Statistical Characterization of a real video on demand service_ user behaviour and streaming-media

Statistical characterization of a real video on demandservice:User behaviour and streaming-media workload analysisRoberto Garcı´a *,Xabiel G.Pan ˜eda,Victor Garcı´a,David Melendi,Manuel Vilas Department of Computer Science,University of Oviedo,Campus de Viesques,sn,33204Xixo´n-Asturies,Spain Received 14September 2006;received in revised form 20December 2006;accepted 10February 2007Available online 25February 2007AbstractThe video on demand service characterized in this article has surpassed 4years of service,and the access log database stores information of more than 160,000reproductions and 900videos.The wide variety of subjects,the range of content lengths and a frequent update of contents (several new videos every day),clearly differentiate this study from other pre-vious research limited to specific users,subjects or environments.We present a statistical study of user behaviour and streaming traffic,analyzing session characteristics,mistaken reproductions,amount of media delivered,number and length of pauses and jumps in the reproduction,popularity and daily access profile.The results of the analysis will allow us to develop simulation models and workload generators to evaluate different scenarios and situations of the service.Ó2007Elsevier B.V.All rights reserved.Keywords:Video on demand;Multimedia streaming;Network traffic measurement;Workload characterization1.IntroductionThe speeds offered by the Internet access providers have been increasing continuously in recent years.As penetration grows,broadband providers in the OECD are increasingly offering voice and video services over their platforms [19].The bandwidth enhancement in subscribers’access capabilities has given rise to the appearance of Internet audio/video services,based mainly on streaming technology.Streaming media has been widely used over the Internet for an increasingly demanding and growing consumer population [35].Thus,the characterization of streaming access workloads has become an essential factor to evaluate the per-formance of these services and their implications on the rest of the services in the network.In this paper,a user behaviour analysis and workload characterization performed on the video on demand service of www.lne.es (La Nueva Espan ˜a Digital )is presented.This digital news service is one of the most suc-cessful in Spain and its video on demand service,called LNE TV ,has interesting characteristics,such as a wide1569-190X/$-see front matter Ó2007Elsevier B.V.All rights reserved.doi:10.1016/j.simpat.2007.02.004*Corresponding author.Tel.:+34985182488;fax:+34985181986.E-mail addresses:garciaroberto@uniovi.es (R.Garcı´a),xabiel@uniovi.es (X.G.Pan ˜eda),victor@uniovi.es (V.Garcı´a),melendi@uniovi.es (D.Melendi),vilasmanuel@uniovi.es (M.Vilas).URL:http://www.it.uniovi.es (R.Garcı´a).Simulation Modelling Practice and Theory 15(2007)672–689R.Garcı´a et al./Simulation Modelling Practice and Theory15(2007)672–689673 variety of subjects,a wide range of content lengths,a continuous content update(several new videos every day),etc.Over a period of4years the access logfiles have been captured and stored in a database to be ana-lyzed.The obtained results have been compared with some previous classical studies,such as[2,6,12,16,33]. The results of the analysis have been used to develop user behaviour and workload traffic models which can be integrated in video on demand service models and load generators with the aim of evaluating future situations in the service.This paper makes the following contributions:A primary contribution of our work is its statistical analysis of user behaviour in a commercial video on demand service over a4-year period.The mistaken reproductions,amount of media delivered,number and length of pauses and jumps in the reproduction,session characteristics,popularity and daily access profile have been characterized.Unlike most previous media workload characterizations,which are focused on an educational content,the workload analyzed consists of accesses to a news site in Spain,containing vari-able-length and variable-subject entertainment content.Thus,the results improve and enrich current knowl-edge of typical media workload patterns.Secondly,a streaming-media workload characterization has been performed,differentiating traffic per pro-tocol,source device,audio/videoflow and workload,during both regular load and buffering time intervals. The service uses RealNetworks technology.Finally,we propose a simulation model of the audio/video on demand service.This model is oriented to services with different types of information and lengths.The rest of the paper is organized as follows:in Section2other related works are analyzed.A general description of the case study is carried out in Section3.The user behaviour analysis and characterization are presented in Section4.A measurement study and characterization of RealMedia streaming traffic is carried out in Section5.In Section6we propose a simulation model of the video on demand service.Finally,con-clusions are presented in Section7.2.Related workVideo on demand analysis is a relatively recentfield in the research world.Video on demand services are not widely deployed on the Internet,and,therefore,analysis studies are not abundant.In spite of that,some interesting papers on streaming service analysis have appeared during the last few years.These papers study different elements of user behaviour,quality of service and content popularity.Elements such as session length,delivered time,user’s interactions,etc.have been studied in[2,6].In[33], the study is centred on the user’s connection,user’s origin,and quality requested.The quality of service has been evaluated in[16,33].Different parameters,such as packet loss,jitter,and user perceived quality have been the main study points of these papers.Another element widely analyzed is document popularity[2,5,6,9,12,30].The distribution of users’repro-ductions among the videos is compared with the Zipf-like distribution and theÉparameter is calculated.Ach-arya et al.[1]characterized user accesses to video objects on the web(VOW),showing that,contrary to previous studies,popularity did not follow a Zipf distribution.Using some of the results obtained in these analyses,other studies and tools have appeared.These were able to simulate real service behaviour and carry out performance evaluations.That is the case of[3]where the server capacity is evaluated using a simulated workload.The work of[13]follows the same line.In their paper a tool for video on demand service analysis using simulated workload is presented[31]presents a character-ization of live streaming workloads from a server used by one of the top ten content providers in Brazil[27] analyzes the live streaming workload from a large content delivery network.They study the workload char-acterizing popularity,arrival process,session duration and transport protocol used.Although the analysis of Internet traffic usually includes streaming multimedia traffic,there are few specific works about RealMedia streaming traffic[17]presents an exhaustive study of RealAudio traffic from a popular Internet audio service.They found that RealAudio traffic is dominated by specific packet sizes[6]analyzes a client-based streaming-media workload generated by clients from the University of Washington connecting to servers in the Internet.Aspects about workload characterization,such as bandwidth utilization,session char-acteristics and server popularity are treated in detail.Further research on RealMedia traffic is presented in[14]. The authors study the traffic both in the application layer and the network layer and the relationship between674R.Garcı´a et al./Simulation Modelling Practice and Theory15(2007)672–689these two layers is investigated.More recently[5],analyzes an enterprise media server workload using access logs from two different servers in HP Laboratories.They divide the media workload properties into two groups:static and temporary,and propose two new metrics to characterize temporary properties of the media workload:newfiles impact metric and life span metric.Video popularity distributions and client behaviour have been treated in this research.A workload generator of streaming media services is presented in[29].This work focuses on the long term behaviour of network services.They proposed a generalized Zipf-like distribution to model the content pop-ularity and partition media workload properties in two groups,static and temporary,as in[5].Daytime access patterns are treated in this paper[10]proposed a new streaming media workload generator,parameterized from the real workload characterizations indicated in[9].3.Case studyThe presented study has been performed on the video on demand service of La Nueva Espan˜a Digital (www.lne.es)which has an important number of accesses and has reached8th position in the ranking of digital news sites in Spain.In2001,www.lne.es presented its video on demand service(LNE TV)developed by the Computer Science Department of the University of Oviedo.The number of visits and the volume of information have risen since then.3.1.Service descriptionThe multimedia section of www.lne.es has an architecture formed by two servers[28].One of them is the streaming server and the other supports the web pages used to access the videos,the analysis system,and a redundant streaming server.The analysis server stores all the modules of the analysis tool[22],including the database,the web server,and data loaders and analyzers.Helix from RealNetworks[24]is the technology used to stream the videos,delivered on demand when a subscriber performs a request.3.2.Content descriptionThe multimedia service contents have been classified into15subsections according to their subject.Some of these subjects are News,Music,Tourism,Science,Cinema,Comedy,Leisure,Sports,Others,etc.The lengths range from30s,the shortest,to more than an hour,the longest.Currently,the service has more than900 videos.3.3.Media server log formatsEvery request processed by the multimedia server is recorded in its logfiles.Logfiles in multimedia services usually provide detailed information about users,resources,requests,and data packet transmission[26].Table 1summarizes the overall statistics of the LNE TV service,obtained by analyzing the logfiles of this service with the tool Fesoria[22].Table1LNE TV service summarized statisticsDuration4Years Different subjects15 Total sessions47,000Storage requirements40GB Total reproductions160,000Bytes transferred188GB Unique clients31,902Bitrate(average)145,170bps Stored videos910Bitrate(standard deviation)84,116bpsR.Garcı´a et al./Simulation Modelling Practice and Theory15(2007)672–689675 er behaviour analysis and characterizationThe analysis presented in this paper has been performed using the information obtained from the logfiles of the multimedia server.Thefinal goal of the analysis is to construct a model of user behaviour.Thus,all the parameters which characterize this behaviour will be evaluated.A preliminary analysis has been presented in [32].Users access the service during a session,which is composed by one or more reproductions with reflection periods(think time)between them(Fig.1a).Each reproduction begins with a play interaction and continues to the end of the audio/video or to a stop interaction.During this time users can perform intermediate inter-actions(pause,forward and backward jumps,play,stop),generating paused periods,when no information is being delivered,and active periods,when users are receiving data.In the model proposed in[10]there is only one reproduction per user session.We have tried to go one step further than this previous definition,analyzing the user behaviour over longer periods,discovering the number of successive reproductions,and the reflection time between them.To distinguish one user session from another,we have considered a minimum gap of30min between two consecutive reproductions,the same as[21,9].Fig.1b shows the conceptual model of the user behaviour in a session,taking into account if the request for a video is correct or invalid,differentiating the characterization for short and long videos depending on the length of the requested video,and taking into account all the available user interactions(pause,jump,play, stop).At the end of a video reproduction the user can request a new video or end the session.The aim of this section is to determine the distributions for all the parameters needed to create an accurate model of user behaviour.Due to the variability in the lengths of the offered videos,we have differentiated the characterization for short(under5min)and long videos,the same as in[16].User behaviour characteristics are evaluated and the most suitable distributions arefitted to the empirically observed data,using maximum likelihood estimators(MLE)[15]and least squares differences.As in[9],in order tofind the distribution that best models the characteristic to be analyzed,we compared the least squares differences of the bestfitted curves for a set of alternative distributions.We also visually compared the curve fittings at the body and at the tail of the measured data,favouring a betterfit for the body or the tail,if nec-essary,in order to analyze its impact on the performance of the service.4.1.Session characteristicsAs Fig.1a shows,there are two essential elements in a session that must be determined:the number of reproductions and the inter-reproduction time(reflection or think time).Regarding the number of reproductions in a session,Fig.2a shows that 47.99%of the sessions present only one reproduction.The concatenation of two Zipf-like distributions,with h 1=1.77(94.28%of the sessions,1–6reproductions)and h 2=3.09(the remaining 5.72%,6to 25reproductions),closely match the empirical values.Fig.2a shows real accesses and estimated values in log–log plots.Zipf-like distribution [36]is represented by the following pdf distribution:p ði Þ¼C i H ;and C ¼1P n i ¼11i Hð1ÞThese results are consistent with [9],where the concatenation of two Zipf-like distributions is suggested as agood approximation model.In order to characterize the inter-reproduction time,a set of alternative distributions (Exponential,Pareto,Normal,Gamma,Weibull)have been tested,obtaining its parameters using MLE [15].By comparing the least squares differences between these best fitted curves and the real data,as in [9],the best adjust is done using a Weibull distribution [15],as is indicated in Fig.2b.Both the body and the tail of the empirical distribution are quite well fitted by the Weibull distribution.Moreover,Kolmogorov-Smirnov and Anderson-Darling tests [15]also demonstrate that Weibull is the most suitable distribution to characterize the inter-reproduction time.These results are consistent with the works in [9,2],where Weibull or Lognormal are the most accurate dis-tributions,depending on file size.The Weibull pdf distribution used in this paper is characterized by the a and b parameters,following the expression:f ðx Þ¼ab x b À1e Àa x bI ð0;1Þðx Þð2ÞFig.2b shows the cumulative distribution of the measured and simulated values,indicating the MLE parameters of the Weibull distribution and the mean of the inter-reproduction times in a session.As Fig.2b indicates,90%of the inter-reproduction times are below 200s,and the heavy-tailed Weibull distribu-tion represents non-zero probabilities of long inter-reproduction times.4.2.Mistaken reproductionsA mistaken reproduction is a user request which generates an error.It can be identified because it generates an entry in the error log or the duration reported in the entry of the access log is 0s.Generally,this type of reproduction is due to the fact that the user does not have the suitable program to watch the video.Despite the fact that Web masters issue warnings and develop plug-in detectors,a significant percentage of the total676R.Garcı´a et al./Simulation Modelling Practice and Theory 15(2007)672–689R.Garcı´a et al./Simulation Modelling Practice and Theory15(2007)672–689677 requests usually fail for this reason.Thus,the mistaken reproductions maintain a percentage of15%during the whole life of the service.4.3.Media delivered per reproductionOne important parameter in the analysis of video on demand services is the quantity of media delivered in a reproduction.Different characterizations have been carried out for short and long videos.Previous studies, like[16],have detected distributions,such as Lognormal and Exponential for short videos(0–5min),and Gamma+Pareto or Gamma+Lognormal for long videos(5–50min)[29]use an aggregated distribution, whose PDF is the sum of the PDFs of a set of normal distributions proportionally.However,the LNE TV service presents the interesting characteristic that a hybrid distribution,i.e.a composition of two distributions,fits well with the length of video reproductions(Fig.3).Using the length reproduction histogram of four representative examples of videos chosen at random (Fig.3),it is possible to appreciate afirst distribution from the(0–10%)interval to(90–100%)and another around100%.The weight of each distribution depends on the success of the analyzed video.Short videos(vid-eos1,2)usually have a second distribution heavier than thefirst,while for long videos(videos3,4)the weight of thefirst distribution is higher than the second.The empirical observed data are modelled,by least squares difference,using the composition of two expo-nential distributions with MLE parameters l1=0.16and l2=0.06,respectively for short videos,and MLE parameters l1=0.2and l2=0.27for long videos.Thefinal formula to calculate the reproduction length distribution is the following:fðxÞ¼pf1ðxÞþð1Àp1Þf2ðxÞð3Þ1f1ðxÞ¼1l1eÀx l10<x<10otherwise(and f2ðxÞ¼1l2eÀxÀ1l2x>10otherwise(where p1is the success probability of the video.The way these distributions combine depends on the type of user who reproduces the video:slightly inter-ested or highly interested.Normally,slightly interested users leave video reproductions during thefirst seconds and,moreover,their interest decreases quickly with time.On the other hand,highly interested users usually watch the entire video,and sometimes make backward jumps to re-watch some parts.In the case of previous studies,such as[6],despite the short length of a considerable amount of the repro-ductions,the majority are between2.5and4.5min.With regard to[16],the main difference can be that LNE TV offers news and entertainment programs,while[16]offered educational information.The visualization of the latter was essential for students to pass their courses,so the behaviour could have been conditioned. 4.4.Reproduction characteristicsThroughout the video reproductions,users can make pauses as well as backward and forward jumps.These user interactions have a significant influence on the on–offperiods in the transmission of media data and the buffer reloads on the clients’side.4.4.1.PausesPauses are important to determine the duration of the on–offperiods in the media ually, reproductions are modelled using on and offperiods to simulate receiving and paused times.Our analysis show that there are relatively few user interactions during the sessions.Fig.4shows the histogram with the percent-ages extracted from LNE TV.In short videos,only4.3%of the reproductions have user interactions,increasing to9.5%in long videos. These results are consistent with[9,2],which report a larger number of user interactions within a video session of longer videos.However,in[9,2]more than20%of the reproductions for longfiles have10or more inter-active requests.This difference is due to the educational content of the analyzed services in[9,2].In the enter-tainment service analyzed in[9],only4%of the reproductions have three or more pauses.A discrete distribution,with the values indicated in Fig.4,has been used to characterize the number of pauses.The length of paused time,which determines the offperiods in the user model,has also been characterized (Fig.5).We have compared the least squares difference of a set of alternative distributions,a Weibull being the most adequate distribution,both for long and short videos.The average value of the pause length is55.73s for short videos and95.09s for long videos.Fig.5shows the cumulative distribution function compared with the measured empirical data and the MLE parameters of the Weibull distribution used in our model.The presence678R.Garcı´a et al./Simulation Modelling Practice and Theory15(2007)672–689R.Garcı´a et al./Simulation Modelling Practice and Theory15(2007)672–689679of heavy-tailed distributions indicates non zero probabilities of long pauses,modelling the real usage of the users in the service.4.4.2.Forward and backward jumpsDespite the fact that forward and backward jumps do not change the on–offperiods of user behaviour,they are important in order to analyze the cache performance in this type of services.The number of forward and backward jumps has been modelled with a Zipf-like distribution,using least squares differences.Real accesses and estimated values for forward jumps are displayed in Fig.6.Similar results are obtained for backward jumps.Table2shows the interactions and the estimated h parameters.The high values of these parameters are due to the fact that most of the sessions do not have forward or backward jumps in the reproduction of the video (97.81%in Fig.6a and96.62%in Fig.6b).In the entertainment service analyzed in[9],asfile size increases the percentage of jumps increases.The percentage of reproductions with jumps in[9]is higher than that found in our analysis.The same results as in[9]were found in the educational workload analyzed in[21].The length of the jumps has also been analyzed and modelled with a Weibull distribution,using MLE and least squares difference.Table3shows the parameters for the heavy-tailed distribution.680R.Garcı´a et al./Simulation Modelling Practice and Theory15(2007)672–689Table2Number of jumps characterizationInteraction Short videos Long videos #Forward jumps Zipf(h=5.80)Zipf(h=5.22) #Backward jumps Zipf(h=4.32)Zipf(h=3.73)Table3Jump length characterizationInteraction Distribution Short videos Long videosForward jump length Weibull a=0.13214Average a=0.10916Averageb=0.5863150.226s b=0.54129113.577s Backward jump length Weibull a=0.08108Average a=0.05959Averageb=0.5985100.097s b=0.59236187.383s These results are consistent with the general rule of[9,21],indicating that the distances of backward jumps are usually longer.In[9],the average jump in distance in either direction is under45s.Our results and the results in[9]are in sharp contrast with the very long average jump distance(over2000s)observed in the MANIC educational system[21].4.5.PopularityAnother important element in the workload study is the popularity of the offered ually, user accesses are compared with a Zipf-like distribution[2,6,9,27,33]with different values of the h parameter. The cacheability of the information will be widely successful when h parameter is high and unproductive when values are close to zero.Yu and Chou[34]propose a caching algorithm to characterize the empirical internal popularity distribution.The performed study has detected that document popularity follows a combination of two Zipf-like distributions with a parameter h1=0.56for98%of the accesses,when the analyzed period is long(Fig.7a).With shorter periods the results present a smaller h1(Fig.7b).The remaining2%of the accesses is considered negligible for caching policies.Previous studies have presented different conclusions[6]detected a low h equal to0.47and[2,9,27]the com-bination of two Zipf-like distributions[29]present a novel generalized Zipf-like distribution to capture the pop-ularity distribution.Fig.7shows,on the left,the results for the total analysis period(4years)and,on the right,for a period of one year.We have used a log–log scale to show the power-law structure of the Zipf law.R.Garcı´a et al./Simulation Modelling Practice and Theory15(2007)672–689681 4.6.Daily access profileAs in[29],the access distribution throughout the day has also been analyzed.There is a period between0 am and10am where the demand is low.The number of accesses increases in the morning to reach the max-imum demand at1pm.From2pm to3pm the number of accesses decreases slightly(this is the typical lunch time in Spain).The number of accesses rises from3pm to8pm,reaching the general maximum at7pm.This temporal evolution is used to model the average number of users who access the service throughout the day.5.Workload characterizationAs the last stage of video on demand service characterization,this section analyzes the basic characteristics of the streaming-media workload.We differentiate the traffic per protocol,device,audio and videoflow,reg-ular load and buffering time.Once the video is coded and stored,the media server transmits the data to the client through the network. Due to the effects of the network,the transmitted information reaches the receiver with a speed different to the playing rate of the client.Besides,the presence of a buffer in the receiver compensates the mismatch between transference and playing rates.Furthermore,the client could request a higher transference rate from the server to avoid the starvation of the receiver buffer.This situation,different from the regular load,is known as buf-fering time.In this paper we differentiate the characterization in regular load and in buffering time intervals.5.1.Streaming-media communicationIn this section we briefly describe the streaming media communications using RealNetworks tools.The media contents are created using RealProducer and stored in RealMedia File Format(RMFF)files.The Real-Time Streaming Protocol(RTSP)controls the streaming session,allowing the client to interact with the server by means of the play,pause,stop,fast-forward and rewind control functions[25].RTSP servers use different transport protocols to send data to RTSP clients.RealNetworks uses Real Data Transport (RDT)protocol as the transport mechanism to deliver data from servers to clients.The decision making pro-cess RealPlayer uses to choose either UDP or TCP as transport protocol to send media data is not publicly documented[8].In our study,we have detected the use of UDP.5.2.RealMedia workloadIn order to characterize the media workload,the most suitable audio/video qualities for the connection char-acteristics of the users that access the service were selected;most of the LNE TV users connect either through ADSL lines or through cable networks.Following the recommendations in[23],qualities of140-32-20and90-16-15have been selected.Thefiles nomenclature follows the format‘‘Total quality(Kbps)–Audio Quality (Kbps)–Frames/s’’.In Table4the most representative information from RMFFfiles is shown[14].Table4Summary information of the RMFFfilesItem140-32-2090-16-15Audio Video Audio Video Total packets53,19137,649Avg.packet size(bytes)465296320288 Max.packet size(bytes)465666320656 Avg.bit rate(bps)32,041107,95916,00074,000 Max.bit rate(bps)32,041107,95916,00074,000 Preroll(ms)185774839606870 Duration(ms)998,504998,502Target frame rate(frames/s)20155.3.Traffic collectionTo produce traces,eachfile is transmitted from the multimedia server exclusively for one user in non-over-lapped experiments.Providing that the main goal of the experiment is to define the traffic introduced in the communications network,traces are captured on the server’s side using Ethereal utility.Ethereal is used to analyze the behaviour of the RTSP and RDT protocols.Hence,we can differentiate audio streams from video streams,analyze control messages(resume,play,pause,buffering)and also obtain the timestamp of the deliv-ered data.Measurements with several qualities of media data indicate that UDP traffic from server to client represents more than96%of the packets and99%of the delivered bytes.Therefore,the consumption of network resources in LNE TV service,using RealNetworks technology,is mainly caused by the UDP-RDT traffic. For this reason we have carried out an accurate analysis to characterize this type of traffic.5.4.Streaming-media traffic characterizationIn this section we analyze and characterize the measured network-layer traffic of the RealMedia streams collected.We have focused the study on the UDP-RDT traffic,differentiating audio streams from video streams by means of the StreamIDfield in the RDT protocol.Moreover,the interaction and synchronism between the different trafficflows has been taken into account in order to construct a model of the media workload.As shown in Fig.8,after the initial negotiations between client and server,TCP traffic is also used to increase the transference rate during the preroll time(initial buffering time to load the buffer of the client [24])and posterior buffering intervals.Fig.8shows the preroll and buffering time intervals delimited by TCP client requests.5.4.1.Analysis summary for the streamsTables5and6show the analysis summary for audio and video streams,with the considered qualities in this research.Packet sizes include IP,UDP and RDT protocol overheads,which constitute38bytes over the media data.During the buffering time intervals the inter-arrival time decreases,approximately four times the usual load intervals,increasing the transference rate to quickly load the receiver buffer.These results are consistent with[7,4]for RealNetworks systems and[18],which work on Windows Stream-ing Media(WSM).We also observed that this increase in the transference rate is only made by reducing the inter-arrival times and not by increasing the sizes of the packets.During the buffering time and regular load intervals the packet size distributions remain invariable.。