March杭州wifi

MileHighWiFi:AFirstLookAtIn-FlightInternetConnectivityJohnP.
RulaAkamai/NorthwesternUniversityjohn.
rula@eecs.
northwestern.
eduJamesNewmanNorthwesternUniversityjnewm@eecs.
northwestern.
eduFabiánE.
BustamanteNorthwesternUniversityfabianb@eecs.
northwestern.
eduArashMolaviKakhkiNortheasternUniversityarash@ccs.
neu.
eduDavidChoffnesNortheasternUniversitychoffnes@ccs.
neu.
eduABSTRACTIn-FlightCommunication(IFC),availableonagrowingnumberofcommercialflights,isoftenreceivedbyconsumerswithbothaweforitsmereavailabilityandharshcriticismforitspoorperformance.
Indeed,IFCprovidesInternetconnectivityinsomeofthemostchallengingconditionswithaircrafttravelingatspeedsinexcessof500mphat30,000feetabovetheground.
Yet,whileexistingservicesdoprovidebasicInternetaccessibility,anecdotalreportsranktheirqualityofserviceas,atbest,poor.
Inthispaper,wepresentthefirstcharacterizationofdeployedIFCsystems.
Usingover45flight-hoursofmeasurements,weprofiletheperformanceofIFCacrossthetwodominantaccesstechnolo-gies–directair-to-groundcommunication(DA2GC)andmobilesatelliteservice(MSS).
WeshowthatIFCQoSisinlargepartde-terminedbythehighlatenciesinherenttoDA2GCandMSS,withRTTsaveraging200msand750ms,respectively,andthatthesehighlatenciesdirectlyimpacttheperformanceofcommonapplicationssuchaswebbrowsing.
WhileeachIFCtechnologyisbasedonwellstudiedwirelesscommunicationtechnologies,ourfindingsrevealthatIFClinksexperiencefurtherdegradedlinkperformancethantheirtechnologicalantecedents.
Wefindmedianlossratesof7%,andnearly40%lossatthe90thpercentileforMSS,6.
8xlargerthanrecentcharacterizationsofresidentialsatellitenetworks.
WeextendourIFCstudyexploringthepotentialofthenewlyreleasedHTTP/2andQUICprotocolsinanemulatedIFCenviron-ment,findingthatQUICisabletoimprovepageloadtimesbyasmuchas7.
9times.
Inaddition,wefindthatHTTP/2'suseofmulti-plexingmultiplerequestsontoasingleTCPconnectionperformsupto4.
8xworsethanHTTP/1.
1whenfacedwithlargenumbersofobjects.
Weusenetworkemulationtoexploreproposedtech-nologicalimprovementstoexistingIFCsystemsfindingthathighlinklosses,andnotbandwidth,accountforthelargestfactorofperformancedegradationwithapplicationssuchaswebbrowsing.
CCSCONCEPTSNetworks→Networkexperimentation;Networkmeasure-ment;Mobilenetworks;ThispaperispublishedundertheCreativeCommonsAttribution4.
0International(CCBY4.
0)license.
AuthorsreservetheirrightstodisseminatetheworkontheirpersonalandcorporateWebsiteswiththeappropriateattribution.
WWW2018,April23–27,2018,Lyon,France2018IW3C2(InternationalWorldWideWebConferenceCommittee),publishedunderCreativeCommonsCCBY4.
0License.
ACMISBN978-1-4503-5639-8/18/04.
https://doi.
org/10.
1145/3178876.
3186057KEYWORDSIn-FlightConnectivity1INTRODUCTIONInjustafewyears,ubiquitousconnectivityhasmovedfromavisionstatementtoanassumedrealityinmuchofthedevelopedworld.
Leveragingthisexpectation,severalairlinesofferin-flightconnec-tivity(IFC)amongtheirextraamenitiesoncommercialflights.
Attheendof2015,72airlineshadalreadyinstalledorannouncedplanstoinstallpassengerconnectivitysystemsonboard,andthenumberofconnectedcommercialaircraftisexpectedtogrow5xoverthe2015-2025period,toreach62%oftheglobalfleet.
In2017,thereareover56airlinesthatofferWiFiasaserviceaccordingtoapopularfrequentflyers'website[5].
Sincefirstappearingonthemarketinlate2004[10],IFChasgrowntobecomeakeyfeatureofflightsformanypassengersandanimportantcomponentofrevenueforairlines[20].
PassengersarereportedtoconsiderIFCwhenmakingtraveldecisions.
AHoney-wellsurveyfoundthat85%ofpassengersusedIFCin2013-2014and66%ofthemselectedflightsbasedonIFCavailability[2].
AccordingtoarecentsurveyfromInmarsat,61%ofpassengersconsiderWiFimoreimportantthanin-flightentertainmentand40%rankitasoneofthetop-3driversforairlinechoice[14].
A2016marketreportfromEuroconsultstatesthattotalrevenuefrompassengerconnec-tivityservicesareexpectedtogrowfrom$700millionin2015tonearly$5.
4billionby2025,a23%compoundannualgrowthrate(CAGR)overthe10-yearperiod[9].
Beyondpassengerconnectivityandairlines'revenue,IFCtechnologiesarebeingproposedasthebasisforfutureiterationsofcriticalaviationinfrastructuresuchasair-trafficmanagementsystems[18,24].
DespitethegrowingimportanceofIFCandthemanyinter-estingchallengesfacedbythistechnology,welackevenabasicunderstandingofcurrentandpotentialperformanceofthediffer-entapproachesinuse.
Wetakeaninitialstepinthisnewdomain,presentingthefirstcharacterizationofdeployedIFCsystemsandevaluatingthepotentialbenefitsofnewprotocolsandtechnolo-giesinthespace.
WefindthatIFCtechnologiesexperiencehigherlatencyvariance,andsignificantlyhigherlossratesthantheirterres-trialcounterparts,andthatnoveloptimizationsareneededtoimproveperformanceinsuchchallengingarea.
Weshowthroughemulatedexperimentsthatprotocoloptimizationprovidesthemostimmedi-atepathforwardforimprovingIFCexperience.
Thispapermakesthefollowingcontributions:Figure1:TechnologyalternativesforIFC.
DirectAir-to-GroundCommunication(DA2GC)utilizescellularconnec-tivitytogroundstations,andMobileSatelliteService(MSS)connectsthroughsatelliterelays.
WeprofiletheperformanceofcompetingIFCtechnologies,acrossthetwodominantaccesstechnologies–directair-to-groundcommunication(DA2GC)andmobilesatelliteservice(MSS)–usingover45flight-hoursofmeasurements,over16flightsandsixdifferentairlines.
Ourcharacterizationincludeslinkproperties,suchaslatency,lossandthroughput,aswellasapplicationperformanceofDA2GCandMSS.
Wefindthatlargelast-milelatenciesofIFCtechnologies–750msforMSSand200msforDA2GC–adverselyimpacttheperformanceofcommonapplicationssuchaswebbrowsing.
Inaddition,therelativelyhighlossratesforIFC,averaging3.
3%forDA2GCand6%forMSS,greatlydeterioratestheperformanceofTCP-basedcommunication.
WefindthatwhileeachIFCtechnologyisbasedonwellstudiedwirelesscommunicationtechnologies,IFClinksex-periencedegradedlinkperformancewellbeyondtheirtech-nologicalantecedents.
Wefindmedianlossratesof7%,andnearly40%lossatthe90thpercentileforMSS.
WepresentananalysisofcachingpoliciesondeployedIFCsystems.
Wefindthatin-flightcachingofDNSandHTTPcanofferlargeperformanceimprovements;however,currentimplementationssufferfromineffectivecachingpolicies.
Usingempiricallyderivedemulationweevaluatetheperfor-manceoftherecentlyreleasedHTTP/2andQUICprotocols.
WeshowthatHTTP/2provideslittleperformancebene-fitunderthehighlatencyandlossconditionsofIFC,andsubstantiallyworsewithlargenumbersofobjects,duetoHTTP/2'sknownHeadofLine(HOL)blockingwithTCP'scongestioncontrol.
QUIC,ontheotherhand,offerssignif-icantimprovementsofuptoa7.
9xreductioninPLTonexistingIFCtechnologies.
Weexplore,throughemulation,thepotentialbenefitsofnext-generationIFC.
WefindthatincreasingbandwidthdoeslittletoimprovePLToverexistingtechnologies.
Inthenextsectionwedescribethetechnologiesbehindtoday'sIFC.
Wepresentourmethodologyanddatasetofin-flightperfor-mancein§3,anddiscussourfindingsin§4and§5.
In§6wediscussouremulationexperimentsandtheirresults.
Finally,webrieflydescriberelatedwork,beforesummarizingandconcludingin§9.
2IN-FLIGHTCOMMUNICATIONIFCsystemscanbedividedintotwomaingroupsbasedontheirunderlyingtechnologies:thecellular-basedDirectAir-To-GroundCommunication(DA2GC)andsatellite-basedMobileSatelliteService(MSS).
DA2GCincludesthecommonlydeployed2/3GtechnologyandnewlyproposedLTE-basedservices,whileMSSoperatesovertheKuandKasatellitebands[7].
Thefollowingparagraphsprovidesomeadditionalbackgroundinformationonthedifferenttechnolo-gies,highlightingtheirinherentdifferences,andsomeofthemainIFCproviderstoday.
2.
1DirectAirToGroundCommunicationDirectAir-To-GroundCommunication(DA2GC)utilizescellulartechnologytolinktheplaneandtheground.
Thesesystemsareimplementedusingthreekeyinfrastructurepieces:theAircraftStation(AS),theGroundStation(GS)andtheDA2GCnetworkcore(Fig.
1).
Theaircraftstationconsistsoftheradioreceiverandtransmitter,aswellasnetworkappliancesforhandlingin-flightentertainmentsystemscommononmanyaircraft.
GroundStationsaretowersthatcommunicatewithpassingflights.
Thesestationsaresimilartocellulartowers,withtheexceptionthattheirradiotransmittersaredirectedupward,andthattheyareplacedwithmuchagreaterdistancesbetween(e.
g.
50to150kmradius).
DA2GCsystemsalsooperatetheirowncorenetworks,analogoustomoderncellularnetworks,thathandleaircraftmobilityandtowerhand-offs.
TrafficfromflightsisreceivedbyeachGS,andtunnelledthroughtotheDA2GC'scorenetworkbeforeegressingintothepublicInternet.
ExistingDA2GCsystemsoperateon2/3Gcellulartechnologiesfortheair-to-groundlink.
AlthoughsystemsusingnewerLTEtechnologyhavebeenproposed[6,8],nonehavebeendeployedasofJune2017.
DA2GCsystemshavebeensuccessfullydeployedinNorthAmer-icaandChina.
IntheU.
S.
,GoGoBiz[12](formerlyAircell)oper-atesanationwidenetworkoftowersprovidingconnectivityinthecontinentalU.
S.
,availablesince2008.
DA2GCsystemshavebeenproposedintheEU,buthaveyettobedeployedasofJune2016[8].
CurrentlyeachDA2GCsystemrunson3Gcellulartechnology,andwhileLTEbasedsystemshavebeendevelopedandtested,theyhaveyettobedeployedcommercially.
2.
2MobileSatelliteServiceMobileSatelliteService(MSS)utilizesgeostationarysatelliterelaystoestablishconnectivitybetweenaircraftandgroundstations.
MSSprovidersoftenleaseafractionoftheavailablebandwidthfromexistinggeostationarysatelliteInternetproviders.
Duetothelargedistancesnecessarytoreachgeostationarysatellites,MSSrequiresprecisedirectionaltransmissiontosuccessfullyachieveconnectiv-ity.
MSS-equippedaircraftaretypicallyoutfittedwithamechanicaldirectionalantenna.
Thus,underturbulentconditions,MSSoftenlosesconnectivityasitsantennalosesitstrackingposition.
WhileMSSconnectivityisnotrestrictedtoareaswithgroundtowers,theyarestillsubjecttogeographiccoverageconstraints,andmustalsoperformconnectivityhand-offs.
Duetothelargedis-tancestraversedbywirelesssignalsinsatellitecommunication,andthelargepath-fadingeffectsoftransmission,satellitetransmissionsaredividedintoseveralbeamsofafewdegreesoflatitudeandlon-gitude.
ThismeansthatMSSaircraftmustalsoperformhandoversastheycrossbetweenindividualbeamboundaries,similartothehandovermadebyDA2GCastheplanetravelsbetweengroundstationboundaries.
MSSisprovidedbyseveralcompanies,withViaSat,PanasonicAviation,Inmarsat,Row44,GoGoandDeutscheTelecomprovidingalargeshareofMSS-basedIFC.
ThemajorityofservicesofferedtodayareKubandservices,availablefromPanasonic,ViaSat,GoGo,andRow44,withemergingKabandsystemsprovidedbycompaniessuchasInmarsat.
3MEASUREMENTMETHODOLOGYWeusedatacollectedfromflightswithInternetconnectivityfromFebruary2015untilMarch2016,usingatestbenchwedevelopedforthiswork.
Ourdatasetconsistsofmorethan45hoursofIFCon16flights(fromsixdifferentairlines)equippedwitheitherMSS(13)andDA2GC(3)technologiesfromfivedifferentIFCproviders.
Thetestbenchconductsaseriesofnetworkmeasurementstochar-acterizetheperformanceandreliabilityofIFCservices,issuingpings,DNSrequests,HTTPrequests,andtraceroutes.
Resultsarerecordedonlocalstorageandtransferredtoouranalysisserversaftertheflight.
Tounderstandthelatencyandloss,wemeasuredpinglatencytowww.
google.
comevery2seconds.
WechoseGooglebecausetheirserversarecloseto(orinside)mostISPs,andtheyofferhighavailability,sotheperformancemeasuredislikelytobeabestcaseforlatencyandloss.
Expandingtherangeofmeasurementtargetsispartoffuturework.
Thetestbenchissuedtraceroutestowww.
google.
comevery3minutes.
Mostofthetraceroutesweperformeddidnotreachtheirdestination,withprobesbeingdroppedafter3-5hops.
1Regardless,informationaboutthefirstfewhopsallowustomeasure802.
11performanceforthein-planewirelesslink.
ForDNS,weperformedqueries,every5minutes,usingthede-faultresolver(obtainedviaDHCP)totheAlexatop100sites.
WechosethetopAlexasitestocharacterizetheperformanceimpactofDNSforthesitesthatusersaremostlikelytovisit.
ForHTTP,thesoftwareissuesGETrequeststoAlexatop100sitesusingPhan-tomJS[23],generatinganHTTPArchive(HAR)foreachsite.
TheHARfilecontainssufficientinformationtoidentifypageloadtimesandotherimportantWebperformancemetrics.
Last,thetestbenchrunsNetworkDiagnosticTests(NDT)[19]every5minutes.
ThetestsweredirectedtothenearestserverthatsupportsNDTtests.
Amajorityofthesetestsfailedduetotimeoutsinsignalingpackets.
Inaddition,wemappedthecollectedIFCmeasurementstoflightlocationusingdataonflightgeographicpositionobtainedfromflightaware.
com;wepartiallyrelyonthiswheninterpretingourresults.
Wedeterminetheaccesstechnologyofeachflightbylookingattheminimumpinglatencyrecordedduringeachmeasurement1Usingdelay-basedinference,webelievefilteringoccursonthegroundstations;wehavenotyetbeenabletovalidatethis.
ProviderASNCarriersEquippedPanasonicAvionicsASN39996,ASN22351UnitedAirlinesGoGoASN11167U.
S.
Airways,DeltaAir-lines,UnitedAirlinesRow44ASN6621SouthwestAirlinesT-MobileASN3320AmericanAirlines,LufthansaAirlinesViaSatASN7155UnitedAirlinesTable1:IFCprovidersinourdataset.
Airlinesoftenusemul-tipleprovidersandtechnologies,acrosstheirfleets.
(a)DA2GC(b)MSSFigure2:Latenciestowww.
google.
comseparatedbyaccesstechnology.
MSSlatenciesaresignificantlylargerduetothelargermilestraveledbypacketstoandfromasatellite.
period.
Inlightofthelargespeed-of-lightdelay(>470ms),weclas-sifiedflightswithminimumlatenciesbelow400msasDA2GC,andthoseaboveasMSS.
Wefoundminimumlatenciesbetween50.
7and93.
2ms,andforMSSflightswefoundminimumlatenciesrangingbetween536.
1and682ms.
3.
1IFCProviderCoverageTodeterminetheIFCproviderforaparticularflight,weusedtheclient'sIPaddressfromourtestbenchclientwhichweperiodicallyrecordedthroughanIPechoservice.
WethenmappedeachIPaddresstoanASNusingpWhoIsdata[1].
WewereabletoidentifyeachIFCproviderbyitsASmapping,inallfinding6uniqueASesforthe5providers.
Table1presentsasummaryofproviders,theirASnumbersandcarriersthatusedthem.
AsmallnumberofIFCoperatorsprovideservicetoamajorityoftheairlineindustry.
TheseprovidersdifferbothinthetechnologytheyusetoprovideIFCservice,andinthecaseofMSS,thesatellitetechnologyused.
ThesetofIFCprovidersincludedinourmeasure-mentdatasetcapturesnearly94%ofoverallIFCmarketshare[22].
ThisincludesGoGoat53.
1%,PanasonicAviationat18%,ViaSatat12.
7%andRow44at10.
3%ofmarketshare.
4IFCLINK-LEVELPERFORMANCEInthisandthenextsectionwepresentourcharacterizationofIFCperformance.
Wefirstdiscusslink-levelpropertiesbeforelookingatapplication-levelperformance.
Giventhedominantroleoftheunderlyingtechnology,weaggregateresultsforDA2GG-andMSS-supportedflights,independentlyofIFCprovidersandairlines.
Figure3:Cabinlatency(802.
11latency)inIFC,measuredbytheRTTtothefirsttraceroutehopforeachflight.
4.
1LatencyWefindthatlatencyislargelydeterminedbytheIFCtechnology,withMSSlatencynearlyanorderofmagnitudelargerthanthatofDA2GC.
Figure2plotsthedistributionofallpingprobestowww.
google.
comforallflightsinourdataset,aggregatedbyIFCtechnology.
ThefigureshowsthelargedisparityinlatencybetweenDA2GCandMSS,withanearly500msmeandistancebetweenthedistributions,andaminimum(average)latenciesforDA2GCrang-ingbetween50-93ms(260-310ms)comparedwithMSSminimum(average)latenciesinthe530-680ms(730-1100ms)range.
Thisisnotsurprisingwhenoneconsidersthatmostcommuni-cationsatellitesareingeostationaryorbit,22,000milesabovetheEarth'ssurface.
Thec-latencytogroundviaasatellite,i.
e.
,thetimeforlighttotravelthefour-legtripfromplanetosatellitetoground(Fig.
1),andbackisnearly500milliseconds.
Comparedtoitsterrestrialcounterpart,wefindthatMSSperfor-manceissignificantlymorevariable.
UtilizingpublicdatafromtheFCCBroadbandAmericastudy[11],fromSeptember2015,wefindthattherttswereontheorderof1.
2to1.
9timeshigher,comparedtothe599-640msonaverageforterrestrialsatellitebroadbandconnections.
Wefindthatthelatencyvariancewassubstantiallyhigher,withaverageterrestrialstandarddeviationsrangingfrom31.
9to43.
1ms,andMSSstandarddeviationsaveraging333ms,andrangingfrom159to707ms.
Giventhedense,confinedspaceofairlinecabins,wewonderedif802.
11latencieswerecontributingtothetaillatencyofIFC.
Wefoundthatthemajorityofthetime,WiFidelaysinthecabintypi-callycontributelittletotheend-to-endlatency,plottedinFigure3.
WecalculatethiscabinlatencybymeasuringtheRTTtothefirsttraceroutehop,assumedtobetheWiFirouter.
Wefindthat,whilethelatencydistributionshowsalongtailwithapproximately5-10%ofthemeasurements,stretchingto10sofmilliseconds(wellbelowthe100sofmillisecondsseeninIFClatencie),fornearly70%ofprobesthecabinWiFicomponentoflatencyisbelow2ms.
4.
2LossWemeasurepacketlossbysendingapingevery2secondsandtakingtheaveragefractionofpingstowww.
google.
comthataredroppedevery100seconds(50pings)WeplotthedistributionofpacketlosspercentageforthetwotechnologiesinFigure4.
TheplotshowstheconsiderablehigherreliabilityofDA2GCoverMSS.
Nearly75%ofDA2GCtestshave0%packetloss,whilefewerthan12%ofMSStestsseethesame.
Inaddition,alloftheDA2GCtestsfallbelow30%packetloss,while10%ofMSStestsFigure4:DistributionofpacketlosspercentageforMSSandDA2GC.
haveover30%loss.
ThisdisparityinpacketlossdemonstratesthechallengesfacedbyMSSpackets,includinghighspeedsandaltitude.
Theimprecisionofdirectionaltransmissiontoasatellite22,000milesawaywhiletravelingathundredsofmilesperhourisalikelycauseofthepacketlossexperiencedonMSSflights.
WhenwecompareMSSperformancetoitsterrestrialcounter-parts,wefindpacketitspacketlossratesare6.
8timeslarger,withterrestrialbroadbandaveraginglossratesof1.
38%,comparedtothe9.
4%averagelossratesfoundinourdataset.
4.
3ThroughputAswithlatency,throughputvariesbasedonthetechnologyusedforcommunication.
AsFigure5(a)shows,downstreamthroughputforDA2GCrangesfrom100kb/sand800kb/s,whilethroughputforMSSisevenmorevariable,butcanachieveuptotwoordersofmagnitudelargerratesthanDA2GC.
Further,theaverageandmediandownloadspeedsforMSSarelargerthanDA2GC.
Ontheotherhand,fornearlyathirdofthesamples,wefindthatMSSofferslowerdownstreamthroughputthanDA2GC.
Infact,MSSnearlyalwaysoffersatleast10kb/s,whileMSSprovidessmallfractionsofdial-upspeedsforasignificantnumberofsamples.
WeseesimilartrendsforupstreamthroughputinFigure5(b).
Themaindistinctionisthatthepeakthroughputandwidthofthethroughputdistributionsarebothsmaller,clearlyindicatingasymmetricbandwidthallocation.
WespeculatethatthelargevarianceinthroughputispartiallyduetothedifferentMSSbandsinuse,suchasKuandKaband,whichhavedifferentthroughputcapabilities.
Additionally,thehighpathlossalongthepathbetweentheplaneandsatellitecancontributetohighlossesthatlimitTCPthroughput.
4.
4GeographicConsiderationsGiventhelargedistancesanddiversegeographycoveredbycom-mercialflights,wewouldexpecttoseeaneffectofgeographiclocationonIFCperformance.
Theextenttowhichgeographyim-pactsnetworkperformancedepends,asexpected,ontheparticularaccesstechnology.
TherelianceofDA2GCongroundstations,forinstance,makesitmoresusceptibletotheparticularcoverageofthesetowersandtovariationsintheunderlyinggeography(e.
g.
,mountains).
Figure6displaysthelatencyandpacketlossexpe-riencedinflight,overlaidontothegeographicflightpathforarepresentativeflightforeachaccesstechnology.
Foragivengeo-graphiccoordinateonthebluecurveinthemap(middle),weplotthelatency(top)andloss(bottom)measuredatthatlocationusing(a)ServertoClient(b)ClienttoServerFigure5:Distributionsoftheserver-to-clientandclient-to-serverthroughput.
DA2GCsystemsproviderelativelyconsistentthroughput,whileMSSsystemsexhibithighlyvariableperformanceandhigherpeakthroughput.
(a)AWE-534(DA2GC)(b)UA-627(MSS)Figure6:Networklatencyandpathlossplottedbylatitudeandshowninrelationtotwoflightpaths.
WeobservegreatercorrelationsbetweengeographiclocationandperformancecharacteristicsforDA2GCthanwedoforMSS.
thesamex-coordinate(i.
e.
,alongaverticallinethatintersectsthelocationonthemap).
Thefigureshowstherelativestabilityofsatelliteservicesovergeographicspace(right),whilecellulartechnologyshowsperfor-mancedegradationandlatencyandpacketlossspikescorrelatedtospecificgeographiclocations(left).
Inparticular,latencyandpacketlossspikewhiletheplanetraversesnorthernNewMexiconeartheArizonaandColoradoborders.
5APPLICATIONPERFORMANCEInthefollowingparagraphswefocusonDNSandHTTPperfor-mance.
5.
1DNSToevaluateDNSperformanceinflight,weusedthelocallycon-figuredresolver(assignedthroughDHCP)torepeatedlyresolvethelistofAlexatop100sites.
Foreachhostname,weperformedtwosequentialresolutionstomeasuretheeffectoflocalresolvercachingonperformance.
ThedistributionofDNSresolutiontimesforthefirstofthetwosequentialresolutionsisshowninFigure7.
Therearetwoclearmodesat10and150msforDA2GCand10and725msforMSS.
Thefirstmodeisgivenbyrequestsserveddirectlyfromthein-flightresolver'scache.
ThesecondmodeisdrivenbycachemissesthatrequirecontactingagroundDNSserver,andisprimarilydeterminedbythelargeaccesslatenciesbetweentheplaneandFigure7:DNSresolutiontimesforthefirstquery.
thegroundserver.
Thefigurealsoshowstheprobabilityofcachehitsonin-flightappliances.
Forinstance,lessthan36.
7%offirstresolutionstotheAlexaTop100werereturnedfromin-planecacheforDA2GCflights,and39.
3%ofresolutionsinthecaseofMSS.
Weexplorein-flightDNScachinginalatersection.
5.
2HTTPWemeasuredtheHTTPperformanceexperiencedbyIFCclients.
WeconductedourHTTPexperimentsusingPhantomJS[23]–afullyfunctionalheadlessbrowser,withfullsupportforJavascriptexecution.
ThenetworkbehaviorfromaPhantomJSshouldcloselyresemblethatofaGUI-drivenbrowsersuchasChromeorFirefox.
WeloadedtheAlexaTop100sitesinorder,repeatingwitha5min.
Figure8:PageloadtimeforAlexaTop100,aggregatedbyIFCtechnology.
(a)DA2GC(b)MSSFigure9:DownloadtimeforHTTPObjectstakenfromtheAlexaTop100sites.
Bothtechnologiesshowasmalldiffer-enceinthetimeittakestodownloada1byteobjectversusa100KBobject.
intervalbetweensubsequenttests.
Foreachpageretrieved,Phan-tomJScreatedanHTTPArchive(HAR)fileforlateranalysis.
FromthesetestswecapturetheonPageLoadeventsfromeachbrowser.
Thelargeaccesslatenciesofeachtechnology,andMSSinpar-ticular,causethesevereperformancedegradationonmodernWebpagescontaininglargenumbersofobjectsanddependencies.
Fig-ure8showstheabovetwoeventsplottedforallpagesrequested,aggregatedbyaccesstechnology.
ForflightsutilizingMSSconnec-tivity,themedianpageloadtimeismorethan30sec,andmorethan12secforDA2GCtechnology.
TheseinflatedloadtimesareduetodelaysfordownloadingindividualHTTPobjectsoneachsite.
Toexplorethis,wecapturethetimingsforeachobjectfromtheHARfilerecordedfromPhantomJSforeachHTTPrequest.
Figure9plotsthetimetoretrieveeachHTTPobjectacrosstheloadedpages,foreachtechnology.
Theplothighlightstheaccess-linkbottleneckforbothtechnolo-gies,evidentwhenconsideringthesmalldifferencebetweenthetimetodownloada1byteobjectversusa100KBobject.
Thisisparticularlyvisiblewithsatellitelinks(Fig.
9b),wheremostobjectfetchtimesarebetween1–10secondsregardlessofsize.
ThefigurealsoshowsalimitedamountofHTTPcachingonin-flightappliances,identifiedbycaseswherethetimetodownloadanobjectfallsbelowtheminimumRTTachievablefromeachaccesslink(500msinsatelliteand50msinDA2GC).
WecombineouranalysisofHTTP(andDNS)cachinginthefollowingsection.
5.
3CachingintheAirThelargelatenciesforallIFCsystemssuggestsapotentialforimprovementfromcachingobjectsonin-flightappliances.
WenowexploretheuseofcachingandpoliciesemployedonIFCsystemsaspartofourpreliminarystudy.
WedetectDNScachingonalloftheflightsinourdataset,andwereabletoexplicitlyverifyHTTPcachingin3flights.
ForDNSrequests,wedetecttheuseofin-flightcachingbycom-paringtheresponsetimeagainsttheminimumpingRTT.
WhileitispossiblethatacachedobjectmayhavealargerdownloadtimethanthisminimumRTTdueto802.
11delays,webelievethatthisscenarioisrareconsideringthelargeRTTsincurredforin-flightcommunication.
ForHTTPrequests,wedetectedcachingforindi-vidualobjectsthroughtheViaHTTPheader,whichsignalstheuseofproxiesinpath,whenaprivateIPaddresswasindicatedasthefinalproxytraversed.
DNSCaching.
WefirstexploretheuseofDNScachinginIFC.
Aspreviouslydescribed,welaunchedtwoback-to-backqueriesimmediatelyforeachhostintheAlexaTop100tocapturethepresenceofIFCDNScaches.
Wedetectin-flightDNScachingincaseswheretheresolutiontimewasbelowtheminimumpinglatencyfoundforeachflight,approximately60msforDA2GCand600msforMSS.
(a)DA2GC(b)MSSFigure10:SequentialDNSquerieslaunchedadjacentlydur-ingmeasurement.
Asignificantportionofqueriesarenotcachedbytheplane'slocalresolver.
Figure10plotstheresultsfromeachsequentialDNSresolution,withtheresolutiontimefromthefirstresponseplottedalongthex-axisandtheresolutiontimefromthesecondresponsealongthey-axis.
Thefiguresdisplayfourcasesofresolverbehavior,separatedintoquadrantsbyeachflight'sminimumpinglatency:(lower-left)bothqueriesservedfromin-flightcaches,(lower-right)cachemissforthefirstandcachehitforthesecondquery,(upper-left)cachehitforthefirstqueryandcachemissforthesecond,(upper-right)cachemissforbothqueries.
WefindthatfirstDNSqueriesarecachedin-flightfor33%and39%ofcasesofDA2GCandMSSflightsrespectively,andsecondquerieswerecachedin91.
4%ofDA2GCtestsand95.
6%ofMSStests.
However,thefigurealsoshowsalargefractioninwhichboththefirstandsecondquerieswerenotcached(7.
7%forMSSand4.
1%forDA2GC),locatedintheupper-rightquadrant.
Whilethisbehaviorhasbeenidentifiedinpreviouswork[3],itwasduetothepresenceofDNSclusterswithoutsharedcaches.
Wedonotbelievethattobethecasehere,andarecontinuingtoinvestigatethesourceofthisbehavior.
(a)DL-2374(DA2GC)(b)SW-2374(MSS)Figure11:HTTPobjectperformanceforobjectscachedin-plane(HIT)andthosenotcachedontheflight(MISS).
HTTPObjectCaching.
DespitethehighaccesslatenciesofIFC,wefoundlittleHTTPobjectcaching.
TodetectcachingforindividualHTTPobjectswerelyontheViaHTTPheader.
Thisheadersignalstheuseofproxiesinpath,arechainedtogetherforallproxiesusedinanHTTPobject'spath.
Proxiesthatidentifytheirpresencedosobyappendinganidentifier,typicallyusingtheproxyIPaddressandproxysoftware(e.
g.
,"squid"),anexampleofwhichisshownbelow.
Via:1.
0172.
19.
134.
2:3128(squid/2.
6.
STABLE14)Althoughitispossiblethatin-pathproxiesaretransparent(i.
e.
,theydonotidentifytheirpresencethroughthisheader),weidenti-fiedexplicitHTTPproxiesinthreeflightsinourdataset:DL-2374,SW-2374,andUAW-534.
InadditiontotheViaheader,manyprox-iesalsoappendstateabouttheircacheoperationsintheHTTPresponsethroughtheX-Cacheheader.
Fromthisheader,wecanidentifywhethertheobjectwasservedfromcacheviathe"HIT"or"MISS"indicators.
Figure11showstheHTTPobjectperformanceforobjectscachedin-plane(HIT)andthosenotcachedontheflight(MISS).
Whileob-jectscachedontheplaneresultinsignificantlybetterperformance(particularlyforMSS),theobjectdownloadtimesarestilllargerthanwouldbeexpectedbasedonWiFilatencyalone.
AninterestingobservationisthatforDL-2374,morethan80%ofthecacheHITstooklongerthan100ms,whichislongerthantheminimumping-basedInternetRTTforthatflight.
Itisunclearwhythisprocesstakessolongforobjectscachedaerially;exploringpotentialreasonsispartofongoingwork.
6COMPARINGIMPROVEMENTSInthissection,weextendourIFCstudyexploringthepotentialofalternativeapplicationprotocolsHTTP2andQUIC[13],andsomeproposedlinktechnologyimprovements.
Forthisanalysis,weuseemulationdrivenbyparametersderivedfromtworepresentativeflightsforeachIFCtechnology.
ForourDA2GCflightwemodeledthelinkwithaveragevaluesforDA2GC[BW=0.
468Mb/s,RTT=262ms,loss=3.
3%],andMSS[BW=1.
89Mb/s,RTT=761ms,loss=6%].
Weusetcandnetem[26]tomodelnetworkconditionsaftertheperformanceofin-flightWi-Fi.
Wedownloadeddifferentnumbersofwebpagesofincreasingsize,suchasapageconsistingofone100KBobjecttoapagecon-sistingoftwo500KBobjects.
Theobjectsaredownloadedoverthreedifferentprotocols:HTTP/1.
1,HTTP/2,QUIC.
Wemeasurethepageloadtime(PLT)ofthewebpagesover10testseachandgraphthethreeprotocolssidebysideforthesimulatednetworkconditionsoftheDA2GCandMSSflights.
6.
1ProtocolEvaluationWeanalyzethebenefitsofadoptingtheHTTP/2andQUICproto-cols,overHTTP/1.
1,tousersintheIFCenvironment.
PriorworkhasshownthatSPDY–theprotocolforwhichHTTP/2isbased–hadmixedperformanceresultsinhighlatencyandlossenviron-ments,duetohead-of-lineblockingofitssingleTCPconnectioninthefaceofdroppedpackets.
Wangetal.
showedthatinhighlatencyandlossenvironments,SPDYoftenperformedworsethantraditionalHTTP[27].
QUICissimilartoSPDYinthatitmultiplexesmanyrequestsintoasingleflow,yetwithlesspotentialforhead-of-lineblockingsincelossishandledseparatelywithineachrequestinthesameflow[17].
QUICcanalsobequickertostartthanTCPduetoits0-RTTfeaturewhichforrepeatedconnectionsdonotrequireaseparatehandshake.
Figure12showstheaverageperformanceforeachprotocoloversimulatednetworkconditionsofcompetingconnectiontechnolo-gies,witherrorbarsrepresentingthestandarddeviationoverthesamples.
Wefindthat,onaverage,QUICperformsbetterthantheothertwoprotocols,particularlywhenloadinglargerobjectssuchasa1MBobject.
Inaddition,QUICperformsbetterthanHTTP/1.
1andHTTP/2inbothofthesimulatedenvironments,capableofim-provingPLTtimesby50%.
AdoptingQUICforIFCsystemswouldimproveusers'QoEacrosstechnologiesandproviders.
6.
2PotentialofTechnologicalImprovementsLast,weevaluatepotentialtechnologicalimprovementsofexistingIFCtechnologies.
PlannedupgradestocurrentIFCinfrastructuresincludetechnologiessuchas2KusatellitesystemsandLTE-basedDA2GC.
Toevaluatethesenewtechnologies,andinanabsenceofactuallinkpropertiesofthesetechnologies,weexploretheparam-eterspaceforIFCtechnologiesbyperformingthesamenetworkemulationusedintheprevioussection,andvaryingasinglelinkparameter–doublingthelinkbandwidth,halvingpacketlossratesandhalvinglinklatencies–tounderstandwhichisthemostimpact-fulinimprovingfutureperformance.
Thisdoublingisderivedfromthenewlyreleased2Kusatellitetechnology,whichexactlydoublethebandwidthofexistingKusystemsthroughchannelbonding.
Figure13showstheperformanceofdifferentwebprotocolsoverthesehypotheticalnetworkconditions.
Ineachrowofthegrid,wemodifyasinglelinkparameterfromtheoriginallinkcharacteristicsusedintheprevioussectionfornetworkemulation–bandwidth,latencyandloss–andcomparetheperformanceofeachprotocoltotheexistingtechnology.
Indescendingroworder,wedoubledthebandwidthto3.
77and0.
936Mbps,decreasedthelatencyto380.
5msand131ms,decreasedthepacketlossto3%and1.
65%forMSSandDA2GCrespectively.
Ouremulationresultsfoundthatdoublingthebandwidthre-sultedinlittle,ifany,reductioninPLT.
ForourMSSemulation,increasingthebandwidthresultedina1.
6%overallreductioninPLTforHTTP,a5.
9%increaseforHTTP2anda0.
9%reductionforQUIC.
ForDA2GC,surprisingly,allthreeprotocolsyieldedslightlyhigherPLTs,onaverage,afterthebandwidthincreased,indicatingother(a)DA2GC.
(b)MSS.
Figure12:ResultsfromrunningHTTP,HTTP/2,andQUICoversimulatednetworkconditions.
Onaverage,objectsloadfasterwithQUICoverbothtechnologies,indicatingitwouldbeaviableoptionforadoptionregardlessofIFCtechnology.
propertiesaremoredeterminantofIFCperformance.
MuchoftheconversationrevolvingaroundIFCtechnologiesanduserqualityofexperiencehascenteredaroundthethroughputofthesetechnolo-gies,yetitisclearfromourexperimentsthatforIFC,improvingthroughputdoeslittletoimproveperformanceofapplicationssuchaswebbrowsing.
Incontrast,halvingthelatencyandlossofeachlinkgreatlyreducedloadtimes,withlossreductionhavingthelargestimpactforbothtechnologies.
ForDA2GC,reducinglatencybyhalfreducesloadtimesanaverageof88.
2%acrossthethreeprotocols,andreduc-ingthelossbyhalfaveragesa92.
5%reduction.
MSSexperiencedsimilarperformanceimprovementsacrossprotocols.
ForMSS,re-ducingthelatencyresultedina38.
2%decreaseinloadtimeforHTTP/1.
1andHTTP/2,whilehalvingthelossrateprovideda51%decreaseforbothHTTPprotocols.
WhilereducingthelatencyandlossforIFCtechnologiesmaybechallenging,theparameterexplorationpresentedallowsustoviewthemajorbottlenecksinexistingperformance.
Forinstance,HTTP/2performedconsistentlyworsethanHTTP/1.
1overbothIFCtechnologies.
ThevastperformanceimprovementsobtainedinthereducedlossenvironmentallowustodeducethatthehighlossofbothIFCtechnologiesisadverselyaffectingTCPperformance,andHTTP/2'smultiplexingofrequestsontoasingleTCPflowonlyexacerbatesthisproblem.
Similarly,thelosssensitivityofTCPpartlyexplainsthevastlyimprovedperformanceofQUICoverMSS,sinceQUICincorporatesimprovedlossrecoverymechanisms[17].
OurresultsmakeitclearthatexistingIFCperformancecanbegreatlyimprovedbyoptimizingotherlayersofthenetworkstack,suckasthetransportlayer,evenwithexistingtechnology.
7DISCUSSIONOurstudyshowsthatwhiletoday'sIFCsystemsprovidesufficientconnectivitytosupportconsumer-gradeInternetservice,thereremainsasignificantsetofchallengesforIFCperformanceandreliabilitytocomeclosetothosefromterrestrialwirelessandfixed-linenetworks.
WhilethephysicallinkremainsthebottleneckinexistingIFCdeployments,wefoundseveralalternativesolutionsthatcanvastlyimproveexistingperformance.
Forexample,themostpopulartrans-portprotocol,TCP,wasnotdesignedforthehighlatenciesandpacketlossratesofsatellitelinks.
However,wefindthatGoogle'sQUICtransportprotocol,whichhandlesvariablelatencyandlossmuchmoregracefullythanTCP,exhibitsgoodPLTperformance(2.
5xfasterthanTCP)inthechallengingIFCenvironment.
Thus,areasonableapproachtoimprovingIFCperformancetodayisforcon-tentproviderstoadoptthistechnology,orevenforthedeploymentofTCP-to-QUICproxiesforthesatellitelink.
WealsofoundthatimprovingDA2GCaccesstechnologieswillimprovePLTs;however,thisisnotthecasefornewerMSSbands.
Further,wenotethattheapplicationlayerprotocolcansignificantlyimpactperformance.
Specifically,HTTP/2overTCPperformssev-eraltimesworsethanHTTP/1.
1.
Thusitisclearthatcareful,end-to-endevaluationsofnewtechnologiesarerequiredtounderstandwhethertheywillyieldgainsintheIFCenvironment.
AsIFCperformanceandreliabilityimprove,thereispotentialtouseittoaugmentandextendexistingair-traffic-control(ATC)andair-traffic-management(ATM)systems[18,24].
Thishastheadvantagesofprovidingahighbandwidth,reliableandsecuredatalinkbetweenaircraft,ATCandcarriers,toimprovetheefficiencyandsafetyofairtravel.
Ourresultsindicatethattoday'sIFCpro-videssufficientcapabilitiestocomplementATC/ATMasdeployedtoday,andwebelievethatimprovementstoconnectivitywillfur-theropenthedoortoinnovationsandimprovementsintheairtravelecosystem.
7.
1ComparingIFCConsideringthevarietyofcarriers,IFCproviders,andaccesstech-nologies,thereisaquestionofhowoneshouldcompareIFCmea-surementstoprovidemeaningfulanalysis.
Aswehaveseen,theparticularsoftheaccesstechnologyplayadominantroleinnetworkcharacteristicsandserviceperformance.
Ouranalysisshowsthatwhileaccesstechnologylargelydeter-mineslinklatencies,throughputisacombinationofaccesstechnol-ogyandIFCproviderpolicies.
Accesstechnologyplaysadominantroleindeterminingthelatencyofin-flightcommunication.
Thedrasticdifferencesbetweenaccesstechnologylimitationsismostlyduetothehighspeed-of-lightdelayforsatellitescommunications.
EachMSSroundtripmusttravelover88,000miles(22,000x4)incurring472msoflighttravelforthislastmilealone.
WefindthroughputismoredependentontheIFCprovider,andtheirunderlyingpolicies,thanonparticularaccesstechnologies.
Accesstechnologydoesplaceupperboundsondatathroughput,however.
DA2GCtechnologiesarecurrentlylimitedto3.
2Mbps,whileKubandhavealimitof10MbpsandKa+bandhavealimitof(a)DA2GC.
(b)MSS.
Figure13:Performanceofthethreedifferentwebprotocolsovervariedemulatednetworkconditions(orange)comparedtoexistingnetworkconditions(blue).
Eachrowmodifiesasingleparameter,doublingbandwidth,halvinglatencyandhalvingloss,foreachIFCtechnology.
Figure14:DownstreamthroughputforeachIFCprovider.
100MbpsinMSS.
Ourresultsshowthatindividualproviderpolicieshaveamuchlargerimpactontheachievedthroughput.
Forinstance,wefoundMSSproviderRow44isunabletoprovidethroughputlargerthan100kbps,mostlikelyduetothrottlingappliedtoeachuser.
ResultsofNDTmeasuredthroughputgroupedbyeachIFCproviderinourdatasetareshowninFigure14.
8RELATEDWORKIn-FlightCommunicationisachallengingandmostlyunexploredareafornetworkingresearchatthetransportlayerandabove.
Ourearlierpositionpaper[24]reportedasubsetofpreliminaryresultsfromourexperiments.
Thisworkrepresents,tothebestofourknowledge,thefirstcomprehensivecharacterizationoftheperformanceofIFCdeployedsystems.
IFChasbecomepossibledueinparttotheadvancementsoflinklayertechnologiesinbothsatellitesystems[7,28]andinDA2GC[6].
Inpreviouswork,weadvocatedforashiftofairtrafficmanagementtoacommonIP-baseddatachanneltosupportflightcommunica-tion,andidentifiedseveralopportunitieswherethiscouldgreatlyincreasethescalabilityoftheglobalairlinesystem[24].
Withsimi-largoals,Ayazetal.
[4]describeaproposeddesignforaIPv6-basedAeronauticalTelecommunicationNetwork.
AnotherlineofresearchexploresthefeasibilityofconnectivitythroughairborneMANETs[15,16,21,25].
Whiletheseproposalshavethepotentialtoaddnewavenuesofconnectivityforaeronau-ticalnetworks,theyhaveyettobeimplemented.
9CONCLUSIONWepresentedthefirstcharacterizationofdeployedIFCsystemsusingover45flight-hoursofmeasurements,over16flightsandsixdifferentairlines.
OuranalysisshowsthatIFCQoSisinlargepartdeterminedbythehighlatenciesinherenttoDA2GCandMSSandthattheselatenciesdirectlyimpacttheperformanceofcommonapplicationssuchaswebbrowsing.
Inaddition,wefoundveryhighlinklossratesforIFC–nearly40%lossatthe90thpercentileforMSS–thatseverelyimpacttheperformanceofTCPandotherloss-basedcongestion-controlprotocols.
Weexploredthepotentialofalternativeprotocolsandupcomingtechnologyimprovements.
Usingempirically-informedemulationwefoundthattherecentlyreleasedHTTP/2protocolperformspoorlyduetotheaforementionedhighloss,whiletherecentQUICprotocoloutperformsHTTP/1.
1overTCPinlargepartduetoitsadvanceddelayinferenceandlossrecoverytechniques.
Thereareanumberoffuturedirectionsweplantoexplore,in-cludingevaluatingthepotentialbenefitsofhybridDA2GC/MSSsystems,alternativecontent-cachingstrategies,andexploringthetechnicalchallengesofsupportingreal-timeand/orhigh-bandwidthconnections,bothforrelayingcriticalflightinformationandsup-portingrichmediaapplications.
ACKNOWLEDGMENTSTheauthorswouldliketothanktheanonymousWWWreviewersfortheirvaluablecommentsandhelpfulsuggestions.
ThisworkwaspartiallyfundedbytheNationalScienceFoundation(CNS-1218287)andaGoogleResearchAward.
AllopinionsexpressedarethoseoftheauthorsanddononecessarilyreflecttheviewofNSForGoogle.
REFERENCES[1]2015.
ThePrefixWhoIsProject.
http://www.
pwhois.
org.
(2015).
http://www.
pwhois.
org[2]HoneywellAerospace.
2014.
InFlightConnectivitySurvey.
https://aerospace.
honeywell.
com/press-release-listing.
(July2014).
[3]BernhardAger,WolfgangMühlbauer,GeorgiosSmaragdakis,andSteveUhlig.
2010.
ComparingDNSResolversintheWild.
InProc.
IMC.
[4]SerkanAyaz,ChristianBauer,ChristianKissling,FrankSchreckenbach,FabriceArnal,CedricBaudoin,KatiaLeconte,MaxEhammer,andThomasGraeupl.
2009.
ArchitectureofanIP-basedAeronauticalNetwork.
InProc.
ofICNS.
IEEE,1–9.
[5]LetsFlyCheaper.
[n.
d.
].
AirlinesThatOfferInflightWiFi-TheDe-finitive2017List.
https://www.
letsflycheaper.
com/blog/airlines-2/airlines-that-offer-inflight-wifi/.
([n.
d.
]).
[6]Shun-PingChen.
2014.
PerfomanceanalysisandoptimizationofDA2GCusingLTEadvancedtechnology.
InProc.
VITAEw.
[7]PaoloChini,GiovanniGiambene,andSastriKota.
2010.
Asurveyonmobilesatel-litesystems.
InternationalJournalofSatelliteCommunicationsandNetworking28,1(2010),29–57.
[8]ElectronicsCommunicationsCommittee.
2014.
BroadbandDirect-Air-to-GroundCommunications(DA2GC).
TechnicalReportECC-214.
CEPT.
[9]Euroconsult.
2016.
PassengerConnectivityServicestoSurpass$5Billionby2025.
http://www.
euroconsult-ec.
com/4_February_2016.
(February2016).
[10]FastCompany.
[n.
d.
].
HowTerribleIn-FlightWi-FiWillFinallyBe-comeAThingOfThePast.
https://www.
fastcompany.
com/3042609/how-terrible-in-flight-wi-fi-will-finally-become-a-thing-of-the-past.
([n.
d.
]).
[11]FCC.
[n.
d.
].
MeasuringBroadbandAmerica.
http://www.
fcc.
gov/measuring-broadband-america.
([n.
d.
]).
[12]GoGo.
[n.
d.
].
GoGoBizWirelessNetworkforAircraft.
https://business.
gogoair.
com/technology/north-american-broadband-network/gogo-biz/.
([n.
d.
]).
[13]Google,QUIC.
[n.
d.
].
QUIC,amultiplexedstreamtransportoverUDP.
https://www.
chromium.
org/quic.
([n.
d.
]).
[14]Inmarsat.
[n.
d.
].
InflightConnectivitySurvey.
https://www.
inmarsat.
com/aviation/commercial-aviation/in-flight-connectivity-survey.
([n.
d.
]).
[15]KimonKarras,TheodoreKyritsis,MassimilianoAmirfeiz,andStefanoBaiotti.
2008.
Aeronauticalmobileadhocnetworks.
InProc.
EuropeanWirelessConference.
[16]RyanW.
Kingsbury.
2009.
MobileAdhocnetworksforoceanicaircraftcommuni-cations.
Ph.
D.
Dissertation.
MassachusettsInstituteofTechnology.
[17]AdamLangley,AlistairRiddoch,AlyssaWilk,AntonioVicente,CharlesKrasic,DanZhang,FanYang,FedorKouranov,IanSwett,JanardhanIyengar,JeffBailey,JeremyDorfman,JimRoskind,JoannaKulik,PatrikWestin,RamanTenneti,RobbieShade,RyanHamilton,VictorVasiliev,Wan-TehChang,andZhongyiShi.
2017.
TheQUICTransportProtocol:DesignandInternet-ScaleDeployment.
InProc.
ACMSIGCOMM.
[18]OliverLückeandErizaHafidFazli.
2011.
ANetworkingTestbedforIPv6-BasedFutureAirTrafficManagement(ATM)Network.
.
InPersonalSatelliteServices.
Springer.
[19]M-Lab.
[n.
d.
].
NDT(NetworkDiagnosticTest).
http://www.
measurementlab.
net/tools/ndt.
([n.
d.
]).
[20]MarketsandMarkets.
2015.
CommercialAviationInFlightEntertainmentandCommunicationsMarket(2012-2017).
http://www.
marketsandmarkets.
com/Market-Reports/in-flight-entertainment-communications-market-860.
html.
(September2015).
[21]DanielMedina,FelixHoffmann,SerkanAyaz,andC-H.
Rokitansky.
2008.
Feasi-bilityofanaeronauticalmobileadhocnetworkoverthenorthatlanticcorridor.
InProc.
ofSECON.
[22]NorthernSkyResearch.
2013.
TheIn-FlightConnectivityMarket.
http://www.
nsr.
com/upload/presentations/NSR_Webinar-_The_In-Flight_Connectivity_Market-_A_Boom_for_the_Satellite_Industry.
pdf.
(October2013).
[23]PhantomJS.
[n.
d.
].
PhantomJS|PhantomJS.
http://phantomjs.
org.
([n.
d.
]).
[24]JohnPRula,FabiánEBustamante,andDavidRChoffnes.
2016.
WhenIPsFly:ACaseforRedefiningAirlineCommunication.
InProc.
ofHotMobile.
[25]EhssanSakhaeeandAbbasJamalipour.
2006.
Theglobalin-flightInternet.
IEEEJSAC24,9(2006),1748–1757.
[26]TheLinuxFoundation,NETEM.
[n.
d.
].
NETEM.
http://www.
linuxfoundation.
org/collaborate/workgroups/networking/netem.
([n.
d.
]).
[27]XiaoSophiaWang,ArunaBalasubramanian,ArvindKrishnamurthy,andDavidWetherall.
2014.
HowspeedyisSPDY.
InProc.
USENIXNSDI.
[28]WilliamWu,EdwardMiller,WilburPritchard,andRaymondPickholtz.
1994.
Mobilesatellitecommunications.
Proc.
oftheIEEE82,9(1994),1431–1448.