Nokia Siemens NetworksHow QoS differentiationenhances theOTT video streaming experienceNetflix over a QoS enabledLTE network
Executive summaryContents3 Why over-the-top video streamingmatters to mobile networks4 What video users expect frommobile networks4 How QoS can be used toenhance the OTT videoexperience on LTE5 The impact of QoS on aNetflix session duringcongestion9 Conclusion11 AbbreviationsCellular networks are moreand more impacted by thegrowing popularity of Over theTop (OTT) video streamingservices. The introduction ofsmartphones and tablets haveopened the use of videostreaming services which waslimited to personal computersand televisions tied to wirelinenetworks in the past.Consumers expect a highquality experience whenwatching OTT video streams,whether it be short formYouTube videos or long formmovies or television programsfrom providers such as Netflix.The transport of videostreaming in addition to all theother types of traffic such asVoIP, email, web browsing,instant messaging and socialmedia can benefit if Quality ofService (QoS) differentiationis used – in particular duringpeak usage times. QoSdifferentiation brings clearimprovements to the customerexperience.Nokia Siemens Networks hastested how QoS mechanismscan be used to improve videostreaming quality in LTEnetworks during congestion.The findings from the testsusing Netflix point to gains forboth maintaining video qualityand video streamingperformance where re-buffering/stalling of the streamis minimized as compared tobest effort transport whichcreates a sub-standard userexperience over a congestedLTE network.QoS mechanisms are aneffective tool for operators toprovide differentiated deliveryof services transporting largepayload content. With QoS theexisting business models forcontent delivery can bechanged. New content deliverybusiness models can becreated which involve theoperator as part of the contentdelivery valuation.
OTT video streaming isdelivered to the end user over aInternet Service Provider (ISP)or wireless operator by a 3rdparty without the operatorcontrolling the service otherthan to provide transport of thecontent. The rapidly growingOTT video market is generallyconsidered to include thestreaming of content such asmovies, television programs,and music videos by contentproviders such as Netflix, Hulu,Amazon Prime, iTunes,YouTube and others. Theseproviders are globally offeringservices to both fixed andwireless mobile customers.Netflix, one of the early OTTvideo service pioneers, hasgrown to 32 million videostreaming users at the end of2012. Many OTT video serviceproviders are experiencingadditional growth driven bysmartphone and tablet usersconsuming video services overwireless networks. In fact,Network World reported thatvideo accounted for half oftotal mobile data traffic in 2012(www.networkworld.com),which increased from 42% in2011 according to Billing andOSS world(http://www.billingworld.com).By 2017, mobile video willrepresent 66% of all mobiledata traffic according to Cisco.The video share on mobiledata traffic will be even higherthan in fixed networks (CiscoVirtual Networking Index2013).Coincident with the growth inOTT video is a change in itscomposition. In the past, thebulk of video traffic wascomposed of short-form,YouTube-ish clips and manymay have tolerated lowresolution video clips thatfrequently froze, especiallywhen the video was for free.However when customers payfor entire movies or episodesand display them on HDscreens, they become verysensitive to anything other thancrystal-clear video and audiofidelity. With video servicestaking an increasing share ofthe users’ screen time onmobile devices, the importanceof high quality is evident.Operators have the opportunityto differentiate themselves fromthe competition and change thegame in content delivery byenhancing the customerexperience of video services.On the other hand scarcenetwork capacity has to beutilized in an economic way.Figure 1. Global mobile data traffic forecast.Source: Cisco Virtual Networking Index 201302.000.0004.000.0006.000.0008.000.00010.000.00012.000.0002012 2013 2014 2015 2016 2017M2MFile SharingDataVideo89%34%55%75%CAGR2012-2017Why over-the-top video streamingmatters to mobile networks
Streaming of video programssuch as movies and televisionover cellular networks inaddition to other types ofapplications like web browsing,social media, email and voiceadds a significant load due tothe associated size of the3GPP (Third GenerationPartnership Project) designedQoS mechanisms (TS 23.203Policy and Charging ControlArchitecture) to allow LTEoperators to manage the qualityof experience for users basedon the application types usedon the network.The nine standardized QoSclass identifiers (QCI) addressprioritized handling and qualityparameters for the manydifferent types of traffic thenetwork has to transport.transported payload coupledwith the need for goodthroughput speeds (figure 2).Moreover, users expect a highlevel of service delivery forvideo without re-buffering,video quality degradation, or3GPP’s QCI concept isillustrated in figure 3.QoS Class Identifier (QCI)values define the level ofservice required by theapplication.There are two broad resourcetype levels: Guaranteed (GBR)and non-Guaranteed (Non-GBR) to support differentservice types. Services usingGBR’s get pre-allocatedcapacity in the network whilenon-GBR services are givenaccess as needed.slow start times which createsthe need for QoS mechanismsto be employed to maintain ahigh quality of experience(QoE).The ability of a cellular networkto provide users with a goodquality of experience for videovaries based on signal strength,interference, and to a largeextent the bandwidth availableduring peak demand periods forthe network. To managecongestion, cellular networksneed to employ moresophisticated trafficmanagement features tomaintain quality for demandingapplications like video while stillbeing able to serve otherapplications which users need.Netflix as an OTT videostreaming service wouldnormally be handled just like allother internet applications overthe network with a QCI9 qualitytreatment which is referredwidely as best effort (BE)similarly to other non-operatorservices such as OTT VoIP.Cellular networks treat besteffort applications with the samepriority and will schedule theassociated transmissionsequally. In normal conditionswhen there is no congestion,best effort treatment worksWhat video users expect from mobilenetworksFigure 2: Non-HD Video average requirementsSource: 4G Americas / Rysavy Research Mobile Broadband Explosion2012 WhitepaperNon-HD Video Throughput (Mbps) MB/HourSmall Screen Video(feature phone)0.2 90Medium Screen Video(smartphone)1.0 450Large Screen Video(tablet)2.0 900How QoS can be used to enhance theOTT video experience on LTE
without noticeable impact onperformance.When congestion in cellsdevelops during peak usagetimes or when there aremultiple subscribers usingbandwidth intense, videostreaming applications thenbest effort is not able toschedule the transmission ofdata often enough or longenough. The lack of schedulingTo analyze the potential impactof QoS on Netflix sessions indifferent conditions, a battery oftests was executed in the LTElab network of Nokia SiemensNetworks with the Netflixapplication running on anAndroid based Smartphone.The test setup is described intime from network resourceswith limited availability resultsin users experiencingproblems with their videostream such as slow start, re-buffering / stalling anddegraded video quality.3GPP does define other QCIlevels which could be appliedto OTT streaming services:QCI 6 for video with bufferedstreams for non-GBR services.Figure 4. Note that real-worldresults may be different fromlab environment results.The main test scenario was touse the Netflix application on asmartphone and observe thevideo performance for anydisruptions like re-buffering orfor variations in displayedVideo streaming applicationsare best supported by a non-GBR resource type due tointermittent usage of networkresources versusconversational video which isalways sending data.video quality, while the networkwas being utilized in thebackground by other users whowere loaded incrementally froma UE (User Equipment)simulator as the testprogressed.Figure 3. 3GPP TS 23.203 standardized QoS Class Identifiers (QCI) characteristics.The impact of QoS on a Netflix sessionduring congestionQCI ResourceTypePriority Packet DelayBudgetPacketError LossRateExample Services1GBR2 100 ms 10-2 Conversational Voice2 4 150 ms 10-3 Conversational Video (Live Streaming)3 3 50 ms 10-3 Real Time Gaming4 5 300 ms 10-6 Non-Conversational Video (BufferedStreaming)5NON-GBR1 100 ms 10-6 IMS Signalling66300 ms 10-6Video (Buffered Streaming)TCP-based (e.g., www, e-mail, chat, ftp,p2p file sharing, progressive video, etc.)7 7 100 ms 10-3 Voice, Video (Live Streaming)Interactive Gaming8 8300 ms 10-6Video (Buffered Streaming)TCP-based (e.g., www, e-mail, chat, ftp,p2p file sharing, progressive video, etc.)9 9
While the network experiencedcongestion generated from theUE simulator, Netflix videoquality performance and thedata rate required to maintainthe stream was noted. TheNetflix video quality wasmeasured based on thefollowing factors to determinethe user experience:• High – the observed quality ofthe video being played isexcellent and the end user hasa very good service experience(no re-buffering)• Medium – the observedquality of the video beingplayed is slightly degraded butthe end user serviceexperience is acceptable andend user would continuewatching the video (some re-buffering)• Low – the observed quality ofthe video is not acceptableand the user would willinglystop the video (significant re-buffering)The test scenario consisted ofdifferent experimentsdescribed in Figure 5. Thesame video was used in all thetest cases. Note that mostOTT video streamingapplications do notcontinuously transmit data;therefore using a GBR bearerwould be wasteful of networkresources. As a result, non-GBRs are better suited to theusage model.Baselining of Netflixperformance in a besteffort networkA baseline test was conductedto understand the ideal behaviorof the Netflix application in acongestion free environmentusing a default bearer withQCI 9 QoS treatment (seeFigure 3). Subsequent testsintroduced congestion anddifferent priority levels.Figure 4. Test setupFigure 5. Description of tests to verify the impact of QoS on Netflix during congestionTest Scenario Test DescriptionBase linea. Netflix application as a best effort user / no congestionb. Netflix application as a best effort user / with congestionHigh PriorityNetflix application as a high priority user / with congestionMedium PriorityNetflix application as a medium priority user / with congestion
Netflix session as a besteffort user, no networkcongestionFrom Figure 6 it can beobserved that• the Netflix application buffer isfilled completely by the initialvideo stream data chunk whichcontained 4 MB of datatransmitted over a period of 5seconds• the initial video stream datachunk contained enough datafor the Netflix application toplay 40 seconds of videobefore the applicationrequested for more data to besent• the Netflix application bufferwas re-filled subsequently byvideo stream data chunkscontaining approximately 1 MBof data transmitted every 15 –20 seconds• the video quality is high andthe stream did not experiencere-buffering.Netflix session as a besteffort user with networkcongestionA second baseline test wasconducted to understand thebehavior of the Netflixapplication in a congestedenvironment using a defaultbearer with QCI 9 QoStreatment (see Figure 3).From Figure 7 it can beobserved that• the Netflix application bufferis never filled to a point that theapplication stops requestingdata due to lower throughputover the entire session starvingthe buffer. There were no gapsin transmission as experiencedin the previous test• Netflix user throughput (inred) decreased as networkcongestion increased• there was a significant amountand frequency of data re-transmissions (green) due topacket loss• as congestion reached itsmaximum level, multiple videostalling / re-buffering eventsoccurred with significantduration• the Netflix video qualitysignificantly decreased from theFigure 6. Netflix throughput, best effort user, no congestion in the networkFigure 7. Netflix throughput, best effort user with congestion in the networkTimeDataRe-TransmissionsThroughputHigh QualityMedium QualityLow Quality
initial start as networkcongestion was increasing.Netflix session as a highpriority user with networkcongestionFigure 8 shows that• the Netflix application bufferfilled twice to a point where theapplication stops (similar togaps in transmission asexperienced in the firstscenario) requesting data dueto the higher availablethroughput during the initialthird of the session ascongestion developed• there was a significantreduction in the amount andfrequency of data re-transmissions (green) due topacket loss• the Netflix application withhigh priority was able to sustaingood video quality at highercongestion levels compared tothe best effort user case withcongestion. Higher priorityincreased the overallthroughput for the userresulting in the Netflixapplication maintaining bettervideo quality•as congestion increased tovery high levels, compared tothe best effort case, it wasobserved that video quality didbecome extremely patchy, butthe video stream did not sufferfrom stalling / re-bufferingwhich had happened in thebest effort congestion case.Netflix Session asmedium priority userwith network congestionThe fourth and final test weconducted to understand thebehavior of the Netflixapplication in a congestedenvironment using a defaultbearer with QCI 6 (see Figure3), medium priority QoStreatment. Does the use ofQCI 6 and medium priorityimprove the user experience incongestion situationscompared to best effort? Canthe application still benefit fromsome level of priority whereminimal extra resources areutilized to assist the session?Figure 9 contains themeasurement. It can be seenthat• the Netflix application bufferfilled three times to a pointwhere the application stopped(similar to gaps in transmissionas experienced in the very firstscenario) requesting data dueto the higher availablethroughput during the initial startof the session and subsequentlyas congestion was developing.The video codec shifted downto a lower quality, therefore asmaller buffer size was needed• the throughput of the Netflixapplication was lower than inthe high priority user case butstill better than in the best effortcase in the congested network• even during congestion thedata download is steady.Retransmission frequency andamount were reducedcompared to the best effortcase in congestion• video quality got extremelypatchy at higher congestionsituations but stalling of videowas minimal compared to thebest effort congestion case• medium priority does benefitthe user experience and pushesthe point where low qualityoccurs to higher levels ofcongestion in comparison to thebest effort congestion case.Figure 8. Netflix throughput, high priority user with congestion in the networkTimeDataRe-TransmissionsThroughputHigh QualityMedium QualityLow QualityBuffer FullBuffer Full
OTT video streaming usage isincreasing and is projected toaccount for vast majority ofdata transmitted over cellularnetworks. As an application,video is very demanding for thenetwork while userexpectations for quality areequally high.Netflix can be considered as aproxy for many OTT videostreaming applications. Netflixsuffers video qualitydegradation and re-bufferingwhen its data is treated like allother best effort data duringperiods of congestion. Othervideo applications whichbehave like Netflix may benefitas well from longer schedulinglength and higher schedulingfrequency in order to preventthe application’s buffer fromexperiencing starvation andlower video quality.Real world results will vary fromcontrolled lab environments;however the measured effectsshould be very similar whenQoS mechanisms are invoked.Nokia Siemens Networks LTElab tests point out that theintroduction of even mediumpriority can positively impact theFigure 10. Test results overviewCase End user experience1. Netflix streaming as a besteffort user / no networkcongestion• HD-like video2. Netflix streaming as a besteffort user / DL throughputcongestion• Video quality degraded gradually• Numerous instances of varying episodes of stalling / re-buffering3. Netflix streaming as a highpriority user / DL throughputcongestion• Video maintained high quality at lower congestion levels• Medium quality at higher congestions levels compared to no priority case• Video became extremely patchy but no stalling observed4. Netflix streaming as a mediumpriority user / DL throughputcongestion• Video degraded gradually at a similar congestion level as in high prioritycase• The throughput experienced by the application was lower and videoquality degraded earlier than the high priority case• Better quality than in no priority case but lower than in high priority caseFigure 9: Netflix throughput, medium priority user with congestion in the networkConclusionTimeDataRe-TransmissionsThroughputHigh QualityMedium QualityLow QualityBuffer Full Buffer FullBuffer Full
Quality of Experience for theend user during congestion byreducing video qualitydegradation and re-bufferingoccurrences. High priorityproduces the best userexperience with better overallvideo quality in congestionconditions as compared tomedium priority but doesrequire the assignment of morenetwork resources to the videostreaming session. Incongested networks videostreaming applications such asNetflix suffer from reducedvideo quality and re-bufferingeven with adaptive codecchanges and applicationbuffering.In a nutshell, our testmeasurements show that theapplication of QoSdifferentiation brings significantimprovements in customerexperience for video streaming.QoS differentiation is astrategic tool for operatorswhich can be used to developnew business models incontent delivery. It is anopportunity for operators toprovide a value add that can bemonetized.
Abbreviations3GPP Third Generation Partnership ProjectBE Best EfforteNB Evolved NodeBGBR Bearer with reserved Bitrate ResourcesISP Internet Service ProviderOTT Over-the-TopLTE Long Term EvolutionNon-GBR Bearer without reserved Bitrate ResourcesQCI Quality Class IndicatorQoE Quality of ExperienceQoS Quality of ServiceUE User Equipment