A Novel Traffic Rate Measurement Algorithm for QoE-Aware Video Admission Control
11 A Novel Traffic Rate Measurement Algorithm forQoE-Aware Video Admission Control
Qahhar Muhammad Qadir, Alexander A. Kist, and Zhongwei Zhang
Abstract —With the inevitable dominance of video traffic on theInternet, providing perceptually good video quality is becoming achallenging task. This is partly due to the bursty nature of videotraffic, changing network conditions and limitations of networktransport protocols. This growth of video traffic has made Qualityof Experience (QoE) of the end user the focus of the researchcommunity. In contrast, Internet service providers are concernedabout maximizing revenue by accepting as many sessions aspossible, as long as customers remain satisfied. However, thereis still no entirely satisfactory admission algorithm for flowswith variable rate. The trade-off between the number of sessionsand perceived QoE can be optimized by exploiting the burstynature of video traffic. This paper proposes a novel algorithmto determine the upper limit of the aggregate video rate thatcan exceed the available bandwidth without degrading the QoEof accepted video sessions. A parameter β that defines theexceedable limit is defined. The proposed algorithm results inaccepting more sessions without compromising the QoE of on-going video sessions. Thus it contributes to the optimization ofthe QoE-Session trade-off in support of the expected growth ofvideo traffic on the Internet. Index Terms —QoE, MBAC, Video, Optimization.
I. I
NTRODUCTION W ITH the inevitable dominance of video traffic on theInternet, it is becoming a challenging task to provideperceptually good video quality. This is partly due to thebursty nature of video traffic, changing network conditions andlimitations of network transport protocols. Cisco predicts that“The sum of all forms of video (TV, video on demand [VoD],Internet, and P2P) will be in the range of 80 to 90 percentof global consumer traffic by 2018” [1]. Over the last decade,efforts have been made to provide Quality of Service (QoS)within the core network by considering objective parameters atthe network layer such as bandwidth, delay and jitter. Diffserv[2] is an example of these paradigms that can support QoS.The research community and Internet Service Providers (ISP)s
Manuscript received August 09, 2014; revised December 19, 2014 andMarch 10, 2015; accepted March 21, 2015. Date of publication March 25,2015.Qahhar Muhammad Qadir and Alexander A. Kist are with the Schoolof Mechanical and Electrical Engineering, University of Southern Queens-land, Toowoomba, QLD 4350, Australia (e-mail: [email protected];[email protected]).Zhongwei Zhang is with the School of Agricultural, Computational andEnvironmental Sciences, University of Southern Queensland, Toowoomba,QLD 4350, Australia (e-mail: [email protected]).© 2015 IEEE. Personal use of this material is permitted. Permission fromIEEE must be obtained for all other uses, in any current or future media,including reprinting/republishing this material for advertising or promotionalpurposes, creating new collective works, for resale or redistribution to serversor lists, or reuse of any copyrighted component of this work in other works.Digital Object Identifier 10.1109/TMM.2015.2416637 have made subjective quality, as perceived by the end user,a main research target. The International TelecommunicationUnion (ITU) defines this parameter as “Quality of Experience”(QoE) [3]. The current design of the Internet has to beenhanced to extend the scope of QoS to consider end-to-endquality, be content-aware and user centric.Admission control is a well known technique to keeptraffic load at acceptable levels and guarantee quality foradmitted sessions via resource reservation. This idea hasbeen adopted in the past in QoS architectures such as inDiffserv. Thus, some sort of explicit admission control isrequired to provide per-session QoE by which the networkhas the right to deny sessions to ensure that the QoE of activesessions is not affected by new sessions. ISP are concernedabout maximizing revenue by accepting as many sessions aspossible. Measurement-Based Admission Control (MBAC) hasbeen proposed as a solution. In contrast to parameter-basedadmission control, it is better suited to video traffic. MBACrelies on the measurement of video characteristics such ascurrent load and peak rate. Different algorithms have beenproposed to estimate network load [4]; however there arealgorithms which rely on the Instantaneous Aggregate ArrivalRate (
IAAR ) for their operations.Despite all the efforts, there is no entirely satisfactoryadmission algorithm for variable rate flows [5]. Admissioncontrol algorithms must not rely on worst-case bounds orinstantaneous video arrival rate, as they do not reflect thebursty characteristic of video traffic. This is due to the factthat the burstiness of video flows can be compensated bythe silence of other flows. The Internet Engineering TaskForce (IETF) has standardized the Pre-Congestion Notification(PCN) based admission control for the Internet [6] whichmerely relies on the calculated rate for a measurement pe-riod. The perceived QoE-Session relationship can be greatlyoptimized by exploiting the bursty nature of video traffic.This paper contributes to the measurement mechanism forQoE-aware admission control. It proposes a novel traffic ratemeasuring algorithm for video admission control mechanisms.The relationship between
IAAR and the proposed rate is es-tablished mathematically. We call the proposed measured rate“Proposed Instantaneous Aggregate Arrival Rate” (
Pro-IAAR )and proposed admission control procedure based on
Pro-IAAR “ Pro-IAAR -Based Measurement Admission Control” (
Pro-IBMAC ). We also call the admission control procedures whichare based on the Calculated Rate (
CalR ) such as PCN “
CalR -Based Admission Control” (
CBAC ).Whereas traffic measurement algorithms and MBAC havebeen widely covered by the research community, to the best a r X i v : . [ c s . N I] A ug of our knowledge this is the first work that includes QoE in thearea of the QoE-Session optimization. The main contributionsof this paper are twofold:1) A novel algorithm for traffic measurement supportedby the mathematical model is proposed. The algorithmmeasures the exceedable video aggregate rate that is ableto keep the video quality unimpaired. The exceedablerate is the total bitrate of enrolled video traffic that canexceed the available link capacity without degradationto the user’s perception of quality.2) Operation of the proposed measurement algorithm isdemonstrated with an implementation in a QoE-awareadmission control procedure for video admission.The remainder of the paper is organized as follows. SectionII presents related work. Assumptions made by this paperare detailed in Section III. Section IV provides a theoreticalbackground of QoE. Section V presents the mathematicalmodel for the proposed algorithm. The simulation setup isexplained and results are presented and discussed in SectionVI. Section VII describes the environment of the subjectivetests and analysis of the collected data. The proposed modelis validated in Section VIII. The paper concludes with SectionIX. II. R ELATED W ORK
MBAC is not a new topic as work has been undertaken sincevideo traffic has emerged on the Internet. It includes two maincomponents: measurements of network load and admissionpolicies. Four MBAC algorithms are presented in [7] basedon Chernoff bounds. A MBAC scheme based on measuredmean and variance of load offered to the cross-protect priorityqueue is proposed in [5].As traffic flow rate is only meaningful if it is associatedwith a corresponding interval length. Network traffic oversome interval has been studied as an essential part of theMBAC functionality. The admission control scheme proposedin [8] estimates the equivalent capacity of a class of aggregatedtraffic based on Hoeffding bounds for controlled-load services.The suitability of the average instead of the instantaneousarrival rate for video streaming admission decisions has beeninvestigated in [9]. An algorithm for MBAC has been intro-duced in [10] that employs adaptive and measured peak rateenvelopes of the aggregate traffic flow to allocate resources formulticlass networks with link sharing. The flow behavior as afunction of interval length can be described by the proposedrate envelope which characterizes the extreme values (maximalrates) of the aggregate flow that can avoid packet loss. As asupporting mechanism in flow and admission control, tech-niques have been developed for estimating available bandwidth[11], [12], [13], [14] and [15].Other studies have compared the performance of MBACalgorithms. The simple sum; a parameter-based admissioncontrol algorithm has been compared to three measurement-based algorithms; the measured sum, acceptance region andequivalent bandwidth based on the link utilization and adher-ence to service commitment [16]. The robustness of [8], [16]and [10] in meeting the QoS target have been compared in [17]. They have been further evaluated based on maximumtolerable packet loss rate and maximum packet queuing delaywithout assuming any explicit knowledge on incoming flowsand on-going traffic [18]. All of the three studied algorithmswere found to meet the first target of maximum tolerablepacket loss rate while only [10] was able to always meetthe second target of maximum packet queuing delay. Theknowledge-base admission control scheme introduced in [19]determines whether to accept a flow based on QoS perfor-mance parameters such as maximum tolerable delay or packetloss rate. The scheme achieves a good trade-off between flowperformance and resource utilization compared to [16] and[10]. In [20] the architecture of centralized, distributed, hybrid,class-based and active/passive MBAC and their limitations onthe quality control of network services have been compared.The efficiency of MBAC algorithms depends on interactionsbetween several time-scales, ranging from the very short timescales to the entire session. Work in [21] has studied howuncertainty in the measurements of MBAC varies with thelength of the observation window and described a method-ology for analyzing measurement errors and performance.The concept of similar flows and adding slack in bandwidthhave been introduced to minimize the probability of falseacceptance. In [22] an implementation-based comparison ofMBAC algorithms has been made using a purpose built testenvironment. It has revealed that there is no single idealMBAC algorithm due to computation overheads, multipletimescales present in both traffic and management and errorresulting from random properties of measurements. These fea-tures dramatically impact the MBAC algorithm’s performance.Work presented in [23] has proposed a delay-aware ad-mission control to guarantee delay bounds for delay sensitiveapplications. The video quality model presented in [24] targetsSkype video calls based on measurement and can be used foruser QoE-aware network provisioning. The model can find theminimum bandwidth needed to accommodate N concurrentSkype video calls with satisfactory Mean Opinion Score(MOS). The study conducted in [25] has investigated the sys-tem architecture, video generation and adaptation, packet lossrecovery and QoE of video-conferencing solutions. Google+,iChat and Skype were all covered in the study. The deliveredquality was measured in terms of end-to-end delay in a widerange of real and emulated network scenarios. The study hasfound that the layered video coding and server architecture(used by Google+ and Skype) can significantly improve userconferencing experiences. Most recently, [26] has proposeda model-based admission control algorithm to predict theQoS metrics based on which and the QoS constraints of theflows, appropriate decision for new flow is taken. The averagenumber of satisfied users was maximized through a QoE-awarescheduling framework by sending a single bit feedback toindicate the satisfaction level [27].As a cutting edge proposed admission control mechanismfor multimedia network, the PCN-based admission control [28]has attracted the attention of researchers. Several modificationsto the PCN algorithm have been proposed in [29]. An exten-sion to the PCN-based admission control system has been pro-posed in [30]. A novel metering algorithm based on a sliding- window, to cope with the bursty nature of video sessions andanother adaptive algorithm to facilitate the configuration ofPCN were proposed in that work.Admission control has also been proposed to better supportapplications with QoS requirements in wireless networks. Theappropriate thresholds for admission decisions were studied by[31]. A flow-level mechanism for multiple antenna equippednodes to maximize flow acceptance and improve networkthroughput has been developed in [32]. A QoE-based admis-sion control for wireless has been proposed in [33] in whichthe access point controls video sessions based on the MOSscores computed by pseudo-subjective quality assessment toolrun on the access point.Most of the MBAC algorithms that have been discussedin the literature are per-aggregate MBAC algorithms. The per-flow MBAC algorithm presented in [34] targets the flow-awarenetwork by adopting dynamic priority scheduling for flowaggregation. A newly admitted flow is given a lower priority bythe proposed algorithm, however its priority is improved whenan existing flow leaves the network. Finally, an enhancementto the MBAC has been proposed to mitigate the impact offair rate degradation and ensure better quality in flow-awarenetwork by [35]. III. A SSUMPTIONS
This paper makes the following assumptions: • Video traffic is the dominant Internet traffic [1]. It is theonly traffic that is subject to admission control. Othertraffic volumes are small in comparison and thereforeonly video traffic is considered. • Video traffic is bursty in nature as video applications inreality send traffic at a very variable rate [21]. • An explicit admission control is required to provide anacceptable level of QoE [21] on bottleneck links. • There is no danger of a “flash crowd”, in which manyadmission requests arrive within the reaction time ofadmission mechanisms, because then they all might getadmitted and so overload the network [28].IV. Q O E; A N EW Q UALITY P ARADIGM
QoE is the quality as experienced by end users. The purposeof introducing QoE is to include all aspects of multimediasystems that are related to media quality. Addressing qualityfrom end user experience or perceived QoE is a relativelynew approach which requires more research in all directionssuch as optimization, assessment, monitoring, managementand prediction. This is due to the emergence of massivevideo services and development of huge number of videocapable devices such as smart phones [1]. Various layers (fromvideo encoding to decoding) and across the access and/orcore networks are involved in providing an end-to-end QoEto end users. Technically, perceived video quality is affectedby the trade-off relationship between encoding redundancy andnetwork impairments. In addition to network parameters suchas bandwidth, delay, packet loss ration, other technical andnon-technical parameters may affect quality [36]. There are different approaches to measure and estimateQoE. Subjective, objective or hybrid approaches are mainlyused for that purpose. Since people have different perceptionsof the same video content, groups of people carry out sub-jective tests by grading the shown sequence. This is time-consuming and costly; however it is worthwhile as real usersare involved in the tests. Objective video quality metricsare often proposed because none of the QoS parameters canprecisely define the QoE of multimedia services [37]. Theseobjective approaches are carried out by the use of algorithmsand formulas. Peak Signal to Noise Ratio (PSNR) and Struc-tural Similarity (SSIM) are two full reference objective videoquality metrics. They compare the original video with received(possibly distorted) video and calculate the MOS value. PSNRis mostly used for its simplicity and good correlation withthe subjective video test result. PSNR tools are available tocalculate the PSNR value. A possible mapping of PSNR toMOS is shown in Table I [38]. However, this is a problematicapproach as PSNR does not directly correspond to MOS [39].On the other hand, SSIM estimates the perceived qualityframe by frame and is considered to have a higher correlationwith subjective quality ratings [40]. The SSIM index assumesthat the human visual system is more oriented towards theidentification of structural information in video sequences. Itproduces a score between 0 and 1 from original and receivedsignals [41]. The third approach is a hybrid between subjectiveand objective methods in which both the technical parametersas well as human rating are taken into account [42] [43].ITU recommends objective modeling of measurable technicalperformance and subjective testing with people [36].
TABLE IP
OSSIBLE
PSNR TO MOS
MAPPING
PSNR MOS Quality >
37 5 Excellent31-37 4 Good25-31 3 Fair20-25 2 Poor <
20 1 Bad
V. M
ODELING
A. Measurement Algorithm
In this section, we describe a new approach to measure traf-fic rate that suits video traffic. For the benefit of comparison,we introduce the traditional approach of traffic measurement
IAAR then present our proposed measurement algorithm
Pro-IAAR . Since, the measurement mechanism is proposed forvideo admission procedures, it will be modeled into an ad-mission control scheme
Pro-IBMAC ; Equation (9).
IAAR atany time t > i > IAAR ( t ) = n (cid:88) i =1 x i ( t ) (1)where x i ( t ) is the instantaneous arrival rate (throughput) ofsession i at time t , and n is the number of sessions. Let x i ( t ) bean independent random variable with minimum rate x mini ( t ) ,peak rate x maxi ( t ) and x mini ( t ) ≤ x i ( t ) ≤ x maxi ( t ) . Further assuming that x i ( t ) is a discrete random variable that takesany set of values from a finite data set x ( t ) , x ( t ) , .... x n ( t ) each of probability p ( t ) , p ( t ) , .... p n ( t ) respectively.A new session will be accepted by CBAC , if the sum of
CalR( τ ) for the time window ( τ ) plus the peak rate of thenew session x new is less or equal to the link’s capacity C l asgiven by Equation (2): CalR ( τ ) + x new ≤ C l (2)In our proposed scheme we consider Pro-IAAR(t) as anadmission parameter instead of
CalR( τ ) . Now we find how Pro-IAAR(t) is related to
IAAR(t) . We utilize the Hoeffdinginequality theorem [44] to develop a model for
Pro-IAAR(t) .The reason behind this approach is that the Hoeffding theoremrelates
IAAR(t) and the average of
IAAR(t) ; µ S ( t ) . It definesthe upper bound of the probability that the sum of n indepen-dent random variables will be greater than the average by n (cid:15) or more for (cid:15) > . Equation (3) quantifies this probabilityrelationship between IAAR(t) and µ S ( t ) . We then developa relationship between Pro-IAAR(t) and
IAAR(t) . Hoeffdingbounds were first used for admission control algorithms in[8].
P r { IAAR ( t ) ≥ µ S ( t ) + n(cid:15) } ≤ γ (3)where γ is given by Equation (4): γ = exp (cid:18) − n (cid:15) (cid:80) ni =1 ( x maxi ( t ) − x mini ( t )) (cid:19) (4) µ S ( t ) is the expectation value of IAAR(t) which is given byEquation (5) in which p i represents the probability the session i is active: µ S ( t ) = EIAAR ( t ) = n (cid:88) i =1 x i ( t ) p i ( t ) (5)The term µ S ( t )+ n(cid:15) in Equation (3) represents the proposed Pro-IAAR(t) at time t which is given by Equation (6) and (cid:15) isgiven by Equation (7): P ro - IAAR ( t ) = µ S ( t ) + n(cid:15) (6) (cid:15) = βµ S ( t ) n − n < β ≤ (7)Parameter β represents how much the total bitrate of en-rolled video traffic can exceed the available link capacity with-out degradation to the user perception quality. It governs thedegree of the efficiency of Pro-IBMAC . Therefore, choosing aproper value for β controls the degree of risk of the admissiondecision as it balances the QoE-Session trade-off relationship.The value of β that optimizes this relationship is referred toas “proposed value” in this paper.The condition (cid:15)> β > n > β >
1, the scope of the proposed schemeis only for 0 < β ≤
1. High values of β within this range lets Pro-IBMAC function similar to traditional admission control mechanisms, while a smaller value leads to accepting moresessions and compromising QoE. We propose a model for β in Section V-B.A new requested session will be accepted by Pro-IBMAC if the condition in Equation (8) meets:
P ro - IAAR ( t ) + x new ≤ C l (8)Substituting Equations (5) and (7) in Equation (6), thenEquation (6) in Equation (8), we get: n (cid:88) i =1 x i ( t ) p i ( t ) { β ( n − } + x new ≤ C l (9)In Equation (9), x new is the required rate of new sessionand Cl is the link capacity. Studies recommend that peak ratebe measured for x new using techniques such as token bucketsand traffic envelopes [8] and [10]. Others compute the peakrate of a new incoming flow by tracking the first A packets ofthe flow and using sliding window [19].In summary, Pro-IBMAC in Equation (9) employs
Pro-IAAR(t) in Equation (6) which is based on the Hoeffdinginequality theorem. The value of γ in Equation (4) specifiesthe level of optimization achieved by considering Pro-IAAR(t) in terms of number of sessions that can be fitted on a particularlink compared to the
CalR( τ ) in Equation (2). B. Proposed Model for β The tuning parameter β affects the operation of the proposedalgorithm. The value can be set to optimize the trade-offrelationship between QoE of enrolled sessions and numberof sessions. In this section, we develop a model for β . Weestimate the value of β using two publicly available videosequences; a 30-seconds clip called Mother And Daughter( MAD ) and a 35-seconds clip called
Paris . These two videosequences are used to validate the proposed β model forvarious video content. Similar short sequences have also beenused for video streaming service and subjective tests [45].While choosing the videos, the following points were takeninto consideration: firstly, long video is not practical forsubjective tests in which subjects evaluate a numbers of videos.Secondly, the aim was to evaluate the admission control-specifically the acceptance/rejection of sessions-and evaluatethe admission rate. Thus the duration of video is not expectedto have effect on the evaluation of the proposed algorithm.The MAD sequence was taken as a slow moving content dueto low motion of its video scenes and
Paris as a fast movingcontent due to fast motion of its video scenes. The consideredcontents were classified into slow and fast based on commonconventions and the size of their encoded frames, as fastercontent has larger frame size. Other studies have classifiedvideo contents in a similar way, e.g. [45]. Details about thevideo sequences are shown in Table II. Other simulationsettings including the coding and network parameters areexplained in Section VI-A.We run extensive simulation to find parameters that poten-tially affect β . C l , n and QoE were found to have impact on β . QoE was measured by simulation which will be explained
TABLE IIV
IDEO SEQUENCE DESCRIPTION
Description Video sequence 1 Video sequence 2Name
MAD Paris
Description A mother and daughter A woman playing with aspeaking, low motion. ball and a man spinning apen continuously, highmotion.Frame Size CIF (352x288) CIF (352x288)Duration(second) 30 35Number of frames 900 1065 in Section VI-A. To understand the impact of any of theseparameters on β , the values of the other two parameters(controlling parameter) were kept fixed. The values of thecontrolling parameters for both sequences are shown in Fig.1, 2 and 3. These figures also show the relationship between β and each of C l , n and QoE respectively.Equation (10) shows the mathematical relationship betweenthe four parameters. However, in this paper we focus on avalue of β that produces excellent quality (MOS=5) only. Thus QoE was not considered as a variable in the proposed modelof β . The exponential relationship between β and QoE shownin Fig. 3 will be included to the model of β in future studiesto provide multi-class MOS. β ∝ QoE, C l n (10)The simulation data was analyzed with 2-way repeatedanalysis of variance (ANOVA) [46] to confirm the significanceof C l and n in modeling of β . Also, it can find the differencebetween means given by the remaining two parameters C l and n . ANOVA let us understand the effect of parameters andtheir interaction on β which will later be used in the regressionmodeling. The ANOVA results are shown in Table III for Fand p-values: the Cumulative Distribution Function (CDF) ofF. Parameter with (p < β . The analysis results indicate that β is affectedby each of C l and n as p-values are 0 and 0.0023 respectively.The result also shows that the combined parameters have nointeraction effect on β because the p-value is 0.6249. This canbe justified by the fact that n is determined by C l ; the highercapacity of the link, the more sessions are accepted. Based onthe value of p in the table, we can conclude that β is affectedmore by C l than by n .The relationship between β , n and C l can be establishedfrom ANOVA analysis and Fig. 1 and 2. We found that thereis a linear relationship between β and C l and a polynomialrelationship between β and n . Finally, the rational modelshown in Equation (11) was formulated to estimate the value of β from the nonlinear regression analysis of the simulation datausing MATLAB. The values of the coefficients of Equation(11) are listed in Table IV and V. As n is determined by thesize of video frames (content dependent), different values forthe model coefficients were found for slow ( MAD sequence)and fast (
Paris sequence) moving contents. The table alsoshows the correlation coefficient (R ) and Root Mean SquaredError (RMSE) of the proposed model for both contents. Fig. 1. β - Link capacity relationshipFig. 2. β - Number of sessions relationshipFig. 3. β - QoE relationship β = α + ( C l δ ∗ n ) (11) TABLE IIIANOVA R
ESULTS FOR M AIN AND I NTERACTION E FFECTS
Source Sum of Degree of Mean F p-Valuesquares freedom SquaresC l l *n 0.00047 2 0.00023 0.51 0.6249TABLE IVC OEFFICIENTS OF β PREDICTION MODEL AND MODEL VALIDATIONCORRELATION COEFFICIENTS - SLOW MOVING CONTENT ( MAD
VIDEOSEQUENCE ) α δ -0.5429 0.9689Adjusted R (Validation) %88.44RMSE (Validation) 0.0149TABLE VC OEFFICIENTS OF β PREDICTION MODEL AND MODEL VALIDATIONCORRELATION COEFFICIENTS - FAST MOVING CONTENT ( Paris
VIDEOSEQUENCE ) α δ -0.1227 1.952Adjusted R (Validation) %90.54RMSE (Validation) 0.0124 The model for β was proposed based on two video se-quences ( MAD and
Paris ), however the methodology is similarand applies to faster moving content, such as sports, of thesame format (CIF). Thus, the model is limited to the videoformat and coding parameters specified in Table II. The modelcan be applied to other formats and coding parameters withdifferent coefficient values. This is because other formatsand/or coding parameters generate different frame sizes andbit rates which control the number of sessions (parameter n in the model) for a specific link capacity (parameter C l in themodel). They only have impact on the value of the coefficientsof the model. The model will be validated by CIF and QCIFvideo formats in Section VIII.VI. R ESULTS AND A NALYSIS
The simulation environment is explained in Section VI-A.Section VI-B compares the proposed
Pro-IBMAC to CBAC interms of MOS and number of the sessions, packet drop ratioand delay. The impact of β on the functionality of Pro-IBMAC is discussed in Section VI-C.
A. Simulation Setup
Since the number of admitted sessions for a specific linkcapacity is the target of this study, only acceptance/rejectionpolicy of admission control was investigated. Queue size andsimulation time were chosen so as not to cause drop due toinsufficient queue length or time. The video format such as CIFor QCIF has impact on the number of admitted sessions due tothe difference in the size of encoded frames. In this paper, CIF(352x288) is assumed for input video as an acceptable video format for most video capable devices such as handsets andmobiles [45]. It is also suitable for videoconferencing systemsdelivered on telephone lines. While modern devices supportmuch higher resolution, CIF makes packet level simulationpractical. A bottleneck link of dumbbell topology over whichthe video sources send to their peer destinations was con-sidered for the implementation of the proposed
Pro-IBMAC scheme. In addition to β , C l was the main variable in thesimulation. Other parameters such as link delay, queue length,packet size were kept fixed. Lost packets were replaced with0 by the etmp4 [39] decoder as a way for coping with thelosses. The values of the simulation parameters and settingsare shown in Table VI. TABLE VIE
NCODER AND N ETWORK S ETTINGS
Parameter ValueEncoder Frame size CIF(352x288)Frame rate 30fpsGroup of picture 30Network C l (Mbps) 22, 24, 30, 36, 39, 40Topology DumbbellPacket size(byte) 1024UDP header size(byte) 8IP Header Size(byte) 20Queue size(packet) 5300Queue management algorithm DroptailQueue discipline FIFO(First In First Out)Simulation time(second) 500 New sessions were requested randomly and continuouslyevery second. They were accepted as long as there was enoughbandwidth on the bottleneck link i.e: Equation (9) was satis-fied. NS-2 [47] was used to measure
CalR( τ ) and Pro-IAAR and implement
CBAC and
Pro-IBMAC . The implementationof the proposed
Pro-IBMAC is summarized in Algorithm 1.
Algorithm 1
Proposed
Pro-IBMAC
Given C l , x new and n for Every video session request do Compute µ S ( t ) f rom Equation (5) Compute β f rom Equation (11)
Compute P ro - IAAR ( t ) f rom Equation (6) if Equation (9) =
T rue then
Request accepted else
Request rejected end ifend for
The time window τ has an impact on the operation ofthe admission control. The smaller the value of τ , the moreconservative the admission control and more sensitive to thetraffic bursts. On the other hand, the larger the value of τ , thesmoother the measured rate and less reactive to the changesin the network load. In practice, τ will be a few seconds [48].In this paper, IAAR(t) was averaged over 1-second.The
MAD video sequence described in Table II was fed tothe NS-2 simulator using EvalVid [39]. Evalvid provides a setof tools to analyze and evaluate video quality by means ofPSNR and MOS metrics. The Evalvid MOS metric (We call it simulated MOS) was used in this paper which calculatesthe average MOS value of all frames for the entire video witha number between 1 and 5, instead of the frame-wise PSNRmetric. The MOS metric represents the impression of the userfor the entire received video and has been widely used byresearch community [49]–[60]. Although the MOS metric doesnot map very well to the subjective impression for a longvideo sequence, it was used for short video sequences (30-35 seconds) in this paper. In addition to the MOS metric,we calculated the Distortion In Interval (DIV) metric [39] torestrict the MOS metric within a fixed interval (30 framesin this paper). This stringent metric calculates the maximumpercentage of received frames with a MOS smaller than thatof the sent frame within a given interval.The efficiency of the proposed
Pro-IBMAC and
CBAC wasevaluated based on MOS, number of sessions, packet dropratio and mean delay. These performance metrics were chosendue to their impact on multimedia traffic. The performance of
Pro-IBMAC was tested to find the maximum number of videosessions on a bottleneck link while keeping the QoE of eachsession at acceptable or required levels. This was comparedto other procedures such as
CBAC . The objective was to seehow
Pro-IBMAC utilizes the available bandwidth compared to
CBAC . Further simulations were used to investigate the effectof parameter β on the performance metrics. B. Pro-IBMAC vs CBACIt has been found that there is a considerable differencebetween the two schemes in terms of the number of acceptedsessions. This is plotted in Fig. 4. The number of admitted ses-sions is always higher for
Pro-IBMAC . The difference betweenthe number of admitted sessions increases with increasingof the link capacity. For example, the number of admittedsessions to 22Mbps link is 15 against 14 for
Pro-IBMAC and
CBAC respectively, whereas it is 30 against 25 in the case of40Mbps link. The main role of any admission control is toensure that the acceptance of a new session does not violatethe QoE of on-going sessions. We computed the MOS ofevery single accepted session for both schemes. We found thatincrease in n does not come at the cost of QoE as all acceptedsessions by Pro-IBMAC and
CBAC were scored MOS 5. Notethat the MOS of video sessions is labeled on the secondaryy-axis in the figure. The value of β that produces this increasein n and guarantees the video quality is also shown in thefigure. This will be further described in Section VI-C. Howeverthis simulation outcome can not be generalized. Pro-IBMAC may not guarantee the same level of QoE as
CBAC in a realimplementation. This is because our proposed scheme is basedon a probabilistic approach therefore, there is a possibility ofthe upper bound to be lower than the bursty instantaneous rate,especially for small τ .Table VII shows mean MOS and DIV. The DIV values (0%)indicate that all received frames have the same MOS as ofthe original frames. It also lists the packet drop ratio of theaccepted sessions of Pro-IBMAC and
CBAC for each link.Since we aim at a β value that doesn’t degrade the MOS ofreceived videos, as mentioned in Section V-B, no packet dropwas expected. +-
22 24 30 36 39 4005101520253035 0510152025Pro-IBMAC CB MOS C l (Mbps) n M O S β =0.96 β =0.95 β =0.94 β =0.84 β =0.83 β =0.87 AC Fig. 4. MOS of the
CBAC and
Pro-IBMAC admitted sessionsTABLE VIIP
ACKET DROP RATIO AND ADMITTED SESSIONS OF
Pro-IBMAC
AND
CBACPro-IBMAC CBAC C l (Mbps) PacketDrop MOS DIV β n n % %22 0 5 0 0.96 15 1424 0 5 0 0.95 17 1530 0 5 0 0.94 21 1936 0 5 0 0.87 26 2339 0 5 0 0.84 29 2540 0 5 0 0.83 30 25 As for the delay, we measured the mean delay using thens-2 trace files for both schemes. Fig. 5 illustrates the CDFof the mean delay for the
Pro-IBMAC and
CBAC sessionsfor 40Mbps link. As shown in Table VII, 30 sessions areaccepted by
Pro-IBMAC for β =0.83 and 25 by CBAC . Moresessions on the same link by
Pro-IBMAC caused higher delaydue to more buffering. Therefore the
Pro-IBMAC sessionsexperienced higher delay compared to the lower delay of the
CBAC sessions. Nevertheless, increase in the delay that comesat the cost of the optimization of QoE-Session can not betolerated by real-time video traffic. For
Pro-IBMAC to beapplicable to realtime traffic, a proper value of β must beselected. Video streaming services can tolerate a delay of 5-seconds [61], thus it can be used within this limit. In futurework, we will further investigate this relationship and developthe model of β to include delay as another variable. C. The Impact of β on Pro-IBMACAs mentioned earlier, parameter β controls the degree ofrisk between the admission decision and QoE of existing ses-sions. Fig. 6 shows IAAR(t) (dash-dot line) and the upper limitof the exceedable aggregate rate (solid line) that allows moresessions (compared to sessions allowed by
IAAR(t) ), withoutQoE degradation of enrolled video sessions. The proposedvalue of β for four scenarios (22, 30, 36 and 40Mbps) is shownin the figure. It can be seen that the lower the value of β , thewider the gap between the two rates. Decreasing β causesincrease in the limit of the exceedable rate. This makes Pro-IBMAC more flexible and it accepts more sessions. This canbe better observed in Fig. 7. It depicts the number of admitted
Fig. 5. CDF of the mean delay of the
CBAC and
Pro-IBMAC sessions sessions for different link scenarios. The solid line shows thenumber of sessions admitted by
CBAC , while the other threelines show sessions admitted by
Pro-IBMAC for three differentvalue of β (0.9, 0.85 and 0.78). For the same link, the linearrelationship between n and C l allows more sessions to beaccepted by lowering the value of β . For instance, for 39Mbpslink, Pro-IBMAC accommodates 27, 28 and 30 sessions for β =0.9, 0.85 and 0.78 respectively compared to 25 sessions ofCBAC. Note that β ≥ Fig. 6. IAAR and upper limit of the exceedable rate for different linkcapacities
However, continuous decreasing of β will degrade the QoEof admitted sessions as more sessions are accepted. Therefore,care is required to fine tune the value of β that optimizesthe operation of Pro-IBMAC . The aim is to accept as manysessions as possible, while keeping the QoE of the sessionsat required levels. As per the proposed model, the valueof β depends on C l , n and required QoE . We investigatedthis further for 22Mbps and 24Mbps links. Fig. 8 shows the
Fig. 7. Admitted sessions of
CBAC and
Pro-IBMAC for different linkcapacities number of MOS 2, 3, 4 and 5 sessions separately as well astotal number of sessions for 22Mbps link. If we consider thatthe required class of QoE is MOS 5, then the proposed valueof β is 0.96, i.e. for β less than 0.96, sessions with multi-MOSlevels exist, while for β ≥ β from 0.96to 0.5 increases the total number of sessions and number ofMOS 3 and 2 sessions while decreasing the number of MOS5 and 4 sessions. Fig. 8. Impact of β on MOS and n , C l =22Mbps In another scenario, we found that the proposed value of β is0.95 for 24Mbps link as shown in Fig. 9. β of 0.95 or greater,maintains the MOS of accepted sessions at 5, while β less than0.95 produces sessions with multi-MOS classes which comesat the cost of the QoE of enrolled sessions. For instance, β of0.8 creates 18 sessions with MOS 4 and 1 session with MOS3. Whilst β of 0.6 changes the number of MOS 4 sessions to5 and MOS 3 sessions to 19 . Note that there are 19 sessionsin total for β =0.8 and 24 sessions for β =0.6. Fig. 8 and 9 also show the DIV values of accepted sessionsat different β values. As the DIV was 0% for sessions withMOS 5 and between 0% and 100% for sessions with MOS < < DIV <
100 denote thatMOS of sessions are less than 5.
Fig. 9. Impact of β on MOS and n , C l =24Mbps Although most real-time applications can tolerate somepacket loss, more than an acceptable level may degrade thequality of received video. As expected, fewer sessions of
CBAC will guarantee no packet loss, in contrast extra addedsessions of
Pro-IBMAC cause packet drop when β is setlower than the proposed value and increases slightly with theincrease of the number of sessions. Table VIII presents thepercentage of the packet drop ratio of the Pro-IBMAC admittedsessions for different value of β for 22Mbps link. The ratioincreases with the decrease of β due to fitting a higher numberof sessions into the same link. The table shows 0.45%, 4.06%and 6.70% drop of the total number of packets for β = 0.89,0.85 and 0.78 respectively. The proposed value of β (0.96)ensures that no packets are dropped as shown in the table. TABLE VIIIP
ACKET DROP RATIO AND ADMITTED SESSION OF
Pro-IBMAC
FORDIFFERENT β , Cl = 22M BPS β Packet Drop Ratio % n Improper values of β not only causes packet drop, but italso degrades the MOS levels (discussed earlier) and increasesthe delay. Fig. 5 demonstrates how a high number of sessionscaused by a low value of β can contribute to the increase of thedelay which can be substantial for a large number of sessions.The disadvantage of lowering the value of β is not onlythat it causes degradation to the MOS level of video sessions,or increase in the delay and packet loss. We observed thatthe decoder takes longer to decode and play back the received video for low value of β , for instance when β =0.6 for 24Mbpslink. The ISP can tune the value of β to control the trade-offbetween providing the required level of QoE and increasingtheir revenue by accommodating as many user sessions aspossible. VII. S UBJECTIVE T ESTS
We performed subjective tests to involve human subjectsin rating the quality of videos. The tests followed the ITU-RBT.500-13 recommendation [62]. The five-grade scale from1 to 5 of the Single Stimulus (SS) adjectival categoricaljudgment method was used in which 1 represents ’bad’ and5 represents ’excellent’ quality. Each video was presented inrandom order and rated individually by 17 subjects one ata time. The number of participants exceeded the minimumrecommended number (15 subjects).As the
MAD sequence was chosen, 48 videos deliveredthrough different link capacities and different values of β shown in Table VII, Fig. 8 and Fig. 9 were used in the tests.They were decoded from the simulations and selected fromFig. 4 (MOS 5), Fig. 8 (MOS 2, 3, 4 and 5) and Fig. 9 (MOS3, 4 and 5). The description of the testing video sequence,coding and network parameters were the same as describedin Tables II and VI. Each video was identified by the MOSvalue calculated with Evalvid, regardless of the capacity of thelink and/or value of β . The aim was to have variety of videoswith different MOS values through changing the capacity ofthe link and value of β . The simulated β and predicted β ofthe testing videos will be plotted in Section VIII.The videos were presented in their original size (352x288),embedded in a separate web page with grey backgroundand rated on the same page. There were two sessions, eachlasting up to 30 minutes with 10 minutes break in between.To stabilize the subjects’ opinion, five dummy videos weredisplayed at the beginning of the session without consideringtheir scores. Prior to the actual rating, the subjects werecarefully introduced to the assessment method, likely qualityartifacts that might be observed, rating scale and timing. Theywere given unrestricted time and the viewing distance wascomfortable.The tests were conducted in a white background laboratoryon 29 inch LCD monitor (Dell P2213) with 1680x1050resolution and 32 bit true color. 5 female and 12 male non-expert observers participated in the tests. All participants wereuniversity students, 1 in the range of 18-25, 7 in the range of26-30 and 9 over 30. At the end of the tests, subjects who weresurveyed on the duration and comfortability of the tests didnot express any concern. The subjects were screened for anypossible outliers following the screening procedure of the SSmethod [62]. Two subjects have been eliminated and their datawere not considered in the analysis. The MOS was calculatedby taking the mean score for each of the videos following theprocedure described in [62].The bar chart in Fig. 10 illustrates the subjective mean MOSof every presented video with the confidence interval. It showsthe mean and range (the upper and lower limits) of MOS givento each video by the subjects. The analysis shows that around
40% of the scores went for a MOS of 3.5. The distribution ofthe scores is plotted in Fig. 11.
Fig. 10. Bar chart of the subjective MOS with confidence interval forindividual videoFig. 11. Bar chart of the percentage of scores of the subjective MOS
VIII. V
ALIDATION OF THE P ROPOSED M ODEL
In this section, the validation of the proposed model of β with simulation results is explained. It also demonstrates thevalidation of the simulated MOS with subjective MOS.The scatter plot in Fig. 12 shows the simulated MOS againstsubjective MOS. Overall, the subjects were irritated by videoimpairments, their scores therefore underestimate the simula-tion scores. The majority of simulated MOS scores seen arehigher than subjective MOS. However, both scores are gettingcloser for less impaired videos (subjective MOS between 4.78-5). These videos were delivered with the proposed valuesof β for each value of C l . Note that as there are about 11overlapping scores within this range, all can not be seen inthe figure. Overlapping of the scores can be further noticed inFig. 10, in which there are 11 scores in the range of 4.78-5. The relationship is nearly linear correlated for videos deliveredwith the proposed value of β that have MOS close to 5. Thisindicates that the model can provide better quality for endusers with the proposed value of β . Fig. 12. Validation of the simulated MOS with subjective MOS β predicted by the model (Equation 11) has been validatedby the one found by simulations. Fig. 13 shows the resulting β ’s scatter point plot of the predicted β against simulated β for slow and fast moving contents separately. As shown inTables IV and V, the model of β suits fast moving contentwith a correlation coefficient of 90.54% compared to 88.44%for slow moving content. This can be also observed in Fig. 13.Thus, the model best suits dynamic content with high variationin bitrate. Note that there were few videos for each value of β plotted in the figure, therefore the number of plotted pointsis less than the number of the testing videos (48). Fig. 13. Validation of the proposed model of β with simulation results As mentioned in Section V-B, the model of β can be appliedto other video formats with different values of coefficients α and δ . It has been validated by QCIF video format usingthe 45-seconds Deadline video sequence of 1374 frames. The model achieved an adjusted R of 83.59% and RMSE of0.0194. The values of α and δ were -0.1323 and 0.4991respectively. IX. C ONCLUSION
We proposed a novel algorithm to find the upper limit of thevideo total rate that can exceed a specific link capacity withoutQoE degradation of ongoing video sessions. A mathematicalmodel for the measurement algorithm was developed andimplemented in an admission control system. Its performancehas been validated by simulating publicly available videosequences and subjective tests. The exceedable limit has beendefined by parameter β in the algorithm. This parametercan be used by ISPs to balance the trade-off between QoEand the number of video sessions. The simulation resultshave shown that the proposed admission control compared tocalculated rate-based admission control optimizes the trade-off relationship between QoE-Session through fine tuning thevalue of β . The proposed algorithm can be applied within thescope of the video format and coding parameters specified inthis paper. In future work, we will further develop the modelof β to include delay as another variable. The calculated MOSwill be compared with SSIM metric. Moreover, an implemen-tation of the proposed scheme in a cross-layer architecture foroptimizing the QoE of video session will be investigated.X. A CKNOWLEDGMENT
The authors would also like to thank volunteers for their par-ticipation in the subjective tests and the anonymous reviewersfor their constructive comments.R
EFERENCES [1] Cisco Documentation, “Cisco visual networking index:Forecast and methodology 2013-2018,” Cisco, Jun. 2014.[2] S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, andW. Weiss,
An Architecture for Differentiated Services ,IETF RFC 2475, Dec. 1998.[3] ITU-T Document FG IPTV-IL-0050, “Definition of Qual-ity of Experience (QoE),” ITU-T, Jan. 2007.[4] L. Breslau, S. Jamin, and S. Shenker, “Comments on theperformance of measurement-based admission control al-gorithms,” in
INFOCOM 2000. Nineteenth Annual JointConference of the IEEE Computer and CommunicationsSocieties. Proceedings. IEEE , vol. 3, 2000, pp. 1233–1242.[5] J. Auge, S. Oueslati, and J. Roberts, “Measurement-based admission control for flow-aware implicit servicedifferentiation,” in
Teletraffic Congress (ITC), 2011 23rdInternational , 2011, pp. 206–213.[6] M. Menth and L. Lehrieder, “Performance of PCN-based admission control under challenging conditions,”
Networking, IEEE/ACM Transactions on , vol. 20, no. 2,pp. 422–435, Apr. 2012.[7] R. J. Gibbens and F. P. Kelly, “Measurement-based con-nection admission control,” in , Jun. 1997. [8] S. Floyd, “Comments on measurement-based admissionscontrol for controlled-load services,” Tech. Rep., 1996.[9] S. Qadir and A. A. Kist, “Video-aware measurement-based admission control,” in
Telecommunication Net-works and Applications Conference (ATNAC), 2013 Aus-tralasian , 2013, pp. 178–182.[10] J. Qiu and E. Knightly, “Measurement-based admissioncontrol with aggregate traffic envelopes,”
Networking,IEEE/ACM Transactions on , vol. 9, no. 2, pp. 199–210,Apr. 2001.[11] S. Y. Nam, S. Kim, and W. Park, “Analysis of minimalbacklogging-based available bandwidth estimation mech-anism,”
Computer Communications , vol. 35, no. 4, pp.431–443, 2012.[12] C. D. Guerrero and M. A. Labrador, “On the applicabilityof available bandwidth estimation techniques and tools,”
Computer Communications , vol. 33, no. 1, pp. 11–22,2010.[13] S. Y. Nam, S. J. Kim, S. Lee, and H. S. Kim, “Estimationof the available bandwidth ratio of a remote link or pathsegments,”
Computer Networks , vol. 57, no. 1, pp. 61–77, 2013.[14] R. Lubben, M. Fidler, and J. Liebeherr, “Stochasticbandwidth estimation in networks with random service,”
Networking, IEEE/ACM Transactions on , vol. 22, no. 2,pp. 484–497, Apr. 2014.[15] B. Cavusoglu and E. A. Oral, “Estimation of availablebandwidth share by tracking unknown cross-traffic withadaptive extended Kalman filter,”
Computer Communi-cations , vol. 47, no. 0, pp. 34–50, 2014.[16] S. Jamin, S. Shenker, and P. Danzig, “Comparisonof measurement-based admission control algorithms forcontrolled-load service,” in
INFOCOM ’97. SixteenthAnnual Joint Conference of the IEEE Computer andCommunications Societies. Driving the Information Rev-olution., Proceedings IEEE , vol. 3, 1997, pp. 973–980.[17] A. Nevin, Y. Jiang, and P. Emstad, “Robustness study ofMBAC algorithms,” in
Computers and Communications,2008. ISCC 2008. IEEE Symposium on , Jul. 2008, pp.1040–1046.[18] D. Ammar, T. Begin, I. Guerin-Lassous, and L. Noirie,“Evaluation and comparison of MBAC solutions,” in
Local Computer Networks (LCN), 2011 IEEE 36th Con-ference on , Oct. 2011, pp. 215–218.[19] ——, “KBAC: Knowledge-based admission control,” in
Local Computer Networks (LCN), 2012 IEEE 37th Con-ference on , Oct. 2012, pp. 537–544.[20] S. Lima, P. Carvalho, and V. Freitas, “Admission controlin multiservice IP networks: Architectural issues andtrends,”
Communications Magazine, IEEE , vol. 45, no. 4,pp. 114–121, 2007.[21] A. Nevin, “The uncertainty of decisions in measurementbased admission control,” Ph.D. dissertation, NorwegianUniversity of Science and Technology, Nov. 2010.[22] A. W. Moore, “Measurement-based management of net-work resources,” Ph.D. dissertation, University of Cam-bridge, Apr. 2002.[23] S. Y. Nam, S. Kim, and D. K. Sung, “Measurement- based admission control at edge routers,” Networking,IEEE/ACM Transactions on , vol. 16, no. 2, pp. 410–423,2008.[24] X. Zhang, Y. Xu, H. Hu, Y. Liu, Z. Guo, and Y. Wang,“Modeling and analysis of Skype video calls: Rate con-trol and video quality,”
Multimedia, IEEE Transactionson , vol. 15, no. 6, pp. 1446–1457, Oct. 2013.[25] Y. Xu, C. Yu, J. Li, and Y. Liu, “Video telephony forend-consumers: Measurement study of Google+, iChat,and Skype,”
Networking, IEEE/ACM Transactions on ,vol. 22, no. 3, pp. 826–839, Jun. 2014.[26] N. Chendeb Taher, Y. Ghamri Doudane, B. El Hassan,and N. Agoulmine, “Towards voice/video applicationsupport in 802.11e WLANs: A model-based admissioncontrol algorithm,”
Computer Communications , vol. 39,pp. 41–53, 2014.[27] G. Lee, H. Kim, Y. Cho, and S.-H. Lee, “QoE-awarescheduling for Sigmoid optimization in wireless net-works,”
Communications Letters, IEEE , vol. 18, no. 11,pp. 1995–1998, Nov. 2014.[28] E. Eardley P., “Pre-Congestion Notification (PCN) archi-tecture,” IETF, RFC 5559, Jun. 2009.[29] S. Latre and, K. Roobroeck, T. Wauters, and F. Turck,“Protecting video service quality in multimedia accessnetworks through PCN,”
Communications Magazine,IEEE , vol. 49, no. 12, pp. 94–101, Dec. 2011.[30] S. Latr´e, B. De Vleeschauwer, W. Van De Meerssche,K. De Schepper, C. Hublet, W. Van Leekwijck, andF. De Turck, “PCN based admission control for auto-nomic video quality differentiation: Design and evalua-tion,”
J. Netw. Syst. Manage. , vol. 19, no. 1, pp. 32–57,Mar. 2011.[31] Y. Xu, J. Deng, and M. Nowostawski, “Quality of servicefor video streaming over multi-hop wireless networks:Admission control approach based on analytical capacityestimation,” in
Intelligent Sensors, Sensor Networks andInformation Processing, 2013 IEEE Eighth InternationalConference on , 2013, pp. 345–350.[32] B. Hamdaoui and P. Ramanathan, “A cross-layer admis-sion control framework for wireless ad-hoc networks us-ing multiple antennas,”
Wireless Communications, IEEETransactions on , vol. 6, no. 11, pp. 4014–4024, Nov.2007.[33] K. Piamrat, A. Ksentini, C. Viho, and J. Bonnin, “QoE-aware admission control for multimedia applications inIEEE 802.11 wireless networks,” in
Vehicular TechnologyConference, 2008. VTC 2008-Fall. IEEE 68th , Sep. 2008,pp. 1–5.[34] Y. Jiang, P. Emstad, A. Nevin, V. Nicola, and M. Fidler,“Measurement-based admission control for a flow-awarenetwork,” in
Next Generation Internet Networks, 2005 ,2005, pp. 318–325.[35] R. Wojcik, J. Domzal, and A. Jajszczvk, “Enhancedmeasurement-based admission control for flow-awarenetworks,” in
Computing, Networking and Communica-tions (ICNC), 2013 International Conference on , 2013,pp. 922–926.[36] P. Brooks and B. Hestnes, “User measures of quality of experience: Why being objective and quantitative isimportant,”
Network, IEEE , vol. 24, no. 2, pp. 8 –13,Mar.-Apr. 2010.[37] S. Latre, P. Simoens, B. D. Vleeschauwer, W. V.de Meerssche, F. D. Turck, B. Dhoedt, P. Demeester,S. V. den Berghe, and E. G. de Lumley, “An autonomicarchitecture for optimizing QoE in multimedia accessnetworks,”
Computer Networks , vol. 53, no. 10, pp.1587–1602, 2009.[38] J.-R. Ohm,
Multimedia Communication Technology .Springer, 2004, vol. 1.[39] J. Gross, J. Klaue, H. Karl, and A. Wolisz, “Cross-layeroptimization of OFDM transmission systems for MPEG-4 video streaming,”
Computer Communications , vol. 27,no. 11, pp. 1044–1055, Jul. 2004.[40] V. Q. E. Group, “VQEG multimedia project: Final re-port,” 2008.[41] Z. Wang, A. Bovik, H. Sheikh, and E. Simoncelli, “Imagequality assessment: from error visibility to structuralsimilarity,”
Image Processing, IEEE Transactions on ,vol. 13, no. 4, pp. 600–612, Apr. 2004.[42] W. Cherif, A. Ksentini, D. Negru, and M. Sidibe, “APSQA: Efficient real-time video streaming QoE tool in afuture media internet context,” in
Multimedia and Expo(ICME), 2011 IEEE International Conference on , Jul.2011, pp. 1–6.[43] K. Piamrat, C. Viho, J. Bonnin, and A. Ksentini, “Qualityof experience measurements for video streaming overwireless networks,” in
Information Technology: NewGenerations, 2009. ITNG ’09. Sixth International Con-ference on , Apr. 2009, pp. 1184–1189.[44] W. Hoeffding, “Probability inequalities for sums ofbounded random variables,”
Journal of the AmericanStatistical Association , vol. 58, no. 301, pp. 13–30, Mar.1963.[45] A. Khan, L. Sun, and E. Ifeachor, “QoE predictionmodel and its application in video quality adaptation overUMTS networks,”
Multimedia, IEEE Transactions on ,vol. 14, no. 2, pp. 431–442, Apr. 2012.[46] R. G. Miller and B. W. Brown,
Beyond ANOVA: basicsof applied statistics , 1st ed, Ed. Chapman & Hall, 1997.[47] NS-2,
The Network Simulator
Wireless Communications, IEEE Transactions on , vol. 9,no. 1, pp. 338–347, Jan. 2010.[50] A. Khan, L. Sun, and E. Ifeachor, “Learning modelsfor video quality prediction over wireless local areanetwork and universal mobile telecommunication systemnetworks,”
Communications, IET , vol. 4, no. 12, pp.1389–1403, Aug. 2010.[51] D. Kim and K. Chung, “A network-aware quality adap-tation scheme for device collaboration service in homenetworks,”
Consumer Electronics, IEEE Transactions on , vol. 58, no. 2, pp. 374–381, May 2012.[52] A. Khan, L. Sun, and E. Ifeachor, “Content clusteringbased video quality prediction model for MPEG4 videostreaming over wireless networks,” in Communications,2009. ICC ’09. IEEE International Conference on , Jun.2009, pp. 1–5.[53] F. Tommasi, C. Melle, and V. D. Luca, “OpenSatRelay-ing: a hybrid approach to real-time audio-video distri-bution over the Internet,”
Journal of Communications ,vol. 9, no. 3, pp. 248–261, 2014.[54] P. Papadimitriou and V. Tsaoussidis, “SSVP: A con-gestion control scheme for real-time video streaming,”
Computer Networks , vol. 51, no. 15, pp. 4377–4395,2007.[55] A. Khan, L. Sun, and E. Ifeachor, “Content-based videoquality prediction for MPEG4 video streaming overwireless networks,”
Journal of Multimedia , vol. 4, no. 4,2009.[56] X. Ma, F. Li, F. Hu, and X. Liu, “A hybrid channelassignment strategy to QoS support of video-streamingover multi-channel ad hoc networks,”
Journal of Systemsand Software , vol. 85, no. 2, pp. 300–308, 2012, spe-cial issue with selected papers from the 23rd BrazilianSymposium on Software Engineering.[57] A. C. C. Aguiar, “Multi-user multi-flow packet schedul-ing for wireless channels,” Ph.D. dissertation, TechnichenUniversit Berlin, 2008.[58] M. Erdelj, “Mobile wireless sensor network architectureapplications to mobile sensor deployment,” Ph.D. disser-tation, Universit des Sciences et Technologie de Lille,Dec. 2013.[59] K. Tan, “Design and implementation of spectrum-awarewireless multimedia communication system,” Ph.D. dis-sertation, University of California, 2013.[60] P. J. P. Escuer, “Analysis and evaluation of in-homenetworks based on HomePlug-AV power line commu-nications,” Ph.D. dissertation, Universidad Politcnica deCartagena, 2014.[61] T. Szigeti and C. Hattingh,
End-to-End QoS NetworkDesign: Quality of Service in LANs, WANs, and VPNs(Networking Technology) . Cisco Press, 2004.[62] ITU-R Recommendation BT.500-13,