Research ArticleSDP-Based Quality Adaptation and Performance Prediction inAdaptive Streaming of VBR Videos

Thoa Nguyen1 Thang Vu1 Nam PhamNgoc1 and Truong Cong Thang2

1Hanoi University of Science and Technology Hanoi Vietnam2The University of Aizu Aizuwakamatsu Japan

Correspondence should be addressed toThoa Nguyen thoanguyenthikimhusteduvn

Received 27 January 2017 Revised 8 May 2017 Accepted 24 May 2017 Published 2 July 2017

Academic Editor Mei-Ling Shyu

Copyright copy 2017 Thoa Nguyen et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Recently various adaptationmethods have been proposed to copewith throughput fluctuations inHTTP adaptive streaming (HAS)However these methods have mostly focused on constant bitrate (CBR) videos Moreover most of them are qualitative in the sensethat performance metrics could only be obtained after a streaming session In this paper we propose a new adaptation method forstreaming variable bitrate (VBR) videos using stochastic dynamic programming (SDP)With this approach the system should havea probabilistic characterization along with the definition of a cost function that is minimized by a control strategy Our solutionis based on a new statistical model where the future streaming performance is directly related to the past bandwidth statistics Wedevelop mathematical models to predict and develop simulation models to measure the average performance of the adaptationpolicy The experimental results show that the prediction models can provide accurate performance prediction which is useful inplanning adaptation policy and that our proposed adaptation method outperforms the existing ones in terms of average qualityand average quality switch

1 Introduction

Nowadays video services are increasingly popular on theInternet According to a recent study and forecast [1] globalInternet video traffic will be 80 of the entire consumerInternet traffic in 2019 Besides HTTP protocol has becomea cost-effective solution for video streaming thanks to theabundance of Web platforms and broadband connections[2 3] Furthermore for interoperability of HTTP streamingin the industry ISOIEC MPEG has developed ldquoDynamicAdaptive Streaming over HTTPrdquo (DASH) [4] as the firststandard for video streaming over HTTP

DASH requires a video to be available in multiple bitratesand split into small segments each containing a few secondsof playtime Based on the current network conditions andterminal capacity the client can adaptively decide a suit-able data rate so that stalling is avoided and the availablebandwidth is best possibly utilized If the video is encodedin only one bitrate either the bitrate is smaller than the

available bandwidth resulting in a smooth playback butsparing resources which could be utilized for a better videoquality or the video bitrate is higher than the availablebandwidth leading to video stalling Thus DASH enablesservice providers to improve resource utilization and qualityof experience (QoE)

So far existing studies have proposed simple heuristics foradapting video at the client These heuristics can be dividedinto two types buffer-based methods and throughput-basedmethodsThepurpose of buffer-basedmethods is tomaintainthe stability of the buffer within a certain range to ensurecontinuous video playback However when the bandwidth isdrastically reduced the buffer-basedmethodsmay cause sud-den change of bitrate [5ndash8] Meanwhile throughput-basedmethods adaptively decide version based on the estimatedthroughputThesemethods are generally able to react quicklyto the throughput variations the streaming quality howevermay be unstable [9]

HindawiAdvances in MultimediaVolume 2017 Article ID 7323681 12 pageshttpsdoiorg10115520177323681

2 Advances in Multimedia

Recently several Markov decision-based methods havebeen proposed to optimize decisionmaking for the streamingclient under time-varying network conditions Howeverthese existing methods mostly focus on constant bitrate(CBR) videos The authors in [10] are the first to propose anadaptation algorithm in which stochastic dynamic program-ming (SDP) is employed to find optimal decision policieswhen streaming VBR videos The segment requests are ruledby the policies which map a control parameter to everypossible state of the system however it is limited to videoswith weak bitrate fluctuations To the extent of the authorsrsquoknowledge in the context of adaptive streaming there havenot been any adaptive streaming methods that could (1)support variable bitrate (VBR) videos with strong bitratefluctuations and (2) predict the streaming performance withdifferent streaming settings in order to select the optimal one

In this paper we tackle these challenges by proposing anadaptation method using stochastic dynamic programmingFirstly we discretize a system including data throughputbuffer level and bitrate of a VBR video to form the systemstates Secondly we define a cost function that takes intoaccount parameters that affect the subjective perceptualquality of users In the cost function the weights are assignedto the difference between data throughput and the bitrateof the next segment the variance of the buffer from itsoptimal value and the quality switch of the video Finally weconstruct an infinite horizon problem (IHP) and solve it tofind the optimal policies for all system states The role of apolicy is mapping the control parameter (ie the version ofthe video) to every possible state of the system This paperis an extended work of our preliminary study in [11] Theextension in this work is multifold First we predicted theCDF of the requested versions in a streaming session sothe maximum version could be decided for the streamingsession Second we predicted CDF of the buffer levels toknow the variance of the buffer level under the fluctuations ofthe network Finally we also evaluated the proposed methodin the online context where the statistics of bandwidth isupdated periodically Besides we compared the performanceprediction results with measurement ones in in both offlineand online contexts

A policy is optimal when it minimizes an average costBased on the obtained policies and the constructed systemmodel we develop mathematical models that could predictthe streaming performance for the new streaming sessionincluding average video version average version switch persegment average buffer and average underflow probabilityExperiments are conducted to verify the mathematical mod-els by comparing the predicted performance obtained fromthe models and the measured performance The proposedmethod is evaluated in two contexts (1) offline context usingstatistics of a history bandwidth trace and (2) online contextusing bandwidth statistics of previous video segments

Thepaper is organized as follows Section 2briefly reviewsrelatedwork Section 3 describes the system and themodelingof the system in detail Section 4 presents the formulation ofthe IHP which is solved by SDPThe performance predictionis given in Section 5 Section 6 presents the experimentalresults and discussions Finally Section 7 concludes ourwork

2 Related Work

In recent years many heuristic adaptation methods foradaptive streaming have been developed (eg [5ndash9 14]) Anextensive evaluation of typical adaptation methods has beencarried out in [15] Though these methods prove to be effec-tive in their specific settings they cannot tell quantitativelythe streaming performance with different system settingsFurthermore most of them focus on CBR videos

For streaming VBR video several adaptation methodshave been proposed based on the sensibility of the receiverrsquosbuffer [6 16] Dubin et al [16] propose an adaptation logicthat supports its bandwidth estimation decisions based onthe client buffer redundancy This method considers thefluctuations of mobile network without emphasizing thecharacteristics of bitrate fluctuation of VBR videos whichleads to the lack of smoothness In [6] a partial-linear trendprediction model is developed to estimate the trend of clientbuffer level variation The client will continue to have thecurrent version for the next segment when the estimatedbuffer level has no significant change The drawback of thismethod is the sudden version switch when the actual bufferlevel drops dramatically In [14] the authors propose anadaptive logic for VBR videos based on bitrate estimationThemethod demonstrates an effective adaptation behavior asit keeps the buffer at a stable and high level however it is stillqualitative and has no mechanism to balance the streamingperformance

Several recent studies have proposedmathematicalmodelbased adaptation methods in streaming video [17ndash21] Amathematical model is proposed in [17] to calculate theunderflow probability of VBR streaming under VBR channelbased on initial delay and maximum buffer However thiswork only considered VBR videos with one version and con-stant play-out curve without developing any adaptive logicMeanwhile Kang et al [18] present a no-reference content-based QoE estimation model for video streaming serviceover wireless networks using neural networks Nonethelessneural networks are computationally complex and requirelarge training data and long training time Besides Xiang etal [19] propose a rate adaptation method using the MarkovDecision Process to obtain an optimal streaming strategyfor VBR videos Nevertheless their proposal is not able toadapt to the real bandwidth changes and has no performanceprediction The prediction of streaming performance wasfirst proposed for streaming CBR videos in [20] and VBRvideos in [21] by Liu et al In their studies a video sessionis divided into subsessions With a given target rebufferingprobability the video bitratethe average video bitrate foreach streaming subsession is predicted when streaming CBRvideosVBR videos respectively The results show that theaverage actual rebuffering probability achieved by thesemethods is reasonably close to the target However they havenot done any assessments in terms of video quality and videoquality switch so the QoE can be affected

Recently SDP has been known as an effective techniquefor solving optimization problems in video streaming [10 2223] For instance the authors in [22] apply SDP to find theoptimal policy for choosing sending rates when streaming

Advances in Multimedia 3

HTTP response

HTTP request

ChannelMedia le

Server Client

Figure 1 General DASH system

on-demand scalable VBR videos over wireless network Nev-ertheless they have not considered the effect of channel-stateaware adaptation Meanwhile Garcıa et al [10] construct aninfinite horizon problem and apply SDP to solve it specificallyfor HTTP adaptive streamingThey propose a channel modelin which transitions are only possible to adjacent stateswith equal probabilities which is only suitable for the stablebandwidth with little fluctuation In addition by observingthe histograms of segment size encoded with VBR theauthors assume that segment size (which is proportional tosegment bitrate) can be modeled through a discrete Gaussiandistribution Then the probability distribution of segmentsizes is used to calculate transition probabilities Howeverthe probability distribution of segment sizes is not takeninto account in the cost function Instead the average bitraterepresenting the bitrate for each version is used This is onlyreasonablewhen the deviations of segment sizes (ie segmentbitrates) are small In another study Xing et al [23] use SDPto find the optimization for Advance Video Coding contentwhen streaming through several wireless connections at thesame time They offer a cost function in terms of QoE butthe computational complexity of their model is significantlycaused by eight system variables in each state

The SDP-based method proposed for HTTP streamingin this paper is different from the previous studies in severalpoints First our method considers an actual time-varyingbandwidth of mobile networks Second the drastic bitratefluctuations of actual VBR videos are effectively supportedThird we develop mathematic models to predict the perfor-mance of a streaming subsession which helps to select themaximum allowed version parameter in advance

3 System Modeling

31 System Overview Figure 1 shows the functional diagramof a general DASH system consisting of a server and aclient The server holds the media files with different qualityversions Each version is further divided into small segmentsThe client has the information about the characteristics andlocations of media segments and can request any of themduring a streaming session For the next segment the clientmakes a decision of what video version to request based oncurrent status of client buffer and data throughput to providethe best streaming experience possible In this paper thesegment selection policy for the client aims at maintainingthe client buffer at a reasonable level while balancing betweenaverage version (ie average quality) and average versionswitch (ie quality variations)

Original bandwidth traceQuantized bandwidth trace

0100020003000400050006000

Bitr

ate (

kbps

)

20 40 60 80 100 120 140 160 180 2000Time (s)

Figure 2 A bandwidth trace obtained from a mobile network [12]

P32

BW1

P11

P21

P31

P12

P13

P23

P22P33

BW2

BW3

Figure 3 General 3-state Markov-chain bandwidth model

In order to apply SDP our system ismodeled as a discrete-time stochastic one Specifically the timeline is divided intotime stages At each stage 119896 the system is represented bya state variable 119904119896 When the next segment is completelydownloaded at stage 119896 + 1 the systemmoves to the state 119904119896+1As the system transits to the next state a certain cost occursBesides the channel the buffer and themedia are discretizedas explained in the following subsection

32 Channel Buffer and Media Model In our work wediscretize the bandwidth range into119882 levels The bandwidthtrace before and after discretization is shown in Figure 2with W = 10 With this level of quantization the quantizedbandwidth covers the original bandwidthwellWe then create119882 different bandwidth states BW119894 (1 le 119894 le 119882) from these119882 bandwidth levels The value of each bandwidth state is theaverage of the maximum value and minimum value of thecorresponding bandwidth level To represent the transitionfrom one bandwidth state to another on a bandwidth statespace we use the Markov chain model which has been usedwidely in previous studies [10 24 25]

Figure 3 presents a general Markov-chain model whichconsists of three bandwidth states Each state is representedby one data throughput valueThere is a transition probabilitywhen the bandwidth moves from one state to another after

4 Advances in Multimedia

02000400060008000

100001200014000

Bitr

ate (

Kbps

)

20 40 60 80 100 120 140 160 180 2000Segment index

V = 9

V = 8

V = 7

V = 6

V = 5

V = 4

V = 3

V = 2

V = 1

Figure 4 Bitrates of different versions of Tokyo Olympic video [13]

each time step Thus by simply extracting the statistics fromthe bandwidth history the transition probability between allbandwidth states is generated

Similar to the bandwidth trace we divide the buffer into119861levels from 0 to 119861119904 with 119861119904 being the buffer size In additionwe denote the video version by 119881 and represent the versionwith lowest quality as V = 1 and the highest quality as V =Vmax assuming that the video has Vmax different versionsThe bitrates of different versions of a VBR video are shown inFigure 4

Because the segment bitrate of a VBR video versionfluctuates very strongly we divide the bitrate of each versioninto 119868 intervals (from interval 1 to interval I) each of whichis represented by its average bitrate value For example if aversion that has the highest bitrate of 5000 kbps is dividedinto 10 bitrate intervals the interval 1 will range from 0 to500 kbps and its representative bitrate will be 250 kbps

We assume that all versions at a bitrate interval representa separate video flow If the current segment bitrate belongs toone interval the next segment bitrate will also belong to thatinterval regardless of segment versionWith this assumptionwe generate 119868 different policy sets for 119868 bitrate intervalsWhena segment is completely downloaded the client measures itsbitrate to find the bitrate interval it belongs toThen the clientwill determine the policy set corresponding to that bitrateinterval

4 Problem Formulation and Solution

41 SystemState With the systembeing discretized above weobserve the system state variable s119896(119887119896 bw119896 V119896) when a videosegment is completely downloaded at stage 119896 Here 119887119896 is thebuffer level representing the number of segments availablein the buffer bw119896 is the bandwidth whose value belongsto BW119894 and V119896 is the version index of the downloadedsegment The case where 119887119896 = 1 corresponds to the bufferunderflow event In each state 119904119896 the system may choose anyaction 119886 For our system an action is basically a decisionabout the version for the next segment As there are 119881maxversions to choose fromwe have totally119881max possible actions

The system then randomly moves into a new state 119904119896+1 at thenext time step resulting in a corresponding cost119862(119904119896 119904119896+1 119886)With each bitrate interval 119894 (1 le 119894 le 119868) we have a system stateset s119894 and a policy set 120583119894 Let119873 be the number of states in s119894and we have

119873 = 119861 lowast119882 lowast 119881max (1)

42 Transition Probabilities Since the system is stochasticwhich means the system outcome of each action 119886 is notdeterministic the state transition probability between everytwo states that depends on action 119886must be constructed Wedenote the probability that state 119904119896 will lead to state 119904119896+1 givenaction 119886 as follows

119875 (119904119896 119904119896+1 119886) = Pr 119904119896+1 | 119904119896 119886 (2)

Due to the independence among (119887119896 bw119896 V119896) we havePr 119904119896+1 | 119904119896 119886 = Pr 119887119896+1 | 119887119896 bw119896 V119896 119886

sdot Pr bw119896+1 | bw119896Pr V119896+1 | 119886 (3)

In the right hand side of (3) the first term can becalculated as follows

Pr 119887119896+1| 119887119896 bw119896 V119896 119886 = 1 119887119896+1 = 1198871198860 otherwise (4)

where 119887119886 is the next buffer level estimated based on thecurrent system status and action 119886 We calculate 119887119886 as follows

119887119886 = 119887119896 + 1 minus [ 119903119886bw119896+1

] (5)

where 119903119886 is the bitrate of the target versionWhen the throughput significantly drops meaning a very

low value of 119887119908119896+1 119887119886 could be lower than zero Howeverat the beginning of a new stage there is one segment beingdownloaded resulting in at least one segment being always inthe buffer Therefore (5) can be modified as follows

119887119886 = max119887119896 + 1 minus [ 119903119886bw119896+1

] 1 (6)

The second term is easily obtained from the bandwidthmodel And the third term can be simply calculated by

Pr V119896+1 | 119886 = 1 V119896+1 = 1198860 otherwise (7)

Thus expression (3) can be simplified as follows

Pr 119904119896+1 | 119904119896 119886

= Pr bw119896+1 | bw119896 119887119896+1 = 119887119886 V119896+1 = 1198860 otherwise

(8)

Advances in Multimedia 5

Input number of states119873probability function 119875cost function 119862

Output optimal policy for each state 120583119904119896119895 = 0 119895 is the iteration index120583119895119904119896 = 0 forall119904119896 isin s 120583119895119904119896 is policy of state 119904119896 at 119895th iterationrepeat119895 = 119895 + 1Policy evaluation compute 120582119895 ℎ119895(119904119896) using the below119873 + 1 equations

ℎ119895 (119904119873) = 0120582119895 + ℎ119895 (119904119896) = 119862 (119904119896 119904119896+1 120583119895119904119896) +

119873sum119903=1

119875 (119904119896 119904119903 120583119895119904119903) ℎ119895 (119904119903)forall119904119896 isin s

Policy improvement find for all 119904119896120583119895+1119904119896 = argmin

119886[119862(119904119896 119904119896+1 119886) +

119873sum119903=1

119875(119904119896 119904119903 119886)ℎ119895(119904119903)]until 120582119895+1 = 120582119895 and ℎ119895+1(119904119896) = ℎ119895(119904119896) forall119904119896 isin s

Algorithm 1 Finding the policy set for one interval

43 Cost Function In this section a cost function is definedto punish the situations that may cause a decrease in usersrsquoQoE We focus on three objective parameters that affectstrongly the subjective perception of the users which arequality level video stalling and quality switch First the costfunction should favor the selected bitrate to be close to thecurrent bandwidth so it punishes the difference Δ119903 betweenthe current bandwidth and the bitrate of the next segmentselected by action 119886 with

Δ119903 = 1003816100381610038161003816bw119896 minus 1199031198861003816100381610038161003816 (9)

Second to prevent video stalling the buffer level should neverbe underflow We define an optimal buffer level 119887opt that is adesired value the client should try to keep during a streamingsession When the buffer level is close to 119887opt the bufferunderflow is avoided Therefore the cost function penalizesthe deviation Δ119887 of the current buffer level from the optimalbuffer level where

Δ119887 = 10038161003816100381610038161003816119887119896 minus 119887opt10038161003816100381610038161003816 (10)

Third in order to reduce the quality switches the cost func-tion should contain the difference Δ119902 between the selectedquality and the last one To punish a QoE reduction becauseof high quality variations we define Δ119902 as follows

Δ119902 = (119886 minus V119896)2 (11)

Let 120572 120573 120574 be the trade-off parameters of three objectsnamely quality level video stalling and quality switchrespectivelyThe cost incurredwhen the system changes fromstate 119904119896 to state 119904119896+1 given action 119886 can be calculated by

119862 (119904119896 119904119896+1 119886) = 120572Δ119903 + 120573Δ119887 + 120574Δ119902 (12)

44 Optimization Solution As the system is discrete and thenumber of states is large we can formulate an infinite horizon

problem For every state 119904119896 the most appropriate action 119886called policy for state 119904119896 has to be decided so that the meancost per state is minimum As mentioned above our systemhas 119868 system state sets corresponding to 119868 bitrate intervalsso we have to find 119868 corresponding policy sets For simplicitywe only present the optimization solution for a general bitrateinterval with the system state set s and a corresponding policyset 120583 Finding optimal solutions for all bitrate intervals willbe done similarly Mathematically we have to minimize 119862119860which is the average cost per state obtained after downloading119871 video segments 119862119860 is calculated as follows

119862119860 = lim119871rarrinfin

1119871119871

sum119897=1

119875 (119904119896 119904119896+1 119886) 119862119897 (119904119896 119904119896+1 119886) (13)

with 119862119897(119904119896 119904119896+1 119886) being the cost incurred after downloadingthe 119897th segment Here 119871 is the number of state transitions andis also the number of video segments in the session Basedon the standard policy iteration algorithm (PIA) [26] wesolve the IHP problem by using an algorithm as presented inAlgorithm 1

Applying PIA of SDP for 119868 bitrate intervals we wouldgenerate 119868 policy sets which serve like a look-up tablemapping each state to an optimal action Thus the client isable to decide an appropriate version for the next segmentbased on the current system condition

Let 119862prob 119862cost be the computational complexity of thecalculation of transition probability and the cost from onestate to the remaining states respectively Let 119862PIA be thecomputational complexity of Algorithm 1 The complexity ofour model is 119862 = 119862prob +119862cost +119862PIA For each interval eachaction and each state we consider the cost and the transitionprobability to119873minus1 remaining states So the complexity of thecalculation of transition probability and the cost is describedas follows

119862prob = 119862cost = 119874 (119868119881max1198732) (14)

6 Advances in Multimedia

Based on [27] we have

119862PIA = 119874 (119868119881119873max) (15)

Therefore the complexity of our model is

119862 = 119874 (2119868119881max1198732 + 119868119881119873max) (16)

5 Performance Prediction

After Section 4 we achieve 119868 policy sets corresponding to 119868intervals of a video bitrate In this section we use theMarkovchain model to predict the streaming performance for asession Similar to Section 44 this section only presents theperformance prediction for a general bitrate interval with asystem state set s and a corresponding policy set 120583 Predictingperformance for all bitrate intervals will be done similarlyAfter carrying out the calculation for all 119868 bitrate intervalswe take the average values as the final results

The key is to determine the average state probability p119886 =[1199011198861 1199011198862 119901119886119873] with 119901119886119899 (1 le 119899 le 119873) being the averageprobability that the system is at the 119899th state throughout thestreaming session The probability p119886 is the average valueof state probabilities after downloading 119897 segments p119897 =[1199011198971 1199011198972 119901119897119873] (1 le 119897 le 119871) Here 119901119897119899 (1 le 119899 le 119873)is the probability that the system is at the 119899th state afterdownloading 119897 segments From the Markov chain theorythe state probability after downloading 119897 + 1 segments canbe computed as the product of the state probability afterdownloading 119897 segments and a transition matrix p119897+1 =p119897PTS Assuming that the initial probability 1199010 is known Thetransition matrix PTS is a119873times119873matrix which represents thetransition probability from state 119904119896 to state 119904119896+1 PTS is definedas follows

PTS = 119875 (119904119896 119904119896+1) = Pr 119904119896+1 | 119904119896 119886 isin 120583 (17)

The average state probability p119886 is calculated as follows

p119886 = sum119871119897=1 p119897PTS119871 (18)

Currently most of the adaptation algorithms developedfor HTTP streaming are qualitative in the sense that the per-formance metrics could only be obtained after the streamingsession In this study the predicted streaming performancecould be calculated based on the average state probabilityand the information inside every state Specifically wemainlyfocus on the following aspects bitrate prediction qualityswitch prediction and buffer prediction

51 Quality Prediction The video quality that the users per-ceive is presented through the selected versionThehigher theversion is selected the better the video quality is perceivedby the users Furthermore setting the maximum version forthe streaming session also affects the perceptual quality of theusers Obviously a very low value of the maximum versionmay result in a poor perceptual quality while a very highvalue may increase the chance of buffer underflow In thissection we predict the quality performance of the streaming

session based on the average version 119860V that is calculatedusing (19) and the cumulative distribution function (CDF)of the versions 119891(V) (V isin [1 119881max]) that is shown in (20)Based on the predicted probability of the versions throughouta streaming session the maximum version could be decidedfor the session

119860V = sum119873119899=1 V119899119901119886119899119873 (19)

119891 (V) = sum119873119899=1 V119899119901119886119899119873 | V = V119899 (20)

52 Quality Switch Prediction Quality switch is an importantfactor affecting the perception of the users The users oftenexpect a smooth playback with the minimum number ofquality switches and small switch amplitude from one seg-ment to the next We can predict the average version switchper segment 119860 sw as follows

119860 sw = sum119873119899=1 1003816100381610038161003816119886119899 minus V1198991003816100381610038161003816 119901119886119899

119873 119886119899 isin 120583 (21)

53 Buffer Prediction Video stalling is one of the importantobjective parameters that affect the subjective perception ofthe users Stalling occurswhen the play-out buffer underrunsTo prevent this event the buffer must be maintained within asafe range In this session we evaluate the buffer performancethrough the average buffer level 119860119887 the CDF of the bufferlevel 119891(119887) (119887 isin [1 119861max]) and the buffer underflowprobability Prund (ie when the system stays at buffer level1) which are described as follows

119860119887 = sum119873119899=1 119887119899119901119886119899119873

119891 (119887) = sum119873119899=1 119887119899119901119886119899119873 | 119887119899 lt 119887

Prund = sum119873119899=1 119901119886119899119873 | 119887119899 = 1

(22)

119860119887 represents the safety of the buffer If 119860119887 is small thebuffer level is often in low levels which may cause playbackinterruptionwhen the current bandwidth drops dramatically119891(119887) reflects the variance of the buffer level under thefluctuations of the network Prund shows the probability thatthe playback would be interrupted in the streaming session

6 Experimental Results

In order to evaluate the proposed system model and per-formance prediction accuracy in this section we perform anumber of experiments in both offline and online contextsand compare the performance predicted results with themeasurement ones We also compare our proposed methodwith two existing ones namely the SDP method presented[10] which could obtain the best performance among the SDPmethods and the bitrate estimation based method presented[14] which is the best among the qualitative methods

Advances in Multimedia 7

61 Experiment Setup For the simulation our test-bed con-sists of a client running Java 80 which implements theadaptation and a server running Apache2 which holds themedia segments The client runs on a Window 7 computerwith an Intel i5-17 GHz CPU and 4GB memory and theserver runs on Ubuntu 1204LTS (with default TCP CUBIC)with 1 G RAM The channel bandwidth is simulated usingDummyNet [28] We use the Tokyo Olympic video from [6]For the video 119881max = 9 and for the bandwidth 119882 = 10Since we measure the buffer size and compute the buffer costin the segment duration unit we implement our adaptationmethod with one setting of the segment duration In ourexperiments we select a segment duration of 2 secondswhich is similar to those of [7 8 10] The impacts of segmentdurations on adaptation performance have been consideredin some recent studies [29 30] Further evaluation of differentsegment durations with a fixed buffer size (in seconds) will bereserved for our future work Maximum buffer level is set to5 segments (ie 10 seconds) and optimal buffer level is set to4 segments (ie 8 seconds)

In our method a streaming provider can adjust thebalance between the requirements for high quality levelpreventing video stalling and reducing quality switchesby changing the trade-off parameters 120572 120574 and 120573 Sinceselecting an optimal combination of trade-off parametersof the cost function involves solving a hard optimizationproblem it will be investigated in our future work In thispaper we select qualitatively the trade-off parameters of costfunction as follows Initially we fix 120573 and select the othertwo parameters Since we want to prioritize requirement ofsmooth quality switch 120574 is selected so that the contributionof the quality switch cost Δ119902 is higher than that of buffercost Δ119887 With parameter 120572 because the bitrate cost Δ119903 can beup to thousands parameter 120572 should be small to reduce thecontribution of the bitrate cost to the overall cost 119862 Basedon our experience good empirical values on parameters 120572 120573and 120574 are 0003 4 and 20 respectively

62 Experimental Results In the first part of the experimentwe evaluate the accuracy of performance prediction in offlinecontext using a given bandwidth trace obtained fromamobilenetwork [12] In this context the number of video segments119871 is 300

Figures 5 and 6 show the predicted performance usingthe formulas presented in Section 5 and themeasured perfor-mance obtained from the experiments when the maximumallowed version is set to 7 8 and 9 These figures point outthat the prediction results are close to the measurement onesWe can see from Figure 5 that when the maximum versionincreases the average version as well as the average versionswitch also increases Figure 6 shows that in both predictionand measurement cases when the maximum version is 7 theunderflowprobability is almost zero and increases very slowlywhen the maximum allowed version increases This analysisimplies that setting themaximumversion to 8 is reasonable interms of balancing between average video quality and qualityswitch meanwhile setting the maximum version to 7 ensuresa very stable streaming experience

0

02

04

06

08

1

Aver

age s

witc

hse

gmen

t

8 97Maximum version

1

3

5

7

9

Vers

ion

Measured avr_switchMeasured avr_version

Predicted avr_switchPredicted avr_version

Figure 5 Predicted performance and measured performance interms of average version and average version switch per segment inoffline context using a given bandwidth trace

0

005

01

Und

er

ow p

roba

bilit

y

0

5

10

Aver

age b

uer

(s)

8 97Maximum version

Measured prob_underowMeasured avr_buer

Predicted prob_underowPredicted avr_buer

Figure 6 Predicted performance and measured performance interms of average buffer and underflow probability in offline contextusing a given bandwidth trace

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

012345678910

Vers

ion

Figure 7 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 9

Figures 7 8 and 9 show the bitrate and version switchbehavior of the proposed method in the three cases ofmaximum version It can be seen very clearly from thesefigures that when the maximum version is reduced thenumber of version switches (or quality changes) decreases

Table 1 shows more detailed statistics of the experimentalresults in three cases of maximum version It can be drawn

8 Advances in Multimedia

Table 1 Compare predicted performance and measured performance in offline context using a given bandwidth trace

Statistics 119881max = 9 119881max = 8 119881max = 7Predicted Measured Predicted Measured Predicted Measured

119860119887(119904) 948 852 959 877 98 968119860119902 764 788 745 773 648 687119860 sw 030 018 014 009 003 002Prundflow 001 000 0006 000 000 000

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

0123456789

Vers

ion

Figure 8 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 8

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

012345678

Vers

ion

Figure 9 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 7

from the table that there is no significant difference betweenthe predicted performance and the measured performance

In the second part of the experiment we use two historybandwidth traces recorded from two previous streamingsessions of the client The CDFs of both bandwidths areshown in Figure 10

Bandwidth 1198871199081 is used for calculating the statisticalmodels and 1198871199082 is used in the simulation for measuringperformance parameters Figures 11 12 and 13 show thebitrate and version switch behavior of the proposed methodin the three cases of maximum version The detailed resultsare listed in Table 2 Based on these figures and tablewe affirm once again that the mathematical performanceprediction model agrees well with the measurement

bw1bw2

0010203040506070809

1

CDF

1000 2000 3000 4000 50000Bandwidth (kbps)

Figure 10 The cumulative distribution function of bw1 and bw2

Segment rpBitrateVersion

100 200 300 400 500 6000Time (s)

01000200030004000500060007000

Bitr

ate (

kbps

)

012345678910

Vers

ion

Figure 11 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 9

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

0123456789

Vers

ion

Figure 12 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 8

Advances in Multimedia 9

Table 2 Compare predicted performance and measured performance in offline context using a history bandwidth trace for statistics

Statistics 119881max = 9 119881max = 8 119881max = 7Predicted Measured Predicted Measured Predicted Measured

119860119887(119904) 952 844 961 885 969 965119860119902 748 792 727 771 680 698119860 sw 031 019 018 011 008 005Prundflow 000 000 000 000 000 000

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

12345678

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 13 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 7

In the third part of the experiment we consider the onlinecontext in which the prediction of the future bandwidth isbased on the statistical parameters of all previous segmentsSpecifically we divide an entire session into chunks each ofwhich has119871 video segmentsWe treat each chunk individuallyas a mini streaming session Assuming that initially we haveenough statistical data to predict the performance of thefirst chunk The prediction of the subsequent chunks willbe done based on the previous chunks At the beginning ofeach chunk the bandwidth statistics are updated leading to arecomputation of the policy set and the performance In theexperiment we set119871 to 100 video segments It can be observedfrom our experiments that it took only several seconds to(re)compute the whole model Therefore the computationaloverhead is about 3 which is appropriate for the onlinecontext Figures 14 15 and 16 show the adaptation bitrateand version switch behavior of the proposed method in thethree cases of maximum version The predicted performanceand the measured performance in the online context in thethree cases are presented in detail in Tables 3 4 and 5It is very clear that in the online context the predictedperformance is also very close to the measured performance

Next we compare our proposed method with the SDPmethod [10] and the bitrate estimation based method [14]using the same simulation settings as in the first part of ourexperimentThe experimental results obtained by simulatingthe SDP method [10] and bitrate estimation based method[14] are shown in Figures 17 and 18 respectively We can seethat both methods provide very fluctuating version switchesThe detailed statistics of these adaptation methods with themaximum version set to 9 and our proposedmethodwith the

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

kbps

)

012345678910

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 14 Experimental results of the proposed method in onlinecontext with Vmax = 9

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

0123456789

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 15 Experimental results of the proposed method in onlinecontext with Vmax = 8

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

12345678

Vers

ion

Figure 16 Experimental results of the proposed method in onlinecontext with Vmax = 7

10 Advances in Multimedia

Table 3 Compare predicted performance and measured performance in online context with 119881max = 9Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 931 896 892 845 936 886119860119902 769 804 789 768 783 772119860 sw 005 009 004 007 002 005Prundflow 000 000 000 000 000 000

Table 4 Compare predicted performance and measured performance in online context with 119881max = 8Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 953 901 942 863 942 896119860119902 756 786 765 763 765 767119860 sw 000 003 001 003 002 004Prundflow 000 000 000 000 000 000

Table 5 Compare predicted performance and measured performance in online context with 119881max = 7Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 977 993 977 952 983 959119860119902 692 700 692 695 692 700119860 sw 000 000 000 002 000 000Prundflow 000 000 000 000 000 000

Table 6 Statistics of different adaptation methods in offline context

Statistics SDP method119881max = 9

Bitrate estimation method119881max = 9

Proposed method119881max = 9

Proposed method119881max = 8

Proposed method119881max = 7

119860119887(119904) 862 885 852 877 968119860119902 740 785 788 773 687119860 sw 076 043 018 009 002Prundflow 000 000 000 000 000

maximumversion set to 7 8 and 9 are shown in Table 6 It canbe seen that the performance of the bitrate estimationmethodis less effective than our method in terms of average versionand average version switch Regarding the SDP method itis evident that the performance of this method is the worstamong all (with low average version and highest averageswitch per segment) This can be expected because thismethod was not originally designed for real VBR videos

The buffer level curves of the three methods in offlinecontext are shown in Figure 19 We can see that all buffercurves imply streaming sessionswithout any freezes for usersAmong these the SDP method and the proposed methodwith 119881max = 9 provide the most unstable buffers while thebitrate estimation based method and the proposed methodwith 119881max = 7 provide the most stable buffers It can beobserved from Figures 17 and 18 and Table 6 that the SDPmethod and bitrate estimation based method result in veryfluctuating version curves and the proposed method with

Segment rpBitrateVersion

010002000300040005000600070008000

100 300 500 600200 4000Time (s)

0246810

Vers

ion

Bitr

ate (

kbps

)

Figure 17 Experimental results of SDP method

119881max = 7 obtains the lowest average quality Therefore fromTable 6 and Figure 19 we can see that the proposedmethod inoffline context with 119881max = 8 or 9 can provide the best per-formance

Advances in Multimedia 11

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

s)

100 200 300 400 500 6000Time (s)

0246810

Vers

ion

Figure 18 Experimental results of bitrate estimation basedmethod

02468

1012

Bue

r (s)

100 300 500 600200 4000Time (s)

SDPMax 9Max 8

Max 7BE

Figure 19 Buffer level curves of SDP method bitrate estimationbased method and the proposed method (Max9 Max8 and Max7)in offline context

7 Conclusion

In this paper we have proposed an adaptation method forHTTP streaming based on stochastic dynamic programmingThe system model was targeted at real bandwidth trace withstrong bitrate fluctuation of VBR videos Furthermore wehave developed a model to predict the system performancewith the aim of choosing the best setting based on theperformance requirements The experimental results haveshown that our method can effectively adapt VBR videos andperform accurate performance prediction which is useful inplanning adaptation policy

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

References

[1] Cisco Systems Inc ldquoNetworking index Forecast and method-ology 2012ndash2017rdquo Cisco Visual May 2013

[2] A C Begen T Akgul and M Baugher ldquoWatching video overthe web part 1 streaming protocolsrdquo IEEE Internet Computingvol 15 no 2 pp 54ndash63 2011

[3] T C Thang Q-D Ho J W Kang and A T Pham ldquoAdaptivestreaming of audiovisual content using MPEG DASHrdquo IEEETransactions on Consumer Electronics vol 58 no 1 pp 78ndash852012

[4] T Stockhammer ldquoDynamic adaptive streaming over HTTPstandards and design principlesrdquo in Proceeding of the 2ndAnnual ACM Multimedia Systems Conference (MMSys rsquo11) pp133ndash144 San Jose CA USA February 2011

[5] K Miller E Quacchio G Gennari and A Wolisz ldquoAdaptationalgorithm for adaptive streaming over HTTPrdquo in Proceedings ofthe 19th International Packet Video Workshop (PV rsquo12) pp 173ndash178 IEEE Munich Germany May 2012

[6] Y Zhou Y Duan J Sun and Z Guo ldquoTowards simple andsmooth rate adaption for VBR video in DASHrdquo in Proceedingsof the IEEE Visual Communications and Image Processing Con-ference (VCIP rsquo14) pp 9ndash12 IEEE Valletta Malta December2014

[7] S Akhshabi S Narayanaswamy A C Begen and C DovrolisldquoAn experimental evaluation of rate-adaptive video players overHTTPrdquo Signal Processing Image Communication vol 27 no 4pp 271ndash287 2012

[8] C Muller S Lederer and C Timmerer ldquoAn evaluation ofdynamic adaptive streaming over HTTP in vehicular environ-mentsrdquo in Proceedings of the 4th Workshop on Mobile Video(MoVid rsquo12) pp 37ndash42 New York NY USA February 2012

[9] C Liu I Bouazizi and M Gabbouj ldquoRate adaptation foradaptive HTTP streamingrdquo in Proceedings of the 2nd AnnualACMMultimedia Systems Conference (MMSys rsquo11) pp 169ndash174San Jose Calif USA February 2011

[10] S Garcıa J Cabrera and N Garcıa ldquoQuality-control algorithmfor adaptive streaming services over wireless channelsrdquo IEEEJournal on Selected Topics in Signal Processing vol 9 no 1 pp50ndash59 2015

[11] A H Duong T Nguyen T Vu T T Do N P Ngoc and T CThang ldquoSDP-based adaptation for quality control in adaptivestreamingrdquo in Proceedings of the International Conference onComputing Management and Telecommunications (ComMan-Tel rsquo15) pp 194ndash199 IEEE DaNang Vietnam December 2015

[12] S Lederer C Mueller C Timmerer C Concolato J Le Feuvreand K Fliegel ldquoDistributedDASHdatasetrdquo in Proceedings of the4th ACMMultimedia Systems Conference (MMSys rsquo13) pp 131ndash135 Oslo Norway March 2013

[13] ldquoH264AVC video trace libraryrdquo httptraceeasasueduh264[14] T CThang H T Le H X Nguyen A T Pham JW Kang and

Y M Ro ldquoAdaptive video streaming over HTTP with dynamicresource estimationrdquo IEEE Trans On consumer electronics vol58 no 1 pp 78ndash85 2012

[15] T C Thang H T Le A T Pham and Y M Ro ldquoAn evaluationof bitrate adaptation methods for HTTP live streamingrdquo IEEEJournal on Selected Areas in Communications vol 32 no 4 pp693ndash705 2014

[16] R Dubin O Hadar and A Dvir ldquoThe effect of client buffer andMBR consideration onDASHAdaptation Logicrdquo in Proceedingsof the 2013 IEEE Wireless Communications and NetworkingConference (WCNC rsquo13) pp 2178ndash2183 IEEE Shanghai ChinaApril 2013

[17] G Liang and B Liang ldquoEffect of delay and buffering on jitter-free streaming over random VBR channelsrdquo IEEE Transactionson Multimedia vol 10 no 6 pp 1128ndash1141 2008

[18] Y Kang H Chen and L Xie ldquoAn artificial-neural-network-based QoE estimation model for Video streaming over wirelessnetworksrdquo in Proceedings of the 2013 IEEECIC InternationalConference on Communications in China (ICCC rsquo13) pp 264ndash269 IEEE Xirsquoan China August 2013

[19] S Xiang L Cai and J Pan ldquoAdaptive scalable video streaminginwireless networksrdquo in Proceedings of the 3rdACMMultimedia

Advances in Multimedia 3

HTTP response

HTTP request

ChannelMedia le

Server Client

Figure 1 General DASH system

on-demand scalable VBR videos over wireless network Nev-ertheless they have not considered the effect of channel-stateaware adaptation Meanwhile Garcıa et al [10] construct aninfinite horizon problem and apply SDP to solve it specificallyfor HTTP adaptive streamingThey propose a channel modelin which transitions are only possible to adjacent stateswith equal probabilities which is only suitable for the stablebandwidth with little fluctuation In addition by observingthe histograms of segment size encoded with VBR theauthors assume that segment size (which is proportional tosegment bitrate) can be modeled through a discrete Gaussiandistribution Then the probability distribution of segmentsizes is used to calculate transition probabilities Howeverthe probability distribution of segment sizes is not takeninto account in the cost function Instead the average bitraterepresenting the bitrate for each version is used This is onlyreasonablewhen the deviations of segment sizes (ie segmentbitrates) are small In another study Xing et al [23] use SDPto find the optimization for Advance Video Coding contentwhen streaming through several wireless connections at thesame time They offer a cost function in terms of QoE butthe computational complexity of their model is significantlycaused by eight system variables in each state

The SDP-based method proposed for HTTP streamingin this paper is different from the previous studies in severalpoints First our method considers an actual time-varyingbandwidth of mobile networks Second the drastic bitratefluctuations of actual VBR videos are effectively supportedThird we develop mathematic models to predict the perfor-mance of a streaming subsession which helps to select themaximum allowed version parameter in advance

3 System Modeling

31 System Overview Figure 1 shows the functional diagramof a general DASH system consisting of a server and aclient The server holds the media files with different qualityversions Each version is further divided into small segmentsThe client has the information about the characteristics andlocations of media segments and can request any of themduring a streaming session For the next segment the clientmakes a decision of what video version to request based oncurrent status of client buffer and data throughput to providethe best streaming experience possible In this paper thesegment selection policy for the client aims at maintainingthe client buffer at a reasonable level while balancing betweenaverage version (ie average quality) and average versionswitch (ie quality variations)

Original bandwidth traceQuantized bandwidth trace

0100020003000400050006000

Bitr

ate (

kbps

)

20 40 60 80 100 120 140 160 180 2000Time (s)

Figure 2 A bandwidth trace obtained from a mobile network [12]

P32

BW1

P11

P21

P31

P12

P13

P23

P22P33

BW2

BW3

Figure 3 General 3-state Markov-chain bandwidth model

In order to apply SDP our system ismodeled as a discrete-time stochastic one Specifically the timeline is divided intotime stages At each stage 119896 the system is represented bya state variable 119904119896 When the next segment is completelydownloaded at stage 119896 + 1 the systemmoves to the state 119904119896+1As the system transits to the next state a certain cost occursBesides the channel the buffer and themedia are discretizedas explained in the following subsection

32 Channel Buffer and Media Model In our work wediscretize the bandwidth range into119882 levels The bandwidthtrace before and after discretization is shown in Figure 2with W = 10 With this level of quantization the quantizedbandwidth covers the original bandwidthwellWe then create119882 different bandwidth states BW119894 (1 le 119894 le 119882) from these119882 bandwidth levels The value of each bandwidth state is theaverage of the maximum value and minimum value of thecorresponding bandwidth level To represent the transitionfrom one bandwidth state to another on a bandwidth statespace we use the Markov chain model which has been usedwidely in previous studies [10 24 25]

Figure 3 presents a general Markov-chain model whichconsists of three bandwidth states Each state is representedby one data throughput valueThere is a transition probabilitywhen the bandwidth moves from one state to another after

4 Advances in Multimedia

02000400060008000

100001200014000

Bitr

ate (

Kbps

)

20 40 60 80 100 120 140 160 180 2000Segment index

V = 9

V = 8

V = 7

V = 6

V = 5

V = 4

V = 3

V = 2

V = 1

Figure 4 Bitrates of different versions of Tokyo Olympic video [13]

each time step Thus by simply extracting the statistics fromthe bandwidth history the transition probability between allbandwidth states is generated

Similar to the bandwidth trace we divide the buffer into119861levels from 0 to 119861119904 with 119861119904 being the buffer size In additionwe denote the video version by 119881 and represent the versionwith lowest quality as V = 1 and the highest quality as V =Vmax assuming that the video has Vmax different versionsThe bitrates of different versions of a VBR video are shown inFigure 4

Because the segment bitrate of a VBR video versionfluctuates very strongly we divide the bitrate of each versioninto 119868 intervals (from interval 1 to interval I) each of whichis represented by its average bitrate value For example if aversion that has the highest bitrate of 5000 kbps is dividedinto 10 bitrate intervals the interval 1 will range from 0 to500 kbps and its representative bitrate will be 250 kbps

We assume that all versions at a bitrate interval representa separate video flow If the current segment bitrate belongs toone interval the next segment bitrate will also belong to thatinterval regardless of segment versionWith this assumptionwe generate 119868 different policy sets for 119868 bitrate intervalsWhena segment is completely downloaded the client measures itsbitrate to find the bitrate interval it belongs toThen the clientwill determine the policy set corresponding to that bitrateinterval

4 Problem Formulation and Solution

41 SystemState With the systembeing discretized above weobserve the system state variable s119896(119887119896 bw119896 V119896) when a videosegment is completely downloaded at stage 119896 Here 119887119896 is thebuffer level representing the number of segments availablein the buffer bw119896 is the bandwidth whose value belongsto BW119894 and V119896 is the version index of the downloadedsegment The case where 119887119896 = 1 corresponds to the bufferunderflow event In each state 119904119896 the system may choose anyaction 119886 For our system an action is basically a decisionabout the version for the next segment As there are 119881maxversions to choose fromwe have totally119881max possible actions

The system then randomly moves into a new state 119904119896+1 at thenext time step resulting in a corresponding cost119862(119904119896 119904119896+1 119886)With each bitrate interval 119894 (1 le 119894 le 119868) we have a system stateset s119894 and a policy set 120583119894 Let119873 be the number of states in s119894and we have

119873 = 119861 lowast119882 lowast 119881max (1)

42 Transition Probabilities Since the system is stochasticwhich means the system outcome of each action 119886 is notdeterministic the state transition probability between everytwo states that depends on action 119886must be constructed Wedenote the probability that state 119904119896 will lead to state 119904119896+1 givenaction 119886 as follows

119875 (119904119896 119904119896+1 119886) = Pr 119904119896+1 | 119904119896 119886 (2)

Due to the independence among (119887119896 bw119896 V119896) we havePr 119904119896+1 | 119904119896 119886 = Pr 119887119896+1 | 119887119896 bw119896 V119896 119886

sdot Pr bw119896+1 | bw119896Pr V119896+1 | 119886 (3)

In the right hand side of (3) the first term can becalculated as follows

Pr 119887119896+1| 119887119896 bw119896 V119896 119886 = 1 119887119896+1 = 1198871198860 otherwise (4)

where 119887119886 is the next buffer level estimated based on thecurrent system status and action 119886 We calculate 119887119886 as follows

119887119886 = 119887119896 + 1 minus [ 119903119886bw119896+1

] (5)

where 119903119886 is the bitrate of the target versionWhen the throughput significantly drops meaning a very

low value of 119887119908119896+1 119887119886 could be lower than zero Howeverat the beginning of a new stage there is one segment beingdownloaded resulting in at least one segment being always inthe buffer Therefore (5) can be modified as follows

119887119886 = max119887119896 + 1 minus [ 119903119886bw119896+1

] 1 (6)

The second term is easily obtained from the bandwidthmodel And the third term can be simply calculated by

Pr V119896+1 | 119886 = 1 V119896+1 = 1198860 otherwise (7)

Thus expression (3) can be simplified as follows

Pr 119904119896+1 | 119904119896 119886

= Pr bw119896+1 | bw119896 119887119896+1 = 119887119886 V119896+1 = 1198860 otherwise

(8)

Advances in Multimedia 5

Input number of states119873probability function 119875cost function 119862

Output optimal policy for each state 120583119904119896119895 = 0 119895 is the iteration index120583119895119904119896 = 0 forall119904119896 isin s 120583119895119904119896 is policy of state 119904119896 at 119895th iterationrepeat119895 = 119895 + 1Policy evaluation compute 120582119895 ℎ119895(119904119896) using the below119873 + 1 equations

ℎ119895 (119904119873) = 0120582119895 + ℎ119895 (119904119896) = 119862 (119904119896 119904119896+1 120583119895119904119896) +

119873sum119903=1

119875 (119904119896 119904119903 120583119895119904119903) ℎ119895 (119904119903)forall119904119896 isin s

Policy improvement find for all 119904119896120583119895+1119904119896 = argmin

119886[119862(119904119896 119904119896+1 119886) +

119873sum119903=1

119875(119904119896 119904119903 119886)ℎ119895(119904119903)]until 120582119895+1 = 120582119895 and ℎ119895+1(119904119896) = ℎ119895(119904119896) forall119904119896 isin s

Algorithm 1 Finding the policy set for one interval

43 Cost Function In this section a cost function is definedto punish the situations that may cause a decrease in usersrsquoQoE We focus on three objective parameters that affectstrongly the subjective perception of the users which arequality level video stalling and quality switch First the costfunction should favor the selected bitrate to be close to thecurrent bandwidth so it punishes the difference Δ119903 betweenthe current bandwidth and the bitrate of the next segmentselected by action 119886 with

Δ119903 = 1003816100381610038161003816bw119896 minus 1199031198861003816100381610038161003816 (9)

Second to prevent video stalling the buffer level should neverbe underflow We define an optimal buffer level 119887opt that is adesired value the client should try to keep during a streamingsession When the buffer level is close to 119887opt the bufferunderflow is avoided Therefore the cost function penalizesthe deviation Δ119887 of the current buffer level from the optimalbuffer level where

Δ119887 = 10038161003816100381610038161003816119887119896 minus 119887opt10038161003816100381610038161003816 (10)

Third in order to reduce the quality switches the cost func-tion should contain the difference Δ119902 between the selectedquality and the last one To punish a QoE reduction becauseof high quality variations we define Δ119902 as follows

Δ119902 = (119886 minus V119896)2 (11)

Let 120572 120573 120574 be the trade-off parameters of three objectsnamely quality level video stalling and quality switchrespectivelyThe cost incurredwhen the system changes fromstate 119904119896 to state 119904119896+1 given action 119886 can be calculated by

119862 (119904119896 119904119896+1 119886) = 120572Δ119903 + 120573Δ119887 + 120574Δ119902 (12)

44 Optimization Solution As the system is discrete and thenumber of states is large we can formulate an infinite horizon

problem For every state 119904119896 the most appropriate action 119886called policy for state 119904119896 has to be decided so that the meancost per state is minimum As mentioned above our systemhas 119868 system state sets corresponding to 119868 bitrate intervalsso we have to find 119868 corresponding policy sets For simplicitywe only present the optimization solution for a general bitrateinterval with the system state set s and a corresponding policyset 120583 Finding optimal solutions for all bitrate intervals willbe done similarly Mathematically we have to minimize 119862119860which is the average cost per state obtained after downloading119871 video segments 119862119860 is calculated as follows

119862119860 = lim119871rarrinfin

1119871119871

sum119897=1

119875 (119904119896 119904119896+1 119886) 119862119897 (119904119896 119904119896+1 119886) (13)

with 119862119897(119904119896 119904119896+1 119886) being the cost incurred after downloadingthe 119897th segment Here 119871 is the number of state transitions andis also the number of video segments in the session Basedon the standard policy iteration algorithm (PIA) [26] wesolve the IHP problem by using an algorithm as presented inAlgorithm 1

Applying PIA of SDP for 119868 bitrate intervals we wouldgenerate 119868 policy sets which serve like a look-up tablemapping each state to an optimal action Thus the client isable to decide an appropriate version for the next segmentbased on the current system condition

Let 119862prob 119862cost be the computational complexity of thecalculation of transition probability and the cost from onestate to the remaining states respectively Let 119862PIA be thecomputational complexity of Algorithm 1 The complexity ofour model is 119862 = 119862prob +119862cost +119862PIA For each interval eachaction and each state we consider the cost and the transitionprobability to119873minus1 remaining states So the complexity of thecalculation of transition probability and the cost is describedas follows

119862prob = 119862cost = 119874 (119868119881max1198732) (14)

6 Advances in Multimedia

Based on [27] we have

119862PIA = 119874 (119868119881119873max) (15)

Therefore the complexity of our model is

119862 = 119874 (2119868119881max1198732 + 119868119881119873max) (16)

5 Performance Prediction

After Section 4 we achieve 119868 policy sets corresponding to 119868intervals of a video bitrate In this section we use theMarkovchain model to predict the streaming performance for asession Similar to Section 44 this section only presents theperformance prediction for a general bitrate interval with asystem state set s and a corresponding policy set 120583 Predictingperformance for all bitrate intervals will be done similarlyAfter carrying out the calculation for all 119868 bitrate intervalswe take the average values as the final results

The key is to determine the average state probability p119886 =[1199011198861 1199011198862 119901119886119873] with 119901119886119899 (1 le 119899 le 119873) being the averageprobability that the system is at the 119899th state throughout thestreaming session The probability p119886 is the average valueof state probabilities after downloading 119897 segments p119897 =[1199011198971 1199011198972 119901119897119873] (1 le 119897 le 119871) Here 119901119897119899 (1 le 119899 le 119873)is the probability that the system is at the 119899th state afterdownloading 119897 segments From the Markov chain theorythe state probability after downloading 119897 + 1 segments canbe computed as the product of the state probability afterdownloading 119897 segments and a transition matrix p119897+1 =p119897PTS Assuming that the initial probability 1199010 is known Thetransition matrix PTS is a119873times119873matrix which represents thetransition probability from state 119904119896 to state 119904119896+1 PTS is definedas follows

PTS = 119875 (119904119896 119904119896+1) = Pr 119904119896+1 | 119904119896 119886 isin 120583 (17)

The average state probability p119886 is calculated as follows

p119886 = sum119871119897=1 p119897PTS119871 (18)

Currently most of the adaptation algorithms developedfor HTTP streaming are qualitative in the sense that the per-formance metrics could only be obtained after the streamingsession In this study the predicted streaming performancecould be calculated based on the average state probabilityand the information inside every state Specifically wemainlyfocus on the following aspects bitrate prediction qualityswitch prediction and buffer prediction

51 Quality Prediction The video quality that the users per-ceive is presented through the selected versionThehigher theversion is selected the better the video quality is perceivedby the users Furthermore setting the maximum version forthe streaming session also affects the perceptual quality of theusers Obviously a very low value of the maximum versionmay result in a poor perceptual quality while a very highvalue may increase the chance of buffer underflow In thissection we predict the quality performance of the streaming

session based on the average version 119860V that is calculatedusing (19) and the cumulative distribution function (CDF)of the versions 119891(V) (V isin [1 119881max]) that is shown in (20)Based on the predicted probability of the versions throughouta streaming session the maximum version could be decidedfor the session

119860V = sum119873119899=1 V119899119901119886119899119873 (19)

119891 (V) = sum119873119899=1 V119899119901119886119899119873 | V = V119899 (20)

52 Quality Switch Prediction Quality switch is an importantfactor affecting the perception of the users The users oftenexpect a smooth playback with the minimum number ofquality switches and small switch amplitude from one seg-ment to the next We can predict the average version switchper segment 119860 sw as follows

119860 sw = sum119873119899=1 1003816100381610038161003816119886119899 minus V1198991003816100381610038161003816 119901119886119899

119873 119886119899 isin 120583 (21)

53 Buffer Prediction Video stalling is one of the importantobjective parameters that affect the subjective perception ofthe users Stalling occurswhen the play-out buffer underrunsTo prevent this event the buffer must be maintained within asafe range In this session we evaluate the buffer performancethrough the average buffer level 119860119887 the CDF of the bufferlevel 119891(119887) (119887 isin [1 119861max]) and the buffer underflowprobability Prund (ie when the system stays at buffer level1) which are described as follows

119860119887 = sum119873119899=1 119887119899119901119886119899119873

119891 (119887) = sum119873119899=1 119887119899119901119886119899119873 | 119887119899 lt 119887

Prund = sum119873119899=1 119901119886119899119873 | 119887119899 = 1

(22)

119860119887 represents the safety of the buffer If 119860119887 is small thebuffer level is often in low levels which may cause playbackinterruptionwhen the current bandwidth drops dramatically119891(119887) reflects the variance of the buffer level under thefluctuations of the network Prund shows the probability thatthe playback would be interrupted in the streaming session

6 Experimental Results

In order to evaluate the proposed system model and per-formance prediction accuracy in this section we perform anumber of experiments in both offline and online contextsand compare the performance predicted results with themeasurement ones We also compare our proposed methodwith two existing ones namely the SDP method presented[10] which could obtain the best performance among the SDPmethods and the bitrate estimation based method presented[14] which is the best among the qualitative methods

Advances in Multimedia 7

61 Experiment Setup For the simulation our test-bed con-sists of a client running Java 80 which implements theadaptation and a server running Apache2 which holds themedia segments The client runs on a Window 7 computerwith an Intel i5-17 GHz CPU and 4GB memory and theserver runs on Ubuntu 1204LTS (with default TCP CUBIC)with 1 G RAM The channel bandwidth is simulated usingDummyNet [28] We use the Tokyo Olympic video from [6]For the video 119881max = 9 and for the bandwidth 119882 = 10Since we measure the buffer size and compute the buffer costin the segment duration unit we implement our adaptationmethod with one setting of the segment duration In ourexperiments we select a segment duration of 2 secondswhich is similar to those of [7 8 10] The impacts of segmentdurations on adaptation performance have been consideredin some recent studies [29 30] Further evaluation of differentsegment durations with a fixed buffer size (in seconds) will bereserved for our future work Maximum buffer level is set to5 segments (ie 10 seconds) and optimal buffer level is set to4 segments (ie 8 seconds)

In our method a streaming provider can adjust thebalance between the requirements for high quality levelpreventing video stalling and reducing quality switchesby changing the trade-off parameters 120572 120574 and 120573 Sinceselecting an optimal combination of trade-off parametersof the cost function involves solving a hard optimizationproblem it will be investigated in our future work In thispaper we select qualitatively the trade-off parameters of costfunction as follows Initially we fix 120573 and select the othertwo parameters Since we want to prioritize requirement ofsmooth quality switch 120574 is selected so that the contributionof the quality switch cost Δ119902 is higher than that of buffercost Δ119887 With parameter 120572 because the bitrate cost Δ119903 can beup to thousands parameter 120572 should be small to reduce thecontribution of the bitrate cost to the overall cost 119862 Basedon our experience good empirical values on parameters 120572 120573and 120574 are 0003 4 and 20 respectively

62 Experimental Results In the first part of the experimentwe evaluate the accuracy of performance prediction in offlinecontext using a given bandwidth trace obtained fromamobilenetwork [12] In this context the number of video segments119871 is 300

Figures 5 and 6 show the predicted performance usingthe formulas presented in Section 5 and themeasured perfor-mance obtained from the experiments when the maximumallowed version is set to 7 8 and 9 These figures point outthat the prediction results are close to the measurement onesWe can see from Figure 5 that when the maximum versionincreases the average version as well as the average versionswitch also increases Figure 6 shows that in both predictionand measurement cases when the maximum version is 7 theunderflowprobability is almost zero and increases very slowlywhen the maximum allowed version increases This analysisimplies that setting themaximumversion to 8 is reasonable interms of balancing between average video quality and qualityswitch meanwhile setting the maximum version to 7 ensuresa very stable streaming experience

0

02

04

06

08

1

Aver

age s

witc

hse

gmen

t

8 97Maximum version

1

3

5

7

9

Vers

ion

Measured avr_switchMeasured avr_version

Predicted avr_switchPredicted avr_version

Figure 5 Predicted performance and measured performance interms of average version and average version switch per segment inoffline context using a given bandwidth trace

0

005

01

Und

er

ow p

roba

bilit

y

0

5

10

Aver

age b

uer

(s)

8 97Maximum version

Measured prob_underowMeasured avr_buer

Predicted prob_underowPredicted avr_buer

Figure 6 Predicted performance and measured performance interms of average buffer and underflow probability in offline contextusing a given bandwidth trace

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

012345678910

Vers

ion

Figure 7 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 9

Figures 7 8 and 9 show the bitrate and version switchbehavior of the proposed method in the three cases ofmaximum version It can be seen very clearly from thesefigures that when the maximum version is reduced thenumber of version switches (or quality changes) decreases

Table 1 shows more detailed statistics of the experimentalresults in three cases of maximum version It can be drawn

8 Advances in Multimedia

Table 1 Compare predicted performance and measured performance in offline context using a given bandwidth trace

Statistics 119881max = 9 119881max = 8 119881max = 7Predicted Measured Predicted Measured Predicted Measured

119860119887(119904) 948 852 959 877 98 968119860119902 764 788 745 773 648 687119860 sw 030 018 014 009 003 002Prundflow 001 000 0006 000 000 000

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

0123456789

Vers

ion

Figure 8 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 8

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

012345678

Vers

ion

Figure 9 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 7

from the table that there is no significant difference betweenthe predicted performance and the measured performance

In the second part of the experiment we use two historybandwidth traces recorded from two previous streamingsessions of the client The CDFs of both bandwidths areshown in Figure 10

Bandwidth 1198871199081 is used for calculating the statisticalmodels and 1198871199082 is used in the simulation for measuringperformance parameters Figures 11 12 and 13 show thebitrate and version switch behavior of the proposed methodin the three cases of maximum version The detailed resultsare listed in Table 2 Based on these figures and tablewe affirm once again that the mathematical performanceprediction model agrees well with the measurement

bw1bw2

0010203040506070809

1

CDF

1000 2000 3000 4000 50000Bandwidth (kbps)

Figure 10 The cumulative distribution function of bw1 and bw2

Segment rpBitrateVersion

100 200 300 400 500 6000Time (s)

01000200030004000500060007000

Bitr

ate (

kbps

)

012345678910

Vers

ion

Figure 11 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 9

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

0123456789

Vers

ion

Figure 12 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 8

Advances in Multimedia 9

Table 2 Compare predicted performance and measured performance in offline context using a history bandwidth trace for statistics

Statistics 119881max = 9 119881max = 8 119881max = 7Predicted Measured Predicted Measured Predicted Measured

119860119887(119904) 952 844 961 885 969 965119860119902 748 792 727 771 680 698119860 sw 031 019 018 011 008 005Prundflow 000 000 000 000 000 000

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

12345678

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 13 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 7

In the third part of the experiment we consider the onlinecontext in which the prediction of the future bandwidth isbased on the statistical parameters of all previous segmentsSpecifically we divide an entire session into chunks each ofwhich has119871 video segmentsWe treat each chunk individuallyas a mini streaming session Assuming that initially we haveenough statistical data to predict the performance of thefirst chunk The prediction of the subsequent chunks willbe done based on the previous chunks At the beginning ofeach chunk the bandwidth statistics are updated leading to arecomputation of the policy set and the performance In theexperiment we set119871 to 100 video segments It can be observedfrom our experiments that it took only several seconds to(re)compute the whole model Therefore the computationaloverhead is about 3 which is appropriate for the onlinecontext Figures 14 15 and 16 show the adaptation bitrateand version switch behavior of the proposed method in thethree cases of maximum version The predicted performanceand the measured performance in the online context in thethree cases are presented in detail in Tables 3 4 and 5It is very clear that in the online context the predictedperformance is also very close to the measured performance

Next we compare our proposed method with the SDPmethod [10] and the bitrate estimation based method [14]using the same simulation settings as in the first part of ourexperimentThe experimental results obtained by simulatingthe SDP method [10] and bitrate estimation based method[14] are shown in Figures 17 and 18 respectively We can seethat both methods provide very fluctuating version switchesThe detailed statistics of these adaptation methods with themaximum version set to 9 and our proposedmethodwith the

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

kbps

)

012345678910

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 14 Experimental results of the proposed method in onlinecontext with Vmax = 9

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

0123456789

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 15 Experimental results of the proposed method in onlinecontext with Vmax = 8

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

12345678

Vers

ion

Figure 16 Experimental results of the proposed method in onlinecontext with Vmax = 7

10 Advances in Multimedia

Table 3 Compare predicted performance and measured performance in online context with 119881max = 9Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 931 896 892 845 936 886119860119902 769 804 789 768 783 772119860 sw 005 009 004 007 002 005Prundflow 000 000 000 000 000 000

Table 4 Compare predicted performance and measured performance in online context with 119881max = 8Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 953 901 942 863 942 896119860119902 756 786 765 763 765 767119860 sw 000 003 001 003 002 004Prundflow 000 000 000 000 000 000

Table 5 Compare predicted performance and measured performance in online context with 119881max = 7Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 977 993 977 952 983 959119860119902 692 700 692 695 692 700119860 sw 000 000 000 002 000 000Prundflow 000 000 000 000 000 000

Table 6 Statistics of different adaptation methods in offline context

Statistics SDP method119881max = 9

Bitrate estimation method119881max = 9

Proposed method119881max = 9

Proposed method119881max = 8

Proposed method119881max = 7

119860119887(119904) 862 885 852 877 968119860119902 740 785 788 773 687119860 sw 076 043 018 009 002Prundflow 000 000 000 000 000

maximumversion set to 7 8 and 9 are shown in Table 6 It canbe seen that the performance of the bitrate estimationmethodis less effective than our method in terms of average versionand average version switch Regarding the SDP method itis evident that the performance of this method is the worstamong all (with low average version and highest averageswitch per segment) This can be expected because thismethod was not originally designed for real VBR videos

The buffer level curves of the three methods in offlinecontext are shown in Figure 19 We can see that all buffercurves imply streaming sessionswithout any freezes for usersAmong these the SDP method and the proposed methodwith 119881max = 9 provide the most unstable buffers while thebitrate estimation based method and the proposed methodwith 119881max = 7 provide the most stable buffers It can beobserved from Figures 17 and 18 and Table 6 that the SDPmethod and bitrate estimation based method result in veryfluctuating version curves and the proposed method with

Segment rpBitrateVersion

010002000300040005000600070008000

100 300 500 600200 4000Time (s)

0246810

Vers

ion

Bitr

ate (

kbps

)

Figure 17 Experimental results of SDP method

119881max = 7 obtains the lowest average quality Therefore fromTable 6 and Figure 19 we can see that the proposedmethod inoffline context with 119881max = 8 or 9 can provide the best per-formance

Advances in Multimedia 11

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

s)

100 200 300 400 500 6000Time (s)

0246810

Vers

ion

Figure 18 Experimental results of bitrate estimation basedmethod

02468

1012

Bue

r (s)

100 300 500 600200 4000Time (s)

SDPMax 9Max 8

Max 7BE

Figure 19 Buffer level curves of SDP method bitrate estimationbased method and the proposed method (Max9 Max8 and Max7)in offline context

7 Conclusion

In this paper we have proposed an adaptation method forHTTP streaming based on stochastic dynamic programmingThe system model was targeted at real bandwidth trace withstrong bitrate fluctuation of VBR videos Furthermore wehave developed a model to predict the system performancewith the aim of choosing the best setting based on theperformance requirements The experimental results haveshown that our method can effectively adapt VBR videos andperform accurate performance prediction which is useful inplanning adaptation policy

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

References

[1] Cisco Systems Inc ldquoNetworking index Forecast and method-ology 2012ndash2017rdquo Cisco Visual May 2013

[2] A C Begen T Akgul and M Baugher ldquoWatching video overthe web part 1 streaming protocolsrdquo IEEE Internet Computingvol 15 no 2 pp 54ndash63 2011

[3] T C Thang Q-D Ho J W Kang and A T Pham ldquoAdaptivestreaming of audiovisual content using MPEG DASHrdquo IEEETransactions on Consumer Electronics vol 58 no 1 pp 78ndash852012

[4] T Stockhammer ldquoDynamic adaptive streaming over HTTPstandards and design principlesrdquo in Proceeding of the 2ndAnnual ACM Multimedia Systems Conference (MMSys rsquo11) pp133ndash144 San Jose CA USA February 2011

[5] K Miller E Quacchio G Gennari and A Wolisz ldquoAdaptationalgorithm for adaptive streaming over HTTPrdquo in Proceedings ofthe 19th International Packet Video Workshop (PV rsquo12) pp 173ndash178 IEEE Munich Germany May 2012

[6] Y Zhou Y Duan J Sun and Z Guo ldquoTowards simple andsmooth rate adaption for VBR video in DASHrdquo in Proceedingsof the IEEE Visual Communications and Image Processing Con-ference (VCIP rsquo14) pp 9ndash12 IEEE Valletta Malta December2014

[7] S Akhshabi S Narayanaswamy A C Begen and C DovrolisldquoAn experimental evaluation of rate-adaptive video players overHTTPrdquo Signal Processing Image Communication vol 27 no 4pp 271ndash287 2012

[8] C Muller S Lederer and C Timmerer ldquoAn evaluation ofdynamic adaptive streaming over HTTP in vehicular environ-mentsrdquo in Proceedings of the 4th Workshop on Mobile Video(MoVid rsquo12) pp 37ndash42 New York NY USA February 2012

[9] C Liu I Bouazizi and M Gabbouj ldquoRate adaptation foradaptive HTTP streamingrdquo in Proceedings of the 2nd AnnualACMMultimedia Systems Conference (MMSys rsquo11) pp 169ndash174San Jose Calif USA February 2011

[10] S Garcıa J Cabrera and N Garcıa ldquoQuality-control algorithmfor adaptive streaming services over wireless channelsrdquo IEEEJournal on Selected Topics in Signal Processing vol 9 no 1 pp50ndash59 2015

[11] A H Duong T Nguyen T Vu T T Do N P Ngoc and T CThang ldquoSDP-based adaptation for quality control in adaptivestreamingrdquo in Proceedings of the International Conference onComputing Management and Telecommunications (ComMan-Tel rsquo15) pp 194ndash199 IEEE DaNang Vietnam December 2015

[12] S Lederer C Mueller C Timmerer C Concolato J Le Feuvreand K Fliegel ldquoDistributedDASHdatasetrdquo in Proceedings of the4th ACMMultimedia Systems Conference (MMSys rsquo13) pp 131ndash135 Oslo Norway March 2013

[13] ldquoH264AVC video trace libraryrdquo httptraceeasasueduh264[14] T CThang H T Le H X Nguyen A T Pham JW Kang and

Y M Ro ldquoAdaptive video streaming over HTTP with dynamicresource estimationrdquo IEEE Trans On consumer electronics vol58 no 1 pp 78ndash85 2012

[15] T C Thang H T Le A T Pham and Y M Ro ldquoAn evaluationof bitrate adaptation methods for HTTP live streamingrdquo IEEEJournal on Selected Areas in Communications vol 32 no 4 pp693ndash705 2014

[16] R Dubin O Hadar and A Dvir ldquoThe effect of client buffer andMBR consideration onDASHAdaptation Logicrdquo in Proceedingsof the 2013 IEEE Wireless Communications and NetworkingConference (WCNC rsquo13) pp 2178ndash2183 IEEE Shanghai ChinaApril 2013

[17] G Liang and B Liang ldquoEffect of delay and buffering on jitter-free streaming over random VBR channelsrdquo IEEE Transactionson Multimedia vol 10 no 6 pp 1128ndash1141 2008

[18] Y Kang H Chen and L Xie ldquoAn artificial-neural-network-based QoE estimation model for Video streaming over wirelessnetworksrdquo in Proceedings of the 2013 IEEECIC InternationalConference on Communications in China (ICCC rsquo13) pp 264ndash269 IEEE Xirsquoan China August 2013

[19] S Xiang L Cai and J Pan ldquoAdaptive scalable video streaminginwireless networksrdquo in Proceedings of the 3rdACMMultimedia

4 Advances in Multimedia

02000400060008000

100001200014000

Bitr

ate (

Kbps

)

20 40 60 80 100 120 140 160 180 2000Segment index

V = 9

V = 8

V = 7

V = 6

V = 5

V = 4

V = 3

V = 2

V = 1

Figure 4 Bitrates of different versions of Tokyo Olympic video [13]

each time step Thus by simply extracting the statistics fromthe bandwidth history the transition probability between allbandwidth states is generated

Similar to the bandwidth trace we divide the buffer into119861levels from 0 to 119861119904 with 119861119904 being the buffer size In additionwe denote the video version by 119881 and represent the versionwith lowest quality as V = 1 and the highest quality as V =Vmax assuming that the video has Vmax different versionsThe bitrates of different versions of a VBR video are shown inFigure 4

Because the segment bitrate of a VBR video versionfluctuates very strongly we divide the bitrate of each versioninto 119868 intervals (from interval 1 to interval I) each of whichis represented by its average bitrate value For example if aversion that has the highest bitrate of 5000 kbps is dividedinto 10 bitrate intervals the interval 1 will range from 0 to500 kbps and its representative bitrate will be 250 kbps

We assume that all versions at a bitrate interval representa separate video flow If the current segment bitrate belongs toone interval the next segment bitrate will also belong to thatinterval regardless of segment versionWith this assumptionwe generate 119868 different policy sets for 119868 bitrate intervalsWhena segment is completely downloaded the client measures itsbitrate to find the bitrate interval it belongs toThen the clientwill determine the policy set corresponding to that bitrateinterval

4 Problem Formulation and Solution

41 SystemState With the systembeing discretized above weobserve the system state variable s119896(119887119896 bw119896 V119896) when a videosegment is completely downloaded at stage 119896 Here 119887119896 is thebuffer level representing the number of segments availablein the buffer bw119896 is the bandwidth whose value belongsto BW119894 and V119896 is the version index of the downloadedsegment The case where 119887119896 = 1 corresponds to the bufferunderflow event In each state 119904119896 the system may choose anyaction 119886 For our system an action is basically a decisionabout the version for the next segment As there are 119881maxversions to choose fromwe have totally119881max possible actions

The system then randomly moves into a new state 119904119896+1 at thenext time step resulting in a corresponding cost119862(119904119896 119904119896+1 119886)With each bitrate interval 119894 (1 le 119894 le 119868) we have a system stateset s119894 and a policy set 120583119894 Let119873 be the number of states in s119894and we have

119873 = 119861 lowast119882 lowast 119881max (1)

42 Transition Probabilities Since the system is stochasticwhich means the system outcome of each action 119886 is notdeterministic the state transition probability between everytwo states that depends on action 119886must be constructed Wedenote the probability that state 119904119896 will lead to state 119904119896+1 givenaction 119886 as follows

119875 (119904119896 119904119896+1 119886) = Pr 119904119896+1 | 119904119896 119886 (2)

Due to the independence among (119887119896 bw119896 V119896) we havePr 119904119896+1 | 119904119896 119886 = Pr 119887119896+1 | 119887119896 bw119896 V119896 119886

sdot Pr bw119896+1 | bw119896Pr V119896+1 | 119886 (3)

In the right hand side of (3) the first term can becalculated as follows

Pr 119887119896+1| 119887119896 bw119896 V119896 119886 = 1 119887119896+1 = 1198871198860 otherwise (4)

where 119887119886 is the next buffer level estimated based on thecurrent system status and action 119886 We calculate 119887119886 as follows

119887119886 = 119887119896 + 1 minus [ 119903119886bw119896+1

] (5)

where 119903119886 is the bitrate of the target versionWhen the throughput significantly drops meaning a very

low value of 119887119908119896+1 119887119886 could be lower than zero Howeverat the beginning of a new stage there is one segment beingdownloaded resulting in at least one segment being always inthe buffer Therefore (5) can be modified as follows

119887119886 = max119887119896 + 1 minus [ 119903119886bw119896+1

] 1 (6)

The second term is easily obtained from the bandwidthmodel And the third term can be simply calculated by

Pr V119896+1 | 119886 = 1 V119896+1 = 1198860 otherwise (7)

Thus expression (3) can be simplified as follows

Pr 119904119896+1 | 119904119896 119886

= Pr bw119896+1 | bw119896 119887119896+1 = 119887119886 V119896+1 = 1198860 otherwise

(8)

Advances in Multimedia 5

Input number of states119873probability function 119875cost function 119862

Output optimal policy for each state 120583119904119896119895 = 0 119895 is the iteration index120583119895119904119896 = 0 forall119904119896 isin s 120583119895119904119896 is policy of state 119904119896 at 119895th iterationrepeat119895 = 119895 + 1Policy evaluation compute 120582119895 ℎ119895(119904119896) using the below119873 + 1 equations

ℎ119895 (119904119873) = 0120582119895 + ℎ119895 (119904119896) = 119862 (119904119896 119904119896+1 120583119895119904119896) +

119873sum119903=1

119875 (119904119896 119904119903 120583119895119904119903) ℎ119895 (119904119903)forall119904119896 isin s

Policy improvement find for all 119904119896120583119895+1119904119896 = argmin

119886[119862(119904119896 119904119896+1 119886) +

119873sum119903=1

119875(119904119896 119904119903 119886)ℎ119895(119904119903)]until 120582119895+1 = 120582119895 and ℎ119895+1(119904119896) = ℎ119895(119904119896) forall119904119896 isin s

Algorithm 1 Finding the policy set for one interval

43 Cost Function In this section a cost function is definedto punish the situations that may cause a decrease in usersrsquoQoE We focus on three objective parameters that affectstrongly the subjective perception of the users which arequality level video stalling and quality switch First the costfunction should favor the selected bitrate to be close to thecurrent bandwidth so it punishes the difference Δ119903 betweenthe current bandwidth and the bitrate of the next segmentselected by action 119886 with

Δ119903 = 1003816100381610038161003816bw119896 minus 1199031198861003816100381610038161003816 (9)

Second to prevent video stalling the buffer level should neverbe underflow We define an optimal buffer level 119887opt that is adesired value the client should try to keep during a streamingsession When the buffer level is close to 119887opt the bufferunderflow is avoided Therefore the cost function penalizesthe deviation Δ119887 of the current buffer level from the optimalbuffer level where

Δ119887 = 10038161003816100381610038161003816119887119896 minus 119887opt10038161003816100381610038161003816 (10)

Third in order to reduce the quality switches the cost func-tion should contain the difference Δ119902 between the selectedquality and the last one To punish a QoE reduction becauseof high quality variations we define Δ119902 as follows

Δ119902 = (119886 minus V119896)2 (11)

Let 120572 120573 120574 be the trade-off parameters of three objectsnamely quality level video stalling and quality switchrespectivelyThe cost incurredwhen the system changes fromstate 119904119896 to state 119904119896+1 given action 119886 can be calculated by

119862 (119904119896 119904119896+1 119886) = 120572Δ119903 + 120573Δ119887 + 120574Δ119902 (12)

44 Optimization Solution As the system is discrete and thenumber of states is large we can formulate an infinite horizon

problem For every state 119904119896 the most appropriate action 119886called policy for state 119904119896 has to be decided so that the meancost per state is minimum As mentioned above our systemhas 119868 system state sets corresponding to 119868 bitrate intervalsso we have to find 119868 corresponding policy sets For simplicitywe only present the optimization solution for a general bitrateinterval with the system state set s and a corresponding policyset 120583 Finding optimal solutions for all bitrate intervals willbe done similarly Mathematically we have to minimize 119862119860which is the average cost per state obtained after downloading119871 video segments 119862119860 is calculated as follows

119862119860 = lim119871rarrinfin

1119871119871

sum119897=1

119875 (119904119896 119904119896+1 119886) 119862119897 (119904119896 119904119896+1 119886) (13)

with 119862119897(119904119896 119904119896+1 119886) being the cost incurred after downloadingthe 119897th segment Here 119871 is the number of state transitions andis also the number of video segments in the session Basedon the standard policy iteration algorithm (PIA) [26] wesolve the IHP problem by using an algorithm as presented inAlgorithm 1

Applying PIA of SDP for 119868 bitrate intervals we wouldgenerate 119868 policy sets which serve like a look-up tablemapping each state to an optimal action Thus the client isable to decide an appropriate version for the next segmentbased on the current system condition

Let 119862prob 119862cost be the computational complexity of thecalculation of transition probability and the cost from onestate to the remaining states respectively Let 119862PIA be thecomputational complexity of Algorithm 1 The complexity ofour model is 119862 = 119862prob +119862cost +119862PIA For each interval eachaction and each state we consider the cost and the transitionprobability to119873minus1 remaining states So the complexity of thecalculation of transition probability and the cost is describedas follows

119862prob = 119862cost = 119874 (119868119881max1198732) (14)

6 Advances in Multimedia

Based on [27] we have

119862PIA = 119874 (119868119881119873max) (15)

Therefore the complexity of our model is

119862 = 119874 (2119868119881max1198732 + 119868119881119873max) (16)

5 Performance Prediction

After Section 4 we achieve 119868 policy sets corresponding to 119868intervals of a video bitrate In this section we use theMarkovchain model to predict the streaming performance for asession Similar to Section 44 this section only presents theperformance prediction for a general bitrate interval with asystem state set s and a corresponding policy set 120583 Predictingperformance for all bitrate intervals will be done similarlyAfter carrying out the calculation for all 119868 bitrate intervalswe take the average values as the final results

The key is to determine the average state probability p119886 =[1199011198861 1199011198862 119901119886119873] with 119901119886119899 (1 le 119899 le 119873) being the averageprobability that the system is at the 119899th state throughout thestreaming session The probability p119886 is the average valueof state probabilities after downloading 119897 segments p119897 =[1199011198971 1199011198972 119901119897119873] (1 le 119897 le 119871) Here 119901119897119899 (1 le 119899 le 119873)is the probability that the system is at the 119899th state afterdownloading 119897 segments From the Markov chain theorythe state probability after downloading 119897 + 1 segments canbe computed as the product of the state probability afterdownloading 119897 segments and a transition matrix p119897+1 =p119897PTS Assuming that the initial probability 1199010 is known Thetransition matrix PTS is a119873times119873matrix which represents thetransition probability from state 119904119896 to state 119904119896+1 PTS is definedas follows

PTS = 119875 (119904119896 119904119896+1) = Pr 119904119896+1 | 119904119896 119886 isin 120583 (17)

The average state probability p119886 is calculated as follows

p119886 = sum119871119897=1 p119897PTS119871 (18)

Currently most of the adaptation algorithms developedfor HTTP streaming are qualitative in the sense that the per-formance metrics could only be obtained after the streamingsession In this study the predicted streaming performancecould be calculated based on the average state probabilityand the information inside every state Specifically wemainlyfocus on the following aspects bitrate prediction qualityswitch prediction and buffer prediction

51 Quality Prediction The video quality that the users per-ceive is presented through the selected versionThehigher theversion is selected the better the video quality is perceivedby the users Furthermore setting the maximum version forthe streaming session also affects the perceptual quality of theusers Obviously a very low value of the maximum versionmay result in a poor perceptual quality while a very highvalue may increase the chance of buffer underflow In thissection we predict the quality performance of the streaming

session based on the average version 119860V that is calculatedusing (19) and the cumulative distribution function (CDF)of the versions 119891(V) (V isin [1 119881max]) that is shown in (20)Based on the predicted probability of the versions throughouta streaming session the maximum version could be decidedfor the session

119860V = sum119873119899=1 V119899119901119886119899119873 (19)

119891 (V) = sum119873119899=1 V119899119901119886119899119873 | V = V119899 (20)

52 Quality Switch Prediction Quality switch is an importantfactor affecting the perception of the users The users oftenexpect a smooth playback with the minimum number ofquality switches and small switch amplitude from one seg-ment to the next We can predict the average version switchper segment 119860 sw as follows

119860 sw = sum119873119899=1 1003816100381610038161003816119886119899 minus V1198991003816100381610038161003816 119901119886119899

119873 119886119899 isin 120583 (21)

53 Buffer Prediction Video stalling is one of the importantobjective parameters that affect the subjective perception ofthe users Stalling occurswhen the play-out buffer underrunsTo prevent this event the buffer must be maintained within asafe range In this session we evaluate the buffer performancethrough the average buffer level 119860119887 the CDF of the bufferlevel 119891(119887) (119887 isin [1 119861max]) and the buffer underflowprobability Prund (ie when the system stays at buffer level1) which are described as follows

119860119887 = sum119873119899=1 119887119899119901119886119899119873

119891 (119887) = sum119873119899=1 119887119899119901119886119899119873 | 119887119899 lt 119887

Prund = sum119873119899=1 119901119886119899119873 | 119887119899 = 1

(22)

119860119887 represents the safety of the buffer If 119860119887 is small thebuffer level is often in low levels which may cause playbackinterruptionwhen the current bandwidth drops dramatically119891(119887) reflects the variance of the buffer level under thefluctuations of the network Prund shows the probability thatthe playback would be interrupted in the streaming session

6 Experimental Results

In order to evaluate the proposed system model and per-formance prediction accuracy in this section we perform anumber of experiments in both offline and online contextsand compare the performance predicted results with themeasurement ones We also compare our proposed methodwith two existing ones namely the SDP method presented[10] which could obtain the best performance among the SDPmethods and the bitrate estimation based method presented[14] which is the best among the qualitative methods

Advances in Multimedia 7

61 Experiment Setup For the simulation our test-bed con-sists of a client running Java 80 which implements theadaptation and a server running Apache2 which holds themedia segments The client runs on a Window 7 computerwith an Intel i5-17 GHz CPU and 4GB memory and theserver runs on Ubuntu 1204LTS (with default TCP CUBIC)with 1 G RAM The channel bandwidth is simulated usingDummyNet [28] We use the Tokyo Olympic video from [6]For the video 119881max = 9 and for the bandwidth 119882 = 10Since we measure the buffer size and compute the buffer costin the segment duration unit we implement our adaptationmethod with one setting of the segment duration In ourexperiments we select a segment duration of 2 secondswhich is similar to those of [7 8 10] The impacts of segmentdurations on adaptation performance have been consideredin some recent studies [29 30] Further evaluation of differentsegment durations with a fixed buffer size (in seconds) will bereserved for our future work Maximum buffer level is set to5 segments (ie 10 seconds) and optimal buffer level is set to4 segments (ie 8 seconds)

In our method a streaming provider can adjust thebalance between the requirements for high quality levelpreventing video stalling and reducing quality switchesby changing the trade-off parameters 120572 120574 and 120573 Sinceselecting an optimal combination of trade-off parametersof the cost function involves solving a hard optimizationproblem it will be investigated in our future work In thispaper we select qualitatively the trade-off parameters of costfunction as follows Initially we fix 120573 and select the othertwo parameters Since we want to prioritize requirement ofsmooth quality switch 120574 is selected so that the contributionof the quality switch cost Δ119902 is higher than that of buffercost Δ119887 With parameter 120572 because the bitrate cost Δ119903 can beup to thousands parameter 120572 should be small to reduce thecontribution of the bitrate cost to the overall cost 119862 Basedon our experience good empirical values on parameters 120572 120573and 120574 are 0003 4 and 20 respectively

62 Experimental Results In the first part of the experimentwe evaluate the accuracy of performance prediction in offlinecontext using a given bandwidth trace obtained fromamobilenetwork [12] In this context the number of video segments119871 is 300

Figures 5 and 6 show the predicted performance usingthe formulas presented in Section 5 and themeasured perfor-mance obtained from the experiments when the maximumallowed version is set to 7 8 and 9 These figures point outthat the prediction results are close to the measurement onesWe can see from Figure 5 that when the maximum versionincreases the average version as well as the average versionswitch also increases Figure 6 shows that in both predictionand measurement cases when the maximum version is 7 theunderflowprobability is almost zero and increases very slowlywhen the maximum allowed version increases This analysisimplies that setting themaximumversion to 8 is reasonable interms of balancing between average video quality and qualityswitch meanwhile setting the maximum version to 7 ensuresa very stable streaming experience

0

02

04

06

08

1

Aver

age s

witc

hse

gmen

t

8 97Maximum version

1

3

5

7

9

Vers

ion

Measured avr_switchMeasured avr_version

Predicted avr_switchPredicted avr_version

Figure 5 Predicted performance and measured performance interms of average version and average version switch per segment inoffline context using a given bandwidth trace

0

005

01

Und

er

ow p

roba

bilit

y

0

5

10

Aver

age b

uer

(s)

8 97Maximum version

Measured prob_underowMeasured avr_buer

Predicted prob_underowPredicted avr_buer

Figure 6 Predicted performance and measured performance interms of average buffer and underflow probability in offline contextusing a given bandwidth trace

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

012345678910

Vers

ion

Figure 7 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 9

Figures 7 8 and 9 show the bitrate and version switchbehavior of the proposed method in the three cases ofmaximum version It can be seen very clearly from thesefigures that when the maximum version is reduced thenumber of version switches (or quality changes) decreases

Table 1 shows more detailed statistics of the experimentalresults in three cases of maximum version It can be drawn

8 Advances in Multimedia

Table 1 Compare predicted performance and measured performance in offline context using a given bandwidth trace

Statistics 119881max = 9 119881max = 8 119881max = 7Predicted Measured Predicted Measured Predicted Measured

119860119887(119904) 948 852 959 877 98 968119860119902 764 788 745 773 648 687119860 sw 030 018 014 009 003 002Prundflow 001 000 0006 000 000 000

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

0123456789

Vers

ion

Figure 8 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 8

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

012345678

Vers

ion

Figure 9 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 7

from the table that there is no significant difference betweenthe predicted performance and the measured performance

In the second part of the experiment we use two historybandwidth traces recorded from two previous streamingsessions of the client The CDFs of both bandwidths areshown in Figure 10

Bandwidth 1198871199081 is used for calculating the statisticalmodels and 1198871199082 is used in the simulation for measuringperformance parameters Figures 11 12 and 13 show thebitrate and version switch behavior of the proposed methodin the three cases of maximum version The detailed resultsare listed in Table 2 Based on these figures and tablewe affirm once again that the mathematical performanceprediction model agrees well with the measurement

bw1bw2

0010203040506070809

1

CDF

1000 2000 3000 4000 50000Bandwidth (kbps)

Figure 10 The cumulative distribution function of bw1 and bw2

Segment rpBitrateVersion

100 200 300 400 500 6000Time (s)

01000200030004000500060007000

Bitr

ate (

kbps

)

012345678910

Vers

ion

Figure 11 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 9

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

0123456789

Vers

ion

Figure 12 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 8

Advances in Multimedia 9

Table 2 Compare predicted performance and measured performance in offline context using a history bandwidth trace for statistics

Statistics 119881max = 9 119881max = 8 119881max = 7Predicted Measured Predicted Measured Predicted Measured

119860119887(119904) 952 844 961 885 969 965119860119902 748 792 727 771 680 698119860 sw 031 019 018 011 008 005Prundflow 000 000 000 000 000 000

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

12345678

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 13 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 7

In the third part of the experiment we consider the onlinecontext in which the prediction of the future bandwidth isbased on the statistical parameters of all previous segmentsSpecifically we divide an entire session into chunks each ofwhich has119871 video segmentsWe treat each chunk individuallyas a mini streaming session Assuming that initially we haveenough statistical data to predict the performance of thefirst chunk The prediction of the subsequent chunks willbe done based on the previous chunks At the beginning ofeach chunk the bandwidth statistics are updated leading to arecomputation of the policy set and the performance In theexperiment we set119871 to 100 video segments It can be observedfrom our experiments that it took only several seconds to(re)compute the whole model Therefore the computationaloverhead is about 3 which is appropriate for the onlinecontext Figures 14 15 and 16 show the adaptation bitrateand version switch behavior of the proposed method in thethree cases of maximum version The predicted performanceand the measured performance in the online context in thethree cases are presented in detail in Tables 3 4 and 5It is very clear that in the online context the predictedperformance is also very close to the measured performance

Next we compare our proposed method with the SDPmethod [10] and the bitrate estimation based method [14]using the same simulation settings as in the first part of ourexperimentThe experimental results obtained by simulatingthe SDP method [10] and bitrate estimation based method[14] are shown in Figures 17 and 18 respectively We can seethat both methods provide very fluctuating version switchesThe detailed statistics of these adaptation methods with themaximum version set to 9 and our proposedmethodwith the

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

kbps

)

012345678910

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 14 Experimental results of the proposed method in onlinecontext with Vmax = 9

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

0123456789

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 15 Experimental results of the proposed method in onlinecontext with Vmax = 8

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

12345678

Vers

ion

Figure 16 Experimental results of the proposed method in onlinecontext with Vmax = 7

10 Advances in Multimedia

Table 3 Compare predicted performance and measured performance in online context with 119881max = 9Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 931 896 892 845 936 886119860119902 769 804 789 768 783 772119860 sw 005 009 004 007 002 005Prundflow 000 000 000 000 000 000

Table 4 Compare predicted performance and measured performance in online context with 119881max = 8Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 953 901 942 863 942 896119860119902 756 786 765 763 765 767119860 sw 000 003 001 003 002 004Prundflow 000 000 000 000 000 000

Table 5 Compare predicted performance and measured performance in online context with 119881max = 7Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 977 993 977 952 983 959119860119902 692 700 692 695 692 700119860 sw 000 000 000 002 000 000Prundflow 000 000 000 000 000 000

Table 6 Statistics of different adaptation methods in offline context

Statistics SDP method119881max = 9

Bitrate estimation method119881max = 9

Proposed method119881max = 9

Proposed method119881max = 8

Proposed method119881max = 7

119860119887(119904) 862 885 852 877 968119860119902 740 785 788 773 687119860 sw 076 043 018 009 002Prundflow 000 000 000 000 000

maximumversion set to 7 8 and 9 are shown in Table 6 It canbe seen that the performance of the bitrate estimationmethodis less effective than our method in terms of average versionand average version switch Regarding the SDP method itis evident that the performance of this method is the worstamong all (with low average version and highest averageswitch per segment) This can be expected because thismethod was not originally designed for real VBR videos

The buffer level curves of the three methods in offlinecontext are shown in Figure 19 We can see that all buffercurves imply streaming sessionswithout any freezes for usersAmong these the SDP method and the proposed methodwith 119881max = 9 provide the most unstable buffers while thebitrate estimation based method and the proposed methodwith 119881max = 7 provide the most stable buffers It can beobserved from Figures 17 and 18 and Table 6 that the SDPmethod and bitrate estimation based method result in veryfluctuating version curves and the proposed method with

Segment rpBitrateVersion

010002000300040005000600070008000

100 300 500 600200 4000Time (s)

0246810

Vers

ion

Bitr

ate (

kbps

)

Figure 17 Experimental results of SDP method

119881max = 7 obtains the lowest average quality Therefore fromTable 6 and Figure 19 we can see that the proposedmethod inoffline context with 119881max = 8 or 9 can provide the best per-formance

Advances in Multimedia 11

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

s)

100 200 300 400 500 6000Time (s)

0246810

Vers

ion

Figure 18 Experimental results of bitrate estimation basedmethod

02468

1012

Bue

r (s)

100 300 500 600200 4000Time (s)

SDPMax 9Max 8

Max 7BE

Figure 19 Buffer level curves of SDP method bitrate estimationbased method and the proposed method (Max9 Max8 and Max7)in offline context

7 Conclusion

In this paper we have proposed an adaptation method forHTTP streaming based on stochastic dynamic programmingThe system model was targeted at real bandwidth trace withstrong bitrate fluctuation of VBR videos Furthermore wehave developed a model to predict the system performancewith the aim of choosing the best setting based on theperformance requirements The experimental results haveshown that our method can effectively adapt VBR videos andperform accurate performance prediction which is useful inplanning adaptation policy

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

References

[1] Cisco Systems Inc ldquoNetworking index Forecast and method-ology 2012ndash2017rdquo Cisco Visual May 2013

[2] A C Begen T Akgul and M Baugher ldquoWatching video overthe web part 1 streaming protocolsrdquo IEEE Internet Computingvol 15 no 2 pp 54ndash63 2011

[3] T C Thang Q-D Ho J W Kang and A T Pham ldquoAdaptivestreaming of audiovisual content using MPEG DASHrdquo IEEETransactions on Consumer Electronics vol 58 no 1 pp 78ndash852012

[4] T Stockhammer ldquoDynamic adaptive streaming over HTTPstandards and design principlesrdquo in Proceeding of the 2ndAnnual ACM Multimedia Systems Conference (MMSys rsquo11) pp133ndash144 San Jose CA USA February 2011

[5] K Miller E Quacchio G Gennari and A Wolisz ldquoAdaptationalgorithm for adaptive streaming over HTTPrdquo in Proceedings ofthe 19th International Packet Video Workshop (PV rsquo12) pp 173ndash178 IEEE Munich Germany May 2012

[6] Y Zhou Y Duan J Sun and Z Guo ldquoTowards simple andsmooth rate adaption for VBR video in DASHrdquo in Proceedingsof the IEEE Visual Communications and Image Processing Con-ference (VCIP rsquo14) pp 9ndash12 IEEE Valletta Malta December2014

[7] S Akhshabi S Narayanaswamy A C Begen and C DovrolisldquoAn experimental evaluation of rate-adaptive video players overHTTPrdquo Signal Processing Image Communication vol 27 no 4pp 271ndash287 2012

[8] C Muller S Lederer and C Timmerer ldquoAn evaluation ofdynamic adaptive streaming over HTTP in vehicular environ-mentsrdquo in Proceedings of the 4th Workshop on Mobile Video(MoVid rsquo12) pp 37ndash42 New York NY USA February 2012

[9] C Liu I Bouazizi and M Gabbouj ldquoRate adaptation foradaptive HTTP streamingrdquo in Proceedings of the 2nd AnnualACMMultimedia Systems Conference (MMSys rsquo11) pp 169ndash174San Jose Calif USA February 2011

[10] S Garcıa J Cabrera and N Garcıa ldquoQuality-control algorithmfor adaptive streaming services over wireless channelsrdquo IEEEJournal on Selected Topics in Signal Processing vol 9 no 1 pp50ndash59 2015

[11] A H Duong T Nguyen T Vu T T Do N P Ngoc and T CThang ldquoSDP-based adaptation for quality control in adaptivestreamingrdquo in Proceedings of the International Conference onComputing Management and Telecommunications (ComMan-Tel rsquo15) pp 194ndash199 IEEE DaNang Vietnam December 2015

[12] S Lederer C Mueller C Timmerer C Concolato J Le Feuvreand K Fliegel ldquoDistributedDASHdatasetrdquo in Proceedings of the4th ACMMultimedia Systems Conference (MMSys rsquo13) pp 131ndash135 Oslo Norway March 2013

[13] ldquoH264AVC video trace libraryrdquo httptraceeasasueduh264[14] T CThang H T Le H X Nguyen A T Pham JW Kang and

Y M Ro ldquoAdaptive video streaming over HTTP with dynamicresource estimationrdquo IEEE Trans On consumer electronics vol58 no 1 pp 78ndash85 2012

[15] T C Thang H T Le A T Pham and Y M Ro ldquoAn evaluationof bitrate adaptation methods for HTTP live streamingrdquo IEEEJournal on Selected Areas in Communications vol 32 no 4 pp693ndash705 2014

[16] R Dubin O Hadar and A Dvir ldquoThe effect of client buffer andMBR consideration onDASHAdaptation Logicrdquo in Proceedingsof the 2013 IEEE Wireless Communications and NetworkingConference (WCNC rsquo13) pp 2178ndash2183 IEEE Shanghai ChinaApril 2013

[17] G Liang and B Liang ldquoEffect of delay and buffering on jitter-free streaming over random VBR channelsrdquo IEEE Transactionson Multimedia vol 10 no 6 pp 1128ndash1141 2008

[18] Y Kang H Chen and L Xie ldquoAn artificial-neural-network-based QoE estimation model for Video streaming over wirelessnetworksrdquo in Proceedings of the 2013 IEEECIC InternationalConference on Communications in China (ICCC rsquo13) pp 264ndash269 IEEE Xirsquoan China August 2013

[19] S Xiang L Cai and J Pan ldquoAdaptive scalable video streaminginwireless networksrdquo in Proceedings of the 3rdACMMultimedia

Advances in Multimedia 5

Input number of states119873probability function 119875cost function 119862

Output optimal policy for each state 120583119904119896119895 = 0 119895 is the iteration index120583119895119904119896 = 0 forall119904119896 isin s 120583119895119904119896 is policy of state 119904119896 at 119895th iterationrepeat119895 = 119895 + 1Policy evaluation compute 120582119895 ℎ119895(119904119896) using the below119873 + 1 equations

ℎ119895 (119904119873) = 0120582119895 + ℎ119895 (119904119896) = 119862 (119904119896 119904119896+1 120583119895119904119896) +

119873sum119903=1

119875 (119904119896 119904119903 120583119895119904119903) ℎ119895 (119904119903)forall119904119896 isin s

Policy improvement find for all 119904119896120583119895+1119904119896 = argmin

119886[119862(119904119896 119904119896+1 119886) +

119873sum119903=1

119875(119904119896 119904119903 119886)ℎ119895(119904119903)]until 120582119895+1 = 120582119895 and ℎ119895+1(119904119896) = ℎ119895(119904119896) forall119904119896 isin s

Algorithm 1 Finding the policy set for one interval

43 Cost Function In this section a cost function is definedto punish the situations that may cause a decrease in usersrsquoQoE We focus on three objective parameters that affectstrongly the subjective perception of the users which arequality level video stalling and quality switch First the costfunction should favor the selected bitrate to be close to thecurrent bandwidth so it punishes the difference Δ119903 betweenthe current bandwidth and the bitrate of the next segmentselected by action 119886 with

Δ119903 = 1003816100381610038161003816bw119896 minus 1199031198861003816100381610038161003816 (9)

Second to prevent video stalling the buffer level should neverbe underflow We define an optimal buffer level 119887opt that is adesired value the client should try to keep during a streamingsession When the buffer level is close to 119887opt the bufferunderflow is avoided Therefore the cost function penalizesthe deviation Δ119887 of the current buffer level from the optimalbuffer level where

Δ119887 = 10038161003816100381610038161003816119887119896 minus 119887opt10038161003816100381610038161003816 (10)

Third in order to reduce the quality switches the cost func-tion should contain the difference Δ119902 between the selectedquality and the last one To punish a QoE reduction becauseof high quality variations we define Δ119902 as follows

Δ119902 = (119886 minus V119896)2 (11)

Let 120572 120573 120574 be the trade-off parameters of three objectsnamely quality level video stalling and quality switchrespectivelyThe cost incurredwhen the system changes fromstate 119904119896 to state 119904119896+1 given action 119886 can be calculated by

119862 (119904119896 119904119896+1 119886) = 120572Δ119903 + 120573Δ119887 + 120574Δ119902 (12)

44 Optimization Solution As the system is discrete and thenumber of states is large we can formulate an infinite horizon

problem For every state 119904119896 the most appropriate action 119886called policy for state 119904119896 has to be decided so that the meancost per state is minimum As mentioned above our systemhas 119868 system state sets corresponding to 119868 bitrate intervalsso we have to find 119868 corresponding policy sets For simplicitywe only present the optimization solution for a general bitrateinterval with the system state set s and a corresponding policyset 120583 Finding optimal solutions for all bitrate intervals willbe done similarly Mathematically we have to minimize 119862119860which is the average cost per state obtained after downloading119871 video segments 119862119860 is calculated as follows

119862119860 = lim119871rarrinfin

1119871119871

sum119897=1

119875 (119904119896 119904119896+1 119886) 119862119897 (119904119896 119904119896+1 119886) (13)

with 119862119897(119904119896 119904119896+1 119886) being the cost incurred after downloadingthe 119897th segment Here 119871 is the number of state transitions andis also the number of video segments in the session Basedon the standard policy iteration algorithm (PIA) [26] wesolve the IHP problem by using an algorithm as presented inAlgorithm 1

Applying PIA of SDP for 119868 bitrate intervals we wouldgenerate 119868 policy sets which serve like a look-up tablemapping each state to an optimal action Thus the client isable to decide an appropriate version for the next segmentbased on the current system condition

Let 119862prob 119862cost be the computational complexity of thecalculation of transition probability and the cost from onestate to the remaining states respectively Let 119862PIA be thecomputational complexity of Algorithm 1 The complexity ofour model is 119862 = 119862prob +119862cost +119862PIA For each interval eachaction and each state we consider the cost and the transitionprobability to119873minus1 remaining states So the complexity of thecalculation of transition probability and the cost is describedas follows

119862prob = 119862cost = 119874 (119868119881max1198732) (14)

6 Advances in Multimedia

Based on [27] we have

119862PIA = 119874 (119868119881119873max) (15)

Therefore the complexity of our model is

119862 = 119874 (2119868119881max1198732 + 119868119881119873max) (16)

5 Performance Prediction

After Section 4 we achieve 119868 policy sets corresponding to 119868intervals of a video bitrate In this section we use theMarkovchain model to predict the streaming performance for asession Similar to Section 44 this section only presents theperformance prediction for a general bitrate interval with asystem state set s and a corresponding policy set 120583 Predictingperformance for all bitrate intervals will be done similarlyAfter carrying out the calculation for all 119868 bitrate intervalswe take the average values as the final results

The key is to determine the average state probability p119886 =[1199011198861 1199011198862 119901119886119873] with 119901119886119899 (1 le 119899 le 119873) being the averageprobability that the system is at the 119899th state throughout thestreaming session The probability p119886 is the average valueof state probabilities after downloading 119897 segments p119897 =[1199011198971 1199011198972 119901119897119873] (1 le 119897 le 119871) Here 119901119897119899 (1 le 119899 le 119873)is the probability that the system is at the 119899th state afterdownloading 119897 segments From the Markov chain theorythe state probability after downloading 119897 + 1 segments canbe computed as the product of the state probability afterdownloading 119897 segments and a transition matrix p119897+1 =p119897PTS Assuming that the initial probability 1199010 is known Thetransition matrix PTS is a119873times119873matrix which represents thetransition probability from state 119904119896 to state 119904119896+1 PTS is definedas follows

PTS = 119875 (119904119896 119904119896+1) = Pr 119904119896+1 | 119904119896 119886 isin 120583 (17)

The average state probability p119886 is calculated as follows

p119886 = sum119871119897=1 p119897PTS119871 (18)

Currently most of the adaptation algorithms developedfor HTTP streaming are qualitative in the sense that the per-formance metrics could only be obtained after the streamingsession In this study the predicted streaming performancecould be calculated based on the average state probabilityand the information inside every state Specifically wemainlyfocus on the following aspects bitrate prediction qualityswitch prediction and buffer prediction

51 Quality Prediction The video quality that the users per-ceive is presented through the selected versionThehigher theversion is selected the better the video quality is perceivedby the users Furthermore setting the maximum version forthe streaming session also affects the perceptual quality of theusers Obviously a very low value of the maximum versionmay result in a poor perceptual quality while a very highvalue may increase the chance of buffer underflow In thissection we predict the quality performance of the streaming

session based on the average version 119860V that is calculatedusing (19) and the cumulative distribution function (CDF)of the versions 119891(V) (V isin [1 119881max]) that is shown in (20)Based on the predicted probability of the versions throughouta streaming session the maximum version could be decidedfor the session

119860V = sum119873119899=1 V119899119901119886119899119873 (19)

119891 (V) = sum119873119899=1 V119899119901119886119899119873 | V = V119899 (20)

52 Quality Switch Prediction Quality switch is an importantfactor affecting the perception of the users The users oftenexpect a smooth playback with the minimum number ofquality switches and small switch amplitude from one seg-ment to the next We can predict the average version switchper segment 119860 sw as follows

119860 sw = sum119873119899=1 1003816100381610038161003816119886119899 minus V1198991003816100381610038161003816 119901119886119899

119873 119886119899 isin 120583 (21)

53 Buffer Prediction Video stalling is one of the importantobjective parameters that affect the subjective perception ofthe users Stalling occurswhen the play-out buffer underrunsTo prevent this event the buffer must be maintained within asafe range In this session we evaluate the buffer performancethrough the average buffer level 119860119887 the CDF of the bufferlevel 119891(119887) (119887 isin [1 119861max]) and the buffer underflowprobability Prund (ie when the system stays at buffer level1) which are described as follows

119860119887 = sum119873119899=1 119887119899119901119886119899119873

119891 (119887) = sum119873119899=1 119887119899119901119886119899119873 | 119887119899 lt 119887

Prund = sum119873119899=1 119901119886119899119873 | 119887119899 = 1

(22)

119860119887 represents the safety of the buffer If 119860119887 is small thebuffer level is often in low levels which may cause playbackinterruptionwhen the current bandwidth drops dramatically119891(119887) reflects the variance of the buffer level under thefluctuations of the network Prund shows the probability thatthe playback would be interrupted in the streaming session

6 Experimental Results

In order to evaluate the proposed system model and per-formance prediction accuracy in this section we perform anumber of experiments in both offline and online contextsand compare the performance predicted results with themeasurement ones We also compare our proposed methodwith two existing ones namely the SDP method presented[10] which could obtain the best performance among the SDPmethods and the bitrate estimation based method presented[14] which is the best among the qualitative methods

Advances in Multimedia 7

61 Experiment Setup For the simulation our test-bed con-sists of a client running Java 80 which implements theadaptation and a server running Apache2 which holds themedia segments The client runs on a Window 7 computerwith an Intel i5-17 GHz CPU and 4GB memory and theserver runs on Ubuntu 1204LTS (with default TCP CUBIC)with 1 G RAM The channel bandwidth is simulated usingDummyNet [28] We use the Tokyo Olympic video from [6]For the video 119881max = 9 and for the bandwidth 119882 = 10Since we measure the buffer size and compute the buffer costin the segment duration unit we implement our adaptationmethod with one setting of the segment duration In ourexperiments we select a segment duration of 2 secondswhich is similar to those of [7 8 10] The impacts of segmentdurations on adaptation performance have been consideredin some recent studies [29 30] Further evaluation of differentsegment durations with a fixed buffer size (in seconds) will bereserved for our future work Maximum buffer level is set to5 segments (ie 10 seconds) and optimal buffer level is set to4 segments (ie 8 seconds)

In our method a streaming provider can adjust thebalance between the requirements for high quality levelpreventing video stalling and reducing quality switchesby changing the trade-off parameters 120572 120574 and 120573 Sinceselecting an optimal combination of trade-off parametersof the cost function involves solving a hard optimizationproblem it will be investigated in our future work In thispaper we select qualitatively the trade-off parameters of costfunction as follows Initially we fix 120573 and select the othertwo parameters Since we want to prioritize requirement ofsmooth quality switch 120574 is selected so that the contributionof the quality switch cost Δ119902 is higher than that of buffercost Δ119887 With parameter 120572 because the bitrate cost Δ119903 can beup to thousands parameter 120572 should be small to reduce thecontribution of the bitrate cost to the overall cost 119862 Basedon our experience good empirical values on parameters 120572 120573and 120574 are 0003 4 and 20 respectively

62 Experimental Results In the first part of the experimentwe evaluate the accuracy of performance prediction in offlinecontext using a given bandwidth trace obtained fromamobilenetwork [12] In this context the number of video segments119871 is 300

Figures 5 and 6 show the predicted performance usingthe formulas presented in Section 5 and themeasured perfor-mance obtained from the experiments when the maximumallowed version is set to 7 8 and 9 These figures point outthat the prediction results are close to the measurement onesWe can see from Figure 5 that when the maximum versionincreases the average version as well as the average versionswitch also increases Figure 6 shows that in both predictionand measurement cases when the maximum version is 7 theunderflowprobability is almost zero and increases very slowlywhen the maximum allowed version increases This analysisimplies that setting themaximumversion to 8 is reasonable interms of balancing between average video quality and qualityswitch meanwhile setting the maximum version to 7 ensuresa very stable streaming experience

0

02

04

06

08

1

Aver

age s

witc

hse

gmen

t

8 97Maximum version

1

3

5

7

9

Vers

ion

Measured avr_switchMeasured avr_version

Predicted avr_switchPredicted avr_version

Figure 5 Predicted performance and measured performance interms of average version and average version switch per segment inoffline context using a given bandwidth trace

0

005

01

Und

er

ow p

roba

bilit

y

0

5

10

Aver

age b

uer

(s)

8 97Maximum version

Measured prob_underowMeasured avr_buer

Predicted prob_underowPredicted avr_buer

Figure 6 Predicted performance and measured performance interms of average buffer and underflow probability in offline contextusing a given bandwidth trace

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

012345678910

Vers

ion

Figure 7 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 9

Figures 7 8 and 9 show the bitrate and version switchbehavior of the proposed method in the three cases ofmaximum version It can be seen very clearly from thesefigures that when the maximum version is reduced thenumber of version switches (or quality changes) decreases

Table 1 shows more detailed statistics of the experimentalresults in three cases of maximum version It can be drawn

8 Advances in Multimedia

Table 1 Compare predicted performance and measured performance in offline context using a given bandwidth trace

Statistics 119881max = 9 119881max = 8 119881max = 7Predicted Measured Predicted Measured Predicted Measured

119860119887(119904) 948 852 959 877 98 968119860119902 764 788 745 773 648 687119860 sw 030 018 014 009 003 002Prundflow 001 000 0006 000 000 000

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

0123456789

Vers

ion

Figure 8 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 8

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

012345678

Vers

ion

Figure 9 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 7

from the table that there is no significant difference betweenthe predicted performance and the measured performance

In the second part of the experiment we use two historybandwidth traces recorded from two previous streamingsessions of the client The CDFs of both bandwidths areshown in Figure 10

Bandwidth 1198871199081 is used for calculating the statisticalmodels and 1198871199082 is used in the simulation for measuringperformance parameters Figures 11 12 and 13 show thebitrate and version switch behavior of the proposed methodin the three cases of maximum version The detailed resultsare listed in Table 2 Based on these figures and tablewe affirm once again that the mathematical performanceprediction model agrees well with the measurement

bw1bw2

0010203040506070809

1

CDF

1000 2000 3000 4000 50000Bandwidth (kbps)

Figure 10 The cumulative distribution function of bw1 and bw2

Segment rpBitrateVersion

100 200 300 400 500 6000Time (s)

01000200030004000500060007000

Bitr

ate (

kbps

)

012345678910

Vers

ion

Figure 11 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 9

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

0123456789

Vers

ion

Figure 12 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 8

Advances in Multimedia 9

Table 2 Compare predicted performance and measured performance in offline context using a history bandwidth trace for statistics

Statistics 119881max = 9 119881max = 8 119881max = 7Predicted Measured Predicted Measured Predicted Measured

119860119887(119904) 952 844 961 885 969 965119860119902 748 792 727 771 680 698119860 sw 031 019 018 011 008 005Prundflow 000 000 000 000 000 000

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

12345678

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 13 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 7

In the third part of the experiment we consider the onlinecontext in which the prediction of the future bandwidth isbased on the statistical parameters of all previous segmentsSpecifically we divide an entire session into chunks each ofwhich has119871 video segmentsWe treat each chunk individuallyas a mini streaming session Assuming that initially we haveenough statistical data to predict the performance of thefirst chunk The prediction of the subsequent chunks willbe done based on the previous chunks At the beginning ofeach chunk the bandwidth statistics are updated leading to arecomputation of the policy set and the performance In theexperiment we set119871 to 100 video segments It can be observedfrom our experiments that it took only several seconds to(re)compute the whole model Therefore the computationaloverhead is about 3 which is appropriate for the onlinecontext Figures 14 15 and 16 show the adaptation bitrateand version switch behavior of the proposed method in thethree cases of maximum version The predicted performanceand the measured performance in the online context in thethree cases are presented in detail in Tables 3 4 and 5It is very clear that in the online context the predictedperformance is also very close to the measured performance

Next we compare our proposed method with the SDPmethod [10] and the bitrate estimation based method [14]using the same simulation settings as in the first part of ourexperimentThe experimental results obtained by simulatingthe SDP method [10] and bitrate estimation based method[14] are shown in Figures 17 and 18 respectively We can seethat both methods provide very fluctuating version switchesThe detailed statistics of these adaptation methods with themaximum version set to 9 and our proposedmethodwith the

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

kbps

)

012345678910

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 14 Experimental results of the proposed method in onlinecontext with Vmax = 9

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

0123456789

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 15 Experimental results of the proposed method in onlinecontext with Vmax = 8

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

12345678

Vers

ion

Figure 16 Experimental results of the proposed method in onlinecontext with Vmax = 7

10 Advances in Multimedia

Table 3 Compare predicted performance and measured performance in online context with 119881max = 9Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 931 896 892 845 936 886119860119902 769 804 789 768 783 772119860 sw 005 009 004 007 002 005Prundflow 000 000 000 000 000 000

Table 4 Compare predicted performance and measured performance in online context with 119881max = 8Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 953 901 942 863 942 896119860119902 756 786 765 763 765 767119860 sw 000 003 001 003 002 004Prundflow 000 000 000 000 000 000

Table 5 Compare predicted performance and measured performance in online context with 119881max = 7Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 977 993 977 952 983 959119860119902 692 700 692 695 692 700119860 sw 000 000 000 002 000 000Prundflow 000 000 000 000 000 000

Table 6 Statistics of different adaptation methods in offline context

Statistics SDP method119881max = 9

Bitrate estimation method119881max = 9

Proposed method119881max = 9

Proposed method119881max = 8

Proposed method119881max = 7

119860119887(119904) 862 885 852 877 968119860119902 740 785 788 773 687119860 sw 076 043 018 009 002Prundflow 000 000 000 000 000

maximumversion set to 7 8 and 9 are shown in Table 6 It canbe seen that the performance of the bitrate estimationmethodis less effective than our method in terms of average versionand average version switch Regarding the SDP method itis evident that the performance of this method is the worstamong all (with low average version and highest averageswitch per segment) This can be expected because thismethod was not originally designed for real VBR videos

The buffer level curves of the three methods in offlinecontext are shown in Figure 19 We can see that all buffercurves imply streaming sessionswithout any freezes for usersAmong these the SDP method and the proposed methodwith 119881max = 9 provide the most unstable buffers while thebitrate estimation based method and the proposed methodwith 119881max = 7 provide the most stable buffers It can beobserved from Figures 17 and 18 and Table 6 that the SDPmethod and bitrate estimation based method result in veryfluctuating version curves and the proposed method with

Segment rpBitrateVersion

010002000300040005000600070008000

100 300 500 600200 4000Time (s)

0246810

Vers

ion

Bitr

ate (

kbps

)

Figure 17 Experimental results of SDP method

119881max = 7 obtains the lowest average quality Therefore fromTable 6 and Figure 19 we can see that the proposedmethod inoffline context with 119881max = 8 or 9 can provide the best per-formance

Advances in Multimedia 11

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

s)

100 200 300 400 500 6000Time (s)

0246810

Vers

ion

Figure 18 Experimental results of bitrate estimation basedmethod

02468

1012

Bue

r (s)

100 300 500 600200 4000Time (s)

SDPMax 9Max 8

Max 7BE

Figure 19 Buffer level curves of SDP method bitrate estimationbased method and the proposed method (Max9 Max8 and Max7)in offline context

7 Conclusion

In this paper we have proposed an adaptation method forHTTP streaming based on stochastic dynamic programmingThe system model was targeted at real bandwidth trace withstrong bitrate fluctuation of VBR videos Furthermore wehave developed a model to predict the system performancewith the aim of choosing the best setting based on theperformance requirements The experimental results haveshown that our method can effectively adapt VBR videos andperform accurate performance prediction which is useful inplanning adaptation policy

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

References

[1] Cisco Systems Inc ldquoNetworking index Forecast and method-ology 2012ndash2017rdquo Cisco Visual May 2013

[2] A C Begen T Akgul and M Baugher ldquoWatching video overthe web part 1 streaming protocolsrdquo IEEE Internet Computingvol 15 no 2 pp 54ndash63 2011

[3] T C Thang Q-D Ho J W Kang and A T Pham ldquoAdaptivestreaming of audiovisual content using MPEG DASHrdquo IEEETransactions on Consumer Electronics vol 58 no 1 pp 78ndash852012

[4] T Stockhammer ldquoDynamic adaptive streaming over HTTPstandards and design principlesrdquo in Proceeding of the 2ndAnnual ACM Multimedia Systems Conference (MMSys rsquo11) pp133ndash144 San Jose CA USA February 2011

[5] K Miller E Quacchio G Gennari and A Wolisz ldquoAdaptationalgorithm for adaptive streaming over HTTPrdquo in Proceedings ofthe 19th International Packet Video Workshop (PV rsquo12) pp 173ndash178 IEEE Munich Germany May 2012

[6] Y Zhou Y Duan J Sun and Z Guo ldquoTowards simple andsmooth rate adaption for VBR video in DASHrdquo in Proceedingsof the IEEE Visual Communications and Image Processing Con-ference (VCIP rsquo14) pp 9ndash12 IEEE Valletta Malta December2014

[7] S Akhshabi S Narayanaswamy A C Begen and C DovrolisldquoAn experimental evaluation of rate-adaptive video players overHTTPrdquo Signal Processing Image Communication vol 27 no 4pp 271ndash287 2012

[8] C Muller S Lederer and C Timmerer ldquoAn evaluation ofdynamic adaptive streaming over HTTP in vehicular environ-mentsrdquo in Proceedings of the 4th Workshop on Mobile Video(MoVid rsquo12) pp 37ndash42 New York NY USA February 2012

[9] C Liu I Bouazizi and M Gabbouj ldquoRate adaptation foradaptive HTTP streamingrdquo in Proceedings of the 2nd AnnualACMMultimedia Systems Conference (MMSys rsquo11) pp 169ndash174San Jose Calif USA February 2011

[10] S Garcıa J Cabrera and N Garcıa ldquoQuality-control algorithmfor adaptive streaming services over wireless channelsrdquo IEEEJournal on Selected Topics in Signal Processing vol 9 no 1 pp50ndash59 2015

[11] A H Duong T Nguyen T Vu T T Do N P Ngoc and T CThang ldquoSDP-based adaptation for quality control in adaptivestreamingrdquo in Proceedings of the International Conference onComputing Management and Telecommunications (ComMan-Tel rsquo15) pp 194ndash199 IEEE DaNang Vietnam December 2015

[12] S Lederer C Mueller C Timmerer C Concolato J Le Feuvreand K Fliegel ldquoDistributedDASHdatasetrdquo in Proceedings of the4th ACMMultimedia Systems Conference (MMSys rsquo13) pp 131ndash135 Oslo Norway March 2013

[13] ldquoH264AVC video trace libraryrdquo httptraceeasasueduh264[14] T CThang H T Le H X Nguyen A T Pham JW Kang and

Y M Ro ldquoAdaptive video streaming over HTTP with dynamicresource estimationrdquo IEEE Trans On consumer electronics vol58 no 1 pp 78ndash85 2012

[15] T C Thang H T Le A T Pham and Y M Ro ldquoAn evaluationof bitrate adaptation methods for HTTP live streamingrdquo IEEEJournal on Selected Areas in Communications vol 32 no 4 pp693ndash705 2014

[16] R Dubin O Hadar and A Dvir ldquoThe effect of client buffer andMBR consideration onDASHAdaptation Logicrdquo in Proceedingsof the 2013 IEEE Wireless Communications and NetworkingConference (WCNC rsquo13) pp 2178ndash2183 IEEE Shanghai ChinaApril 2013

[17] G Liang and B Liang ldquoEffect of delay and buffering on jitter-free streaming over random VBR channelsrdquo IEEE Transactionson Multimedia vol 10 no 6 pp 1128ndash1141 2008

[18] Y Kang H Chen and L Xie ldquoAn artificial-neural-network-based QoE estimation model for Video streaming over wirelessnetworksrdquo in Proceedings of the 2013 IEEECIC InternationalConference on Communications in China (ICCC rsquo13) pp 264ndash269 IEEE Xirsquoan China August 2013

[19] S Xiang L Cai and J Pan ldquoAdaptive scalable video streaminginwireless networksrdquo in Proceedings of the 3rdACMMultimedia

6 Advances in Multimedia

Based on [27] we have

119862PIA = 119874 (119868119881119873max) (15)

Therefore the complexity of our model is

119862 = 119874 (2119868119881max1198732 + 119868119881119873max) (16)

5 Performance Prediction

After Section 4 we achieve 119868 policy sets corresponding to 119868intervals of a video bitrate In this section we use theMarkovchain model to predict the streaming performance for asession Similar to Section 44 this section only presents theperformance prediction for a general bitrate interval with asystem state set s and a corresponding policy set 120583 Predictingperformance for all bitrate intervals will be done similarlyAfter carrying out the calculation for all 119868 bitrate intervalswe take the average values as the final results

The key is to determine the average state probability p119886 =[1199011198861 1199011198862 119901119886119873] with 119901119886119899 (1 le 119899 le 119873) being the averageprobability that the system is at the 119899th state throughout thestreaming session The probability p119886 is the average valueof state probabilities after downloading 119897 segments p119897 =[1199011198971 1199011198972 119901119897119873] (1 le 119897 le 119871) Here 119901119897119899 (1 le 119899 le 119873)is the probability that the system is at the 119899th state afterdownloading 119897 segments From the Markov chain theorythe state probability after downloading 119897 + 1 segments canbe computed as the product of the state probability afterdownloading 119897 segments and a transition matrix p119897+1 =p119897PTS Assuming that the initial probability 1199010 is known Thetransition matrix PTS is a119873times119873matrix which represents thetransition probability from state 119904119896 to state 119904119896+1 PTS is definedas follows

PTS = 119875 (119904119896 119904119896+1) = Pr 119904119896+1 | 119904119896 119886 isin 120583 (17)

The average state probability p119886 is calculated as follows

p119886 = sum119871119897=1 p119897PTS119871 (18)

Currently most of the adaptation algorithms developedfor HTTP streaming are qualitative in the sense that the per-formance metrics could only be obtained after the streamingsession In this study the predicted streaming performancecould be calculated based on the average state probabilityand the information inside every state Specifically wemainlyfocus on the following aspects bitrate prediction qualityswitch prediction and buffer prediction

51 Quality Prediction The video quality that the users per-ceive is presented through the selected versionThehigher theversion is selected the better the video quality is perceivedby the users Furthermore setting the maximum version forthe streaming session also affects the perceptual quality of theusers Obviously a very low value of the maximum versionmay result in a poor perceptual quality while a very highvalue may increase the chance of buffer underflow In thissection we predict the quality performance of the streaming

session based on the average version 119860V that is calculatedusing (19) and the cumulative distribution function (CDF)of the versions 119891(V) (V isin [1 119881max]) that is shown in (20)Based on the predicted probability of the versions throughouta streaming session the maximum version could be decidedfor the session

119860V = sum119873119899=1 V119899119901119886119899119873 (19)

119891 (V) = sum119873119899=1 V119899119901119886119899119873 | V = V119899 (20)

52 Quality Switch Prediction Quality switch is an importantfactor affecting the perception of the users The users oftenexpect a smooth playback with the minimum number ofquality switches and small switch amplitude from one seg-ment to the next We can predict the average version switchper segment 119860 sw as follows

119860 sw = sum119873119899=1 1003816100381610038161003816119886119899 minus V1198991003816100381610038161003816 119901119886119899

119873 119886119899 isin 120583 (21)

53 Buffer Prediction Video stalling is one of the importantobjective parameters that affect the subjective perception ofthe users Stalling occurswhen the play-out buffer underrunsTo prevent this event the buffer must be maintained within asafe range In this session we evaluate the buffer performancethrough the average buffer level 119860119887 the CDF of the bufferlevel 119891(119887) (119887 isin [1 119861max]) and the buffer underflowprobability Prund (ie when the system stays at buffer level1) which are described as follows

119860119887 = sum119873119899=1 119887119899119901119886119899119873

119891 (119887) = sum119873119899=1 119887119899119901119886119899119873 | 119887119899 lt 119887

Prund = sum119873119899=1 119901119886119899119873 | 119887119899 = 1

(22)

119860119887 represents the safety of the buffer If 119860119887 is small thebuffer level is often in low levels which may cause playbackinterruptionwhen the current bandwidth drops dramatically119891(119887) reflects the variance of the buffer level under thefluctuations of the network Prund shows the probability thatthe playback would be interrupted in the streaming session

6 Experimental Results

In order to evaluate the proposed system model and per-formance prediction accuracy in this section we perform anumber of experiments in both offline and online contextsand compare the performance predicted results with themeasurement ones We also compare our proposed methodwith two existing ones namely the SDP method presented[10] which could obtain the best performance among the SDPmethods and the bitrate estimation based method presented[14] which is the best among the qualitative methods

Advances in Multimedia 7

61 Experiment Setup For the simulation our test-bed con-sists of a client running Java 80 which implements theadaptation and a server running Apache2 which holds themedia segments The client runs on a Window 7 computerwith an Intel i5-17 GHz CPU and 4GB memory and theserver runs on Ubuntu 1204LTS (with default TCP CUBIC)with 1 G RAM The channel bandwidth is simulated usingDummyNet [28] We use the Tokyo Olympic video from [6]For the video 119881max = 9 and for the bandwidth 119882 = 10Since we measure the buffer size and compute the buffer costin the segment duration unit we implement our adaptationmethod with one setting of the segment duration In ourexperiments we select a segment duration of 2 secondswhich is similar to those of [7 8 10] The impacts of segmentdurations on adaptation performance have been consideredin some recent studies [29 30] Further evaluation of differentsegment durations with a fixed buffer size (in seconds) will bereserved for our future work Maximum buffer level is set to5 segments (ie 10 seconds) and optimal buffer level is set to4 segments (ie 8 seconds)

In our method a streaming provider can adjust thebalance between the requirements for high quality levelpreventing video stalling and reducing quality switchesby changing the trade-off parameters 120572 120574 and 120573 Sinceselecting an optimal combination of trade-off parametersof the cost function involves solving a hard optimizationproblem it will be investigated in our future work In thispaper we select qualitatively the trade-off parameters of costfunction as follows Initially we fix 120573 and select the othertwo parameters Since we want to prioritize requirement ofsmooth quality switch 120574 is selected so that the contributionof the quality switch cost Δ119902 is higher than that of buffercost Δ119887 With parameter 120572 because the bitrate cost Δ119903 can beup to thousands parameter 120572 should be small to reduce thecontribution of the bitrate cost to the overall cost 119862 Basedon our experience good empirical values on parameters 120572 120573and 120574 are 0003 4 and 20 respectively

62 Experimental Results In the first part of the experimentwe evaluate the accuracy of performance prediction in offlinecontext using a given bandwidth trace obtained fromamobilenetwork [12] In this context the number of video segments119871 is 300

Figures 5 and 6 show the predicted performance usingthe formulas presented in Section 5 and themeasured perfor-mance obtained from the experiments when the maximumallowed version is set to 7 8 and 9 These figures point outthat the prediction results are close to the measurement onesWe can see from Figure 5 that when the maximum versionincreases the average version as well as the average versionswitch also increases Figure 6 shows that in both predictionand measurement cases when the maximum version is 7 theunderflowprobability is almost zero and increases very slowlywhen the maximum allowed version increases This analysisimplies that setting themaximumversion to 8 is reasonable interms of balancing between average video quality and qualityswitch meanwhile setting the maximum version to 7 ensuresa very stable streaming experience

0

02

04

06

08

1

Aver

age s

witc

hse

gmen

t

8 97Maximum version

1

3

5

7

9

Vers

ion

Measured avr_switchMeasured avr_version

Predicted avr_switchPredicted avr_version

Figure 5 Predicted performance and measured performance interms of average version and average version switch per segment inoffline context using a given bandwidth trace

0

005

01

Und

er

ow p

roba

bilit

y

0

5

10

Aver

age b

uer

(s)

8 97Maximum version

Measured prob_underowMeasured avr_buer

Predicted prob_underowPredicted avr_buer

Figure 6 Predicted performance and measured performance interms of average buffer and underflow probability in offline contextusing a given bandwidth trace

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

012345678910

Vers

ion

Figure 7 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 9

Figures 7 8 and 9 show the bitrate and version switchbehavior of the proposed method in the three cases ofmaximum version It can be seen very clearly from thesefigures that when the maximum version is reduced thenumber of version switches (or quality changes) decreases

Table 1 shows more detailed statistics of the experimentalresults in three cases of maximum version It can be drawn

8 Advances in Multimedia

Table 1 Compare predicted performance and measured performance in offline context using a given bandwidth trace

Statistics 119881max = 9 119881max = 8 119881max = 7Predicted Measured Predicted Measured Predicted Measured

119860119887(119904) 948 852 959 877 98 968119860119902 764 788 745 773 648 687119860 sw 030 018 014 009 003 002Prundflow 001 000 0006 000 000 000

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

0123456789

Vers

ion

Figure 8 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 8

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

012345678

Vers

ion

Figure 9 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 7

from the table that there is no significant difference betweenthe predicted performance and the measured performance

In the second part of the experiment we use two historybandwidth traces recorded from two previous streamingsessions of the client The CDFs of both bandwidths areshown in Figure 10

Bandwidth 1198871199081 is used for calculating the statisticalmodels and 1198871199082 is used in the simulation for measuringperformance parameters Figures 11 12 and 13 show thebitrate and version switch behavior of the proposed methodin the three cases of maximum version The detailed resultsare listed in Table 2 Based on these figures and tablewe affirm once again that the mathematical performanceprediction model agrees well with the measurement

bw1bw2

0010203040506070809

1

CDF

1000 2000 3000 4000 50000Bandwidth (kbps)

Figure 10 The cumulative distribution function of bw1 and bw2

Segment rpBitrateVersion

100 200 300 400 500 6000Time (s)

01000200030004000500060007000

Bitr

ate (

kbps

)

012345678910

Vers

ion

Figure 11 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 9

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

0123456789

Vers

ion

Figure 12 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 8

Advances in Multimedia 9

Table 2 Compare predicted performance and measured performance in offline context using a history bandwidth trace for statistics

Statistics 119881max = 9 119881max = 8 119881max = 7Predicted Measured Predicted Measured Predicted Measured

119860119887(119904) 952 844 961 885 969 965119860119902 748 792 727 771 680 698119860 sw 031 019 018 011 008 005Prundflow 000 000 000 000 000 000

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

12345678

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 13 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 7

In the third part of the experiment we consider the onlinecontext in which the prediction of the future bandwidth isbased on the statistical parameters of all previous segmentsSpecifically we divide an entire session into chunks each ofwhich has119871 video segmentsWe treat each chunk individuallyas a mini streaming session Assuming that initially we haveenough statistical data to predict the performance of thefirst chunk The prediction of the subsequent chunks willbe done based on the previous chunks At the beginning ofeach chunk the bandwidth statistics are updated leading to arecomputation of the policy set and the performance In theexperiment we set119871 to 100 video segments It can be observedfrom our experiments that it took only several seconds to(re)compute the whole model Therefore the computationaloverhead is about 3 which is appropriate for the onlinecontext Figures 14 15 and 16 show the adaptation bitrateand version switch behavior of the proposed method in thethree cases of maximum version The predicted performanceand the measured performance in the online context in thethree cases are presented in detail in Tables 3 4 and 5It is very clear that in the online context the predictedperformance is also very close to the measured performance

Next we compare our proposed method with the SDPmethod [10] and the bitrate estimation based method [14]using the same simulation settings as in the first part of ourexperimentThe experimental results obtained by simulatingthe SDP method [10] and bitrate estimation based method[14] are shown in Figures 17 and 18 respectively We can seethat both methods provide very fluctuating version switchesThe detailed statistics of these adaptation methods with themaximum version set to 9 and our proposedmethodwith the

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

kbps

)

012345678910

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 14 Experimental results of the proposed method in onlinecontext with Vmax = 9

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

0123456789

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 15 Experimental results of the proposed method in onlinecontext with Vmax = 8

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

12345678

Vers

ion

Figure 16 Experimental results of the proposed method in onlinecontext with Vmax = 7

10 Advances in Multimedia

Table 3 Compare predicted performance and measured performance in online context with 119881max = 9Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 931 896 892 845 936 886119860119902 769 804 789 768 783 772119860 sw 005 009 004 007 002 005Prundflow 000 000 000 000 000 000

Table 4 Compare predicted performance and measured performance in online context with 119881max = 8Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 953 901 942 863 942 896119860119902 756 786 765 763 765 767119860 sw 000 003 001 003 002 004Prundflow 000 000 000 000 000 000

Table 5 Compare predicted performance and measured performance in online context with 119881max = 7Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 977 993 977 952 983 959119860119902 692 700 692 695 692 700119860 sw 000 000 000 002 000 000Prundflow 000 000 000 000 000 000

Table 6 Statistics of different adaptation methods in offline context

Statistics SDP method119881max = 9

Bitrate estimation method119881max = 9

Proposed method119881max = 9

Proposed method119881max = 8

Proposed method119881max = 7

119860119887(119904) 862 885 852 877 968119860119902 740 785 788 773 687119860 sw 076 043 018 009 002Prundflow 000 000 000 000 000

maximumversion set to 7 8 and 9 are shown in Table 6 It canbe seen that the performance of the bitrate estimationmethodis less effective than our method in terms of average versionand average version switch Regarding the SDP method itis evident that the performance of this method is the worstamong all (with low average version and highest averageswitch per segment) This can be expected because thismethod was not originally designed for real VBR videos

The buffer level curves of the three methods in offlinecontext are shown in Figure 19 We can see that all buffercurves imply streaming sessionswithout any freezes for usersAmong these the SDP method and the proposed methodwith 119881max = 9 provide the most unstable buffers while thebitrate estimation based method and the proposed methodwith 119881max = 7 provide the most stable buffers It can beobserved from Figures 17 and 18 and Table 6 that the SDPmethod and bitrate estimation based method result in veryfluctuating version curves and the proposed method with

Segment rpBitrateVersion

010002000300040005000600070008000

100 300 500 600200 4000Time (s)

0246810

Vers

ion

Bitr

ate (

kbps

)

Figure 17 Experimental results of SDP method

119881max = 7 obtains the lowest average quality Therefore fromTable 6 and Figure 19 we can see that the proposedmethod inoffline context with 119881max = 8 or 9 can provide the best per-formance

Advances in Multimedia 11

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

s)

100 200 300 400 500 6000Time (s)

0246810

Vers

ion

Figure 18 Experimental results of bitrate estimation basedmethod

02468

1012

Bue

r (s)

100 300 500 600200 4000Time (s)

SDPMax 9Max 8

Max 7BE

Figure 19 Buffer level curves of SDP method bitrate estimationbased method and the proposed method (Max9 Max8 and Max7)in offline context

7 Conclusion

In this paper we have proposed an adaptation method forHTTP streaming based on stochastic dynamic programmingThe system model was targeted at real bandwidth trace withstrong bitrate fluctuation of VBR videos Furthermore wehave developed a model to predict the system performancewith the aim of choosing the best setting based on theperformance requirements The experimental results haveshown that our method can effectively adapt VBR videos andperform accurate performance prediction which is useful inplanning adaptation policy

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

References

[1] Cisco Systems Inc ldquoNetworking index Forecast and method-ology 2012ndash2017rdquo Cisco Visual May 2013

[2] A C Begen T Akgul and M Baugher ldquoWatching video overthe web part 1 streaming protocolsrdquo IEEE Internet Computingvol 15 no 2 pp 54ndash63 2011

[3] T C Thang Q-D Ho J W Kang and A T Pham ldquoAdaptivestreaming of audiovisual content using MPEG DASHrdquo IEEETransactions on Consumer Electronics vol 58 no 1 pp 78ndash852012

[4] T Stockhammer ldquoDynamic adaptive streaming over HTTPstandards and design principlesrdquo in Proceeding of the 2ndAnnual ACM Multimedia Systems Conference (MMSys rsquo11) pp133ndash144 San Jose CA USA February 2011

[5] K Miller E Quacchio G Gennari and A Wolisz ldquoAdaptationalgorithm for adaptive streaming over HTTPrdquo in Proceedings ofthe 19th International Packet Video Workshop (PV rsquo12) pp 173ndash178 IEEE Munich Germany May 2012

[6] Y Zhou Y Duan J Sun and Z Guo ldquoTowards simple andsmooth rate adaption for VBR video in DASHrdquo in Proceedingsof the IEEE Visual Communications and Image Processing Con-ference (VCIP rsquo14) pp 9ndash12 IEEE Valletta Malta December2014

[7] S Akhshabi S Narayanaswamy A C Begen and C DovrolisldquoAn experimental evaluation of rate-adaptive video players overHTTPrdquo Signal Processing Image Communication vol 27 no 4pp 271ndash287 2012

[8] C Muller S Lederer and C Timmerer ldquoAn evaluation ofdynamic adaptive streaming over HTTP in vehicular environ-mentsrdquo in Proceedings of the 4th Workshop on Mobile Video(MoVid rsquo12) pp 37ndash42 New York NY USA February 2012

[9] C Liu I Bouazizi and M Gabbouj ldquoRate adaptation foradaptive HTTP streamingrdquo in Proceedings of the 2nd AnnualACMMultimedia Systems Conference (MMSys rsquo11) pp 169ndash174San Jose Calif USA February 2011

[10] S Garcıa J Cabrera and N Garcıa ldquoQuality-control algorithmfor adaptive streaming services over wireless channelsrdquo IEEEJournal on Selected Topics in Signal Processing vol 9 no 1 pp50ndash59 2015

[11] A H Duong T Nguyen T Vu T T Do N P Ngoc and T CThang ldquoSDP-based adaptation for quality control in adaptivestreamingrdquo in Proceedings of the International Conference onComputing Management and Telecommunications (ComMan-Tel rsquo15) pp 194ndash199 IEEE DaNang Vietnam December 2015

[12] S Lederer C Mueller C Timmerer C Concolato J Le Feuvreand K Fliegel ldquoDistributedDASHdatasetrdquo in Proceedings of the4th ACMMultimedia Systems Conference (MMSys rsquo13) pp 131ndash135 Oslo Norway March 2013

[13] ldquoH264AVC video trace libraryrdquo httptraceeasasueduh264[14] T CThang H T Le H X Nguyen A T Pham JW Kang and

Y M Ro ldquoAdaptive video streaming over HTTP with dynamicresource estimationrdquo IEEE Trans On consumer electronics vol58 no 1 pp 78ndash85 2012

[15] T C Thang H T Le A T Pham and Y M Ro ldquoAn evaluationof bitrate adaptation methods for HTTP live streamingrdquo IEEEJournal on Selected Areas in Communications vol 32 no 4 pp693ndash705 2014

[16] R Dubin O Hadar and A Dvir ldquoThe effect of client buffer andMBR consideration onDASHAdaptation Logicrdquo in Proceedingsof the 2013 IEEE Wireless Communications and NetworkingConference (WCNC rsquo13) pp 2178ndash2183 IEEE Shanghai ChinaApril 2013

[17] G Liang and B Liang ldquoEffect of delay and buffering on jitter-free streaming over random VBR channelsrdquo IEEE Transactionson Multimedia vol 10 no 6 pp 1128ndash1141 2008

[18] Y Kang H Chen and L Xie ldquoAn artificial-neural-network-based QoE estimation model for Video streaming over wirelessnetworksrdquo in Proceedings of the 2013 IEEECIC InternationalConference on Communications in China (ICCC rsquo13) pp 264ndash269 IEEE Xirsquoan China August 2013

[19] S Xiang L Cai and J Pan ldquoAdaptive scalable video streaminginwireless networksrdquo in Proceedings of the 3rdACMMultimedia

Advances in Multimedia 7

61 Experiment Setup For the simulation our test-bed con-sists of a client running Java 80 which implements theadaptation and a server running Apache2 which holds themedia segments The client runs on a Window 7 computerwith an Intel i5-17 GHz CPU and 4GB memory and theserver runs on Ubuntu 1204LTS (with default TCP CUBIC)with 1 G RAM The channel bandwidth is simulated usingDummyNet [28] We use the Tokyo Olympic video from [6]For the video 119881max = 9 and for the bandwidth 119882 = 10Since we measure the buffer size and compute the buffer costin the segment duration unit we implement our adaptationmethod with one setting of the segment duration In ourexperiments we select a segment duration of 2 secondswhich is similar to those of [7 8 10] The impacts of segmentdurations on adaptation performance have been consideredin some recent studies [29 30] Further evaluation of differentsegment durations with a fixed buffer size (in seconds) will bereserved for our future work Maximum buffer level is set to5 segments (ie 10 seconds) and optimal buffer level is set to4 segments (ie 8 seconds)

In our method a streaming provider can adjust thebalance between the requirements for high quality levelpreventing video stalling and reducing quality switchesby changing the trade-off parameters 120572 120574 and 120573 Sinceselecting an optimal combination of trade-off parametersof the cost function involves solving a hard optimizationproblem it will be investigated in our future work In thispaper we select qualitatively the trade-off parameters of costfunction as follows Initially we fix 120573 and select the othertwo parameters Since we want to prioritize requirement ofsmooth quality switch 120574 is selected so that the contributionof the quality switch cost Δ119902 is higher than that of buffercost Δ119887 With parameter 120572 because the bitrate cost Δ119903 can beup to thousands parameter 120572 should be small to reduce thecontribution of the bitrate cost to the overall cost 119862 Basedon our experience good empirical values on parameters 120572 120573and 120574 are 0003 4 and 20 respectively

62 Experimental Results In the first part of the experimentwe evaluate the accuracy of performance prediction in offlinecontext using a given bandwidth trace obtained fromamobilenetwork [12] In this context the number of video segments119871 is 300

Figures 5 and 6 show the predicted performance usingthe formulas presented in Section 5 and themeasured perfor-mance obtained from the experiments when the maximumallowed version is set to 7 8 and 9 These figures point outthat the prediction results are close to the measurement onesWe can see from Figure 5 that when the maximum versionincreases the average version as well as the average versionswitch also increases Figure 6 shows that in both predictionand measurement cases when the maximum version is 7 theunderflowprobability is almost zero and increases very slowlywhen the maximum allowed version increases This analysisimplies that setting themaximumversion to 8 is reasonable interms of balancing between average video quality and qualityswitch meanwhile setting the maximum version to 7 ensuresa very stable streaming experience

0

02

04

06

08

1

Aver

age s

witc

hse

gmen

t

8 97Maximum version

1

3

5

7

9

Vers

ion

Measured avr_switchMeasured avr_version

Predicted avr_switchPredicted avr_version

Figure 5 Predicted performance and measured performance interms of average version and average version switch per segment inoffline context using a given bandwidth trace

0

005

01

Und

er

ow p

roba

bilit

y

0

5

10

Aver

age b

uer

(s)

8 97Maximum version

Measured prob_underowMeasured avr_buer

Predicted prob_underowPredicted avr_buer

Figure 6 Predicted performance and measured performance interms of average buffer and underflow probability in offline contextusing a given bandwidth trace

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

012345678910

Vers

ion

Figure 7 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 9

Figures 7 8 and 9 show the bitrate and version switchbehavior of the proposed method in the three cases ofmaximum version It can be seen very clearly from thesefigures that when the maximum version is reduced thenumber of version switches (or quality changes) decreases

Table 1 shows more detailed statistics of the experimentalresults in three cases of maximum version It can be drawn

8 Advances in Multimedia

Table 1 Compare predicted performance and measured performance in offline context using a given bandwidth trace

Statistics 119881max = 9 119881max = 8 119881max = 7Predicted Measured Predicted Measured Predicted Measured

119860119887(119904) 948 852 959 877 98 968119860119902 764 788 745 773 648 687119860 sw 030 018 014 009 003 002Prundflow 001 000 0006 000 000 000

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

0123456789

Vers

ion

Figure 8 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 8

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

012345678

Vers

ion

Figure 9 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 7

from the table that there is no significant difference betweenthe predicted performance and the measured performance

In the second part of the experiment we use two historybandwidth traces recorded from two previous streamingsessions of the client The CDFs of both bandwidths areshown in Figure 10

Bandwidth 1198871199081 is used for calculating the statisticalmodels and 1198871199082 is used in the simulation for measuringperformance parameters Figures 11 12 and 13 show thebitrate and version switch behavior of the proposed methodin the three cases of maximum version The detailed resultsare listed in Table 2 Based on these figures and tablewe affirm once again that the mathematical performanceprediction model agrees well with the measurement

bw1bw2

0010203040506070809

1

CDF

1000 2000 3000 4000 50000Bandwidth (kbps)

Figure 10 The cumulative distribution function of bw1 and bw2

Segment rpBitrateVersion

100 200 300 400 500 6000Time (s)

01000200030004000500060007000

Bitr

ate (

kbps

)

012345678910

Vers

ion

Figure 11 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 9

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

0123456789

Vers

ion

Figure 12 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 8

Advances in Multimedia 9

Table 2 Compare predicted performance and measured performance in offline context using a history bandwidth trace for statistics

Statistics 119881max = 9 119881max = 8 119881max = 7Predicted Measured Predicted Measured Predicted Measured

119860119887(119904) 952 844 961 885 969 965119860119902 748 792 727 771 680 698119860 sw 031 019 018 011 008 005Prundflow 000 000 000 000 000 000

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

12345678

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 13 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 7

In the third part of the experiment we consider the onlinecontext in which the prediction of the future bandwidth isbased on the statistical parameters of all previous segmentsSpecifically we divide an entire session into chunks each ofwhich has119871 video segmentsWe treat each chunk individuallyas a mini streaming session Assuming that initially we haveenough statistical data to predict the performance of thefirst chunk The prediction of the subsequent chunks willbe done based on the previous chunks At the beginning ofeach chunk the bandwidth statistics are updated leading to arecomputation of the policy set and the performance In theexperiment we set119871 to 100 video segments It can be observedfrom our experiments that it took only several seconds to(re)compute the whole model Therefore the computationaloverhead is about 3 which is appropriate for the onlinecontext Figures 14 15 and 16 show the adaptation bitrateand version switch behavior of the proposed method in thethree cases of maximum version The predicted performanceand the measured performance in the online context in thethree cases are presented in detail in Tables 3 4 and 5It is very clear that in the online context the predictedperformance is also very close to the measured performance

Next we compare our proposed method with the SDPmethod [10] and the bitrate estimation based method [14]using the same simulation settings as in the first part of ourexperimentThe experimental results obtained by simulatingthe SDP method [10] and bitrate estimation based method[14] are shown in Figures 17 and 18 respectively We can seethat both methods provide very fluctuating version switchesThe detailed statistics of these adaptation methods with themaximum version set to 9 and our proposedmethodwith the

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

kbps

)

012345678910

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 14 Experimental results of the proposed method in onlinecontext with Vmax = 9

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

0123456789

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 15 Experimental results of the proposed method in onlinecontext with Vmax = 8

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

12345678

Vers

ion

Figure 16 Experimental results of the proposed method in onlinecontext with Vmax = 7

10 Advances in Multimedia

Table 3 Compare predicted performance and measured performance in online context with 119881max = 9Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 931 896 892 845 936 886119860119902 769 804 789 768 783 772119860 sw 005 009 004 007 002 005Prundflow 000 000 000 000 000 000

Table 4 Compare predicted performance and measured performance in online context with 119881max = 8Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 953 901 942 863 942 896119860119902 756 786 765 763 765 767119860 sw 000 003 001 003 002 004Prundflow 000 000 000 000 000 000

Table 5 Compare predicted performance and measured performance in online context with 119881max = 7Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 977 993 977 952 983 959119860119902 692 700 692 695 692 700119860 sw 000 000 000 002 000 000Prundflow 000 000 000 000 000 000

Table 6 Statistics of different adaptation methods in offline context

Statistics SDP method119881max = 9

Bitrate estimation method119881max = 9

Proposed method119881max = 9

Proposed method119881max = 8

Proposed method119881max = 7

119860119887(119904) 862 885 852 877 968119860119902 740 785 788 773 687119860 sw 076 043 018 009 002Prundflow 000 000 000 000 000

maximumversion set to 7 8 and 9 are shown in Table 6 It canbe seen that the performance of the bitrate estimationmethodis less effective than our method in terms of average versionand average version switch Regarding the SDP method itis evident that the performance of this method is the worstamong all (with low average version and highest averageswitch per segment) This can be expected because thismethod was not originally designed for real VBR videos

The buffer level curves of the three methods in offlinecontext are shown in Figure 19 We can see that all buffercurves imply streaming sessionswithout any freezes for usersAmong these the SDP method and the proposed methodwith 119881max = 9 provide the most unstable buffers while thebitrate estimation based method and the proposed methodwith 119881max = 7 provide the most stable buffers It can beobserved from Figures 17 and 18 and Table 6 that the SDPmethod and bitrate estimation based method result in veryfluctuating version curves and the proposed method with

Segment rpBitrateVersion

010002000300040005000600070008000

100 300 500 600200 4000Time (s)

0246810

Vers

ion

Bitr

ate (

kbps

)

Figure 17 Experimental results of SDP method

119881max = 7 obtains the lowest average quality Therefore fromTable 6 and Figure 19 we can see that the proposedmethod inoffline context with 119881max = 8 or 9 can provide the best per-formance

Advances in Multimedia 11

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

s)

100 200 300 400 500 6000Time (s)

0246810

Vers

ion

Figure 18 Experimental results of bitrate estimation basedmethod

02468

1012

Bue

r (s)

100 300 500 600200 4000Time (s)

SDPMax 9Max 8

Max 7BE

Figure 19 Buffer level curves of SDP method bitrate estimationbased method and the proposed method (Max9 Max8 and Max7)in offline context

7 Conclusion

In this paper we have proposed an adaptation method forHTTP streaming based on stochastic dynamic programmingThe system model was targeted at real bandwidth trace withstrong bitrate fluctuation of VBR videos Furthermore wehave developed a model to predict the system performancewith the aim of choosing the best setting based on theperformance requirements The experimental results haveshown that our method can effectively adapt VBR videos andperform accurate performance prediction which is useful inplanning adaptation policy

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

References

[1] Cisco Systems Inc ldquoNetworking index Forecast and method-ology 2012ndash2017rdquo Cisco Visual May 2013

[2] A C Begen T Akgul and M Baugher ldquoWatching video overthe web part 1 streaming protocolsrdquo IEEE Internet Computingvol 15 no 2 pp 54ndash63 2011

[3] T C Thang Q-D Ho J W Kang and A T Pham ldquoAdaptivestreaming of audiovisual content using MPEG DASHrdquo IEEETransactions on Consumer Electronics vol 58 no 1 pp 78ndash852012

[4] T Stockhammer ldquoDynamic adaptive streaming over HTTPstandards and design principlesrdquo in Proceeding of the 2ndAnnual ACM Multimedia Systems Conference (MMSys rsquo11) pp133ndash144 San Jose CA USA February 2011

[5] K Miller E Quacchio G Gennari and A Wolisz ldquoAdaptationalgorithm for adaptive streaming over HTTPrdquo in Proceedings ofthe 19th International Packet Video Workshop (PV rsquo12) pp 173ndash178 IEEE Munich Germany May 2012

[6] Y Zhou Y Duan J Sun and Z Guo ldquoTowards simple andsmooth rate adaption for VBR video in DASHrdquo in Proceedingsof the IEEE Visual Communications and Image Processing Con-ference (VCIP rsquo14) pp 9ndash12 IEEE Valletta Malta December2014

[7] S Akhshabi S Narayanaswamy A C Begen and C DovrolisldquoAn experimental evaluation of rate-adaptive video players overHTTPrdquo Signal Processing Image Communication vol 27 no 4pp 271ndash287 2012

[8] C Muller S Lederer and C Timmerer ldquoAn evaluation ofdynamic adaptive streaming over HTTP in vehicular environ-mentsrdquo in Proceedings of the 4th Workshop on Mobile Video(MoVid rsquo12) pp 37ndash42 New York NY USA February 2012

[9] C Liu I Bouazizi and M Gabbouj ldquoRate adaptation foradaptive HTTP streamingrdquo in Proceedings of the 2nd AnnualACMMultimedia Systems Conference (MMSys rsquo11) pp 169ndash174San Jose Calif USA February 2011

[10] S Garcıa J Cabrera and N Garcıa ldquoQuality-control algorithmfor adaptive streaming services over wireless channelsrdquo IEEEJournal on Selected Topics in Signal Processing vol 9 no 1 pp50ndash59 2015

[11] A H Duong T Nguyen T Vu T T Do N P Ngoc and T CThang ldquoSDP-based adaptation for quality control in adaptivestreamingrdquo in Proceedings of the International Conference onComputing Management and Telecommunications (ComMan-Tel rsquo15) pp 194ndash199 IEEE DaNang Vietnam December 2015

[12] S Lederer C Mueller C Timmerer C Concolato J Le Feuvreand K Fliegel ldquoDistributedDASHdatasetrdquo in Proceedings of the4th ACMMultimedia Systems Conference (MMSys rsquo13) pp 131ndash135 Oslo Norway March 2013

[13] ldquoH264AVC video trace libraryrdquo httptraceeasasueduh264[14] T CThang H T Le H X Nguyen A T Pham JW Kang and

Y M Ro ldquoAdaptive video streaming over HTTP with dynamicresource estimationrdquo IEEE Trans On consumer electronics vol58 no 1 pp 78ndash85 2012

[15] T C Thang H T Le A T Pham and Y M Ro ldquoAn evaluationof bitrate adaptation methods for HTTP live streamingrdquo IEEEJournal on Selected Areas in Communications vol 32 no 4 pp693ndash705 2014

[16] R Dubin O Hadar and A Dvir ldquoThe effect of client buffer andMBR consideration onDASHAdaptation Logicrdquo in Proceedingsof the 2013 IEEE Wireless Communications and NetworkingConference (WCNC rsquo13) pp 2178ndash2183 IEEE Shanghai ChinaApril 2013

[17] G Liang and B Liang ldquoEffect of delay and buffering on jitter-free streaming over random VBR channelsrdquo IEEE Transactionson Multimedia vol 10 no 6 pp 1128ndash1141 2008

[18] Y Kang H Chen and L Xie ldquoAn artificial-neural-network-based QoE estimation model for Video streaming over wirelessnetworksrdquo in Proceedings of the 2013 IEEECIC InternationalConference on Communications in China (ICCC rsquo13) pp 264ndash269 IEEE Xirsquoan China August 2013

[19] S Xiang L Cai and J Pan ldquoAdaptive scalable video streaminginwireless networksrdquo in Proceedings of the 3rdACMMultimedia

8 Advances in Multimedia

Table 1 Compare predicted performance and measured performance in offline context using a given bandwidth trace

Statistics 119881max = 9 119881max = 8 119881max = 7Predicted Measured Predicted Measured Predicted Measured

119860119887(119904) 948 852 959 877 98 968119860119902 764 788 745 773 648 687119860 sw 030 018 014 009 003 002Prundflow 001 000 0006 000 000 000

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

0123456789

Vers

ion

Figure 8 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 8

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

012345678

Vers

ion

Figure 9 Experimental results of proposed method in offlinecontext using a given bandwidth trace with Vmax = 7

from the table that there is no significant difference betweenthe predicted performance and the measured performance

In the second part of the experiment we use two historybandwidth traces recorded from two previous streamingsessions of the client The CDFs of both bandwidths areshown in Figure 10

Bandwidth 1198871199081 is used for calculating the statisticalmodels and 1198871199082 is used in the simulation for measuringperformance parameters Figures 11 12 and 13 show thebitrate and version switch behavior of the proposed methodin the three cases of maximum version The detailed resultsare listed in Table 2 Based on these figures and tablewe affirm once again that the mathematical performanceprediction model agrees well with the measurement

bw1bw2

0010203040506070809

1

CDF

1000 2000 3000 4000 50000Bandwidth (kbps)

Figure 10 The cumulative distribution function of bw1 and bw2

Segment rpBitrateVersion

100 200 300 400 500 6000Time (s)

01000200030004000500060007000

Bitr

ate (

kbps

)

012345678910

Vers

ion

Figure 11 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 9

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

0123456789

Vers

ion

Figure 12 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 8

Advances in Multimedia 9

Table 2 Compare predicted performance and measured performance in offline context using a history bandwidth trace for statistics

Statistics 119881max = 9 119881max = 8 119881max = 7Predicted Measured Predicted Measured Predicted Measured

119860119887(119904) 952 844 961 885 969 965119860119902 748 792 727 771 680 698119860 sw 031 019 018 011 008 005Prundflow 000 000 000 000 000 000

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

12345678

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 13 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 7

In the third part of the experiment we consider the onlinecontext in which the prediction of the future bandwidth isbased on the statistical parameters of all previous segmentsSpecifically we divide an entire session into chunks each ofwhich has119871 video segmentsWe treat each chunk individuallyas a mini streaming session Assuming that initially we haveenough statistical data to predict the performance of thefirst chunk The prediction of the subsequent chunks willbe done based on the previous chunks At the beginning ofeach chunk the bandwidth statistics are updated leading to arecomputation of the policy set and the performance In theexperiment we set119871 to 100 video segments It can be observedfrom our experiments that it took only several seconds to(re)compute the whole model Therefore the computationaloverhead is about 3 which is appropriate for the onlinecontext Figures 14 15 and 16 show the adaptation bitrateand version switch behavior of the proposed method in thethree cases of maximum version The predicted performanceand the measured performance in the online context in thethree cases are presented in detail in Tables 3 4 and 5It is very clear that in the online context the predictedperformance is also very close to the measured performance

Next we compare our proposed method with the SDPmethod [10] and the bitrate estimation based method [14]using the same simulation settings as in the first part of ourexperimentThe experimental results obtained by simulatingthe SDP method [10] and bitrate estimation based method[14] are shown in Figures 17 and 18 respectively We can seethat both methods provide very fluctuating version switchesThe detailed statistics of these adaptation methods with themaximum version set to 9 and our proposedmethodwith the

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

kbps

)

012345678910

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 14 Experimental results of the proposed method in onlinecontext with Vmax = 9

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

0123456789

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 15 Experimental results of the proposed method in onlinecontext with Vmax = 8

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

12345678

Vers

ion

Figure 16 Experimental results of the proposed method in onlinecontext with Vmax = 7

10 Advances in Multimedia

Table 3 Compare predicted performance and measured performance in online context with 119881max = 9Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 931 896 892 845 936 886119860119902 769 804 789 768 783 772119860 sw 005 009 004 007 002 005Prundflow 000 000 000 000 000 000

Table 4 Compare predicted performance and measured performance in online context with 119881max = 8Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 953 901 942 863 942 896119860119902 756 786 765 763 765 767119860 sw 000 003 001 003 002 004Prundflow 000 000 000 000 000 000

Table 5 Compare predicted performance and measured performance in online context with 119881max = 7Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 977 993 977 952 983 959119860119902 692 700 692 695 692 700119860 sw 000 000 000 002 000 000Prundflow 000 000 000 000 000 000

Table 6 Statistics of different adaptation methods in offline context

Statistics SDP method119881max = 9

Bitrate estimation method119881max = 9

Proposed method119881max = 9

Proposed method119881max = 8

Proposed method119881max = 7

119860119887(119904) 862 885 852 877 968119860119902 740 785 788 773 687119860 sw 076 043 018 009 002Prundflow 000 000 000 000 000

maximumversion set to 7 8 and 9 are shown in Table 6 It canbe seen that the performance of the bitrate estimationmethodis less effective than our method in terms of average versionand average version switch Regarding the SDP method itis evident that the performance of this method is the worstamong all (with low average version and highest averageswitch per segment) This can be expected because thismethod was not originally designed for real VBR videos

The buffer level curves of the three methods in offlinecontext are shown in Figure 19 We can see that all buffercurves imply streaming sessionswithout any freezes for usersAmong these the SDP method and the proposed methodwith 119881max = 9 provide the most unstable buffers while thebitrate estimation based method and the proposed methodwith 119881max = 7 provide the most stable buffers It can beobserved from Figures 17 and 18 and Table 6 that the SDPmethod and bitrate estimation based method result in veryfluctuating version curves and the proposed method with

Segment rpBitrateVersion

010002000300040005000600070008000

100 300 500 600200 4000Time (s)

0246810

Vers

ion

Bitr

ate (

kbps

)

Figure 17 Experimental results of SDP method

119881max = 7 obtains the lowest average quality Therefore fromTable 6 and Figure 19 we can see that the proposedmethod inoffline context with 119881max = 8 or 9 can provide the best per-formance

Advances in Multimedia 11

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

s)

100 200 300 400 500 6000Time (s)

0246810

Vers

ion

Figure 18 Experimental results of bitrate estimation basedmethod

02468

1012

Bue

r (s)

100 300 500 600200 4000Time (s)

SDPMax 9Max 8

Max 7BE

Figure 19 Buffer level curves of SDP method bitrate estimationbased method and the proposed method (Max9 Max8 and Max7)in offline context

7 Conclusion

In this paper we have proposed an adaptation method forHTTP streaming based on stochastic dynamic programmingThe system model was targeted at real bandwidth trace withstrong bitrate fluctuation of VBR videos Furthermore wehave developed a model to predict the system performancewith the aim of choosing the best setting based on theperformance requirements The experimental results haveshown that our method can effectively adapt VBR videos andperform accurate performance prediction which is useful inplanning adaptation policy

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

References

[1] Cisco Systems Inc ldquoNetworking index Forecast and method-ology 2012ndash2017rdquo Cisco Visual May 2013

[2] A C Begen T Akgul and M Baugher ldquoWatching video overthe web part 1 streaming protocolsrdquo IEEE Internet Computingvol 15 no 2 pp 54ndash63 2011

[3] T C Thang Q-D Ho J W Kang and A T Pham ldquoAdaptivestreaming of audiovisual content using MPEG DASHrdquo IEEETransactions on Consumer Electronics vol 58 no 1 pp 78ndash852012

[4] T Stockhammer ldquoDynamic adaptive streaming over HTTPstandards and design principlesrdquo in Proceeding of the 2ndAnnual ACM Multimedia Systems Conference (MMSys rsquo11) pp133ndash144 San Jose CA USA February 2011

[5] K Miller E Quacchio G Gennari and A Wolisz ldquoAdaptationalgorithm for adaptive streaming over HTTPrdquo in Proceedings ofthe 19th International Packet Video Workshop (PV rsquo12) pp 173ndash178 IEEE Munich Germany May 2012

[6] Y Zhou Y Duan J Sun and Z Guo ldquoTowards simple andsmooth rate adaption for VBR video in DASHrdquo in Proceedingsof the IEEE Visual Communications and Image Processing Con-ference (VCIP rsquo14) pp 9ndash12 IEEE Valletta Malta December2014

[7] S Akhshabi S Narayanaswamy A C Begen and C DovrolisldquoAn experimental evaluation of rate-adaptive video players overHTTPrdquo Signal Processing Image Communication vol 27 no 4pp 271ndash287 2012

[8] C Muller S Lederer and C Timmerer ldquoAn evaluation ofdynamic adaptive streaming over HTTP in vehicular environ-mentsrdquo in Proceedings of the 4th Workshop on Mobile Video(MoVid rsquo12) pp 37ndash42 New York NY USA February 2012

[9] C Liu I Bouazizi and M Gabbouj ldquoRate adaptation foradaptive HTTP streamingrdquo in Proceedings of the 2nd AnnualACMMultimedia Systems Conference (MMSys rsquo11) pp 169ndash174San Jose Calif USA February 2011

[10] S Garcıa J Cabrera and N Garcıa ldquoQuality-control algorithmfor adaptive streaming services over wireless channelsrdquo IEEEJournal on Selected Topics in Signal Processing vol 9 no 1 pp50ndash59 2015

[11] A H Duong T Nguyen T Vu T T Do N P Ngoc and T CThang ldquoSDP-based adaptation for quality control in adaptivestreamingrdquo in Proceedings of the International Conference onComputing Management and Telecommunications (ComMan-Tel rsquo15) pp 194ndash199 IEEE DaNang Vietnam December 2015

[12] S Lederer C Mueller C Timmerer C Concolato J Le Feuvreand K Fliegel ldquoDistributedDASHdatasetrdquo in Proceedings of the4th ACMMultimedia Systems Conference (MMSys rsquo13) pp 131ndash135 Oslo Norway March 2013

[13] ldquoH264AVC video trace libraryrdquo httptraceeasasueduh264[14] T CThang H T Le H X Nguyen A T Pham JW Kang and

Y M Ro ldquoAdaptive video streaming over HTTP with dynamicresource estimationrdquo IEEE Trans On consumer electronics vol58 no 1 pp 78ndash85 2012

[15] T C Thang H T Le A T Pham and Y M Ro ldquoAn evaluationof bitrate adaptation methods for HTTP live streamingrdquo IEEEJournal on Selected Areas in Communications vol 32 no 4 pp693ndash705 2014

[16] R Dubin O Hadar and A Dvir ldquoThe effect of client buffer andMBR consideration onDASHAdaptation Logicrdquo in Proceedingsof the 2013 IEEE Wireless Communications and NetworkingConference (WCNC rsquo13) pp 2178ndash2183 IEEE Shanghai ChinaApril 2013

[17] G Liang and B Liang ldquoEffect of delay and buffering on jitter-free streaming over random VBR channelsrdquo IEEE Transactionson Multimedia vol 10 no 6 pp 1128ndash1141 2008

[18] Y Kang H Chen and L Xie ldquoAn artificial-neural-network-based QoE estimation model for Video streaming over wirelessnetworksrdquo in Proceedings of the 2013 IEEECIC InternationalConference on Communications in China (ICCC rsquo13) pp 264ndash269 IEEE Xirsquoan China August 2013

[19] S Xiang L Cai and J Pan ldquoAdaptive scalable video streaminginwireless networksrdquo in Proceedings of the 3rdACMMultimedia

Advances in Multimedia 9

Table 2 Compare predicted performance and measured performance in offline context using a history bandwidth trace for statistics

Statistics 119881max = 9 119881max = 8 119881max = 7Predicted Measured Predicted Measured Predicted Measured

119860119887(119904) 952 844 961 885 969 965119860119902 748 792 727 771 680 698119860 sw 031 019 018 011 008 005Prundflow 000 000 000 000 000 000

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

12345678

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 13 Experimental results of proposed method in offlinecontext using a history bandwidth trace with Vmax = 7

In the third part of the experiment we consider the onlinecontext in which the prediction of the future bandwidth isbased on the statistical parameters of all previous segmentsSpecifically we divide an entire session into chunks each ofwhich has119871 video segmentsWe treat each chunk individuallyas a mini streaming session Assuming that initially we haveenough statistical data to predict the performance of thefirst chunk The prediction of the subsequent chunks willbe done based on the previous chunks At the beginning ofeach chunk the bandwidth statistics are updated leading to arecomputation of the policy set and the performance In theexperiment we set119871 to 100 video segments It can be observedfrom our experiments that it took only several seconds to(re)compute the whole model Therefore the computationaloverhead is about 3 which is appropriate for the onlinecontext Figures 14 15 and 16 show the adaptation bitrateand version switch behavior of the proposed method in thethree cases of maximum version The predicted performanceand the measured performance in the online context in thethree cases are presented in detail in Tables 3 4 and 5It is very clear that in the online context the predictedperformance is also very close to the measured performance

Next we compare our proposed method with the SDPmethod [10] and the bitrate estimation based method [14]using the same simulation settings as in the first part of ourexperimentThe experimental results obtained by simulatingthe SDP method [10] and bitrate estimation based method[14] are shown in Figures 17 and 18 respectively We can seethat both methods provide very fluctuating version switchesThe detailed statistics of these adaptation methods with themaximum version set to 9 and our proposedmethodwith the

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

kbps

)

012345678910

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 14 Experimental results of the proposed method in onlinecontext with Vmax = 9

Segment rpBitrateVersion

0100020003000400050006000

Bitr

ate (

kbps

)

0123456789

Vers

ion

100 200 300 400 500 6000Time (s)

Figure 15 Experimental results of the proposed method in onlinecontext with Vmax = 8

Segment rpBitrateVersion

0500

100015002000250030003500400045005000

Bitr

ate (

kbps

)

100 200 300 400 500 6000Time (s)

12345678

Vers

ion

Figure 16 Experimental results of the proposed method in onlinecontext with Vmax = 7

10 Advances in Multimedia

Table 3 Compare predicted performance and measured performance in online context with 119881max = 9Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 931 896 892 845 936 886119860119902 769 804 789 768 783 772119860 sw 005 009 004 007 002 005Prundflow 000 000 000 000 000 000

Table 4 Compare predicted performance and measured performance in online context with 119881max = 8Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 953 901 942 863 942 896119860119902 756 786 765 763 765 767119860 sw 000 003 001 003 002 004Prundflow 000 000 000 000 000 000

Table 5 Compare predicted performance and measured performance in online context with 119881max = 7Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 977 993 977 952 983 959119860119902 692 700 692 695 692 700119860 sw 000 000 000 002 000 000Prundflow 000 000 000 000 000 000

Table 6 Statistics of different adaptation methods in offline context

Statistics SDP method119881max = 9

Bitrate estimation method119881max = 9

Proposed method119881max = 9

Proposed method119881max = 8

Proposed method119881max = 7

119860119887(119904) 862 885 852 877 968119860119902 740 785 788 773 687119860 sw 076 043 018 009 002Prundflow 000 000 000 000 000

maximumversion set to 7 8 and 9 are shown in Table 6 It canbe seen that the performance of the bitrate estimationmethodis less effective than our method in terms of average versionand average version switch Regarding the SDP method itis evident that the performance of this method is the worstamong all (with low average version and highest averageswitch per segment) This can be expected because thismethod was not originally designed for real VBR videos

The buffer level curves of the three methods in offlinecontext are shown in Figure 19 We can see that all buffercurves imply streaming sessionswithout any freezes for usersAmong these the SDP method and the proposed methodwith 119881max = 9 provide the most unstable buffers while thebitrate estimation based method and the proposed methodwith 119881max = 7 provide the most stable buffers It can beobserved from Figures 17 and 18 and Table 6 that the SDPmethod and bitrate estimation based method result in veryfluctuating version curves and the proposed method with

Segment rpBitrateVersion

010002000300040005000600070008000

100 300 500 600200 4000Time (s)

0246810

Vers

ion

Bitr

ate (

kbps

)

Figure 17 Experimental results of SDP method

119881max = 7 obtains the lowest average quality Therefore fromTable 6 and Figure 19 we can see that the proposedmethod inoffline context with 119881max = 8 or 9 can provide the best per-formance

Advances in Multimedia 11

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

s)

100 200 300 400 500 6000Time (s)

0246810

Vers

ion

Figure 18 Experimental results of bitrate estimation basedmethod

02468

1012

Bue

r (s)

100 300 500 600200 4000Time (s)

SDPMax 9Max 8

Max 7BE

Figure 19 Buffer level curves of SDP method bitrate estimationbased method and the proposed method (Max9 Max8 and Max7)in offline context

7 Conclusion

In this paper we have proposed an adaptation method forHTTP streaming based on stochastic dynamic programmingThe system model was targeted at real bandwidth trace withstrong bitrate fluctuation of VBR videos Furthermore wehave developed a model to predict the system performancewith the aim of choosing the best setting based on theperformance requirements The experimental results haveshown that our method can effectively adapt VBR videos andperform accurate performance prediction which is useful inplanning adaptation policy

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

References

[1] Cisco Systems Inc ldquoNetworking index Forecast and method-ology 2012ndash2017rdquo Cisco Visual May 2013

[2] A C Begen T Akgul and M Baugher ldquoWatching video overthe web part 1 streaming protocolsrdquo IEEE Internet Computingvol 15 no 2 pp 54ndash63 2011

[3] T C Thang Q-D Ho J W Kang and A T Pham ldquoAdaptivestreaming of audiovisual content using MPEG DASHrdquo IEEETransactions on Consumer Electronics vol 58 no 1 pp 78ndash852012

[4] T Stockhammer ldquoDynamic adaptive streaming over HTTPstandards and design principlesrdquo in Proceeding of the 2ndAnnual ACM Multimedia Systems Conference (MMSys rsquo11) pp133ndash144 San Jose CA USA February 2011

[5] K Miller E Quacchio G Gennari and A Wolisz ldquoAdaptationalgorithm for adaptive streaming over HTTPrdquo in Proceedings ofthe 19th International Packet Video Workshop (PV rsquo12) pp 173ndash178 IEEE Munich Germany May 2012

[6] Y Zhou Y Duan J Sun and Z Guo ldquoTowards simple andsmooth rate adaption for VBR video in DASHrdquo in Proceedingsof the IEEE Visual Communications and Image Processing Con-ference (VCIP rsquo14) pp 9ndash12 IEEE Valletta Malta December2014

[7] S Akhshabi S Narayanaswamy A C Begen and C DovrolisldquoAn experimental evaluation of rate-adaptive video players overHTTPrdquo Signal Processing Image Communication vol 27 no 4pp 271ndash287 2012

[8] C Muller S Lederer and C Timmerer ldquoAn evaluation ofdynamic adaptive streaming over HTTP in vehicular environ-mentsrdquo in Proceedings of the 4th Workshop on Mobile Video(MoVid rsquo12) pp 37ndash42 New York NY USA February 2012

[9] C Liu I Bouazizi and M Gabbouj ldquoRate adaptation foradaptive HTTP streamingrdquo in Proceedings of the 2nd AnnualACMMultimedia Systems Conference (MMSys rsquo11) pp 169ndash174San Jose Calif USA February 2011

[10] S Garcıa J Cabrera and N Garcıa ldquoQuality-control algorithmfor adaptive streaming services over wireless channelsrdquo IEEEJournal on Selected Topics in Signal Processing vol 9 no 1 pp50ndash59 2015

[11] A H Duong T Nguyen T Vu T T Do N P Ngoc and T CThang ldquoSDP-based adaptation for quality control in adaptivestreamingrdquo in Proceedings of the International Conference onComputing Management and Telecommunications (ComMan-Tel rsquo15) pp 194ndash199 IEEE DaNang Vietnam December 2015

[12] S Lederer C Mueller C Timmerer C Concolato J Le Feuvreand K Fliegel ldquoDistributedDASHdatasetrdquo in Proceedings of the4th ACMMultimedia Systems Conference (MMSys rsquo13) pp 131ndash135 Oslo Norway March 2013

[13] ldquoH264AVC video trace libraryrdquo httptraceeasasueduh264[14] T CThang H T Le H X Nguyen A T Pham JW Kang and

Y M Ro ldquoAdaptive video streaming over HTTP with dynamicresource estimationrdquo IEEE Trans On consumer electronics vol58 no 1 pp 78ndash85 2012

[15] T C Thang H T Le A T Pham and Y M Ro ldquoAn evaluationof bitrate adaptation methods for HTTP live streamingrdquo IEEEJournal on Selected Areas in Communications vol 32 no 4 pp693ndash705 2014

[16] R Dubin O Hadar and A Dvir ldquoThe effect of client buffer andMBR consideration onDASHAdaptation Logicrdquo in Proceedingsof the 2013 IEEE Wireless Communications and NetworkingConference (WCNC rsquo13) pp 2178ndash2183 IEEE Shanghai ChinaApril 2013

[17] G Liang and B Liang ldquoEffect of delay and buffering on jitter-free streaming over random VBR channelsrdquo IEEE Transactionson Multimedia vol 10 no 6 pp 1128ndash1141 2008

[18] Y Kang H Chen and L Xie ldquoAn artificial-neural-network-based QoE estimation model for Video streaming over wirelessnetworksrdquo in Proceedings of the 2013 IEEECIC InternationalConference on Communications in China (ICCC rsquo13) pp 264ndash269 IEEE Xirsquoan China August 2013

[19] S Xiang L Cai and J Pan ldquoAdaptive scalable video streaminginwireless networksrdquo in Proceedings of the 3rdACMMultimedia

10 Advances in Multimedia

Table 3 Compare predicted performance and measured performance in online context with 119881max = 9Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 931 896 892 845 936 886119860119902 769 804 789 768 783 772119860 sw 005 009 004 007 002 005Prundflow 000 000 000 000 000 000

Table 4 Compare predicted performance and measured performance in online context with 119881max = 8Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 953 901 942 863 942 896119860119902 756 786 765 763 765 767119860 sw 000 003 001 003 002 004Prundflow 000 000 000 000 000 000

Table 5 Compare predicted performance and measured performance in online context with 119881max = 7Statistics Chunk 1 Chunk 2 Chunk 3

Predicted Measured Predicted Measured Predicted Measured119860119887(119904) 977 993 977 952 983 959119860119902 692 700 692 695 692 700119860 sw 000 000 000 002 000 000Prundflow 000 000 000 000 000 000

Table 6 Statistics of different adaptation methods in offline context

Statistics SDP method119881max = 9

Bitrate estimation method119881max = 9

Proposed method119881max = 9

Proposed method119881max = 8

Proposed method119881max = 7

119860119887(119904) 862 885 852 877 968119860119902 740 785 788 773 687119860 sw 076 043 018 009 002Prundflow 000 000 000 000 000

maximumversion set to 7 8 and 9 are shown in Table 6 It canbe seen that the performance of the bitrate estimationmethodis less effective than our method in terms of average versionand average version switch Regarding the SDP method itis evident that the performance of this method is the worstamong all (with low average version and highest averageswitch per segment) This can be expected because thismethod was not originally designed for real VBR videos

The buffer level curves of the three methods in offlinecontext are shown in Figure 19 We can see that all buffercurves imply streaming sessionswithout any freezes for usersAmong these the SDP method and the proposed methodwith 119881max = 9 provide the most unstable buffers while thebitrate estimation based method and the proposed methodwith 119881max = 7 provide the most stable buffers It can beobserved from Figures 17 and 18 and Table 6 that the SDPmethod and bitrate estimation based method result in veryfluctuating version curves and the proposed method with

Segment rpBitrateVersion

010002000300040005000600070008000

100 300 500 600200 4000Time (s)

0246810

Vers

ion

Bitr

ate (

kbps

)

Figure 17 Experimental results of SDP method

119881max = 7 obtains the lowest average quality Therefore fromTable 6 and Figure 19 we can see that the proposedmethod inoffline context with 119881max = 8 or 9 can provide the best per-formance

Advances in Multimedia 11

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

s)

100 200 300 400 500 6000Time (s)

0246810

Vers

ion

Figure 18 Experimental results of bitrate estimation basedmethod

02468

1012

Bue

r (s)

100 300 500 600200 4000Time (s)

SDPMax 9Max 8

Max 7BE

Figure 19 Buffer level curves of SDP method bitrate estimationbased method and the proposed method (Max9 Max8 and Max7)in offline context

7 Conclusion

In this paper we have proposed an adaptation method forHTTP streaming based on stochastic dynamic programmingThe system model was targeted at real bandwidth trace withstrong bitrate fluctuation of VBR videos Furthermore wehave developed a model to predict the system performancewith the aim of choosing the best setting based on theperformance requirements The experimental results haveshown that our method can effectively adapt VBR videos andperform accurate performance prediction which is useful inplanning adaptation policy

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

References

[1] Cisco Systems Inc ldquoNetworking index Forecast and method-ology 2012ndash2017rdquo Cisco Visual May 2013

[2] A C Begen T Akgul and M Baugher ldquoWatching video overthe web part 1 streaming protocolsrdquo IEEE Internet Computingvol 15 no 2 pp 54ndash63 2011

[3] T C Thang Q-D Ho J W Kang and A T Pham ldquoAdaptivestreaming of audiovisual content using MPEG DASHrdquo IEEETransactions on Consumer Electronics vol 58 no 1 pp 78ndash852012

[4] T Stockhammer ldquoDynamic adaptive streaming over HTTPstandards and design principlesrdquo in Proceeding of the 2ndAnnual ACM Multimedia Systems Conference (MMSys rsquo11) pp133ndash144 San Jose CA USA February 2011

[5] K Miller E Quacchio G Gennari and A Wolisz ldquoAdaptationalgorithm for adaptive streaming over HTTPrdquo in Proceedings ofthe 19th International Packet Video Workshop (PV rsquo12) pp 173ndash178 IEEE Munich Germany May 2012

[6] Y Zhou Y Duan J Sun and Z Guo ldquoTowards simple andsmooth rate adaption for VBR video in DASHrdquo in Proceedingsof the IEEE Visual Communications and Image Processing Con-ference (VCIP rsquo14) pp 9ndash12 IEEE Valletta Malta December2014

[7] S Akhshabi S Narayanaswamy A C Begen and C DovrolisldquoAn experimental evaluation of rate-adaptive video players overHTTPrdquo Signal Processing Image Communication vol 27 no 4pp 271ndash287 2012

[8] C Muller S Lederer and C Timmerer ldquoAn evaluation ofdynamic adaptive streaming over HTTP in vehicular environ-mentsrdquo in Proceedings of the 4th Workshop on Mobile Video(MoVid rsquo12) pp 37ndash42 New York NY USA February 2012

[9] C Liu I Bouazizi and M Gabbouj ldquoRate adaptation foradaptive HTTP streamingrdquo in Proceedings of the 2nd AnnualACMMultimedia Systems Conference (MMSys rsquo11) pp 169ndash174San Jose Calif USA February 2011

[10] S Garcıa J Cabrera and N Garcıa ldquoQuality-control algorithmfor adaptive streaming services over wireless channelsrdquo IEEEJournal on Selected Topics in Signal Processing vol 9 no 1 pp50ndash59 2015

[11] A H Duong T Nguyen T Vu T T Do N P Ngoc and T CThang ldquoSDP-based adaptation for quality control in adaptivestreamingrdquo in Proceedings of the International Conference onComputing Management and Telecommunications (ComMan-Tel rsquo15) pp 194ndash199 IEEE DaNang Vietnam December 2015

[12] S Lederer C Mueller C Timmerer C Concolato J Le Feuvreand K Fliegel ldquoDistributedDASHdatasetrdquo in Proceedings of the4th ACMMultimedia Systems Conference (MMSys rsquo13) pp 131ndash135 Oslo Norway March 2013

[13] ldquoH264AVC video trace libraryrdquo httptraceeasasueduh264[14] T CThang H T Le H X Nguyen A T Pham JW Kang and

Y M Ro ldquoAdaptive video streaming over HTTP with dynamicresource estimationrdquo IEEE Trans On consumer electronics vol58 no 1 pp 78ndash85 2012

[15] T C Thang H T Le A T Pham and Y M Ro ldquoAn evaluationof bitrate adaptation methods for HTTP live streamingrdquo IEEEJournal on Selected Areas in Communications vol 32 no 4 pp693ndash705 2014

[16] R Dubin O Hadar and A Dvir ldquoThe effect of client buffer andMBR consideration onDASHAdaptation Logicrdquo in Proceedingsof the 2013 IEEE Wireless Communications and NetworkingConference (WCNC rsquo13) pp 2178ndash2183 IEEE Shanghai ChinaApril 2013

[17] G Liang and B Liang ldquoEffect of delay and buffering on jitter-free streaming over random VBR channelsrdquo IEEE Transactionson Multimedia vol 10 no 6 pp 1128ndash1141 2008

[18] Y Kang H Chen and L Xie ldquoAn artificial-neural-network-based QoE estimation model for Video streaming over wirelessnetworksrdquo in Proceedings of the 2013 IEEECIC InternationalConference on Communications in China (ICCC rsquo13) pp 264ndash269 IEEE Xirsquoan China August 2013

[19] S Xiang L Cai and J Pan ldquoAdaptive scalable video streaminginwireless networksrdquo in Proceedings of the 3rdACMMultimedia

Advances in Multimedia 11

Segment rpBitrateVersion

01000200030004000500060007000

Bitr

ate (

s)

100 200 300 400 500 6000Time (s)

0246810

Vers

ion

Figure 18 Experimental results of bitrate estimation basedmethod

02468

1012

Bue

r (s)

100 300 500 600200 4000Time (s)

SDPMax 9Max 8

Max 7BE

Figure 19 Buffer level curves of SDP method bitrate estimationbased method and the proposed method (Max9 Max8 and Max7)in offline context

7 Conclusion

In this paper we have proposed an adaptation method forHTTP streaming based on stochastic dynamic programmingThe system model was targeted at real bandwidth trace withstrong bitrate fluctuation of VBR videos Furthermore wehave developed a model to predict the system performancewith the aim of choosing the best setting based on theperformance requirements The experimental results haveshown that our method can effectively adapt VBR videos andperform accurate performance prediction which is useful inplanning adaptation policy

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

References

[1] Cisco Systems Inc ldquoNetworking index Forecast and method-ology 2012ndash2017rdquo Cisco Visual May 2013

[2] A C Begen T Akgul and M Baugher ldquoWatching video overthe web part 1 streaming protocolsrdquo IEEE Internet Computingvol 15 no 2 pp 54ndash63 2011

[3] T C Thang Q-D Ho J W Kang and A T Pham ldquoAdaptivestreaming of audiovisual content using MPEG DASHrdquo IEEETransactions on Consumer Electronics vol 58 no 1 pp 78ndash852012

[4] T Stockhammer ldquoDynamic adaptive streaming over HTTPstandards and design principlesrdquo in Proceeding of the 2ndAnnual ACM Multimedia Systems Conference (MMSys rsquo11) pp133ndash144 San Jose CA USA February 2011

[5] K Miller E Quacchio G Gennari and A Wolisz ldquoAdaptationalgorithm for adaptive streaming over HTTPrdquo in Proceedings ofthe 19th International Packet Video Workshop (PV rsquo12) pp 173ndash178 IEEE Munich Germany May 2012

[6] Y Zhou Y Duan J Sun and Z Guo ldquoTowards simple andsmooth rate adaption for VBR video in DASHrdquo in Proceedingsof the IEEE Visual Communications and Image Processing Con-ference (VCIP rsquo14) pp 9ndash12 IEEE Valletta Malta December2014

[7] S Akhshabi S Narayanaswamy A C Begen and C DovrolisldquoAn experimental evaluation of rate-adaptive video players overHTTPrdquo Signal Processing Image Communication vol 27 no 4pp 271ndash287 2012

[8] C Muller S Lederer and C Timmerer ldquoAn evaluation ofdynamic adaptive streaming over HTTP in vehicular environ-mentsrdquo in Proceedings of the 4th Workshop on Mobile Video(MoVid rsquo12) pp 37ndash42 New York NY USA February 2012

[9] C Liu I Bouazizi and M Gabbouj ldquoRate adaptation foradaptive HTTP streamingrdquo in Proceedings of the 2nd AnnualACMMultimedia Systems Conference (MMSys rsquo11) pp 169ndash174San Jose Calif USA February 2011

[10] S Garcıa J Cabrera and N Garcıa ldquoQuality-control algorithmfor adaptive streaming services over wireless channelsrdquo IEEEJournal on Selected Topics in Signal Processing vol 9 no 1 pp50ndash59 2015

[11] A H Duong T Nguyen T Vu T T Do N P Ngoc and T CThang ldquoSDP-based adaptation for quality control in adaptivestreamingrdquo in Proceedings of the International Conference onComputing Management and Telecommunications (ComMan-Tel rsquo15) pp 194ndash199 IEEE DaNang Vietnam December 2015

[12] S Lederer C Mueller C Timmerer C Concolato J Le Feuvreand K Fliegel ldquoDistributedDASHdatasetrdquo in Proceedings of the4th ACMMultimedia Systems Conference (MMSys rsquo13) pp 131ndash135 Oslo Norway March 2013

[13] ldquoH264AVC video trace libraryrdquo httptraceeasasueduh264[14] T CThang H T Le H X Nguyen A T Pham JW Kang and

Y M Ro ldquoAdaptive video streaming over HTTP with dynamicresource estimationrdquo IEEE Trans On consumer electronics vol58 no 1 pp 78ndash85 2012

[15] T C Thang H T Le A T Pham and Y M Ro ldquoAn evaluationof bitrate adaptation methods for HTTP live streamingrdquo IEEEJournal on Selected Areas in Communications vol 32 no 4 pp693ndash705 2014

[16] R Dubin O Hadar and A Dvir ldquoThe effect of client buffer andMBR consideration onDASHAdaptation Logicrdquo in Proceedingsof the 2013 IEEE Wireless Communications and NetworkingConference (WCNC rsquo13) pp 2178ndash2183 IEEE Shanghai ChinaApril 2013

[17] G Liang and B Liang ldquoEffect of delay and buffering on jitter-free streaming over random VBR channelsrdquo IEEE Transactionson Multimedia vol 10 no 6 pp 1128ndash1141 2008

[18] Y Kang H Chen and L Xie ldquoAn artificial-neural-network-based QoE estimation model for Video streaming over wirelessnetworksrdquo in Proceedings of the 2013 IEEECIC InternationalConference on Communications in China (ICCC rsquo13) pp 264ndash269 IEEE Xirsquoan China August 2013

[19] S Xiang L Cai and J Pan ldquoAdaptive scalable video streaminginwireless networksrdquo in Proceedings of the 3rdACMMultimedia

12 Advances in Multimedia

Systems Conference (MMSys rsquo12) pp 167ndash172 New York NYUSA February 2012

[20] Y Liu and J Y B Lee ldquoOn adaptive video streaming withpredictable streaming performancerdquo in Proceedings of the 2014IEEE Global Communications Conference (GLOBECOM rsquo14)pp 1164ndash1169 IEEE Austin TX USA December 2014

[21] S Akhshabi A C Begen and C Dovrolis ldquoAn experimentalevaluation of rate-adaptation algorithms in adaptive streamingover HTTPrdquo in Proceedings of the 2nd Annual ACMMultimediaSystemsConference (MMSys rsquo11) pp 157ndash168 San Jose CAUSAFebruary 2011

[22] G Ji and B Liang ldquoStochastic rate control for scalable VBRvideo streaming over wireless networksrdquo in Proceedings of the2009 IEEE Global Telecommunications Conference (GLOBE-COM rsquo09) Honolulu HI USA December 2009

[23] M Xing S Xiang and L Cai ldquoRate adaptation strategy forvideo streaming overmultiple wireless access networksrdquo in Pro-ceedings of the 2012 IEEE Global Communications Conference(GLOBECOM rsquo12) pp 5745ndash5750 IEEE Anaheim CA USADecember 2012

[24] Z Ye R EL-Azouzi T Jimenez and Y Xu ldquoComputing TheQuality of Experience in Network Modeled by a MarkovModulated Fluid Modelrdquo 2014

[25] Y Xu E Altman R El-Azouzi M Haddad S Elayoubiand T Jimenez ldquoAnalysis of buffer starvation with applicationto objective QoE optimization of streaming servicesrdquo IEEETransactions on Multimedia vol 16 no 3 pp 813ndash827 2014

[26] H Thomas E Cormen Charles L Ronald and S CliffordIntroduction to Algorithms MIT Press and McGraw-Hill 3rdedition 2009

[27] Y Mansour and S Singh ldquoOn the complexity of policy itera-tionrdquo UAI pp 401ndash408 1999

[28] L Rizzo ldquoDummynet a simple approach to the evaluationof networkprotocolsrdquo ACM Computer Communication Reviewvol 27 no 1 pp 31ndash41 1997

[29] D M Nguyen L B Tran H T Le N P Ngoc and T CThangldquoAn evaluation of segment duration effects in HTTP adaptivestreaming over mobile networksrdquo in Proceedings of the 2ndNational Foundation for Science and Technology DevelopmentConfernce on Information and Computer Science (NICS rsquo15) pp248ndash253 Ho Chi Minh City Vietnam September 2015

[30] L Koskimies T Taleb and M Bagaa ldquoQoE estimation-basedserver benchmarking for virtual video delivery platformrdquo inProceeding of IEEE International Conference on Communica-tions (ICC rsquo17) Paris France May 2017

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of