Académique Documents
Professionnel Documents
Culture Documents
Proceedings of MoMM2008
Antonio Liotta
Department of Electrical Engineering Eindhoven University of Technology, P.O.Box 513 5600 MB Eindhoven, The Netherlands
fagbom@essex.ac.uk ABSTRACT
The streaming of multimedia contents (e.g., Mobile TV) is a bandwidth intensive service. The network operators aim is to provide an acceptable user experience at minimal network resource usage. It is important from the network operators perspective to be aware of: 1) the thresholds at which the users perception of service quality becomes unacceptable; and 2) the degree of influence of each of the Quality of Service (QoS) parameters on the user perception. However, very little is known about the formal methods to optimize the use of QoS mechanisms in relation to the users Quality of Experience (QoE). In this paper, we explain how the users QoE can be captured. A statistical modelling technique is employed which, correlates QoS parameters with estimates of QoE perceptions and identifies the degree of influence of each QoS parameters on the user perception. The network operator can apply this information to efficiently, and accurately undertake network dimensioning and service provisioning strategies. This proposed methodology is applied to demonstrate QoE management strategies, thus paving the way towards QoE-aware QoS management.
a.liotta@tue.nl
services, there are several parameters, which affect the perceptual visual quality of the displayed multimedia content. These parameters can be grouped under Application-level Quality of Service (AQoS) - which deals with parameters such as content resolution, frame rate and codec type; and Network-level Quality of Service (NQoS) - which deals with parameters such as bandwidth, jitter and packet loss. Traditionally, content delivery systems primarily focus on managing the network parameters, but fail to consider the users perception of service quality at the application level. Although the maximum bitrate may be fixed at a certain value by the access network, this does not necessarily fix the quality of a multimedia content at say high, for example. This is because there are numerous ways the multimedia content could have been encoded. A QoE-aware QoS management process aims at maximizing the user experience whilst at the same time, reducing network costs. However, very little work has been done on the formal methods to be used in optimizing the use of QoS mechanisms in relation to the QoE achievable at the application level. At this level, QoS is driven by human perception of audio and video. Understanding the user requirements at the application level could allow for better control mechanisms that will further lead to better network resource utilization, and increasing user satisfaction. This work extends previous findings reported in [1], [2], [3], and [4], which focused mainly on the collection of QoE data from subjective studies. Herein, we propose a new QoE management methodology, which shows how QoE data may be used for the benefit of network operators and service providers. The paper is organized as follows: Section 2 discusses related works in the fields of QoS management and the users perception of service quality. Section 3 introduces our QoE framework and details the modelling technique employed to perform QoE management. The results obtained from the statistical modelling technique and the QoE management strategies are presented in Section 4. Final remarks are given in Section 5.
General Terms
Management.
Keywords
QoE, QoS, QoE management.
1. INTRODUCTION
The ability to manage service quality has become an essential part of the service delivery chain. The users expectations of service quality have increased over the years, and it is particularly difficult to meet their requirements in the context of multimedia service provisioning. In multimedia streaming
2. RELATED WORKS
Traditional QoS management strategies generally focus on allocating network resources to requested services, with minimal consideration of the user perception of service quality. This often leads to the resource over-provisioning, which can be very expensive. Without considering the users perception of service quality, it cannot be guaranteed that the user will be satisfied with the service. There are a few studies on QoE [21], [26] and QoE frameworks [24] [6] in the public domain. The authors in [21] presented a model for measuring the QoE of mobile multimedia services, taking into account both measurable and non-measurable parameters. The measurable parameters were those related to the
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. MoMM 2008, November 2426, 2008, Linz, Austria. Copyright 2008 ACM 978-1-60558-269-6/08/0011 $5.00.
111
Proceedings of MoMM2008
MoMM 2008
technological aspects of the service, whilst the non-measurable parameters were related to the users perception of the service, their expectations, and behavior. The aim in [21] was to produce quantifiable quality metrics for QoE evaluation. To obtain further understanding of the QoE aspect of multimedia streaming, studies have been carried out at the application and network levels. The studies at the application level mainly focused on the effects of encoding parameters on perceptual quality. Apteker et al [5] studied the effects of frame rate on the users acceptability of video quality. The influences of audio and video on the perceived audiovisual quality are given in [14], [27] and [11]. The integration of audio and video quality on perceived audiovisual quality tends to be content dependent [11], where for less complex scenes (e.g. head and shoulder content), audio quality is slightly weighted higher than video quality. By contrast, for high motion content, video quality is weighted significantly higher than audio quality. Other authors [15], [7] and [25] have focused on how network-level parameters (e.g., packet loss, delay, jitter and bandwidth) affect the users perceived quality of the multimedia content. Adding a playout buffer at the receiver can mitigate the effects of jitter and end-to-end delay in multimedia streaming services. Buffering mechanisms allow for a smooth viewing experience but incur additional start-up delays to the streaming application. Depending on the size of the buffer, the playout delay is typically between 5 and 15 seconds which is acceptable for most video streaming applications [23]. Playout buffers also make the retransmission of lost packets possible. Packet loss tends to be the dominant factor that affects perceptual multimedia quality. It has been found that jitter is nearly as important as packet loss in influencing perceptual video quality [7]. Studies by [16], [19], and [10] focused on the combined effects that the AQoS and NQoS parameters had on perceived video quality. Findings obtained from the literature review provide an indication as to the factors, which have major impacts on the users QoE. The main contribution of this paper is a practical demonstration on how to control AQoS parameters for QoE management.
Figure 1. QoE management framework 4. Carry out subjective assessments: After completing step 3, a subjective assessment is performed. There are several subjective assessment quality methods in the recommendations [12] and [13]. Subjective assessments are carried out using the largest possible sample population. Each subject is presented with the generated QoS test cases along with an automated feedback mechanism, which enables the correlation of the users responses with the generated QoS test cases being evaluated. Care is taken to ensure that subjects are not aware of which QoS parameters they are evaluating. 5. Analysis of results: At this stage, a screening process is carried out to remove unreliable results. The ITU [12] provides procedures for this screening process. Then statistical analysis is conducted to determine whether results are statistically significant; if not, more subjective assessments are conducted until this objective is achieved. Statistical analysis is carried out to determine the minimum user acceptability thresholds, i.e. the point where the users change their opinion of service quality. 6. Statistical modelling technique: The ultimate aim of steps 1 through 5 is to come up with statistically reliable data in order to build a predictive model which correlates QoS parameters (objective parameters) directly with estimates of QoE perceptions (subjective parameters). An effective way for achieving this is to employ Discriminant Analysis in determining whether the chosen predictors (i.e., the QoS parameters) lead to reliable users perception rating. This technique is discussed in detail in the next section. 7. QoE management strategy: Both the analytical results and the prediction models derived at step 5 and step 6, respectively, can be used to realize a variety of management strategies that are aimed at controlling QoS parameters to guarantee a satisfactory level of user experience. Examples of QoE management strategies are illustrated in section 4.
112
MoMM 2008
Proceedings of MoMM2008
each quantitative predictor has on predicting the group membership. Wilkss lambda (): The Wilkss lambda is used to test if the discriminant function is significant (i.e. how accurate the prediction will be). The Wilkss lambda values vary between 0 and 1. Values that are near zero denote high discrimination. The significance of lambda is determined through the p parameter (or sig. for significance), similar to ANOVAs significance. Standardized discriminant coefficients: The standardized discriminant coefficient indicates the relative importance of the quantitative predictors in predicting the qualitative attribute. The relative importance is obtained by examining the magnitude of the standardized coefficients (ignoring the sign). The larger the magnitude, the greater is that variables contribution.
f km
Where
[1]
f km = the predicted discriminant score for case m in the group k X ikm = the value of the quantitative predictors X i for case m in
group k
ui
The classification process is based on either the use of the discriminant functions or the linear combination of the quantitative predictors. These linear combinations of the quantitative predictors are called classification functions and their coefficients are referred to as classification function coefficients. The general formula is as follows:
hk
Where
[2]
hk bki Xi
= the coefficient for variable i in the equation corresponding = the value of the quantitative predictors for variable i
to group k There are as many classification functions as there are groups. For example, a two-group membership with two quantitative predictors, Equation 2 can be expanded to:
h1 h2
[3]
A case is classified as belonging to the group for which it has the highest classification score. Then, it is possible to classify new cases into group membership based on their classification scores. We explain some of the other statistics produced whilst performing discriminant analysis. These statistics helps to determine how accurate the prediction is, and how much influence
113
Proceedings of MoMM2008
MoMM 2008
Table 1. Discriminant analysis showing Wilkss lambda value and its significance Content types Wilks lambda () Significance (p) News 0.760 0.000 Romance movie 0.760 0.000 Cartoon 0.747 0.000 Comedy 0.780 0.000 Music video 0.678 0.000 Top gear 0.742 0.000 Action movie 0.516 0.000 Music concert 0.468 0.000 Cricket 0.481 0.000 Football 0.578 0.000
4. RESULTS
Table 1 shows the Wilkss lambda test of significance of the discriminant functions obtained for each content type separately. The p-values are highly significant (p < 0.001), indicating that for each content type, the means of the discriminant function scores are dissimilar between the two respective qualitative group responses. This leads us to the conclusion that the functions are discriminating well for all of the content types. Tables 2a and 2b show the standardized discriminant function coefficients for each content type. For content types in Table 2a, the relative contribution of the encoding bitrate is more important than that of the frame rate, whilst discriminant functions of content types in Table 2b are more influenced by the frame rate. Table 2. The standardized coefficients of predictors for content types. Table 2(a) Content types News Romance movie Cartoon Comedy Top gear Table 2(b) Content types Music video Action movie Music concert Cricket Football Standardized Coefficients for the predictors Encoding bitrate Frame rate 0.153 -0.938 -0.924 -0.325 -0.267 0.851 1.871 1.854 1.306 1.255 Standardized Coefficients for the predictors Encoding bitrate Frame rate 1.430 1.430 1.189 1.362 0.690 -0.449 -0.449 -0.196 -0.376 0.317
Figure 3 illustrates the derived classification functions for two content types 2 . The derived classification functions coefficients vary considerably depending on the type of content. For a given combination of encoding parameters (case), the classification score (h) is computed for each classification function. The case is classified as belonging to the quality response for which it has the highest classification score. The network operator can predict the user perception of QoE, based on the application-level parameters and then adjust the allocation of network resources accordingly. News
h( Acceptable) h(Unacceptable)
Action movie
5.699 ( 0.080 u Video bitrate ) (1.613 u FR ) 4.223 ( 0.104 u Video bitrate) (1.730 u FR )
h( Acceptable) h(Unacceptable)
9.735 ( 0.111 u Video bitrate) ( 2.411 u FR ) 3.492 ( 0.087 u Video bitrate ) (1.613 u FR )
Figure 3. Classification functions based on content types. The prediction model of user perception correctly classifies 76.9% of the responses in our sample 3 . To assess the predictive accuracy of this model in a new sample, it was estimated that if the leave-one-out technique 4 were used, an equivalent proportion of cases would be accurately classified.
Due to space limitation, we only show the classification functions for two content types. Taking the average of correctly classified cases across the content types Classification functions are derived based on all cases except one, and then only the omitted case is reclassified. This procedure is repeated N times, and it provides as estimate of how well the classifications should predict with a new sample.
The results from Table 2 provide invaluable insights towards network resource dimensioning strategies. It shows how AQoS control parameter mechanisms can be applied to deliver the best possible video quality experience based on the type of content. These can be used to implement network control strategies and provide different levels of service quality (e.g., suitable charging scheme).
114
MoMM 2008
Proceedings of MoMM2008
Bandwidth (kbps)
With QoE
Figure 4. Basic view of a video streaming system. The streaming server used is Darwin Streaming Server [8], which is an open source RTP/RTSP streaming server developed by Apple. The streaming server and the client were connected by a dedicated 100Mb/s Ethernet LAN. The tcpdump monitoring tool was running on the streaming server in order to capture and record the traffic generated by a video content and the server. This was repeated on several trials to obtain accurate results. The tcpflw software developed by [20] was used to analyze the recorded traffic to determine the network bandwidth that is being utilized based on the content type. Figure 5 shows the network bandwidth being utilized by the Action movie content type. This Action movie (without QoE-based control) was encoded at 25fps, with audio and video bitrate of 12.2kbps and 384kbps, respectively. By contrast, if we employ our QoE-aware approach, we can achieve network utilization efficiently and accurately. We can strategically degrade the quality of the video by knowing the degree of influence of each of the encoding parameter on the end user perception. For example, in Section 4 (Table 2), the end user perception of video quality for the Action movie is more influenced by the frame rate. Guided by this information, the encoding bitrate of the Action movie can be gradually reduced to a threshold of 128kbps at 20fps. As seen, the network bandwidth now being utilized by the Action movie (with QoE-aware) is significantly reduced (see Figure 5). Assuming a network LAN of 10Mbps, and having a T1 connection (1.55Mbps) is used, about 25% of the T1 connection is being utilized by the without QoEaware method. The QoE-aware scheme uses only 10% of the T1 connection, thus achieving approximately 15% gains of network resources. In addition, we will still be sure that the users QoE is not affected at all.
500 Without QoE
150
175
200
Time (seconds)
Figure 6. Bandwidth utilization for news content As noticed, we achieve smaller (only 6%) gains of network resources. This scenario can be used for example, to service a user requesting for higher quality of video (the user will incur some additional charges). When used in this context, charging will be based on how QoE is maintained.
5. CONCLUSION
In this paper, we propose a QoE management methodology, which aims at maximizing the users QoE whilst at the same time preserving network resources. Discriminant Analysis, a statistical modeling technique, has been employed to implement the proposed QoE management methodology. We illustrate how this modeling technique identifies the degree of each AQoS parameters on the user perception, and in predicting the user perception of service quality. The prediction model will allow network operators to anticipate the users experience and then allocate network resources accordingly. Our methodology provides the network operator with the ability to implement different QoE management strategies. Some examples of these QoE management strategies, which could be realized were demonstrated in this paper. Our case study illustrates how network operators can maximize the utilization of their network resources (bandwidth) and provide different levels of service quality. These have the advantages of accommodating more users in their networks, increasing their revenue gains and differentiating them from their competitors. A promising approach would be to combine QoE awareness with policy-based management, pursuing a variety of adaptation strategies. We are currently looking at combining business policies (or goals) with user-defined policies and network QoS mechanisms to realize advanced service adaptation in heterogeneous systems.
Bandwidth (kbps)
With QoE
6. ACKNOWLEDGMENTS
150 175 200
Time (seconds)
Figure 5. Bandwidth utilization for Action movie Similarly, we illustrate how this methodology can be used to provide different levels of service quality (policy-based charging schemes). The News content (without QoE-based control) is encoded at 25fps with audio and video bitrate of 12.2kbps and 384kbps, respectively. We know also from section 4 (Table 2), that the end user perception of video quality is more influenced by the encoding bitrate. The Frame rate of the News content can be gradually reduced to a threshold of 8fps, encoded at a bitrate of 256kbps. The bandwidth utilization for the News content is depicted in Figure 6.
The authors would like to thank Mr. Georges Sumo Ngaha for his help with the statistical analysis and Dr Marcos Paredes Farrera for his technical advice in writing this paper. Our thanks to ACM SIGCHI for allowing us to modify templates they had developed.
7. REFERENCES
1. Agboma, F. and Liotta, A.: "Addressing User Expectations in Mobile Content Delivery". Mobile Information Systems, 2007. 3: p. 153-164. 2. Agboma, F. and Liotta, A.: "QoE for Mobile TV Services", in Multimedia Transcoding in Mobile and Wireless Networks. 2008.
115
Proceedings of MoMM2008
MoMM 2008
3. Agboma, F. and Liotta, A.: "User Centric Assessment of Mobile Content Delivery". in Proc. of 4th International Conferences on Advances in Mobile Computing and Multimedia 2006. 4. Agboma, F., Smy, M., and Liotta, A.: "QoE Analysis of a Peerto-Peer Television System". in Proc. of IADIS Telecommunications Networks and Systems. 2008. 5. Apteker, R.T., Fisher, J.A., Kisimov, V.S., and Neishlos, H.: "Video acceptability and frame rate". Multimedia, IEEE, 1995. 2(3): p. 32-40. 6. Bradeanu, O., Munteanu, D., Rincu, I., and Geanta, F.: "Mobile Multimedia End-User Quality of Experience Modeling". in Digital Telecommunications, , 2006. ICDT '06. International Conference on. 2006. 7. Claypool, M., &Tanner,J.: "The effects of Jitter on the Perceptual Quality of Video", in Proc. of ACM Multimedia. 1999. p. 115-118. 8. Darwin Streaming Server. [cited; Available from: http://developer.apple.com/opensource/server/streaming/index.ht ml. 9. Fechner, G.T.: "Elements of Psychophysics". 1966: Translated by Alder,H.E. Edited: Howes,D.H., Boring,E.G Published by Holt Rinehart & Winston. 10. Frank, P. and Incera, J.: "A Neural Network Based Test Bed for Evaluating the Quality of Video Streams in IP Networks". in Proc. of Electronics, Robotics and Automative Mechanics Conference 2006. 11. Hands, D.S.: "A Basic Multimedia Quality Model". IEEE Transactions on Multimedia, 2004. 6(6): p. 806-816. 12. ITU-R: " Recommendation BT500, Methodology for the Subjective Assessment of the Quality of Television Pictures". 2002. 13. ITU-T: " Recommendation P.910, Subjective Video Quality Assessment Methods for Multimedia Applications". 1999. 14. Joly, A., Montard, N., and Buttin, M.: "Audio-visual quality and interactions between television audio and video". in Signal Processing and its Applications, Sixth International, Symposium on. 2001. 2001. 15. Jumisko-Pyykk, S., Vadakital,V., Liinasuo,M., & Hannuksela,M.M.: "Acceptance of Audiovisual Quality in Erroneous Television Sequences over a DVB-H Channel", in Proc. of Workshop in Video Processing and Quality Metrics for Consumer Electronics. 2006. 16. Khan, A., Sun, L., and Ifeachor, E.: "An ANFIS-based Hybrid Video Quality Prediction Model for Video Streaming over
Wireless Networks". in Proc. of 2nd IEEE Conference and Exhibition on Next Generation Mobile Application, Services and Technologies. 2008. 17. Klecka, W.R.: "Discriminant Analysis". Quantitative Applications in the Social Sciences. Vol. 19. 1980: Sage Publications, Inc. 71. 18. Lubin, A.: "Linear and Non-Linear Discriminating Functions". British Journal of Psychology, Statistical Section, 1950. 3: p. 90-103. 19. Mohamed, S. and Rubino, G.: "A Study of Real-Time Packet Video Quality Using Random Neural Networks". IEEE Transactions on Circuits and Systems for Video Technology, 2002. 12(12): p. 1071-1083. 20. Paredes-Farrera, M.: "Packet-by-Packet of Video Traffic Dynamics on IP Networks, PhD Thesis", in Department of Computing and Electronic Systems. 2005, University of Essex: Colchester. 21. Perkis, A., Munkeby, S., and Hillestad, O.I.: "A Model for Measuring Quality of Experience", in Signal Processing Symposium Proc. of the 7th Nordic. 2006. p. 198-201. 22. Richard, A.: "Continuous Qulaity Assessment of DigitallyCoded Television Pictures, PhD Thesis", in Department of Electronic Systems Engineering, 1996, University of Essex: Colchester. 23. Salkintzis, A.S. and Passas, N.: "Emerging Wireless Multimedia Services and Technologies". 2005: John Wiley & Sons. 24. Siller, M. and Woods, J.: "Using an agent based platform to map quality of service to experience in conventional and active networks". Communications, IEE Proceedings-, 2006. 153(6): p. 828-840. 25. Teyeb, O., Srensen, T.B., Mogensen, P., and Wigard, J.: "Evaluation of Packet Service Performance in UMTS and Heterogeneous Networks", in Proc. of ACM Workshop on Quality of Service & Security for Wireless and Mobile Networks. 2006. p. 95-102. 26. Venkataraman, M., Sengupta, S., Chatterjee, M., and Neogi, R.: "Towards a Video QoE Definition in Converged Networks". in Proc.of Second International Conference on Digital Telecommunications, 2007. ICDT '07. . 2007. 27. Winkler, S. and Faller, C.: "Perceived Audiovisual Quality of Low-Bitrate Multimedia Content". IEEE Transactions on Multimedia, 2006. 8(5): p. 973-980.
116