DNSwww.cnnic.cn

0018-9545(c)2016IEEE.
Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seehttp://www.
ieee.
org/publications_standards/publications/rights/index.
htmlformoreinformation.
Thisarticlehasbeenacceptedforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.
Contentmaychangepriortofinalpublication.
Citationinformation:DOI10.
1109/TVT.
2016.
2581020,IEEETransactionsonVehicularTechnologyAbstract—Withthedevelopmentofmodernhigh-speedvehiclesandmobilecommunicationsystems,thereisastrongdemandforoperatorstoprovidestableandhigh-qualityonboardInternetservices.
However,thevehicle-to-groundlinksmaysufferfromseveralproblems,suchasinsufficientbandwidthandlonground-triptime.
Althoughmultiplewirelessnetworks,suchasWiFi,3Gand4G,maybeavailablealongatrack,onboardusershardlyhavegoodexperiencesiftheycanonlyvisittheInternetviaonewirelessnetwork,whichmaybesubjecttocoveragegapsandsignalattenuation.
Inthispaper,thelinkqualitiesofmultipleexisting3Gand4Gtechnologiesarefirstmeasuredinatypicalhigh-speedenvironment.
These3G/4Gnetworksarecandidatesforvehicle-to-groundcommunication.
Then,aconcurrentmultipathtransmissionschemetogetherwithanetworkadaptiveschedulingalgorithmisproposed.
Theschemeisindependentoftheprotocolstacksothatitiseasytodeploy.
ItalsoprovidestransparentInternetservicesforuserswithoutrequiringtheparticipationoftheuserdevicesinanymultipathsignaling.
Meanwhile,theschedulingalgorithmworksatthenetworklayer,insteadofthetransportlayer,tomeetthediversetransmissionrequirementsoftheconnection-orientedandconnectionlessuserapplications.
Thealgorithmcaneffectivelyaggregatethebandwidthofmultipleavailablewirelesslinks,aswellasavoidthereorderingofpacketsbasedonboththepracticaltracingdatabasesandactivepathmonitoring.
AnalysisandexperimentsshowthattheproposedalgorithmcanprovidebetteronboardInternetserviceswithlowercacherequirementsthantheEarliestDeliveryPathFirst(EDPF)andWeightedRoundRobinschedulingalgorithms,intermsofbandwidthimprovementandpacketdisorderreduction.
IndexTerms—mobileInternet,vehicles,schedule,multiplepathsCopyright(c)2015IEEE.
Personaluseofthismaterialispermitted.
However,permissiontousethismaterialforanyotherpurposesmustbeobtainedfromtheIEEEbysendingarequesttopubs-permissions@ieee.
org.
ThisworkwassupportedinpartbytheFundamentalResearchFundsfortheCentralUniversitiesunderGrant2014JBM004and2015JBM001,bytheNationalNaturalScienceFoundationofChinaunderGrant61271202,bytheNational973ProgramofChinaunderGrant2013CB329100,bytheUSNationalScienceFoundationunderGrantCNS-1065444,aswellasbytheUSArmyResearchOfficeunderGrantWF911NF-14-1-0518.
PingDongiswithBeijingJiaotongUniversity,Beijing100044,China(e-mail:pdong@bjtu.
edu.
cn).
BinSongiswithXidianUniversity,Xi'an710071,China(e-mail:bsong@mail.
xidian.
edu.
cn).
(Correspondingauthor)HongkeZhangiswithBeijingJiaotongUniversity,Beijing100044,China(e-mail:hkzhang@bjtu.
edu.
cn).
XiaojiangDuiswithTempleUniversity,Philadelphia,PA19122,USA(e-mail:xjdu@temple.
edu).
I.
INTRODUCTIONToday,withtherapiddevelopmentofmodernhigh-speedvehicles,providingstableandhigh-qualitymobileInternetservicesonboardisbecomingarisingandinspiringtrendforbothpassengersandoperators.
Vehicleswithincreasedspeeds,suchasHigh-SpeedTrains(HSTs)thatexceed300km/h,arebecomingpopularamonglong-distancetravelersbecauseoftheirconvenient,stable,andrelativelyspaciousenvironment.
Nowadays,China,Japan,USA,Germany,andFrancearerapidlydeployingtheirnationalhigh-speedrailnetworks[1].
Accordingtoanofficialstatistic[2]publishedbytheNationalRailwayAdministrationinApril2015,thenumberofpassengershasreached2.
357billionin2014inChina.
InUSA,aplan[3]fora17,000milenationalhigh-speedrailsystemhasbeenoutlinedforcompletionby2030.
Meanwhile,withtheboomofmobilecommunicationtechnologies,suchasWiFiandLong-TermEvolution(LTE),peoplehavebecomeaccustomedtousingvariousmobileapplications.
Areport[4]fromtheChinaInternetNetworkInformationCentershowsthatattheendof2014,theamountofmobilenetworkusershasreached557million,and85.
8%ofthemobileusersvisittheInternetbytheirsmartmobileterminals.
AreportfromWiFiAlliance,thecertificationauthorityofWiFidevices,showsthatabout4.
5billionWiFiproductsareinuseallovertheworldtoday[5].
Therefore,thereisastrongdemandtoprovidehigh-qualityonboardInternetservicestothepassengersonhigh-speedvehicles.
However,thereareseveralspecialissuesassociatedwithhigh-speedmobility[6]suchas:highpenetrationlossesofsignals,severeDopplershift,andfrequenthandoffs.
Alloftheseissuesarefacedbythemobilenetworksystemsforthehigh-speedvehicles.
TheseproblemsmayleadtotheinstabilityofpopularInternetservicesandresultinpoorQualityofService(QoS).
Althoughmultiplewirelessnetworks,suchasWiFi,3G,4G,maybeavailablealongatrack,unfortunately,duetothelimitationsofuserdevices,onboardusershardlyhavegoodexperiencesiftheirdevicescanonlyvisittheInternetviaonewirelessnetwork(e.
g.
,laptopswithWiFi,whichmaybesubjecttocoveragegapsandsignalattenuation).
Tosolvetheseproblems,someproposedonboardnetworkschemesdesignedmobilerouterswithmultipleinterfaces.
TheseinterfacesmaysupportvariouswirelesstechnologiesImprovingOnboardInternetServicesforHigh-SpeedVehiclesbyMultipathTransmissioninHeterogeneousWirelessNetworksPingDong,Member,IEEE,BinSong*,HongkeZhang,XiaojiangDu,SeniorMember,IEEE0018-9545(c)2016IEEE.
Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seehttp://www.
ieee.
org/publications_standards/publications/rights/index.
htmlformoreinformation.
Thisarticlehasbeenacceptedforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.
Contentmaychangepriortofinalpublication.
Citationinformation:DOI10.
1109/TVT.
2016.
2581020,IEEETransactionsonVehicularTechnology(e.
g.
,WiFi,cellularnetworkandsatellitetechnology).
Seamlesshandover[7]extendsNEMO[8]byusingmultipleinterfacesofthesametechnology(LTE)tosupporthigh-speedmobility.
Dynamicnetworkselection[9]wasproposedtoswitchfromapoorperformancenetworktoabetterone.
However,mostoftheresearchinthisdomainhasbeenconfinedtosingleinterfaceusageatanygiventime,whileotherinterfacesworkasbackupsorassistancesforfasthandover.
Theuseofmultiplewirelessinterfacessimultaneouslyopensanewwayofsolvingsomeproblemsinhigh-speedscenariosandcanbringsomenewandinterestingopportunities,suchasbandwidthaggregation,seamlesshandover,andreliability.
Aconcurrentmultipathtransmissionschemecansatisfythedesireofoperatorstoachievestableandhigh-qualityvehicle-to-groundcommunications.
Asurveyfoundthat78%ofthebusinesstravelersaskedsaidtheywoulduseWirelessLocalAreaNetwork(WLAN)ifitwasavailable[10].
However,untilrecently,passengersonhigh-speedvehicles,suchasbullettrains,havenotbeenabletoachievestablehigh-speedInternetservices.
Themultipathtransmissionschemecanprovideafaster,morereliableInternetserviceforonboardusers.
Italsoallowsthevehicleoperatorstodeliverinformationtothepassengers.
Inadditiontothesebenefits,broadbandaccessonvehiclesisalsoimperativeinvehiclecontrolandsafetybyallowinganoperationscentertocarryoutfaultdiagnostics,videosurveillance,andsoon.
Inthispaper,weintroduceaschemeofconcurrentmultipathtransmissionbasedonmultipleinterfacesforhigh-speedmobility.
Anetworkadaptivemultipathschedulingalgorithmisalsoproposed.
Ourmaincontributionsareasfollows:(1)Somepracticalexperimentswereperformedtoevaluatethecharacteristicsofmultiple3G/4Gnetworks,i.
e.
,EV-DO,FDD-LTE/HSPA+andTD-LTE,underdifferentstatuses(staticandhigh-speedmobility).
(2)Amultipathtransmissionarchitecture,togetherwithanetworkadaptivemultipathschedulingalgorithm,isproposed,whichprovidestransparentInternetservicesfortheusersthroughmultiplewirelesstechnologieswithoutrequiringtheparticipationoftheuserdevicesinanymultipathsignaling.
(3)Theproposedalgorithmcanadapttothechangingofpathconditions,andachievegoodbandwidthaggregationwhileavoidingthereorderofpacketswithlowercacherequirementthanEarliestDeliveryPathFirst(EDPF)andWeightedRoundRobinschedulingalgorithms.
Theremainderofthispaperisorganizedasfollows.
SomerelatedworkisintroducedinSectionII.
Theresultsofourpracticalexperimentsonhigh-speedtrainsareshowninSectionIII.
Amultipathtransmissionarchitecture,togetherwithanetworkadaptivemultipathschedulingalgorithm,isproposedinSectionIV.
SomeofthepropertiesofouralgorithmareanalyzedinSectionV.
SimulationsandexperimentalevaluationsareprovidedinSectionVIfollowedbytheconclusionandAppendixAwherealemmaisprovenindetail.
II.
RELATEDWORKSomecontributionswereproposedtotakeadvantageofmultipleinterfacesandmultiplepaths.
Theyworkatdifferentprotocollayersandrequiredifferentchangestodifferentnetworkelements.
Webrieflydiscussthesolutionsrelatedtohigh-speedmobilityinthissection.
A.
Link-layerSolutionsAschemebasedonmultipleinterfaceswasadoptedearlyforvehicle-to-groundcommunicationbyFastTrainHosts(FIFTH)project[11]in2003.
ItprovidesInternetservicesfortrainsbysatellitesandfillsthecoveragegapsbyWiFi.
However,theexpensiveprice,limitedbandwidthandlongpropagationdelaylimitthelarge-scaleapplicationofasatellite-basedapproach.
Comparedtotheliterature'swork,ourmethodemploysmultiplewirelesstechnologies,andselectsall(orsome)oftheinterfacestoworksimultaneouslyforvehicle-to-groundcommunication.
WiththepopularityofWLANtechnology,alink-layerdesigncalledinformationraining[12]wasproposedin2005,whichusesanumberofrepeatersplacedalongatrackandmultipleantennasinstalledontheroofofavehicletoenhancethethroughputonboard.
Anoptimalantennaassignmentstrategy[13]wasproposedforemployingmaximumratiocombining.
AlongwiththeriseofWirelessMetropolitanAreaNetworks(WirelessMAN),theideaofinformationrainingwasextendedin[14]tosupportnetworkcodingthroughIEEE802.
16j.
However,itwillproduceahighdeploymentcosttoprovideseamlesscoveragebytheinformationraininginfostationsalongatrack.
Otherwise,coveragegapswillmaketheresourceallocationchallenging.
Incontrast,ourmethodfocusesonmakingfulluseofthevariousexistinginfrastructurestoimprovetheonboardInternetservices.
J.
Zhang,etal[15]proposedamulti-system-basedaccessarchitectureforhigh-speedtrains.
ThearchitectureintroducesAccessServiceGateways(ASGs)inagroundnetworktosupportmultiplewirelesssystems.
ItalsodeploysRadioAntennaUnits(RAUs)alongarailtrackthatserveasaground-to-trainnetwork.
However,theproposedarchitecturewillproduceahighdeploymentcostforboththegroundnetworkandtheground-to-trainnetwork.
Incontrast,ourmethoddoesnotrequirethedeploymentofnewinfrastructures.
Recently,withtherapiddevelopmentof4Gand5Gtechniques,whichmayprovidehigherbandwidthandlowertransmissionlatencyatlesscost,thecellular-basedsolutionsaregeneratingmoreinterest.
MEN-NEMO[7]proposedanLTEfemtocell-basedschemethatusesmultipleinterfacestosupportseamlessmobilityforhigh-speedrailsystems.
DifferentfromMEN-NEMO,ourmethoddoesnotrequireanymodificationsoftheexistingLTEnetworkssothatitiseasytodeploy.
Overall,toachievemultipathtransmissioninheterogeneouswirelessnetworksforhigh-speedvehicles,link-layerapproachesareinfeasiblebecausethevariousnetworksinquestionmaybelongtodifferentproviders.
Comparedtothelink-layerliteratures,ourschemeworksatthenetworklayersothatitcanavoidthecomplexinterworkingatthelinklayerandtakeadvantageofavarietyofnetworkresourcesmanagedbydifferentoperators.
B.
Network-layerSolutions0018-9545(c)2016IEEE.
Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seehttp://www.
ieee.
org/publications_standards/publications/rights/index.
htmlformoreinformation.
Thisarticlehasbeenacceptedforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.
Contentmaychangepriortofinalpublication.
Citationinformation:DOI10.
1109/TVT.
2016.
2581020,IEEETransactionsonVehicularTechnologySeveralstandardizedprotocolsworkingatthenetworklayer(or3.
5Layer)cansupportmulti-homing,suchasHIP[16],Shim6[17]andLISP[18].
Theseprotocolsarelocator/identifierseparationsolutions,whichenablecontinuityofcommunicationsacrossIPaddresschangesinmobilescenarios.
However,theyarenotdesignedforconcurrentmultipathtransmission,norsuitableforfast-movingvehicles.
Incontrast,weaimatconcurrentmultipathtransmissionforhigh-speedvehiclesinheterogeneousnetworks.
IETFMultipleInterfaces(MIF)workinggroupisnowworkingonthestandards[19]fornodeswithmultipleinterfaces.
However,itmainlyfocusesontheconflictingconfigurationproblemsassociatedwithdifferentnetworks,suchasDNSserverandHTTPproxy,butnotthemultipathtransmissionissues.
MAR[20]introducedawirelessmulti-homeddevicesimilartoourmobileroutertoexploitwirelessdiversity.
However,differentfromMAR,ourmobilerouterisnotconfiguredastheuserdevices'defaultDNSserverthatmayoverlycomplicatethemobilerouter'sdesign.
Moreover,ourmobilerouterdoesnotworkasasource-NATboxforuserpackets,instead,userpacketsareroutedtoanaccessrouterthroughlogictunnels.
Inaddition,withoutdesigningorspecifyinganyschedulingalgorithm,MARisassumedtorelyoncustom-builtschedulingprotocols,whilewedesignaschedulingalgorithmadaptingtowirelesspaths.
AnIP-in-IPencapsulationmethod[21][22]wasproposedtoprovidebandwidthaggregationbysplittingadataflowacrossmultipleinterfacesattheIPlevel.
However,itsimplyassumesthattheeffectivebandwidthofapathisconstantthroughoutthelifetimeofaTCPconnection.
Italsoassumesallthepacketssentalongapathhavethesamesize.
Theseassumptionsdonotholdinhigh-speedmobilityscenarios.
EDPF[23]alsousesIP-in-IPencapsulationtotunnelpacketsandproposesabandwidthaggregationschemeforthereal-timeapplications.
Itschedulesthepacketsbasedontheirestimateddeliverytime.
However,EDPFignoresthewirelesspropagationdelay,instead,theyonlyconsiderthewirelinedelayonthepathsfromtheproxytothebasestations,whichisnotaccurateinhigh-speedmobilityscenarios.
Incontrast,formoreaccurateschedulinginhigh-speedmobilityscenarios,ouralgorithmconsidersthetotaldelayonthepathsfromthemobileroutertotheaccessrouter,andtakesthevariablepathdelaysintoconsiderationwhenshowingthepropertiesofouralgorithm.
Furthermore,ouralgorithmemploysreschedulingmechanisms(alsocalledcompensationmechanisms)thatautomaticallyadapttotheunexpectedchangesoflinkparameters.
Severalliteratures[24][25]haveshownthatthehistoricaltracingdatacanaccuratelyforecastthewirelessnetworkcapabilitiesandisofgreatvalueforintelligentscheduling.
BreadCrumbs[24]predictsthechangesofnetworkswithpastobservationsofwirelessnetworkcapabilities.
Itevaluatedtheefficacyoftheforecastswithseveralweeksofreal-worldusage,andfoundthattheschemecanprovidesignificantlyimprovedperformance.
YJun,etal[25]foundthatthepastbandwidthinformationisagoodindicatoroftheactualbandwidthexperiencedatagivenlocation.
Theystoredpastbandwidthdataintheformofbandwidthmapsanddemonstratedtheusefulnessofthesemapswithtwocasestudies(i.
e.
,adaptivemultimediaandmobilevehicularInternet).
Resultsshowedthattheuseofpastlocation-specificbandwidthknowledgecansignificantlyimprovetheQoSofmultimediaapplicationsinhigh-speedmobility.
However,weightedroundrobinschedulingisadoptedby[25],whichmayresultinthedisorderofpackets.
Inaddition,nomethodisdesignedtodealwithunexpectedchannelfluctuations.
Incontrast,ourschedulingalgorithmselectsthepaththatcandeliverapacketfastest.
Reschedulingmechanismsarealsodesignedtodealwiththeunexpectedchannelfluctuations.
Tomeetwiththespecialrequirementsinhigh-speedmobilityscenarios,weextendEDPFbyaddingtheparameterofwirelesstransmissiondelayintoouralgorithm.
Inaddition,ouralgorithmautomaticallyadaptstothepathstatus,intermsofbandwidthanddelay,accordingtobothhistoricaltracingdatabasesandactivepathmonitoring.
Moreover,areschedulingmechanismisdesignedtodealwithunexpectedchannelfluctuations.
C.
Transport-layerSolutionsAlthoughstandardizedSCTP[26][27]supportsmulti-homing,itdoesnottransmitdataovermultiplepathssimultaneously;instead,itselectsaprimarydatatransmissionpathwithenablingtransparentfail-overbetweenredundantnetworkpaths.
Someapproachesextendthemulti-homingcapabilityofSCTPtoaggregatingthebandwidthofmultiplepaths,includingWiMP-SCTP[28]andcmpSCTP[29].
However,mostofthesenon-standardizedextensionsofSCTPrequiresignificantlychangingtheprotocolstacksofbothendpoints,whichmakesitdifficultforthemtobeacceptedandrapidlydeployed.
Inaddition,theexistenceofvarioustypesofmiddleboxesgreatlyconstrainsthedeployabilityofSCTP.
Incontrast,ourmethodisindependentoftheprotocolstackandtransparenttothetransport-layerprotocols.
Recently,amultipathextensiontoTCP,namedMPTCP[30],hasbeenstandardizedtotransmitdataonmultiplepathssimultaneously.
Asdeclaredbyitsarchitecturalguidelines[31],MPTCPneedstomeetthegoalsofimprovedreliabilityandthroughput.
Itsperformancebenefitswereverifiedin[32],andtheresultsshowthatMPTCPachievesbothgoodthroughputandfairness.
However,toachievemultipathtransmissionforhigh-speedvehicles,MPTCPmaynotbeasuitablesolutionforthefollowingreasons:1)Variousapplicationsondifferentuserterminalssharethemultiplevehicle-to-groundpaths.
Someoftheapplications,however,arenotbasedontheconnection-orientedserviceprovidedbyMPTCP,suchasDomainNameSystem(DNS)andmostvoice/videoservices.
Thus,itisunreasonabletoselectMPTCPasthemultipathschemesinceallapplicationswillbeforcedtouseaconnection-orientedservice.
2)Theretransmissiontimermayexpirewhenthereisahighpathdelay.
Asaresult,MTCPusuallyreducesitscongestionwindowto1andentersaslowtransmissionphase.
Consequently,allthepacketsofallapplicationswillbeblockedbecausetheysharethevehicle-to-groundpaths.
Incontrast,anetwork-layerschemeis0018-9545(c)2016IEEE.
Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seehttp://www.
ieee.
org/publications_standards/publications/rights/index.
htmlformoreinformation.
Thisarticlehasbeenacceptedforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.
Contentmaychangepriortofinalpublication.
Citationinformation:DOI10.
1109/TVT.
2016.
2581020,IEEETransactionsonVehicularTechnologymoreflexiblebecauseitsbesteffortdeliveryservicecanfullyutilizethelimitedresourcesofthevehicle-to-groundpaths.
3)Inpracticethough,theexistenceofvarioustypesofmiddleboxesgreatlyconstrainsthedeployabilityofMPTCP[33],especiallyinmobilityscenarioswherethemiddleboxesonthevehicle-to-groundpathsvarywiththelocationofamovingvehicle.
Thus,weemployanetwork-layermultipathtransmissionscheme,whichismoreflexibleforuserapplicationstochoosetheirowntransport-layerprotocols.
MPTCPthroughputisknowntodecreasesignificantlyunderthecircumstancesofpathheterogeneity,especiallywhentherearesomebottleneckpaths[33][34].
Thus,severalalgorithmswereproposedrecentlytoimproveMPTCP'sperformance.
Thesealgorithmscanberoughlydividedintotwocategories:1)AlgorithmscloselyrelatedtoMPTCP,whichimprovetheMPTCPcongestioncontrolscheme.
2)Moregeneralalgorithms,whicharenotlimitedtouseintransportlayer.
Thefirstcategoryincludescongestionwindowadaptionalgorithm[34]andfairnessdesignwithcongestionbalancing[35].
TheytrytosolvethethroughputdecreaseproblemofMPTCP,buttheyarenotaimedatthehigh-speedmobilityscenario.
TheywillnotbefurtherdiscussedbecauseMPTCPitselfisnotselectedtobethevehicle-to-groundsolutionasdiscussedbefore.
ThesecondcategoryincludesRoundRobin(RR)schedulingalgorithmanditsvariants,suchasLoadSharingforSCTP(LS-SCTP)[36]andDelay-AwarePacketScheduling(DAPS)[37].
LS-SCTP[36]supportsweightedroundrobinandassignsdatatoeachpathaccordingtotheratio(congestionwindow/roundtriptime).
DAPS[37]distributesdatatotwodifferentpathsdependingontheratioand.
However,anidealratioishardtoachievebecausetheratioislimitedbythesize.
Nopacketcanbedeliveredonapathiftheisfull,evenifthepathmaydeliverthesubsequentpacketearliesttothedestination.
Moreover,largevariationofend-to-enddelaywillmakeLS-SCTPandDAPSfragile.
Incontrast,weemployanetwork-layerschedulingalgorithmthatcanadjustthesizeofsendingbuffersasneeded.
Moreover,wedesignareschedulingmechanismtodealwithunexpecteddelayfluctuations.
III.
MEASUREMENTSOF3G/4GWIRELESSNETWORKCHARACTERISTICSINHIGH-SPEEDSCENARIOFig.
1showstheexperimentalenvironment.
TwotypesofourmobileroutersareshowninFig.
2.
Wemeasuredthenetworkcharacteristicsof3Gand4Gtechnologiesbyestablishingcommunicationsbetweentheclients,themobilerouter,andtheserver[38].
InourexperimentsshowninFig.
1,WiFiwasusedinsideofatrainfordatatransmissionbetweentheuserdevicesandthemobilerouter,while3G/4Gtechnologieswereusedforvehicle-to-groundcommunications.
WiFicommunicationinsideofthetrainmaysufferfromperformancedegradationwhentheWiFilinkissharedbymultipleusers.
ThisproblemismainlycausedbythecollisionavoidanceschemeofWiFi,butitisbeyondthescopeofthispapersincewefocusonimprovingvehicle-to-groundStaticinLabCampusNetworkWiFiMobileRouterHigh-SpeedTrainChinaUnicomFDD-LTE/HSPA+ChinaTelecomEVDOServerinouruniversityClient1Client2ChinaMobileTD-LTEFig.
1.
ExperimentaltopologyFig.
2.
Mobilerouterscommunicationbymakingfulluseofvariousexistinginfrastructures.
Asacandidateforvehicle-to-groundcommunication,WiFicanprovidehighbandwidth.
However,WiFiistypicallyusedasgapfillersduetothehandoverissuesandhighdeploymentcosts[6].
Inpractice,WiFiismoresuitabletobedeployedatthestationswherevehiclesslowdown,ratherthantheroadsidewherevehiclesareunderhigh-speedmobility.
OurmobilerouterhasbothexternalWiFiand3G/4Ginterfacesforvehicle-to-groundcommunication,andcanutilizealltheseinterfacessimultaneouslybasedontheproposedmethodwhenatrainisnearastation.
However,noWiFiAPcouldbeaccessedbytheexternalWiFiinterfaceswhenthetrainwasunderhigh-speedmobility.
Thus,only3G/4Gnetworkcharacteristicswerecomparedbetweenstationaryandhigh-speedstatus,asshowninthefollowing.
TherearethreemajormobilenetworkprovidersinChina,i.
e.
,ChinaTelecom,ChinaUnicomandChinaMobile.
Today,EV-DO(3G),FDD-LTE/HSPA+(4G/3G)andTD-LTE(4G)arethemainmobiletechnologiesadoptedbytheirnetworks,respectively.
Weevaluatedallthesetechnologiesinourexperiments.
Themobilerouter,inwhichpluggedmultipleMiniPCI-E3G/4Gcards,wasplacedonahigh-speedtrainandcouldvisittheInternetthroughmultiplenetworkssimultaneously.
Inparticular,themobilerouterinstalledUbuntu12.
04LTSwithLinuxkernel3.
6.
0.
WvDialwasusedtomake3G/4GconnectionstotheInternet.
All3G/4Gcardswereattheirdefaultconfigurationsbecauseweaimedatmakingfulluseoftheexistinginfrastructures.
Iptrafwasusedtomonitorthenetworkparameters.
Forproducinghistoricaltracingdatabases,anextensiontoIptrafwasprogrammedtosupportMySQL.
NotethatforChinaUnicom,thecardmaychangeitsmodebetweenFDD-LTEandHSPA+automaticallyduetotheincompletecoverageofFDD-LTE.
EachinterfacewasallocatedIPaddressesdynamicallywhenthetrainmovedalong0018-9545(c)2016IEEE.
Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seehttp://www.
ieee.
org/publications_standards/publications/rights/index.
htmlformoreinformation.
Thisarticlehasbeenacceptedforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.
Contentmaychangepriortofinalpublication.
Citationinformation:DOI10.
1109/TVT.
2016.
2581020,IEEETransactionsonVehicularTechnologyFig.
3.
RSSIinastaticstateFig.
4.
RSSIonahigh-speedtrainFig.
5.
RTTinastaticstateFig.
6.
RTTonahigh-speedtrain86%12%2%1Mb/s14%38%16%13%11%8%2.
5Mb/s33%16%5%1%27%6Mb/sEV-DOFDD-LTE/HSPA+TD-LTEFig.
7.
Bandwidthdistributiononahigh-speedtraintherail.
TheserverwasplacedintheInformationCenterofBeijingJiaotongUniversityandcouldvisittheInternetthroughthecampusnetwork.
Duringtheexperiments,werecordedtheReceivedSignalStrengthIndicator(RSSI),Round-TripTime(RTT)andbandwidthofdifferent3G/4Gnetworks.
Forcomparison,weperformedtheexperimentsbothinourlaboratoryinastaticstateandonahigh-speedtrainwiththespeedof300km/h.
TheRSSIswereobtainedbysendingATcommandstothe3G/4Gcardspluggedintothemobilerouterandclients.
ForEV-DO,FDD-LTE/HSPA+andTD-LTE,theATcommandsreturnedvaluesbetween0and31,whichisarelativequalityofthereceivedsignalanditisuniquetoeachcellulartechnology.
Forthe3G/4Gcardsinuse,thepowerlevelindBmiscalculatedbyrssi2113.
Forexample,10correspondsto-93dBm.
Weusethesevaluestoevaluatethestabilityofeachtechnologyunderdifferentstatuses.
FromFig.
3,wecanobservethattheRSSIsofallthe3G/4Gnetworkskeepstableatarelativelyhighlevelinastaticstate.
However,asshowninFig.
4,theRSSIsfluctuateremarkablyandcanevenbe0attimes.
Fortunately,wecanalsoobservethatdifferent3G/4Gnetworkscancomplementeachotherwellatsomegeographiccoordinates.
RTTwasobtainedbyamodified"ping"program.
WemodifiedthesourcecodesbyaddingatimestampinICMPmessages,andrecordedtheRTTinaMySQLdatabaseevery5seconds.
Fig.
5showstheRTTbetweenthestaticclientandtheserver.
TheRTTsarerelativelylow(around100ms~200ms)andstableforallthenetworksinastaticstate,whilefromFig.
6wecanconcludethatinthehigh-speedenvironment,theRTTsbecomemuchlargerandunstable.
BandwidthwasevaluatedusingthetooliPerf[39],whichisdevelopedforactiveTABLEI.
COMPARISONSBETWEENSTATICANDHIGH-SPEEDEVIRONMENTSFOREV-DO,FDD-LTE/HSPA+ANDTD-LTEParameters3G/4GHigh-SpeedTrainStaticAverageRSSIEV-DO17.
730.
8FDD-LTE/HSPA+15.
622.
6TD-LTE18.
322.
7AverageRTTEV-DO1255ms90msFDD-LTE/HSPA+763ms155msTD-LTE1163ms151msAverageBandwidthEV-DO127kbps1.
5MbpsFDD-LTE/HSPA+1.
8Mbps12MbpsTD-LTE2.
1Mbps15MbpsmeasurementsofthemaximumachievablebandwidthonIPnetworks.
Fig.
7showsthebandwidthdistributionofEV-DO,FDD-LTE/HSPA+andTD-LTEobtainedonthetrain.
TableIshowstheaverageRSSI,RTT,andbandwidthofEV-DO,FDD-LTE/HSPA+andTD-LTEintheexperiments.
Althoughtheresultsobservedalongdifferenttrailsmaydifferfromeachotherduetothedifferentnetworkcoverageandgeographicalenvironments,ourresultsreflectthehighlyvariabilityofwirelesslinksinhighlymobilescenarios.
Severalliteratures[24][25]haveshownthatthehistoricaltracingdatacanforecastthewirelessnetworkcapabilitiesandisofgreatvalueforintelligentscheduling.
Weevaluatedtheefficacyoftheseforecasts.
Fig.
8showsthebandwidthontwotrips.
Wecanfindasimilarconclusionas[25],i.
e.
,thepastbandwidthinformationisagoodindicatoroftheactualbandwidthexperiencedatagivenlocation.
Forexample,thereisasharpbandwidthdecreasefromlocation8tolocation9.
Thus,toforecastthebandwidthatlocation9ontrip2,theresultwillbeinaccurateifthebandwidthobservationatlocation8isused.
Incontrast,theresultwillbemoreaccurateifthebandwidthobservationatlocation9ontrip1isused,becausethetwobandwidthobservationsatlocation9onthetwotrips01000200030004000102030EV-DOInastaticstate:2014/12/1615:2001000200030004000102030RSSIFDD-LTE/HSPA+01000200030004000102030elapsedtime(seconds)TD-LTE01000200030004000020EV-DOhigh-speedrail:2014/11/0508:3001000200030004000020RSSIFDD-LTE/HSPA+01000200030004000020elapsedtime(seconds)TD-LTE010002000300040000500EV-DOInastaticstate:2014/12/1615:20010002000300040000500RTT(ms)FDD-LTE/HSPA+010002000300040000500elapsedtime(seconds)TD-LTE0100020003000400005000EV-DOhigh-speedrail:2014/11/0508:300100020003000400005000RTT(ms)FDD-LTE/HSPA+0100020003000400005000elapsedtime(seconds)TD-LTE0018-9545(c)2016IEEE.
Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seehttp://www.
ieee.
org/publications_standards/publications/rights/index.
htmlformoreinformation.
Thisarticlehasbeenacceptedforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.
Contentmaychangepriortofinalpublication.
Citationinformation:DOI10.
1109/TVT.
2016.
2581020,IEEETransactionsonVehicularTechnologyFig.
8.
PastversuspresentbandwidthFig.
9.
PastversuspresentRTTaremuchcloser.
TheRTTsontwotripsillustratedinFig.
9showasimilartrend.
Toconclude,therearetwokeyobservationstoguideusinthedesignofamultipathmethodforvehiclesinhigh-speedmobilityscenarios:First,comparedtostationaryvehicles,vehiclesmovingatahighspeedfaceamorechallengingwirelessenvironment.
Thewirelesscharacteristics,intermsofRSSI,RTT,andbandwidth,varywidelywithlocation.
Thenetworkparametersexperiencedatalocationcannotaccuratelyindicatethoseofthenextlocation.
Thus,mostschedulingalgorithmsbasedonlinkprobewillbefragile.
Second,thefixedrouteofavehiclegivesusanopportunitytopredictthenetworkcharacteristicsbasedonhistoricaldata,i.
e.
,thepasttriptellsusmorethanthepreviouslocationonthecurrenttrip[24][25].
Althoughitisimpossibletopredictthenetworkcharacteristicswithouterrorbythehistoricaltracingdatabases,theforecastbythismeansisofgreathelpforanefficientscheduling.
Thus,wedesignourmultipathmethod,aswellasourschedulingalgorithm,basedonthetwoobservations.
Furthermore,reschedulingmechanisms(alsocalledcompensationmechanisms)aredesignedtodealwithunexpectedchannelfluctuations.
IV.
NETWORKADAPTIVEMULTIPATHTRANSMISSIONInthissection,wefirstpresentanoverviewoftheproposedmultipathtransmissionarchitectureforhigh-speedvehiclesinheterogeneouswirelessnetworks.
Then,anetworkadaptiveschedulingalgorithmisintroducedtomeettherequirementsintermsofaggregatingbandwidthandavoidingthedisorderofpackets.
InternetUsersDataNIFControlsignalWirelessNetworksNIFNIFHSPA+TD-LTEEVDOWiFiMobileRouterAccessRouterDataTunnelsFig.
10.
OverviewoftheconcurrentmultipathtransmissionschemeA.
MultipathArchitecturewithLogicTunnelsFig.
10showsanoverviewofthemultipathtransmissionarchitecture.
ItisproposedtoprovidemorestableInternetservicesfortheonboardusersbymultipleavailablewirelesstechnologies,withoutrequiringtheparticipationoftheuserdevicesinanymultipathsignaling.
Inthearchitecture,oneormoremobileroutersaredeployedonahigh-speedvehicle.
Moreover,amobileroutercanconnecttotheInternetbyaccessingmultiplewirelessnetworksthatareavailablealongatrack,includingWiFi,3G,4G,andsatellite.
ThemobilerouterthenprovidestransparentInternetservicesforonboardusers.
Anaccessrouterisadevicelocatedontheground.
ItvisitstheInternetviawiredlinks,andmaintainslogictunnelswithmultiplemobilerouters.
Thelogictunnels,whichareestablishedbetweenamobilerouterandanaccessrouter,providetransparentconcurrentmultipathdatatransmissionforonboardusers.
Ateachavailableinterfaceofamobilerouter,alogictunnelisestablished.
Thenumberofavailableinterfacesmaychangewiththemovingofthevehicle.
Thenumberoflogictunnelswillchangecorrespondingly.
Alogictunnelataninterfacewillberemovediftheinterfacebecomesinactive.
Meanwhile,anewtunnelwillbeestablishedifaninterfacebecomesactive,e.
g.
,gettinganewIPaddress.
Thelogictunnelsareindependentofeachother.
Eachtunnelsimplyforwardsthepacketsinitssendingqueue.
Theproposedschedulingalgorithmisresponsibleforputtingpacketsintothesendingqueuesofalllogictunnels.
TheproposedschedulingalgorithmreceivesauserpacketattheIPlayerwhenthepacketarrivesatthemobilerouter.
ThenthealgorithmputstheIPpacketintothesendingqueueoftheselectedtunnel.
Afterwards,thetunnelencapsulatesthepacketandforwardsittotheaccessrouter.
Attheaccessrouter,theencapsulatedpacketisfirstlyputintothereceivingbufferforreordering.
Then,thepacketisdecapsulated,andtheoriginalIPpacketisreturnedtotheIPlayeroftheprotocolstack.
Afterthat,theprotocolstackforwardstheoriginalIPpackettoitsdestination.
Inthearchitecture,schedulingisakeymethodtoimprovethenetworkefficiency.
Fortheonboardapplications,especiallythosebasedonconnection-orientedprotocols(e.
g.
,TCP),theschedulingalgorithmnotonlyimprovesthetotalavailablebandwidthbyaggregatingthebandwidthofmultipleinterfaces,butalsoavoidsunnecessarypacketreorderingcausedbydynamicpathcharacteristics.
0510152025010002000300040005000LocationBandwidth(kb/s)Trip1Trip2Location8Location90510152025020040060080010001200140016001800LocationRTT(ms)Trip1Trip20018-9545(c)2016IEEE.
Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seehttp://www.
ieee.
org/publications_standards/publications/rights/index.
htmlformoreinformation.
Thisarticlehasbeenacceptedforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.
Contentmaychangepriortofinalpublication.
Citationinformation:DOI10.
1109/TVT.
2016.
2581020,IEEETransactionsonVehicularTechnologyFig.
11.
SimplifiedviewofthemultipathnetworkbetweenthemobilerouterandtheaccessrouterinheterogeneouswirelessnetworksTABLEII.
SUMMARYOFIMPORTANTSYMBOLSSymbolsMeaningofthesymbolsApathbetweenawirelessinterfaceofamobilerouterandawiredinterfaceofanaccessrouterTheone-waydelayofpathattimeTheavailablebandwidthofpathattimeQueueingtimeonpath.
Overaperiodoftimeafter,thesendingqueueonpathwillbeemptyTotalsizeofpacketsinthesendingqueueonpathattimeThearrivaltimeofthexthpacketatthemobilerouterThearrivaltimeofpacketxattheaccessrouteronpathTheleavingtimeofthexthpacketfromtheaccessrouter_Thedurationthatpacketxwaitsinthesendingqueueonpathwhenarrivingat_Packetsendingdurationatpathwhenthesendingbegunat_One-waytransmissiondelayatThenormalizeaveragebandwidthofpath.
ThesizeofthexthpacketarrivingatthemobilerouterThebufferingtimeofthexthpacketattheaccessrouterThemaximalacceptabledelayofthepathsThehighestbandwidthofthepathsThelowestbandwidthofthepathsThemaximumpacketsizeNext,weintroduceaschedulingalgorithmthatmeetsourrequirements.
B.
NetworkAdaptiveMultipathSchedulingAlgorithmToprovidetransparentmultipathInternetservicesfortheusersonboard,manychallengesexist.
Amajoroneishowtoschedulepacketsreasonablyonmultipleunstablepaths,asdiscussedinsectionIII.
Ifthepacketsbelongingtothesameflowarescheduledondifferentpaths,ahighpacketout-of-orderislikelytooccurwithoutagoodschedulingalgorithm.
EDPF[23]canguaranteethepacketsarriveinorderwhenthepacketsareofthesamesize.
However,itdoesnottakeintoaccountthedelayonwirelesslinks.
OuralgorithmisavariationofEDPF.
Weextenditbyaddingthewirelesstransmissiondelayintothealgorithm.
Furthermore,basedonourpracticalexperimentsonthehigh-speedtrain,wegotthetracingdatabasesaboutthedelayandbandwidthofvariouswirelessnetworksalongaspecifictrail.
Dependingontherecordednetworkconditions,wecandividethedatabaseintomini-timeslotsduringwhichthedelayandbandwidthremainfairlyconstant.
Wealsomonitorthewirelesslinkconditionsandperformreschedulingincaseofunexpectedchangesofpathparameters,intermsofbandwidthanddelay.
Thefundamentalthoughtofouralgorithmisto(1)takeaccountofthequeueingtimeatthemobilerouter,bandwidthoftheinterfaces,anddelaysofthepathsbetweenthemobileAlgorithm1:NetworkadaptivemultipathschedulingalgorithmInput:Output:forwardingpathforeachpacket;1:initialize0=0,=0;2:updateandpertimeslot;3:4:while(anewpacketarrives)do5:calculation;6:packetisputinthesendingqueueonpath,andwaitstobeforwarded;7:endwhile8:9:foranypath10:monitoringthepathparameters,periodically;11:if|Δ||ΔB|12:Δ;Δ;13:foranypacketinthesendingqueues14:re-calculation;15:packetisputinthesendingqueueonpath,andwaitstobeforwarded;16:endfor17:endif18:endforFig.
12.
Distributionofthetimepointsrouterandtheaccessrouter,(2)basedonthetracingdatabasesandthemonitoringofpathparameters,selectthepaththattakestheshortesttimetodeliverapackettotheaccessrouter,and(3)reschedulethepacketsinsendingqueuesonceachangingofpathconditionsisdetected.
ThenetworkbetweenthemobilerouterandtheaccessroutercanbesimplifiedasshowninFig.
11.
ThemeaningofsymbolscanbefoundinTABLEII.
ThenetworkadaptivemultipathschedulingalgorithmisdetailedinAlgorithm1.
Thecalculationsinstep5andstep14areshowninEq.
(1).
Phase1:PathselectionatthearrivalofapacketAsdepictedfromstep4tostep7,Algorithm1selectstheforwardingpathforpacketatitsarrival.
Instep5,argmin,1(1)whereisthenumberofavailablepaths.
isthearrivaltimeofpacketattheaccessrouteronpath.
Tomakeitclear,thetransmissionprocedureofthexthpacketisillustratedinFig.
12.
Letbethearrivaltimeofpacketatthemobilerouter.
Then,theestimatedarrivaltimeofpacketattheaccessrouteronpathisevaluatedas,where___where_isthedurationthatpacketwaitsinthesendingqueueonpath.
Inotherwords,theforwardingofthexthpacketbeginsat_.
_isthedurationofthe0018-9545(c)2016IEEE.
Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seehttp://www.
ieee.
org/publications_standards/publications/rights/index.
htmlformoreinformation.
Thisarticlehasbeenacceptedforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.
Contentmaychangepriortofinalpublication.
Citationinformation:DOI10.
1109/TVT.
2016.
2581020,IEEETransactionsonVehicularTechnologypacketsendingtimeonpath,i.
e.
istheone-waytransmissiondelayfromthemobileroutertotheaccessrouterattime,i.
e.
,_.
Instep5,thevaluesofparameterforcalculating_,_,and_aredifferentfromeachother.
Inaddition,andareupdatedpertimeslotalongwiththemovingofavehicleinstep2.
Sincethevaluesofandarethedynamicpathparametersstoredinthetracingdatabasesforafixedtrack,calculationinstep5willautomaticallychoosetheappropriatevaluesaccordingtothespeedandlocationofthevehicle.
Asimpleexampleisthat,assumingapacketarrivedatamobilerouterat,ifitswaitingtimeinqueueis_,then,itssendingdurationis_,whereisthesendingstarttime,_.
Afterthat,itsone-waydelay_,where_.
Thus,thepacket'sarrivaltimeattheaccessrouterwouldbeestimatedasThisisreasonablebecausemaynotbeequalto,andmaynotbeequalto.
However,existingresearch[23][37]estimateandbyand.
Asshownin[25]andourobservationsinSectionIII,theseestimationsareinaccurate,especiallyinhigh-speedmobilescenarioswhereandchangeovertime.
Oursolutionistousethehistoricaltracedata.
Thereasonabilityofusinghistorytopredictthefuturehasbeenprovedby[24]and[25]usingreal-worldusage.
Furthermore,wealsoshowitsvaliditybyourexperiments.
ThisisoneofthesignificantdifferencesbetweenouralgorithmandEDPF[23],whichdoesnotconsideravariablewhencalculatingthedeliverytimeofapacket.
Aftermakingadecision,thealgorithmupdates_to__,i.
e.
,the1thforwardingatpathcanbeginonlyafterthexthpackethasbeensentout.
Phase2:ReschedulingwhenlinkstatusdriftsawayThereisapossibilitythattheestimatedpathparameterswilldriftawayfromthetruevalues.
InPhase1oftheproposedalgorithm,thepathestimationsarebasedonthetracingdatabases.
Iftheestimationsareinaccurate,theselectedpathmightnotbetheappropriateonetodeliverthepacket.
Hence,itisreasonabletorequiresomecompensationmethodstodealwiththeparameterdrift.
Therearetwocompensationmethodsdealingwiththeinaccurateestimation.
Theyareasfollows:(1)Thereceivingbufferinanaccessroutercanreordertheout-of-orderpacketscausedbyrandomfluctuationsofpathparameters.
(2)Thereschedulingmechanismcandealwiththesignificantchangesinlinkcapabilities.
Forinstance,newbasestationscouldleadtoasignificantbandwidthgrowthandpopularsocialeventscouldleadtoalongerpathdelayduetonetworkcongestion.
Todealwithunexpectedparameterfluctuations,thereschedulingmechanismperiodicallymonitorsthecurrentparametersandinstep10,andcomparescurrentparameterswithhistoricalparametersinstep11.
Reschedulingwillbeperformediftheforecastoforisobviouslyincorrect,i.
e.
,ΔandΔwillbecompensatedtoand,respectively,instep12.
Afterthat,re-calculationfromstep13tostep16willbeperformedinordertodeliverthedelayedpacketsinqueuesearliesttotheaccessrouter.
NotethatthereschedulingisperformedonlywhenΔorΔexceedsaspecifiedthresholdor.
Otherwise,thereceivingbufferintheaccessroutercanworkasthecompensationmethodtoreorderpackets.
Duetothereasonthattheonboardusersusuallyinitiatethecommunications,suchassendinganemailorbrowsingwebsites,ourexplanation,sofar,focusesonlyonthetransmissionfromthemobileroutertotheaccessrouter.
Itisalsofitforthecasethattherolesofthemobilerouterandtheaccessrouterareinterchanged.
V.
PROPERTIESANALYSISInthesubsequentsections,theeffectivenessofthealgorithmisanalyzed,intermsofthelengthofsendingqueues,thesizeofreceivingbufferforreorderingpackets,andtheabilityofbandwidthaggregation.
A.
DifferencebetweenLengthofSendingQueuesIfthepathsareofconstantbandwidthanddelayandifthepacketsareofthesamesize,wecaneasilyderivefromouralgorithmthatallthepacketscanarriveattheaccessrouterinorder.
Incontrast,ifthepathsareofvariablebandwidthanddelay,and(or)thepacketsareofvariablesize,itiscrucialforthealgorithmtoschedulethepacketsamongthepathsappropriately.
Foranytwopathsand,wecansaythealgorithmperformsagoodschedulingifthemaximumdifferencebetweenthelengthsoftheirsendingqueuesandisnotthefunctionofthetotalnumberofpackets.
Inthisanalysis,wefurtherdefineasthemaximalacceptabledelayofthepaths.
If,weregardthepathasunavailable.
Nonewpacketswillbescheduledonpathifsincepathisnotsuitabletotransmitpacketsatthismoment.
Wedefineandasthehighestandlowestbandwidthofthepaths,respectively.
Sincethebandwidthofanyofthewirelesslinksfluctuatesinhigh-speedscenarios,weregardthepathasunavailableif.
Wedefineasthelargestpacketsize.
Inthefollowing,wefirstpresentalemmathatishelpfultogetthepropertiesofthealgorithm.
Wejustpresentthelemmahereforbetterreadability,whilethedetailedproofcanbefoundinAppendixA.
Lemma1.
Foranytwopaths,andanytime,andarethequeueingtimeonthetwopaths,if,then,.
Next,wewillshowthatthedifferencebetweenthenormalizedlengthsofsendingqueuesisnotrelatedtothenumberofpackets,whereasitisafunctionofforRoundRobinschedulingalgorithmanditsvariants.
Letbethenumberofavailablepaths,foranytwopathsand,therelationbetweennormalizedandnormalized0018-9545(c)2016IEEE.
Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seehttp://www.
ieee.
org/publications_standards/publications/rights/index.
htmlformoreinformation.
Thisarticlehasbeenacceptedforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.
Contentmaychangepriortofinalpublication.
Citationinformation:DOI10.
1109/TVT.
2016.
2581020,IEEETransactionsonVehicularTechnologyis,where,isthenormalizeaveragebandwidthoflinkbetweentimeand.
Proof.
FromLemma1,welearnthatcannotexceed,i.
e.
,,,.
Thus,,,Inthefollowing,weanalyzethebehavioroftheschedulingschemesbasedonRoundRobin.
Assumingthereareavailablepaths0,1,,1.
Atanytime,thetotalnumberofpacketsfortransmissionis.
Thus,anypacketwillbescheduledonpath.
Thetotalsizeofpacketsinthesendingqueueonpathattimeis0,whereisthearrivaltimeofpacket,andisthesizeofpacket.
Then,wecanconcludethat,foranytwopathsand,isafunctionof,andcangrowwithoutbound.
Totakeasimpleexample,ifthesizesofpacketsscheduledontwopathsalternatebetween1000bytesand100bytes,whilethetwopathshavethesamebandwidth,thedifferencebetweenthelengthsofthetwosendingqueuesgrowswiththenumberofpackets.
ForWeightedRoundRobin,thevariablepacketsize,variabledelay,andvariablebandwidthwillmakethedifferencebetweenthequeuelengthgrowwithoutbound.
WeevaluatethepropertiesofthealgorithmsinSectionVIthroughsimulations.
B.
TheSizeofReceivingBufferforReorderingPacketsWhenareceivingbufferforreorderingpacketsisusedattheaccessrouter,letbethearrivaltimeofthexthpacketattheaccessrouter,andbetheleavingtimeofthexthpacketfromtheaccessrouter.
Then,asshowninFig.
12,,whereisthebufferingtimeofpacketattheaccessrouter.
Forouralgorithm,thebuffersizeneededattheaccessrouterforsendingthepacketsinorderoutoftheaccessrouterisatmost∑2,,.
Proof.
Atanytime,foranypacketinthesendingqueuesatthemobilerouter,letthemaximumdeliverytimetotheaccessrouterbe,whereisthelastpacketbeingdeliveredandthedeliverypathis.
Ifanypacketarrivedatthemobilerouterafterwasdeliveredbefore,itwouldneedtobebufferedattheaccessrouter.
Apacketthatarrivesafterpacketmayonlybedeliveredbeforeonanypath,.
Let,whereisthearrivaltimeofthelastpacketinthequeueonpath.
Thus,onpath,alimitedbitsnotexceeding,,,maybedeliveredbefore.
FromLemma1,,2Thus,thetotalbuffersizeneededis∑,,,∑,,2∑2,,∑2,,Wecanseethatthebuffersizeforreorderingpacketsisalsonotrelatedtothetotalnumberofpackets,whileRoundRobindoesnotholdthisproperty.
C.
TheAbilitytoAggregateBandwidthAsinglelinkwiththeaggregatebandwidth∑iscalledAggregatedSingleLink(ASL)iftheavailablebandwidthofpathis.
AschedulingalgorithmwouldbebetterifitsperformancewereclosertoASL.
Inthissection,wewillshowthat,whenthedelaysofwirelesslinksareinconsiderationandreschedulingisperformed,thedifferencebetweentotalthroughputservedbytheproposedalgorithmandASLin0,isupperboundedby∑foranytime.
Proof.
Assumingthereareavailablepaths,foranypaththeavailablebandwidthis.
AssumingASLfinishessendingallthepacketsattime,therearetwopossiblecases:First,oneormorepathsareidleat.
Inthiscase,ouralgorithmgetsthemaximumdifferencewithASLifonlyonepathisidle.
Atthesametime,pathshouldbewiththelowestbandwidthamongallthepaths.
FromLemma1,welearnthatcannotexceedsforanypath,.
Thus,theupperboundis:∑,∑,∑,∑,∑Second,nopathisidleat.
Inthiscase,letτbethetimewhenallthepathsbecomeactive.
Afterτ,allthepathsarebusyforwardingdatapackets.
Thus,thedifferenceoftotalbandwidthbetweenthealgorithmandASLisderivedfromtheperiod0,τ.
And,thefirstcaseboundsthedifferencewithASLatτby∑.
Wecanconcludethatthisupperboundisstillnotrelatedtothetotalnumberofpackets,whereastheschedulingalgorithmsbasedonRoundRobinis.
VI.
SIMULATIONANDEVALUATIONInthissection,weimplementthemultipathtransmissionscheme,andassesstheperformanceadvantagestheproposedschedulingalgorithmhasbycomparingwithAggregatedSingleLink,EDPF,andWeightedRoundRobinschedulingalgorithms.
0018-9545(c)2016IEEE.
Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seehttp://www.
ieee.
org/publications_standards/publications/rights/index.
htmlformoreinformation.
Thisarticlehasbeenacceptedforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.
Contentmaychangepriortofinalpublication.
Citationinformation:DOI10.
1109/TVT.
2016.
2581020,IEEETransactionsonVehicularTechnologyWiredNerworkNIFMobileRouterAccessRouterTunnelIPpacketOtherMobileRoutersNIFDataLayerControlLayerMobilityManagementMultipathManagementDataLayerRouterManagementMobilityManagementMultipathManagementTunnelModesTunnelModesTunnelsSchedulingMonitoringMobilenetworkonboardkernelTunnelModesNetfilter_HOOKPathStatusInternetIPpacketPathStatusIPpacketDataLayerUserSpaceTracingdataFig.
13.
FunctionalmodulesA.
ControlLayerImplementationFig.
13introducesourimplementation.
Inthecontrollayer,themobilitymanagementmodulemaintainstheavailableIPaddressesallocatedbymultiplewirelessnetworks.
Inaddition,amobilerouterwillinformtheaccessrouterofthenewassignedaddresseswheneverahandoveroccurs.
WhenaPR-SCTP[27]tunnelisadopted,theaccessroutercanbeinformedofthenewaddressesbySCTPmessages;whilewhenaUDPorIP-in-IPtunnelisadopted,MobileIPoraself-definemessagecanbeusedtonotifytheaccessrouteraboutthechangingofaddresses.
Themultipathmanagementmoduleisresponsibleforestablishinglogictunnelsonwirelesslinks.
Itselectsandchangesthemodesoflogictunnelsaccordingtodifferentlinkconditions.
Italsochangestheparametersofthelogictunnels,e.
g.
,lifetimeparameterforPR-SCTP,accordingtothepathstatus.
Fourkindsoftunnelmodesaresupported:(a)IP-in-IPTunnel,whichisthesimplestandcommonestmethodtoestablishthelogictunnels.
(b)PR-SCTPTunnel,whichcanprovidepartialtransmissionreliabilityfordataonanunstablewirelesslink.
(c)UDPTunnel,whichcanmaximizethetransmissionofdatawithoutconsideringthelossofpackets.
WhenUDPtunnelsareapplied,theupperlayerapplicationsareresponsibleforthereliabilityofdata.
TheadvantageofaUDPtunnelisthatitcanmaximizethesurvivalofthetunnelintheatrociouswirelessenvironmentswhileprovidingsomeflexibilitiesthatIP-in-IPTunnelcannotachieve.
(d)MixedMode,whichestablishesPR-SCTPtunnelsonsomelinksandUDPtunnelsontheotherssothatthetwokindsoftunnelscanplaytotheirstrengths.
B.
DataLayerImplementationInthedatalayer,ourschedulingalgorithmdeliverspacketstodifferenttunnels.
Then,thepacketsareencapsulatedinthelogicaltunnelsandsenttotheaccessrouter.
Althoughseveralmodesoflogictunnelcanbechosen,ourschedulingalgorithmworksattheIPlayerandistransparentforthelogictunnels.
Fig.
14.
BandwidthshowedbygoogleearthFig.
15.
BandwidthaggregationonmultiplepathsAsshowninFig.
13,theimplementationoftheproposedschedulingalgorithmcanbedividedintotwomodules,oneforthekernelprocessingfunctions,theotherfortheuserspacefunctions.
Inparticular,netfilterqueueisusedintheLinuxkerneltointerceptIPpacketsbyregisteringaHOOKfunction.
Then,theschedulingmoduleintheuserspacegetstheIPpackets,makesdecisionsaccordingtothepathstatus,anddistributestheIPpacketstotheselectedtunnels.
Thus,ourschedulingschemeistransparenttotheprotocolstackandeasytodeploy.
C.
ExperimentsonaHigh-SpeedTrainWeperformedsomeexperimentsonahigh-speedtraintoevaluatetheproposedmultipathscheme.
Intheexperiments,thedatawasgeneratedbyrealapplications.
AllthedatawascapturedandstoredinaMySqldatabase.
Intheexperiment,twoEV-DOcards,twoFDD-LTE/HSPA+cardsandtwoTD-LTEcardswereinstalledinamobilerouter.
Thelogictunnelsaggregatedtheavailablebandwidthofthesixwirelesslinksandprovidedthetotalbandwidthtotheonboarddevices.
Fig.
14illustratesthebandwidthbygoogleearth.
Thegreenpoint,thebluepointandpinkpointrepresentthebandwidthover5Mbps,between(2Mbps,5Mbps),andbetween(0.
1Mbps,1Mbps),respectively.
ThebandwidthofeachofthesixlinksandthetotalbandwidthareshowninFig.
15.
Wecanobservethenumberofavailablepathsandtheavailablebandwidthoneachpathatdifferenttimes.
Wecanconcludethatourconcurrentmultipathtransmissionschemeworkswellinapracticeenvironment,andcanmakefulluseofvariousexistinginfrastructures.
Furthercomparisonswithotherschedulingalgorithmsarepresentedinthenextsubsection.
010002000300040005000051015Averagebandwidth(Mb/s)TotalEVDO-1EVDO-2FDD-LTE-1FDD-LTE-2TD-LTE-1TD-LTE-20100020003000400050000510Numberofavailablelinkstime(Seconds)0018-9545(c)2016IEEE.
Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seehttp://www.
ieee.
org/publications_standards/publications/rights/index.
htmlformoreinformation.
Thisarticlehasbeenacceptedforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.
Contentmaychangepriortofinalpublication.
Citationinformation:DOI10.
1109/TVT.
2016.
2581020,IEEETransactionsonVehicularTechnologyFig.
16.
Theabilitytoaggregatebandwidth:receivingbuffersize:15000bytesFig.
17.
Theabilitytoaggregatebandwidth:receivingbuffersize:9000bytesFig.
18.
PacketorderofarrivalattheaccessrouterfordifferentalgorithmsFig.
19.
QueueingtimesandtheirdifferenceofouralgorithmFig.
20.
QueueingtimesandtheirdifferenceofEDPFFig.
21.
QueueingtimesandtheirdifferenceofWRRD.
SimulationsWeevaluatedtheperformanceoftheproposedalgorithmtoshowitsadvantages.
SomesimulationswereperformedinMatlabtocomparewithAggregatedSingleLink,EDPF,andWeightedRoundRobin,intermsoftheabilityofbandwidthaggregation,thequeueingtime,andthelengthofsendingqueues.
Thesimulatingnetworktopologycomprisedamobilerouterandanaccessrouter.
Twodisjointedpathswereestablishedbetweenthetwodevices.
Tomakethesimulationsmorepractical,realtracingdatabaseswereusedasthebandwidthanddelayparametersofthetwopaths.
Hence,thebandwidthanddelayvariedovertimejustasthemobilerouterwasunderahigh-speedmobility.
Forpath0,thebandwidthanddelaywereroughlyintherangeof0.
9-4Mb/sand200-1500ms,respectively.
Forpath1,thebandwidthanddelaywereroughlyintherangeof0.
5-1.
8Mb/sand80-1300ms,respectively.
Thesizeofpacketswasrandomlybetween1000bytesand1500bytes.
(a)TheabilitytoaggregatebandwidthFirst,theabilitytoaggregatebandwidthwasevaluated.
AggregatedSingleLinkisanidealsinglelinkwiththeaggregatedbandwidth∑whereisthebandwidthofpath.
Itrepresentstheupperboundofaggregatedbandwidth.
TheefficiencyofaschedulingalgorithmcanbeshownbycomparingwithAggregatedSingleLink.
ForWeightedRoundRobin,thepacketsweredistributedtotwodifferentpaths,dependingontheratio,i.
e.
,apathwithlongerRTTwouldbedistributedlesspackets.
Fig.
16andFig.
17showtheabilitytoaggregatebandwidthbydifferentalgorithmswiththereceivingbuffersize15000bytesand9000bytes,respectively.
WecanobservethatouralgorithmhasabetterperformancethanEDPFandWRR.
Inparticular,theperformancesofEDPFandWRRbecomeworsewiththedecreaseofthereceivingbuffersize.
ThereasoncanbefoundinFig.
18,whichshowsthepacketorderofarrivalattheaccessrouterfordifferentalgorithms.
ForAggregatedSingleLink,allthepacketsarrivedattheaccessrouterinordersinceAggregatedSingleLinkistheidealcase.
Forotheralgorithms,packetdisorderoccurs,i.
e.
,somepacketswithsmallersequencenumbersarrivelaterthanthosewithbiggersequencenumbers.
FromFig.
18wecanobservethatourproposedalgorithmgeneratedlessout-of-orderpacketsthanEDPFandWRRsothatitrequiredasmallerreceivingbuffersize,whiletheperformancesofEDPFandWRRdecreasedwhentherewasnotenoughbuffersizeforthemtoreordertheout-of-orderpackets.
(b)DifferencebetweenqueueingtimesInordertoevaluatethedifferencebetweenthequeueingtimesforouralgorithm,EDPFandWRR,wetracedthesendingqueuesonbothpath0andpath1.
Fig.
19,whichiswithdoubleyaxes,illustratesthequeueingtimesonbothpath0andpath1whenouralgorithmwasadopted.
Fig.
19alsoshowsthedifferencebetweenqueueingtimeonthetwopaths.
JustasLemma1shows,thetimedifferencebetweenthetwoqueuesisupperbounded.
Fig.
20showsEDPF'squeueingtimedifferenceontwopaths,whichissmallerthanthatofouralgorithm.
However,thisisbecauseEDPFignoresthevariablepathdelayonwirelesslinks,whichleadstothedisorderofpacketsattheaccessrouter.
Fig.
21showsthequeueingtimesofWRR.
Wecanobservethatthetimedifferencebetweenthetwoqueuesriseswiththenumberofpackets.
Themainreasonisthatthevariablepath050010001500012345x105Time(ms)Throughput(Bytes)AggregatedSingleLinkProposedAlgorithmEDPFWeightedRoundRobin050010001500012345x105Time(ms)Throughput(Bytes)AggregatedSingleLinkProposedAlgorithmEDPFWeightedRoundRobin05101520051015202530OrderofarrivalattheaccessrouterPacketsequencenumberAggregatedSingleLinkProposedAlgorithmEDPFWeightedRoundRobin0501001502000200400600800100012001400QueueingTime(ms)Packets050100150200050100150200250300350DifferencebetweenQueueTimeQueueingtimeonpath0Queueingtimeonpath1Diffenence050100150200010002000QueueingTime(ms)Packets050100150200050100DifferencebetweenQueueTimeQueueingtimeonpath0Queueingtimeonpath1Difference0501001502000100020003000QueueingTime(ms)Packets050100150200050010001500DifferencebetweenQueueTimeQueueingtimeonpath0Queueingtimeonpath1Differece0018-9545(c)2016IEEE.
Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seehttp://www.
ieee.
org/publications_standards/publications/rights/index.
htmlformoreinformation.
Thisarticlehasbeenacceptedforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.
Contentmaychangepriortofinalpublication.
Citationinformation:DOI10.
1109/TVT.
2016.
2581020,IEEETransactionsonVehicularTechnologyFig.
22.
LengthofsendingqueuesandtheirdifferenceofouralgorithmFig.
23.
LengthofsendingqueuesandtheirdifferenceofEDPFFig.
24.
LengthofsendingqueuesandtheirdifferenceofWRRparametersmakeitdifficulttoschedulepacketswell.
(c)DifferencebetweenthenormalizedqueuelengthsWealsocomparedthelengthofsendingqueuesbetweenouralgorithm,EDPFandWRR.
Fig.
22showsthenormalizedqueuelengthsonpath0andpath1,andthedifferencebetweenthenormalizedqueuelengthsforouralgorithm.
JustasshowninSectionIV,thedifferencebetweenthenormalizedqueuelengthsontwopathsisupperbounded.
Fig.
23illustratesnormalizedqueuelengthsandtheirdifferenceforEDPF.
Thedifferencebetweenthenormalizedqueuelengthsissmallerthanthatofouralgorithm.
However,thisisalsobecauseEDPFignoresthevariablepathdelayonwirelesslinks,whichleadstomoreseriouspacketdisorderattheaccessrouterthanouralgorithm.
Fig.
24showsthosewithWRR.
WecanconcludethatthedifferenceofthequeuelengthswithWRRiswithnoupperbound.
Infact,itisafunctionofthetotalpacketnumber.
CONCLUSIONInthispaper,somepracticalmeasurementsoftheexisting3Gand4Gmobilenetworksareperformedinatypicalhigh-speedenvironmentonahigh-speedtrain.
Then,aconcurrentmultipathtransmissionscheme,togetherwithanetworkadaptiveschedulingalgorithm,isproposed,whichprovidestransparentandeffectiveInternetservicesfortheonboardusersbymultipleavailablewirelessnetworks.
Someofthesuperiorpropertiesofouralgorithmareanalyzed,intermsofthelengthofsendingqueues,thebuffersizeforreorderingandtheabilitytoaggregatebandwidth.
SimulationsandexperimentsshowthattheschemecanimproveQoSofInternetservicesonboardbyincreasingtheavailablebandwidthandavoidingthedisorderofpackets.
APPENDIXALemma1.
Foranytwopaths,andanytime,andarethequeueingtimeonthetwopaths,if,then,.
Proof.
Themethodofinductionisusedtoprovethislemma.
Wefirstshowthatthelemmaholdsatthetime0whenthefirstpacketarrives.
Wethenassumethatthelemmaholdswhenpacket1,2,,1arrives,andfurtherprovethelemmaholdsatwhenthethpacketarrives.
Basis.
Forthealgorithm,thefirstpacketwillbescheduledonthepaththatcandeliveritthefastesttotheaccessrouter.
Thus,,whereisthequeueingtimeonpathwhenthefirstpacketjustentersthequeue,i.
e.
,.
isthesizeofthefirstpacket.
Atthesametime,0becausethefirstpacketdoesnotenterthequeueonpath.
Notethatthesendingqueuesonallpathsareemptybeforethearrivingofthefirstpacket.
Thus,thelemmaholdswhenthefirstpacketjustentersthequeueonpathattime.
Foranytimebeforethearrivalofthesecondpacketat,decreaseslinearlywithbecauseisalinearfunctionofforanypath.
Inductivestep.
Assumethatthelemmaholdsforpackets1,2,,1.
Letbethepathchosenfortransmissionofpacket.
AccordingtoEq.
(1),,i.
e.
,whereisthetimejustbefore(atpacketentersthequeue).
isthequeueingtimejustbeforepacketentersthequeue.
Thus,Since,,(2)Thentwocasesariseasfollows:Case1.
.
AccordingtoEq.
(2),Case2.
.
Sincethelemmaholdsattime,wehave.
Since,fromtheaboveinequalitywegetThus,thelemmaholdsattime.
Asinthebasis,decreaseslinearlywithatanytimebefore.
Thus,thelemmafollows.
REFERENCES05010015020001234x105Lengthofqueues(Bytes)Packets05010015020002468x104DifferencebetweenQueuesLengthPath0Path1Difference050100150200024x105LengthofQueues(Bytes)Packets050100150200012x104DifferencebetweenQueuesPath0Path1Difference0501001502000510x105LengthofQueues(Bytes)Packets050100150200024x105DifferencebetweenQueuesPath0Path1Difference0018-9545(c)2016IEEE.
Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seehttp://www.
ieee.
org/publications_standards/publications/rights/index.
htmlformoreinformation.
Thisarticlehasbeenacceptedforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.
Contentmaychangepriortofinalpublication.
Citationinformation:DOI10.
1109/TVT.
2016.
2581020,IEEETransactionsonVehicularTechnology[1]J.
Wang,H.
Zhu,andN.
J.
Gomes,"Distributedantennasystemsformobilecommunicationsinhighspeedtrains,"IEEEJ.
SelectedAreasinCommn.
,vol.
30,no.
4,pp.
675-683,May2012.
[2]NationalRailwayAdministrationofthePeople'sRepublicofChina,"2014railwaystatisticalbulletin,"http://www.
nra.
gov.
cn/fwyd/zlzx/hytj/201504/t20150427_13281.
htm,Apr.
2015.
[3]USHighSpeedRailAssociation,"UShighspeedrailnetworkmap,"http://www.
ushsr.
com/ushsrmap.
html,2015.
[4]ChinaInternetNetworkInformationCenter,"35thChinaInternetnetworkdevelopmentstatestatisticalreport,"https://www.
cnnic.
cn/hlwfzyj/hlwxzbg/201502/P020150203551802054676.
pdf,Feb.
2015.
[5]Wi-Fi.
org,"TotalWi-Fideviceshipmentstosurpasstenbillionthismonth,"http://www.
wi-fi.
org/news-events/newsroom/total-wi-fi-device-shipments-to-surpass-ten-billion-this-month,January5,2015.
[6]D.
T.
FokumandV.
S.
Frost,"AsurveyonmethodsforbroadbandInternetaccessontrains,"IEEECommun.
SurveysTuts.
,vol.
12,no.
2,pp.
171–185,Apr.
2010.
[7]C.
Lee,M.
Chuang,andetal,"Seamlesshandoverforhigh-speedtrainsusingfemtocell-basedmultipleegressnetworkInterfaces,"IEEETrans.
onWirelessCommun.
vol.
13,no.
12,pp.
6619-6628,2014.
[8]V.
Devarapalli,R.
Wakikawa,A.
Petrescu,andP.
Thubert,"Networkmobility(NEMO)basicsupportprotocol,"IETFRFC3963,2005.
[9]M.
Biagi,S.
Rinauro,S.
Colonnese,G.
Scarano,andR.
Cusani,"DynamicnetworkselectioninV2Isystems,"inProc.
IEEEVehicularNetworkingConference(VNC),2013,pp.
190-193.
[10]BBCNews,"Wi-FiMayTemptTravellers,"http://http://news.
bbc.
co.
uk/2/hi/technology/3729583.
stm,2004.
[11]X.
Liang,F.
L.
C.
Ong,P.
M.
Chan,R.
E.
Sheriff,andP.
Conforto,"MobileInternetaccessforhigh-speedtrainsviaheterogeneousnetworks,"inProc.
14thIEEEProceedingsonPersonal,IndoorandMobileRadioCommun.
,vol.
1,2003,pp.
177-181.
[12]D.
HoandS.
Valaee,"Informationrainingandoptimallink-layerdesignformobilehotspots,"IEEETrans.
MobileComputing,vol.
4,no.
3,pp.
271-284,2005.
[13]D.
Huang,P.
Fan,andK.
B.
Letaief,"Anoptimalantennaassignmentstrategyforinformationraining,"IEEETrans.
WirelessCommun.
,vol.
7,no.
4,pp.
1134–1139,2008.
[14]C.
Sue,S.
Sorour,Y.
Yuk,andS.
Valaee,"Networkcodedinformationrainingoverhigh-speedrailthroughIEEE802.
16j,"inProc.
IEEEPIMRC,2009,pp.
1138–1142.
[15]J.
Zhang,Z.
Tan,Z.
Zhong,andY.
Kong,"Amulti-modemulti-bandandmulti-system-basedaccessarchitectureforhigh-speedrailways,"inProc.
IEEEVTC-Fall,Ottawa,Canada,Sep.
2010,pp.
1–5.
[16]R.
MoskowitzandP.
Nikander,"Hostidentityprotocol(HIP)architecture,"IETFRFC4423,2006.
[17]E.
NordmarkandM.
Bagnulo,"Shim6:level3multihomingshimprotocolforIPv6,"IETFRFC5533,2009.
[18]D.
Farinacci,V.
Fuller,D.
MeyerandD.
Lewis,"Locator/IDseparationprotocol(LISP),"IETFRFC6830,2013.
[19]D.
Anipko,"MultipleProvisioningDomainArchitecture,"IETFRFC7556,2015.
[20]P.
Rodriguez,R.
Chakravorty,J.
Chesterfield,I.
Pratt,andS.
Banerjee,"MAR:acommuterrouterinfrastructureforthemobileInternet,"inProc.
2ndinternationalconferenceonMobilesystems,applications,andservices,2004,pp.
217-230.
[21]D.
S.
Phatak,andT.
Goff,"AnovelmechanismfordatastreamingacrossmultipleIPlinksforimprovingthroughputandreliabilityinmobileenvironments,"inProc.
INFOCOM,2002,vol.
2,pp.
773-781.
[22]D.
S.
Phatak,T.
Goff,andJ.
Plusquellic,"IP-in-IPtunnelingtoenablethesimultaneoususeofmultipleIPinterfacesfornetworklevelconnectionstriping,"ComputerNetworks,vol.
43,no.
6,pp.
787-804,2003.
[23]K.
ChebroluandR.
R.
Rao,"Bandwidthaggregationforreal-timeapplicationsinheterogeneouswirelessnetworks,"IEEETrans.
onMobileComput.
,vol.
5,no.
4,pp.
388-403,2006.
[24]A.
J.
NicholsonandB.
D.
Noble,"Breadcrumbs:forecastingmobileconnectivity,"inProc.
ACMMobiCom,2008,pp.
46-57.
[25]J.
Yao,S.
S.
Kanhere,andM.
Hassan,"ImprovingQoSinhigh-speedmobilityusingbandwidthmaps,"IEEETrans.
onMobileComputing,vol.
11,no.
4,2012,pp.
603-617.
[26]R.
Stewart,"Streamcontroltransmissionprotocol,"IETFRFC4960,2007.
[27]R.
Stewart,M.
Ramalho,Q.
Xie,andetal,"streamcontroltransmissionprotocol(SCTP)partialreliabilityextension,"IETFRFC3758,2004.
[28]C.
-M.
HuangandC.
-H.
Tsai,"WiMP-SCTP:Multi-pathtransmissionusingstreamcontroltransmissionprotocol(SCTP)inwirelessnetworks,".
inProc.
AINAWorkshops,2007,pp.
209–214.
[29]J.
Liao,J.
Wang,andX.
Zhu,"cmpSCTP:AnextensionofSCTPtosupportconcurrentmulti-pathtransfer,"inProc.
ICC,2008,pp.
5762-5766.
[30]A.
Ford,C.
Raiciu,M.
HandleyandO.
Bonaventure,"TCPextensionsformultipathoperationwithmultipleaddresses,"IETFRFC6824,2013.
[31]A.
Ford,C.
Raiciu,M.
Handley,S.
BarreandJ.
Iyengar,"ArchitecturalguidelinesforMultipathTCPdevelopment,"IETFRFC6182,2011.
[32]C.
Raiciu,S.
Barre,C.
Pluntke,A.
Greenhalgh,D.
Wischik,andMarkHandley,"Improvingdatacenterperformanceandrobustnesswithmultipathtcp,"ACMSIGCOMMComputerCommunicationReview,vol.
41,no.
4,2011,pp.
266-277.
[33]C.
Raiciu,C.
Paasch,S.
Barre,A.
Ford,M.
Honda,F.
Duchene,andM.
Handley,"HowhardcanitbedesigningandimplementingadeployablemultipathTCP,"inProc.
9thUSENIXconferenceonNetworkedSystemsDesignandImplementation,2012,pp.
29-29.
[34]D.
Zhou,W.
Song,andM.
Shi,"GoodputimprovementformultipathTCPbycongestionwindowadaptationinmulti-radiodevices,"inProc.
CCNC,2013,pp.
508-514.
[35]J.
Zhao,C.
Xu,J.
Guan,andH.
Zhang,"AfluidmodelofmultipathTCPalgorithm:Fairnessdesignwithcongestionbalancing,"inProc.
ICC,2015,pp.
6965-6970.
[36]P.
Amer,M.
Becke,andetal,"LoadSharingfortheStreamControlTransmissionProtocol(SCTP),"draft-tuexen-tsvwg-sctp-multipath-11,2015.
[37]N.
Kuhn,E.
Lochin,A.
Mifdaoui,G.
Sarwar,O.
Mehani,andR.
Boreli,"DAPS:Intelligentdelay-awarepacketschedulingformultipathtransport,"inProc.
ICC,2014,pp.
1222-1227.
[38]P.
Dong,X.
Du,T.
Zheng,H.
Zhang,"ImprovingQoSonhigh-speedvehiclebymultipathtransmissionbasedonpracticalexperiment,"inProc.
2015IEEEVehicularNetworkingConference(VNC),2015,pp.
32-35.
[39]iPerf/iPerf3,"iPerf-Thenetworkbandwidthmeasurementtool,"https://iperf.
fr/.
PingDongiscurrentlyanassociateprofessoratBeijingJiaotongUniversity.
Dr.
DongreceivedhisPh.
D.
degreeinCommunicationsandInformationSystemfromBeijingJiaotongUniversity,Beijing,China,in2009.
HeworkedasavisitingscholarinTempleUniversity,USA,in2015.
HisresearchinterestsincludemobileInternet,software-definednetworkingandnetworksecurity.
Hehaspublishedmorethan40researchpapersintheseareas.
BinSongiscurrentlyaProfessorintheSchoolofTelecommunicationsEngineeringatXidianUniversity.
HeisalsotheassociatedirectorofStateKeyLaboratoryofIntegratedServicesNetworks.
HereceivedtheB.
S.
,M.
S.
,andPh.
D.
degreesincommunicationandinformationsystemsfromXidianUniversity,Xi'an,China,in1996,1999,and2002,respectively.
Hisresearchinterestsandareasofpublicationincludewirelessnetworks,computernetworks,andvideocompressionandtransmissiontechnologies.
Hehasauthoredover50journalpapersorconferencepapersand30patents.
0018-9545(c)2016IEEE.
Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seehttp://www.
ieee.
org/publications_standards/publications/rights/index.
htmlformoreinformation.
Thisarticlehasbeenacceptedforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.
Contentmaychangepriortofinalpublication.
Citationinformation:DOI10.
1109/TVT.
2016.
2581020,IEEETransactionsonVehicularTechnologyHongkeZhangiscurrentlyaprofessorintheSchoolofElectronicandInformationEngineeringatBeijingJiaotongUniversity.
HecurrentlydirectsaNationalEngineeringLaboratoryonNextGenerationInternetinChina.
Dr.
ZhangreceivedhisPh.
D.
degreeinfromUniversityofElectronicScienceandTechnologyofChina,Chengdu,China,in1992.
Hisresearchhasresultedinmanyresearchpapers,books,patents,systemsandequipmentintheareasofcommunications,computernetworks.
Hehasservedontheeditorialboardofseveralinternationaljournals.
Xiaojiang(James)DuiscurrentlyanassociateprofessorintheDepartmentofComputerandInformationSciencesatTempleUniversity.
Dr.
DureceivedhisB.
E.
degreefromTsinghuaUniversity,China,in1996,andhisM.
S.
andPh.
D.
degreesfromtheUniversityofMaryland,CollegePark,in2002and2003,respectively,allinelectricalengineering.
Hisresearchinterestsaresecurity,systems,wirelessnetworks,andcomputernetworks.
Dr.
Duhaspublishedover180journalandconferencepapersintheseareas,andhasbeenawardedmorethan$5MinresearchgrantsfromtheU.
S.
NationalScienceFoundation,ArmyResearchOffice,AirForce,NASA,andAmazon.
Heservesontheeditorialboardofthreeinternationaljournals.