccessexceedsRSUcostforbasecaseassumptions,RSUdeploymentforInternetaccessdoesincreasesocialwelfare.

0 0.1 0.2 0.3 0.4 0.5

Bene

fit a

nd C

ost N

PV (U

SD/k

m2 )

30K

60K

90K

120K

0.2 0.4 0.6 0.8 1

Cost: RSU+OBU+Spectrum

Cost: RSU+OBU

Benefit

Cost: RSU

Page 26: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

26

NPV per Vehicle OBU (USD)

Figure11.BenefitandcostforvaryingNPVperOBU(andotherparametersatbasecasevalues),andoptimalRSUquantity(1RSU/km2)

4.5 InternetTrafficperVehicleItisimportanttoconsiderdifferentvaluesfordataratepervehicle,bothbecausethereareuncertaintiesinanypredictionoffuturedatarate,andbecausedatarateisgenerallyexpectedtoincreaserapidlyovertime(Cisco2015;Clarke2014).ThisSectionexaminestheeffectoftrafficpervehicleonbenefitandcost.

Figure12showsthroughputasafunctionoftrafficpervehicle,assumingthepopulationdensity,quantityofvehicles,penetration,unitcostsandspectrumparametersareheldconstantinthebasecasevaluesforallvaluesoftrafficconsidered.Foranincreaseinthetrafficpervehicle,DSRCthroughputincreases,thoughwithadecreasingmarginalgain.Figure12suggeststhatDSRCthroughputisstillgrowingfortrafficpervehicleashighasfourtimesthebasecasevalue.Iftrafficpervehicleincreasesovertime,thenDSRCthroughputislikelytoincreaseovertimeaswelleveniftrafficgrowsasmuchthewiderangeshowninFigure12,underbasecasevaluesfortheotherassumptions.

50 100 150 200 250 300 350

Bene

fit a

nd C

ost N

PV (U

SD/k

m2 )

0K

30K

60K

90K

120K

150K

Cost: RSU+OBU

Cost: RSU+OBU+Spectrum

Cost: RSU

Benefit

Page 27: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

27

Average Traffic per Vehicle (kbps)

Figure12.AveragetrafficofferedandoffloadrateatapeakhourforvaryingtrafficperDSRC-equippedvehicleontheroad,optimalRSUquantityateachpoint(0.9to2RSUs/km2)

Figure13showsbenefitandcostsasafunctionoftrafficpervehicle,assumingthepopulationdensity,quantityofvehicles,penetration,unitcostsandspectrumparametersareheldconstantinthebasecasevaluesforallvaluesoftrafficconsidered.BenefitofInternetaccessandthecostofRSUsdependonthequantityofRSUsforeachvalueoftrafficpervehicle,whichischosentomaximizethedifferencebetweenbenefitandRSUcost,asexplainedfurtherinSection4.2.Forhigherdatarates,bothbenefitandRSUcostarehigheraswell,andthedifferencebetweenthemisalsohigherthanwithlowerdatarates.InpreviousSectionsitisshownthatbenefitexceedsRSUcostforlocationswithpopulationdensityabove4000peopleperkm2,withthebasecaseassumptionoftrafficpervehicle.SinceFigure13showsthatthedifferencebetweenbenefitofInternetaccessandRSUcostincreaseswithtrafficpervehicle,andiftrafficwillincreaseovertimeassomepredict,thenbenefitwouldexceedRSUcostinlocationswithpopulationdensitiesbelow4000peopleperkm2overtime,i.e.inareaswhichpopulationdensitiesthatwerenothighenoughtoresultinsignificantoffloadsoonafterthemandateiseffective.

Figure13alsoshowsthat,underthebasecasescenariofortheotherassumptions,benefitofInternetaccessexceedsRSUcostforatrafficpervehicleabove250kbpsatpeakhours.Thiscorrespondstoamonthlyusageof3GBpervehicle(also,underbasecaseassumptions).Thus,deployingRSUswouldstillresultinbenefitexceedingRSUcostsoonafterthemandatebecomeseffectiveinthedensely-populatedurbanarearepresentedbyourbasecaseifdatarateisabouthalfwhatsomearecurrentlypredicting.

TheaveragedatarateofaDSRC-equippedvehiclemayalsoexceedtheaveragedatarateofallvehiclesifvehicleownerspurchaseOBUsvoluntarily,ratherthanonlyinresponsetoamandate.Theownerswhoadoptvoluntarilywouldbetheoneswhobenefitthemost.IfownersarechargedforInternetservicebasedonusage,thenmoreownersofvehicleswithhighervolumesofInternettrafficwouldoptin,andaveragedataratescouldbemuchgreaterthanthebasecase.Forexample,abuscompanyofferingInternetservicefor

0 200 400 600 800 1000 1200 1400

Peak

-hou

r Tra

ffic

(Mbp

s/km

2 )

0

16

32

48

64

80

Offloaded

Offered

Page 28: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

28

passengers(suchastheoneinPorto)mightvoluntarilyinstallOBUsassoonasRSUsareoperatingbecausethebuscompanyexpectsadataratepervehiclethatiswellaboveaverage,andcarryingthattrafficoveracellularnetworkwouldbeexpensive.Thus,foragivenOBUpenetrationrate,benefitofInternetaccesswillexceedcostsatalowerpopulationdensityifthereisasignificantlevelofvoluntaryadoptionofOBUs.

Average Traffic per Vehicle (kbps)

Figure13.BenefitandcostforvaryingvaluesoftrafficperDSRC-equippedvehicleontheroad(andotherparametersatbasecasevalues),andoptimalRSUquantityateachpoint

4.6 CostperRoadsideUnit(RSU)IfthecostperRSUislowerthaninthebasecase,thenitmaybeworthwhiletodeploymoreRSUstoincreasetotalthroughput.Ontheotherhand,ifRSUsaresignificantlymoreexpensivethaninthebasecase,thenthatmaypreventdeploymentandresultinnobenefitatall.ThisSectionexaminestheeffectofRSUunitcostsontotalbenefitandcost.

Figure14andFigure15showthroughput,benefitandcostsasafunctionofRSUunitcost.Thebasecasevaluesofpopulationdensity,thequantityofvehicles,penetration,traffic,OBUandmacrocellularunitcosts,andspectrumparametersareassumed.ThequantityofinfrastructureforeachvalueofRSUunitcostischosentomaximizethedifferencebetweenbenefitandRSUcost,asexplainedfurtherinSection4.2.ThecostperRSUaffectsthatoptimalquantityofRSUs,whichinfluencesDSRCthroughput.ThisisshowninFigure14:ifthecostperRSUislowerthanthebasecasevalue($14,000),thentheDSRCthroughputishigherandviceversa.However,evenwiththatvariationinthroughput,Figure15showsthatthetotalbenefitandcostresultsarerobusttoawidevariationofcostsperRSU.Evenifthiscostis30%higher(orlower)thanthebasecase,benefitofInternetaccesswillstillexceedtotalRSUcost.

However,thatresultmightchangeifthecostperRSUisradicallydifferentthanthebasecase.Forexample,ifRSUsaredeployedbybusinessesinplacesthatrequireexpensivepolesorlackofaccesstocommercialpowerorcommunications,thenthecostperRSU

0 200 400 600 800 1000 1200 1400

Bene

fit a

nd C

ost N

PV (U

SD/k

m2 )

0K

50K

100K

150K

200K

Cost: RSU+OBU+Spectrum

Cost: RSU+OBU

Cost: RSU

Benefit

Page 29: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

29

mightbemuchhigher,andFigure15showsthatbenefitofInternetaccessislowerwhentotalRSUcostifitsunitcostishigherthan$20,000perRSUandotherassumptionsareatbasecase.Ontheotherhand,ifthedecisiontodeployRSUsaremadebyamunicipalitythatalreadyhaspole,energyandbackhaulinfrastructureavailable,costperRSUmaybelow,andRSUdeploymentmightbebeneficialevenforlessdenselypopulatedcitiesthanthe“threshold”densityshowninSection4.2forbasecaseassumptions,aslongasspectrumandOBUcostsaresunkunderamandate.

NPV per RSU (USD)

Figure14.AverageoffloadrateatapeakhourforvaryingPVperRSU,andoptimalRSUquantityateachpoint(1.3to0.8RSU/km2)

5K 10K 15K 20K 25K

Peak

-hou

r Tra

ffic

(Mbp

s/km

2 )

0

4

8

12

16

20

Offloaded

Offered

Page 30: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

30

NPV per RSU (USD)

Figure15.BenefitandcostforvaryingPVperRSU(andotherparametersatbasecasevalues),andoptimalRSUquantityateachpoint

4.7 MacrocellularCostsMacrocellularcostsareexpectedtoinfluencebenefitandcostsintheoppositewayastheDSRCunitcostsanalyzedinthepreviousSection.Figure16showsbenefitandcostsasafunctionoftheunitcostpermacrocellulartower,forwhichisassumedthebasecasevaluesofpopulationdensity,thequantityofvehicles,penetration,traffic,DSRCcostsandspectrumparameters.IftheNPVofthecostpermacrocellulartowerishigherthanthebasecaseassumption,thenbenefitofInternetaccessexceedsRSUcostinlesspopulatedareasthaninthebasecasescenario.Ontheotherhand,ifmacrocellularcostislowerthaninthebasecase,thanthebenefitmightbelowerthaninthebasecase.However,Figure16showsthatthefindingsinpreviousSectionsdonotchangesubstantiallyifthecostpermacrocellulartowerchangesoverarangeof20%beloworabovethebasecasevalue.

5K 10K 15K 20K 25K

Bene

fit a

nd C

ost N

PV (U

SD/k

m2 )

0K

50K

100K

150K

200K

Cost: RSU+OBU+Spectrum

Benefit

Cost: RSU+OBU

Cost: RSU

Page 31: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

31

NPV per Macrocell Tower (USD)

Figure16.BenefitandcostforvaryingNPVpermacrocellulartower,andoptimalRSUquantity(1RSU/km2)

Bandwidth per Macrocellular Carrier (MHz)

Figure17.Benefitandcostforvaryingbandwidthavailableformacrocells,andoptimalRSUquantityateachpoint(1RSU/km2)

Benefitsandcostsmayalsobeinfluencediftheamountofspectrumacarrierhasavailablevariesfromthebasevalue.Figure17showsbenefitandcostsasafunctionofthebandwidthavailablepercarrier,andbasecasevaluesofpopulationdensity,quantityofvehicles,penetration,traffic,unitcostsandDSRCspectrum,andindicatesthatbenefitofInternetaccessexceedsRSUcostifasmuchas20%morebandwidthpercarrierisinuse.Spectrumholdingsforcellularservicemayincreaseovertimeaslongasthegrowing

600K 700K 800K 900K

Bene

fit a

nd C

ost N

PV (U

SD/k

m2 )

0K

50K

100K

150K

200K

Benefit

Cost: RSU+OBU+Spectrum

Cost: RSU+OBU

Cost: RSU

50 60 70 80 90

Bene

fit a

nd C

ost N

PV (U

SD/k

m2 )

0K

50K

100K

150K

200K

Benefit

Cost: RSU+OBU+Spectrum

Cost: RSU+OBU

Cost: RSU

Page 32: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

32

demandformobileInternettriggerdecisionstoreallocatespectrumfromotherusestocellular–forexample,in2010theU.S.NationalBroadbandPlanrecommendedincreasingtheamountofspectrumavailableforbroadbandby500MHz.However,spectrumreallocationsarenotfrequentandtakeyearstobecomeeffective–65MHzwereauctionedin2015,beingthefirstsignificantadditiontomobilespectrumsince2008intheU.S.(Clarke2014).Therefore,overagivenperiodtheamountofspectrummayincreaselessthantherapidgrowthexpectedfortrafficpervehicle(whichincreasesbenefit,asshowninSection4.5),whichsuggeststhatthegrowthincellularspectrumisnotlikelytochangeourestimatesthatbenefitofInternetaccessexceedsRSUcostforbasecasevaluesoftheotherassumptions.Theaverageamountpercarriermayalsoincreaseifthereisareductioninthenumberofcarrierscompetinginaregion.However,Figure17showsthisisalsounlikelytobelargeenoughtochangeourconclusionswithinaplanninghorizon.

5 ConclusionsInthispaperweanalyzebenefitsandcostsofInternetaccessthroughDSRC.WefindthatiftherehasalreadybeenamandatetodeployDSRCinnewvehicles,thenthedeploymentofRSUsforInternetaccessincreasessocialwelfarefordenseurbanareaswhenOBUpenetrationisrepresentativeofafewyearsafteramandatebecomeseffectiveandpeak-hourInternettrafficpervehicleiscompatiblewithforecastsforthenextyears,andevenifthoseRSUsarenotsharedwithsafetyorotherapplications.Moreover,RSUdeploymentislikelytobecomewelfareenhancinginthefutureformanyless-populatedareasaswell,aslongaspenetrationorInternettrafficincreasesovertime.

BenefitisdefinedasthecostsavingsfromreducingthenumberofmacrocellulartowersthatwouldotherwisebeneededtocarrytrafficwhichisoffloadedthroughDSRC,andthecostsarethoseofDSRCRSUs,spectrumandOBUs.UnderamandatetodeployOBUs,ourresultsshowthatOBUcostismuchgreaterthanspectrumandRSUcosts,andOBUcostaloneexceedsthebenefitofInternetaccessthroughDSRC.Thus,ifDSRChadnosafetybenefitswhatsoever,thenmandatingOBUsandallocatingspectrumforthoseOBUswoulddecreasesocialwelfare.

However,ithasbeenestimatedthatanOBUmandatewillaccruesignificantroadsafetybenefits(Hardingetal.2014),whichhasmotivatedtheallocationofDSRCspectrumandthepossibilityofamandatetodeployDSRCinallnewvehiclesintheU.S.Ifthismandateoccurs,thenthedecisionofwhethertouseDSRCnetworksforInternetaccessbecomesadecisionaboutwhethertodeployroadsideinfrastructurethatcanserveasagatewaytotheInternet.Forthisdecision,bothOBUandspectrumcostswouldbesunk,andifbenefitofInternetaccessexceedsRSUcost,thenadecisiontodeployRSUinfrastructurewouldincreasesocialwelfare.OurresultsshowthatbenefitdoesexceedRSUcostunderbasecaseassumptions,whichcorrespondtodenseurbanareas.

Benefitandcostsarebothaffectedbypopulationdensity.Ifallelseisequal,benefitofInternetaccessthroughDSRCminusthecostofRSUsisgreaterwhenpopulationdensityisgreater.Withbasecaseassumptions,benefitexceedsRSUcostinlocationswithpopulationdensityabove4000peopleperkm2,i.e.onlyinfairlydenselypopulatedurbanareas.However,thisshouldchangeovertime.UnderanOBUmandate,thevolumeoftrafficpervehicleandOBUpenetrationarebothlikelytoriserapidlybeyondourbaselineassumptionsinthecomingyears,andourresultsshowthateitherofthesechangeswould

Page 33: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

33

increasebenefitofInternetaccessminusRSUcosts.Thus,ifallassumptionsareclosetobasecasevaluesexceptOBUpenetrationandtrafficpervehicle,thenbenefitwillexceedthecostofRSUsinregionswithlowerandlowerpopulationdensitiesovertime,andthedeploymentofRSUswillbecomesocial-welfare-enhancingovermoreofthecountry.However,therewillremainareaswheredeploymentofRSUsdoesnotenhancesocialwelfare,includingthoseruralareaswherepopulationdensityissolowthatcellularnetworksarenotcapacity-limited,i.e.theyhaveexcesscapacityanddon’tneedoffload.

RSUcostalsoaffectswhetherdeploymentofRSUswouldincreasesocialwelfare,andRSUcostvariesfromcommunitytocommunity.Forexample,allelsebeingequal,benefitsofInternetaccessthroughDSRCminusRSUcostswillbelowerwheretheproviderhastoacquireinfrastructure(poles,backhaul,etc.)thanwhereRSUsaredeployedbyamunicipalitythatalreadyhasinfrastructureavailable,orwherepartoftheRSUcostisincurredforanotherpurpose,e.g.agivenRSUissharedforsafetyandInternettraffic.

Likeanymodelofacomplexsystem,ouranalysisisbasedonanumberofsimplifyingassumptions,someofwhichwemayexplorefurtherinfutureresearch,suchasthevariabilityintrafficpervehicleandamongvehicletypes,andthedynamicsoftraffic,penetrationandcostsovertime.However,theconclusionthatbenefitexceedsRSUcostinurbanareasbutislowerthanthesumofRSU,spectrumandOBUcostsissufficientlyrobustthatasmallchangeofaround20%inanyoftheseassumptionswouldnotchangeit.Ifrealitydiffersfromthebasecaseevenmorethanthis,thisismostlikelyeitherbecauseofourassumptionaboutamandateorourassumptionaboutmobiletrafficlevels.TheU.S.Dept.ofTransportationhasnotmadeafinaldecisionaboutamandate,andwhateveritdecides,othercountriesmaydecidedifferently.IfOBUsarenotmandated,thenpenetrationcouldbelowerthanthebasecase.Moreover,ownersofvehicleswillopttopurchaseOBUsiftheirindividualbenefitexceedstheirindividualcosts.CarsthatwereequippedwithDSRCOBUsbecauseofthishighbenefitmaydifferfromcarsthatwereequippedwithDSRCduetoablanketmandate,anditisthelatterthatarebestreflectedinourbaselineassumptions.Thecomparisonbetweenbenefitsandcostswithoutamandateisasubjectforfuturework.

Theotherassumptionthatmayvarydramaticallyfromthebasecaseassumptionisthetrafficpervehicle.Throughputratefromvehiclesandothermobiledevicesobviouslydependsontheamountoftrafficflowingtoandfrommobiledevices.Whiletherehavebeenprominentpredictionsthatdataratesassociatedwithmobiledeviceswillincreaserapidlyandexponentially(Cisco2015),andproductsareemergingthatwouldgeneratethistraffic,theactualdemandisunknown.Ifdataratesaresubstantiallyhigherorlowerthanourbaselinesestimateof5GBpermonthpervehicle,thenthepopulationdensityrequiredforthebenefitofInternetaccessthroughDSRCtoexceedthecostofRSUsmaybemoreorlessthanourestimated4000peopleperkm2,respectively.

6 AcknowledgmentsThisworkissupportedundertheCMU-PortugalPartnership(scholarshipSFRH/BD/51564/2011),thePortugalFoundationfortheScienceandTechnology(ref.UID/EEA/50008/2013),andtheFutureCitiesProject(EuropeanCommissionEUFP7undergrantnumber316296).TheauthorsalsothanktheInstitutodeTelecomunicações–Porto,VeniamNetworks,themunicipalityofPortoandSTCP,forthedataandsupportprovided.

Page 34: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

34

7 ReferencesAlotaibi,Mohammed,JonM.Peha,andMarvinA.Sirbu.2015.“ImpactofSpectrum

AggregationTechnologyandSpectrumAllocationonCellularNetworkPerformance.”InIEEEConferenceonDynamicSpectrumAccessNetworks(DySPAN).Stockholm,Sweden.

Ameixieira,Carlos,AndréCardote,FilipeNeves,RuiMeireles,SusanaSargento,LuísCoelho,JoãoAfonso,etal.2014.“HarborNet:AReal-WorldTestbedforVehicularNetworks.”IEEECommunicationsMagazine(September):1–6.http://arxiv.org/abs/1312.1920.

Asplund,Henrik,AndrésAlayónGlazunov,AndreasFMolisch,KlausIPedersen,andMartinSteinbauer.2006.“TheCOST259DirectionalChannelModel–PartII:Macrocells.”IEEETransactionsonWirelessCommunications5(12):3434–3450.

AT&T.2015.“AT&TWi-Fi.”http://www.att.com/gen/general?pid=5949.

Bai,Fan,DanielD.Stancil,andHariharanKrishnan.2010.“TowardUnderstandingCharacteristicsofDedicatedShortRangeCommunications(DSRC)fromaPerspectiveofVehicularNetworkEngineers.”InProceedingsoftheSixteenthAnnualInternationalConferenceonMobileComputingandNetworking-MobiCom’10,329.NewYork,NewYork,USA:ACMPress.doi:10.1145/1859995.1860033.

Balasubramanian,Aruna,RatulMahajan,andArunVenkataramani.2010.“AugmentingMobile3GUsingWiFi.”InProceedingsofthe8thInternationalConferenceonMobileSystems,Applications,andServices-MobiSys’10,209.NewYork,NewYork,USA:ACMPress.doi:10.1145/1814433.1814456.

Balasubramanian,Aruna,RatulMahajan,ArunVenkataramani,BrianNeilLevine,andJohnZahorjan.2008.“InteractiveWifiConnectivityforMovingVehicles.”InACMSIGCOMMComputerCommunicationReview,38:427.doi:10.1145/1402946.1403006.http://portal.acm.org/citation.cfm?doid=1402946.1403006.

Boban,Mate,TiagoT.V.Vinhoza,MichelFerreira,JoaoBarros,andOzanK.Tonguz.2011.“ImpactofVehiclesasObstaclesinVehicularAdHocNetworks.”IEEEJournalonSelectedAreasinCommunications29(January):15–28.doi:10.1109/JSAC.2011.110103.

Bychkovsky,Vladimir,BretHull,AllenMiu,HariBalakrishnan,andSamuelMadden.2006.“AMeasurementStudyofVehicularInternetAccessUsingInSituWi-FiNetworksCategoriesandSubjectDescriptors.”InMobiCom’06Proceedingsofthe12thAnnualInternationalConferenceonMobileComputingandNetworking,50–61.ACM.doi:10.1145/1161089.1161097.

Calcev,George,DmitryChizhik,GBo,StevenHoward,HowardHuang,AchillesKogiantis,FAndreas,ArisLMoustakas,DougReed,andHaoXu.2007.“AWidebandSpatialChannelModelforSystem-WideSimulations.”IEEETransactionsonVehicular

Page 35: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

35

Technology56(2):1–15.

Campolo,Claudia,andAntonellaMolinaro.2013.“MultichannelCommunicationsinVehicularAdHocNetworks:ASurvey.”IEEECommunicationsMagazine51(5)(May):158–169.doi:10.1109/MCOM.2013.6515061.http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=6515061.

Cardote,Andre,FilipeNeves,SusanaSargento,andPeterSteenkiste.2012.“AStatisticalChannelModelforRealisticSimulationinVANET.”In2012IEEEVehicularNetworkingConference(VNC),48–55.Ieee.doi:10.1109/VNC.2012.6407444.

Cisco.2015.“CiscoVisualNetworkingIndex:GlobalMobileDataTrafficForecastUpdate,2014–2019.”

Clarke,RichardN.2014.“ExpandingMobileWirelessCapacity:TheChallengesPresentedbyTechnologyandEconomics.”TelecommunicationsPolicy(38)(December):693.doi:10.1016/j.telpol.2013.11.006.http://linkinghub.elsevier.com/retrieve/pii/S0308596113001900.

Comcast.2013.“ComcastUnveilsPlansforMillionsofXfinityWiFiHotspots.”PressRelease,June10.http://corporate.comcast.com/news-information/news-feed/comcast-unveils-plans-for-millions-of-xfinity-wifi-hotspots-through-its-home-based-neighborhood-hotspot-initiative-2.

———.2015.“XFINITYWiFi.”http://www.xfinity.com/wifi/default.htm.

DeutscheTelekom.2013.“ConnectedCarsGetBigDataRolling.”http://www.telekom.com/media/media-kits/179806.

Engebretson,Joan.2012.“ClearwireControlWouldFurtherStrengthenSoftbank/Sprint.”Telecompetitor.http://www.telecompetitor.com/clearwire-control-would-further-strengthen-softbank-sprint/.

Eriksson,Jakob,HariBalakrishnan,andSamuelMadden.2008.“Cabernet:VehicularContentDeliveryUsingWiFi.”InMobiCom’08Proceedingsofthe14thACMInternationalConferenceonMobileComputingandNetworking,199–210.ACMNewYork,NY,USA.doi:10.1145/1409944.1409968.

EuropeanCommission.2012.EUTransportinFigures-StatisticalPocketbook2012.doi:10.2832/52252.

FederalCommunicationsCommission.2010a.“ABroadbandNetworkCostModel.”

———.2010b.ConnectingAmerica:TheNationalBroadbandPlan.USA.doi:10.1002/yd.20038.http://download.broadband.gov/plan/national-broadband-plan.pdf.

Goldstein,Phil.2015.“In2015,HowMuchLTESpectrumDoVerizon,AT&T,T-MobileandSprintHave--andWhere?”FierceWireless.http://www.fiercewireless.com/special-

Page 36: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

36

reports/2015-how-much-lte-spectrum-do-verizon-att-t-mobile-and-sprint-have-and-wher?confirmation=123.

GoverningInstitute.2015.“PopulationDensityforU.S.CitiesMap.”Governing.AccessedAugust5.http://www.governing.com/gov-data/population-density-land-area-cities-map.html.

Hallahan,Ryan,andJonM.Peha.2009.“TheBusinessCaseofaNationwideWirelessNetworkThatServesBothPublicSafetyandCommercialSubscribers.”In37thTelecommunicationsPolicyResearchConference.

———.2011.“CompensatingCommercialCarriersforPublicSafetyUse:PricingOptionsandtheFinancialBenefitsandRisks.”In39thTPRCTelecommunicationsPolicyResearchConference.Washington,D.C.

Harding,John,GregoryPowell,RebeccaYoon,JoshuaFikentscher,CharleneDoyle,DanaSade,MikeLukuc,JimSimons,andJingWang.2014.“Vehicle-to-VehicleCommunications:ReadinessofV2VTechnologyforApplication.”WashingtonD.C.

IEEE.2010a.“802.11p-2010-IEEEStandardforInformationTechnology--LocalandMetropolitanAreaNetworks--SpecificRequirements--Part11:WirelessLANMediumAccessControl(MAC)andPhysicalLayer(PHY)SpecificationsAmendment6:WirelessAccessinVehicularEn.”IEEE.doi:10.1109/IEEESTD.2010.5514475.

———.2010b.IEEEStandardforWirelessAccessinVehicularEnvironments(WAVE)-NetworkingServices.Vol.2010.

INSEE.2013.“ParisRésuméStatistique.”http://www.statistiques-locales.insee.fr/Fiches%5CRS%5CDEP%5C75%5CCOM%5CRS_COM75056.pdf.

InstitutoNacionaldeEstatistica.2011.“StatisticsPortugal.”https://www.ine.pt/xportal/xmain?xpid=INE&xpgid=ine_indicadores&indOcorrCod=0006044&contexto=bd&selTab=tab2&xlang=en.

Kenney,JohnB.2011.DedicatedShort-RangeCommunications(DSRC)StandardsintheUnitedStates.ProceedingsoftheIEEE.Vol.99.doi:10.1109/JPROC.2011.2132790.http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=5888501.

Kerans,Andrew,DatVo,PhillipConder,andSnezanaKrusevac.2011.“PricingofSpectrumBasedonPhysicalCriteria.”InIEEEInternationalSymposiumonDynamicSpectrumAccessNetworks,DySPAN2011,223–230.IEEE.doi:10.1109/DYSPAN.2011.5936210.

Lacage,Mathieu,andThomasR.Henderson.2006.“YetAnotherNetworkSimulator.”InProceedingfromthe2006WorkshoponNs-2:TheIPNetworkSimulator-WNS2’06,12.NewYork,NewYork,USA:ACMPress.doi:10.1145/1190455.1190467.

Lee,Joohyun,YungYi,SongChong,andYoungmiJin.2014.“EconomicsofWiFiOffloading:TradingDelayforCellularCapacity.”IEEETransactionsonWirelessCommunications13

Page 37: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

37

(3).

Lee,Kyunghan,InjongRhee,JoohyunLee,SongChong,andYungYi.2010.“MobileDataOffloading:HowMuchCanWiFiDeliver?”InThe6thInternationalConferenceonEmergingNetworkingEXperimentsandTechnologies(CoNEXT),1.ACMPress.doi:10.1145/1921168.1921203.

Li,Fan,andYuWang.2007.“RoutinginVehicularAdHocNetworks:ASurvey.”IEEEVehicularTechnologyMagazine.doi:10.1109/MVT.2007.912927.

Markendahl,Jan.2011.“MobileNetworkOperatorsandCooperation:ATele-EconomicStudyofInfrastructureSharingandMobilePaymentServices.”KTH.

Markendahl,Jan,andÖstenMäkitalo.2010.“AComparativeStudyofDeploymentOptions,CapacityandCostStructureforMacrocellularandFemtocellNetworks.”InIEEEInternationalSymposiumonPersonal,IndoorandMobileRadioCommunications,PIMRC,145–150.doi:10.1109/PIMRCW.2010.5670351.

Mecklenbrauker,ChristophF.,AndreasF.Molisch,JohanKaredal,FredrikTufvesson,AlexanderPaier,LauraBernadó,ThomasZemen,OliverKlemp,andNicolaiCzink.2011.“VehicularChannelCharacterizationandItsImplicationsforWirelessSystemDesignandPerformance.”ProceedingsoftheIEEE99(7).

Meinilä,Juha,PekkaKyösti,TommiJämsä,andLassiHentilä.2009.“WINNERIIChannelModels.”InRadioTechnologiesandConceptsforIMT-Advanced,editedbyMartinDöttling,WernerMohr,andAfifOsseiran.Wiley.

Meireles,Rui.2015.“LeveragingDiversityandSpatialConnectivityinMulti-HopVehicularNetworks.”CarnegieMellonUniversity,UniversityofPorto.

Meireles,Rui,PeterSteenkiste,andJoaoBarros.2012.“DAZL:Density-AwareZone-BasedPacketForwardinginVehicularNetworks.”InIEEEVehicularNetworkingConference,VNC,234–241.doi:10.1109/VNC.2012.6407437.

Moore,AndrewW.2001.“K-MeansandHierarchicalClustering.”http://www-2.cs.cmu.edu/~awm/tutorials/kmeans.html.

Murray,David,MichaelDixon,andTerryKoziniec.2007.“ScanningDelaysin802.11Networks.”InThe2007InternationalConferenceonNextGenerationMobileApplications,ServicesandTechnologies(NGMAST2007),2–7.IEEE.

Mussabbir,QaziBouland,andWenbingYao.2007.“OptimizedFMIPv6UsingIEEE802.21MIHServicesinVehicularNetworks.”IEEETransactionsonVehicularTechnology56(6):3397–3407.

Newman,Stagg.2008.“TestimonybeforetheFCCEnBancHearing,PublicSafetyInteroperableCommunicationsandthe700MHzDBlockProceeding,July30,2008.”https://transition.fcc.gov/realaudio/presentations/2008/073008/newman.pdf.

Page 38: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

38

“Ns-3NetworkSimulator.”2015.AccessedJuly10.https://www.nsnam.org/.

OfficeofManagementandBudget.1992.“CircularNo.A-94Revised|TheWhiteHouse.”TheWhiteHouse.https://www.whitehouse.gov/omb/circulars_a094/.

Peha,JonM.2013.“CellularCompetitionandtheWeightedSpectrumScreen.”TPRC41ResearchConferenceonCommunication,InformationandInternetSecurity:1–23.doi:http://dx.doi.org/10.2139/ssrn.2241276.

Puchalsky,Christopher.2005.“ComparisonofEmissionsfromLightRailTransitandBusRapidTransit.”TransportationResearchRecord:JournaloftheTransportationResearchBoard.

Rao,Ashwin,ArnaudLegout,Yeon-supLim,DonTowsley,ChadiBarakat,andWalidDabbous.2011.“NetworkCharacteristicsofVideoStreamingTraffic.”InProceedingsoftheSeventhConferenceonEmergingNetworkingExperimentsandTechnologies-CoNEXT’11,1–12.NewYork,NewYork,USA:ACMPress.doi:10.1145/2079296.2079321.http://dl.acm.org/citation.cfm?doid=2079296.2079321.

Sandvine.2014.“GlobalInternetPhenomenaReport1H2014.”

Santos,Adella,NancyMcGuckin,HikariYukikoNakamoto,DanielleGray,andSusanLiss.2011.“SummaryofTravelTrends:2009NationalHouseholdTravelSurvey.”

Sesia,S,IToufik,andMBaker.2011.LTE-TheUMTSLongTermEvolution:FromTheorytoPractice.2nded.Wiley.

Sheikh,MuhammadUsman.2014.“AspectsofCapacityEnhancementTechniquesinCellularNetworks.”TampereUniversityofTechnology.

SprintNextelCorporation.2011.“PetitiontoDeny.”

Tan,Nicholas,andJonM.Peha.2015.“MeasuresofSpectrumHoldingsThatReflectMarketShareandConcentrationAmongCellularCarriers.”In43rdTelecommunicationsPolicyResearchConference(TPRC).Arlington,VA.

TokyoMetropolitanGovernment.2014.“TokyoStatisticalData.”toukei.metro.tokyo.jp/tnenkan/2008/tn08qa020400.xls.

U.S.DepartmentofTransportation.2015.“PlanningfortheFutureofTransportation:ConnectedVehiclesandITS.”doi:10.1016/S0001-2092(07)69737-3.http://www.its.dot.gov/press/2015/its_future_cv.htm.

U.S.FederalCommunicationsCommission.CodeofFederalRegulations,Title47,Part90§90.371DedicatedShortRangeCommunicationsService.

UKOfficeforNationalStatistics.2012.“RegionalProfiles:KeyStatistics-London.”http://ons.gov.uk/ons/rel/regional-trends/region-and-country-profiles/key-

Page 39: Comparison between Benefits and Costs of Offload of Mobile ...users.ece.cmu.edu/~peha/Vehicular_Network_Offloads... · data collected from an actual vehicular network that is operating

39

statistics-and-profiles---august-2012/key-statistics---london--august-2012.html.

UnitedStatesCensusBureau.2015.“QuickFacts-UnitedStates.”http://www.census.gov/quickfacts/table/PST045214/00.

Uzcategui,R.,andG.Acosta-Marum.2009.“Wave:ATutorial.”IEEECommunicationsMagazine47(5)(May):126–133.doi:10.1109/MCOM.2009.4939288.

Wang,Zhe,andMahbubHassan.2008.“HowMuchofDsrcIsAvailableforNon-SafetyUse?”InProceedingsoftheFifthACMInternationalWorkshoponVehiculArInter-NETworking-VANET’08,23.NewYork,NewYork,USA:ACMPress.doi:10.1145/1410043.1410049.http://portal.acm.org/citation.cfm?doid=1410043.1410049.

Wannstrom,Jeanette,andKeithMallinson.2014.“HetNet/SmallCells.”3GPP.http://www.3gpp.org/hetnet.

Wile,Rob.2014.“TheAverageAgeofUSVehiclesIsataRecordHighof11.4Years.”BusinessInsider.http://www.businessinsider.com/the-average-age-of-us-vehicles-stays-at-record-high-of-114-years-2014-6.

Wisitpongphan,Nawaporn,FanBai,PriyanthaMudalige,VarshaSadekar,andOzanK.Tonguz.2007.“RoutinginSparseVehicularAdHocWirelessNetworks.”IEEEJournalonSelectedAreasinCommunications25(8)(October).doi:10.1109/JSAC.2007.071005.

Wright,James,J.KyleGarrett,ChristopherJ.Hill,GregoryD.Krueger,JulieH.Evans,ScottAndrews,ChristopherK.Wilson,RajatRajbhandari,andBrianBurkhard.2014.“NationalConnectedVehicleFieldInfrastructureFootprintAnalysis.”USDOT-FHWA.doi:DTFH61-11-D-00008.

Zambelli,Alex.2009.“IISSmoothStreamingTechnicalOverview.”MicrosoftCorporation(March).

Zeadally,Sherali,RayHunt,Yuh-ShyanChen,AngelaIrwin,andAamirHassan.2010.“VehicularAdHocNetworks(VANETS):Status,Results,andChallenges.”TelecommunicationSystems(December9).doi:10.1007/s11235-010-9400-5.

Zhao,Xiongwen,T.Rautiainen,K.Kalliola,andP.Vainikainen.2006.“Path-LossModelsforUrbanMicrocellsat5.3GHz.”IEEEAntennasandWirelessPropagationLetters5(1)(December):152–154.doi:10.1109/LAWP.2006.873950.http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=1624453.