SVC-DASH-M: Salable Video Coding Dynami Adaptive Streaming Over HTTP Using Multiple Connetions Samar Ibrahim, Ahmed H. Zahran and Mahmoud H. Ismail Department of Eletronis and Eletrial Communiations, Faulty of Engineering, Cairo University, Giza, Egypt. {ahzahran, mhismail}@ieee.org Abstrat Dynami adaptive streaming over HTTP (DASH) has gained a signifiant momentum for multimedia streaming due to its ability in rossing firewalls and availability of infrastruture. In the mean time salable video oding (SVC) is building a similar momentum as it enables effiient media storage and ahing. In this work, we identify the main omponents of adaptive SVC-DASH lient and propose an streaming heuristi over dynami multiple onnetions. The proposed algorithm is experimentally tested under different onnetion and link onfigurations. Our results show that the algorithm suessfully ahieves interruption free streaming under all the tested bandwidth and link onfigurations. Additionally, the usage of multiple onnetions results in notieable improvements in the ahieved streaming quality for large link delays. I. INTRODUCTION Dynami adaptive streaming over HTTP (DASH) [] has gained a signifiant momentum and is widely adopted by many standardization bodies suh as open IPTv and GPP. Generally, HTTP streaming has several attrative features that helped in speeding up its aeptane. Avoiding firewall and NAT traversal issues represents a major advantage for HTTP streaming over other traditional streaming tehniques (e.g., RTP/UDP). Another advantage of using HTTP is the availability of existing HTTP ahe infrastrutures on the Internet, thus relieving not only the server load but also reduing the overall uplink traffi towards the ahe. The main drawbaks of using HTTP streaming inlude higher end-toend delays in the ommuniation due to using transmission ontrol protool (TCP) and possible redued link utilization due to TCP dynamis, but this an be ompensated using the lient-buffer. Adaptive streaming tehniques are ommonly used to aommodate variations in the bandwidth to avoid annoying streaming interruptions due to buffer under-run. In HTTP streaming, the media is subdivided into short duration hunks, ommonly known as segments. The granularity of streaming adaptation depends on the segment size and the adopted enoding tehnique. If advaned video oding (AVC) based on H./MPEG- Part standard is used, the server would ontain multiple versions of eah segment enoded at different qualities (resolution, frame rates, piture quality or This work is supported by the National Teleommuniation Regulation Authority (NTRA) of Egypt ombination of them). Hene, in AVC-DASH, deisions are typially taken at the segment end. Salable video oding (SVC) [] represents an extension to AVC and it allows enoding a video segment into N layers inluding a base layer and numerous enhanement layers. The aumulation of more layers generally leads to improving the video quality. Hene, in SVC-DASH, the granularity of quality ontrol extends both horizontally over time and vertially over layers. In this paper, we propose and evaluate the performane of an SVC-DASH lient over multiple parallel onnetions (SVC-DASH-M) using a real testbed. In SVC-DASH-M, the quality of the requested layer-segment depends on the amount of buffered media and the network onditions. Additionally, it adopts a quality seletion poliy that is onservative for the frequeny of video quality shifts. Our experimental results indiate that SVC-DASH-M suessfully adapts to different operating onditions inluding various link delays and bandwidths. Our results also bring important insights regarding HTTP server onfiguration and HTTP performane. The rest of the paper is organized as follows: in Setion II, we present a brief overview for bakground and related work. The design of SVC-DASH-M is presented in Setion III followed by its performane evaluation in Setion IV. Finally, our onlusions and diretions for future work are presented in Setion V. II. BACKGROUND In HTTP streaming, the ontent is requested using the HTTP GET request/response dialog that is typially servied over TCP. Hene, HTTP streaming is onsidered a pull protool in whih the media is requested from the server aording to the streaming lient instrutions. The details of the video information based on whih the lient takes the deision are typially obtained from a media presentation desription (MPD) file that inludes web links to the media files with orresponding enoding details of eah file ontent. Many studies fous on the design and performane evaluation of streaming over HTTP. These studies onsider different types of video enoding (AVC[] and SVC [], []), different onnetion onfigurations (persistent versus non-persistent [] and single versus multiple onnetions), different streaming modes(real-time versus stored) and various segment durations.
Most of these studies fous on AVC-streaming over a single persistent onnetion. More reently, several studies have shown that SVC-based streaming over HTTP has several merits and it started to get more attention. In [], it is shown that SVC-DASH has several advantages inluding the possibility of serving a larger number of users with different equipment apabilities as well as having a higher ahing effiieny. In [], the authors assert that SVC-DASH not only needs less buffering requirements in omparison to AVC-DASH but it also improves the quality of experiene (QoE) for the viewer. In [], the authors ompare the performane of different AVC-based and SVCbased heuristis using NS simulations with NS radle. They also show that AVC performs better under high latenies while SVC better adapts to sudden and temporary bandwidth flutuations when a single onnetion is used. The work in [] analytially investigates the optimal seletion poliy for layer segment when SVC is used. The onduted simulations show that a vertial poliy is optimal under fixed bandwidth while a diagonal seletion poliy is optimal for variable rate. A few works onsidered using multiple onnetions in HTTP streaming. In [], the authors present a rate-adaptive algorithm for AVC-based video streaming over two parallel onnetions. Their simulations show that parallel onnetions outperforms the serial segment fething method in ahievable media bit-rates but is slightly inferior in buffer underflow frequeny. In [], a ursor-based SVC lient heuristi Live TV setting is onsidered where the lient side buffer is limited. The authors also exploit using a fixed number of parallel onnetions and pipelined downloading of segment layers to overome the high end-to-end delay issues. In our study, we investigate the usage of dynami multiple onneitons for SVC-based streaming. III. SVC-DASH-M CLIENT DESIGN Typially, the first step performed by an HTTP streaming lient is to download the MPD file and parse it to obtain the video information. For SVC-DASH, every video segment is split into multiple layers per segment depending on the enoding struture. The ontent of eah layer segment is stored as a separate file on the HTTP server. Hene, the MPD information inludes: segment ID, layer ID, layer-segment URL, layer-segment duration, and layer size. Note that the segment ID an be used to identify the timing information of the transmitted segment. One the MPD information is reeived,svc-dash-m operation is based on three interating omponents inluding a layer-segment requester, a onnetion manager, and enhanement layer seletion poliy. Our layer-segment requester initially gets the first n min base layer-segments, where n min represents the minimum number of onnetions. Upon the reeption of a omplete layer-segment on onnetion, SVC-DASH-M onsiders a level-based layer-segment requester that determines the quality of the requested layer-segment as shown in Algorithm. The details of reqlayseg() is shown in Table I. The algorithm defines two appliation buffer-level thresholds denoted as B min and B target. B min represents a low threshold for the Algorithm Level-based Layer-Segment Requester Algorithm. bl = getbufflevel(); //buffer level if bl <= B min //request base layer only reqlayseg(b,onn); elseif B min < bl <= B target //alternate base and enhanement //requests reqlayseg(a,onn); else // bl > B Target //request enhanement layers only reqlayseg(e,onn); Algorithm Connetion Management Algorithm. µ = SD SFT ; if µ > r nxt = estimatenextrate(lstype); ǫ = (r nxt /r prev ); reqnextlayseg(lstype, onn); if (n < n max && µ > ǫ) reqnextlayseg(lstype, -); elseif µ < if n > n min lose(onn); else reqnextlayseg(lstype, onn); data to be maintained in the buffer to aommodate network ondition variations while B target represents a target buffer level that the appliation should be operating around. The defined poliy in Algorithm targets improving the streaming quality at high buffer levels through downloading enhanement layers. Additionally, the layer-segment requester poliy maintains the user experiene quality at low buffer-level through avoiding streaming interruptions by fousing on the base layersegments. The onnetion manager ontrols the dynamis of the used onnetions in SVC-DASH-M. In our algorithm, the number of opened onnetions, denoted as n, is lowerbounded by n min and upper-bounded by n max. The number of used onnetions varies when a layer-segment is reeived. Let denoted the onnetion over whih the layer-segment is reeived. Tne onnetion manager operation is shown in Algorithm. The details of the used funtions are presented in Table I. The streaming performane is evaluated based on the segment duration-feth ratio µ = SD SFT, where SD represents the duration of the reeived segment and SFT represents the duration over whih that segment is fethed. Note that large values of µ indiate that the available bandwidth is large and vie verse. Hene, a new layer-segment is requested over. An additional segment may be requested over a new onnetion if n < n max and there exists suffiient bandwidth for down-
getbufflevel( ); reqlayseg(lstype, onn) estimatenextrate(lstype) returns the buffer level in se. requests a layer-segment from the server over onnetion onn. lstype ould be B for base layer, E for enhanement layer, or A to alternate between base and enhanement layers. given the layer-segment type, this funtion returns the data rate required for downloading the next two segments to be requested. reqnextlayseg(lstype, onn); given the layer-segment type, this funtion requests the next layer-segment to be requested over onn. If onn is set to -, then a new onnetion is opened to request the seleted layer-segment. Table I: Algorithm Funtions. Figure : Testbed Arhiteture. loading two segments. The bandwidth suffiieny ondition is satisfied if the ratio between the required bandwidth for downloadingthetwoseletedlayer-segments(denotedasr nxt ) to the estimated onnetion bandwidth (denoted as r prev ) is less than the segment duratoin-feth ratio. To this end, it is worth noting that all the deision parameters are available or are readily measured at the appliation layer and do not inur any additional overhead to the ommuniation proess. The third design omponent for SVC-DASH-M is layersegment seletion poliy. Generally, this poliy should ensure that any seleted layer-segment is expeted to be reeived before its playout time. In our lient, this poliy is implemented in reqnextlayseg(lstype, onn), whih follows a horizontal seletion poliy by whih no layer-segment from layer n is requested unless all the layer-segments of layer n have been already downloaded. This poliy aims to redue the quality shifts in the reeived video. Other possible poliies inlude vertial poliy, whih is greedy in nature as it downloads all enhanement layer-segments for segment i before downloading enhanement layer-segments for segment i +. In between the two extremes, other poliies may be developed. Irrespetive of the hosen poliy, estimating an aurate expeted reeption time for the seleted layersegment represents a hallenging design issue beause TCP performane depends on several fators inluding the link parameters (e.g., bandwidth, delay, and reliability) and TCP parameters (e.g., initial window size and ongestion ontrol algorithm). In our algorithm, we onsider that a newly opened onnetion would equally share the estimated bandwidth of onnetion. This design hoie is motivated by the wellknown friendly behavior of TCP. Investigating other seletion poliies and developing aurate SFT estimator is indeed worth investigating and is onsidered a future work. A. Testbed Setup IV. PERFORMANCE EVALUATION Our evaluation is based on experimentation over a testbed shown in Figure. In our evaluation, we used the forty-fourseond Paris.yuv obtained from Arizona State University video library []. The video is proessed using the Joint Salable Video Model (JSVM) [] as follows. First, the video is slied into two-seond segments using JSVM downsampler. Eah segment is then enoded to ten salable layers inluding spatial (original CIF and a downsampled QCIF), temporal and quality salability. The enoder produes one. file per segment. We developed additional ode to extrat different portions orresponding to different layers in eah segments using layer information that are obtained from JSVM bitstreamextrator. The orresponding MPD file is then ompiled based on the extrated layer-segment information. Finally, the MPD file together with the layer segment files are uploaded to our HTTP server, a laptop running Ubuntu.. Traffi ontrol tools [] are employed to emulate different bandwidth and delay values. Netem is used to set the delay and delay standard deviation (assumed % of the delay). Additionally, token buket filter (tbf) is used to set the link bandwidthto valuesof interest. For the propersetup of tbf, we set the limit Bw HZ and buffer(burst) = Bw HZ, where Bw represents the emulated bandwidth and HZ represents the Linux kernel timer interrupt frequeny. Additionally, one may need to reset satter-gather, tp-segmentation-offload, generisegmentation-offload and generi-reeive-offload using ethtool to avoid paket dropping. Our lient is implemented using C to onnet and feth the MPD file and video layer-segments aording to the presented algorithm. The lient assumes a playout lateny of six seonds and a target level of ten seonds. If the buffer is depleted during the streaming, the appliation re-enters a buffering mode until it seures a sixseond portion in the buffer again. Wireshark is used to trae the transmitted pakets to estimate some performane metris. Our key performane indiies (KPIs) inlude the number of interrupts (n i ), the number of opened onnetions (n ), the number of losed onnetions due to low bandwidth (n ), the appliation downloaded data (d v ), the time at whih the appliation stops downloading more layer-segments (t d ), the appliation goodput (γ), and the average quality (q). The appliation goodput here is estimated as the ratio between the total transmitted video layer-segment sizes and the time spent in transmitting this amount of data. These KPIs are evaluated for different link bandwidth and delay onfigurations. B. Performane Results Figure plots the average reeived quality for the streamed video versus different bandwidth onfigurations for different min-max onnetion onfigurations and ms link delay.
AVG Quality (,) (,) (,) (,) (,) (,) (,) BW-Quality Rate in Kbps Figure : Segment Quality Versus Bandwidth for HTTP.. Quality (,) (,) (,) (,) (,) (,) Segment number Figure : Segment Quality versus segment number for HTTP.. Note that the average quality is estimated over all reeived segments. Intuitively, inreasing the bandwidth improves the quality of the reeived video for all onnetion onfigurations. Additionally, one an notie that the onnetion onfiguration has no signifiant impat on the reeived quality at a speifi onfiguration. Furthermore, the figure shows that a single onnetion represents a good hoie for the tested link onfigurations espeially at low bandwidth values. Figure shows the ahieved quality for eah segment onsidering different onnetion onfigurations for an average link-delay of ms and a bandwidth of Mbps. The figure provides us with some insights on the quality dynamis versus time. The figure shows that all the onnetion onfigurations managed to download the highest quality towards the video session end. However, the main differene lies in the the starting dynamis. At the beginning of the streaming, the adopted onservative approah for layer-segment request implies requesting base layer-segments only and limits the benefit from the available bandwidth to download enhanement layersegments. Note that this onservative approah helps reduing the initial playout lateny, whih is another important QoS metri. One the minimum amount of media is seured in the buffer, the lient starts to download enhanement layers for the subsequent segments. The figure also points out that the onnetion onfiguration has a diret impat on the streaming quality of the first few segments. Clearly, using a limited number of onnetions speeds the proess of quality improvement at the beginning. This result an be interpreted by the (,) n....... n....... d v (MB)....... t d (se)....... γ (kbps)....... q....... (a) kbps Bandwidth. n....... n....... d v (MB)....... t d (se)....... γ (Mbps)....... q....... (b) Bandwidth Mbps. Table II: KPIs for Bandwidth kbps and Mbps. TCP friendly behavior that splits the bandwidth among ative onnetions. Hene, as the lient opens more onnetions at the beginning, the bandwidth splitting leads to extended download time of every layer-segment. Consequently, suh behavior prevents requesting higher quality layer-segments based on the adopted horizontal enhanement layer-segment request. Table II shows some of the several KPIs for kbps and Mbps links with ms link delay. As a general observation, the onnetion onfiguration has limited impat on all the estimated KPIs exept for n and n. For any onfiguration, the algorithm opens more onnetions as the upper-bound inreases. Typially opening these onnetions is assoiated with a signaling overhead for a onnetion-oriented protool suh as TCP. However, many of these onnetions are also losed by the algorithm as their segment feth time is larger than the segment duration, espeially at low bandwidth. As an example, approximately onnetions out of opened onnetion for (,) onfiguration are losed. This indiates that the upper bound of the onnetions should be arefully hosen to math the operating onditions. Another n max in some onfigurations espeially those with small n max and high bandwidth. For example with (,) onfiguration, three onnetions are opened instead of one onnetion. The reason for this strange behavior is that the default server onfiguration implies that no more than objets should be transferred over the same persistent onnetion. Note that the number of ommuniated objets would typially sale with the number of enoded layers and video duration for a speifi segment size. Our results also indiate that under all the tested onfigurations, the algorithm manages to adapt the streaming to different bandwidths and no interrupts are enountered. Also, as the link bandwidth inreases, the appliation goodput (γ) inreases. This inrease is due to the ability to download more enhanement layer-segments and the typial drop of the download time (t d ). Consequently, a higher average quality is observed as the bandwidth inreases. Figure plots the average video quality versus the average link delay for different onnetion onfigurations assuming interesting observation is that n gets larger than n max
AVG Quality Delay-Quality Delay in mse (,) (,) (,) (,) (,) (,) (,) Figure : Quality versus Link Delay for HTTP.. n....... n....... d v (MB)....... t d (se)....... γ (kbps)....... q....... (a) Link Delay ms. n....... n....... d v (MB)...... t d (se)....... γ (kbps)....... q....... (b) Link Delay ms. Table III: Key performane indiies for different link delays. a bandwidth of Mbps. It is important to note that we set the variane of the link delay to % of the average value. The results indiate that the quality signifiantly drops as the delay inreases. A signifiant drop in the quality is notied as the delay inreases beyond ms. The figure also indiates that using parallel onnetions pays off for large link delays and variane. The (,) onfiguration generally shows the best performane for different link delays. Table III shows the KPIs for ms and ms link delays. Similar to the previous results, the obtained results also show that the algorithm suessfully streams the media without interruptions for different link delays. On the ontrary, the results show notieable variations for all the estimated KPIs for the same link onfiguration under different, n max ) values. The results also indiate that the number of losed onnetions due to limited bandwidth inreases with the inrease in delay for the onfiguration with high maximum onnetion bound. The appliation goodput also signifiantly deteriorates as the link delay inreases. Additionally, the estimated goodput shows signifiant improvement when two onnetions are used in omparison to the serial download. Additional goodput gain is notied between (,) and (,) for different link delays. V. CONCLUSIONS AND FUTURE WORK SVC brings promising benefits to DASH inluding improved adaptive streaming and storage espeially for highly heterogeneous networks. In this work, we presented the different key design omponents for SVC-DASH lient onsidering multiple onnetions. Additionally, we presented SVC-DASH- M as a multiple-onnetion threshold-based lient with horizontal layer-segment seletion poliy. Our experimental results show that using multiple onnetions has limited impat on the streaming performane for low link delays. On the ontrary, it shows a notieable improvement in the streamed quality for links with large delays. As a future work, we are interested in developing aurate estimators for layer-segment feth time, investigating and developing new layer-segment seletion poliies, and developing riteria for determining the optimal onnetion onfiguration parameters. REFERENCES [] T. Stokhammer, Dynami adaptive streaming over http : Standards and design priniples, in Pro. of the Seond Annual ACM Conferene on Multimedia Systems (MMSys ). New York, NY, USA: ACM,, pp.. [] H. Shwarz, D. Marpe, and T. Wiegand, Overview of the Salable Video Coding Extension of the H./AVC Standard, IEEE Trans. on Ciruits and Systems for Video Tehnology, vol., no., pp.,. [] Xiaoling Qiu et. al., Optimizing HTTP-based Adaptive Video Streaming for wireless aess networks, in Pro. rd IEEE International Conferene on Broadband Network and Multimedia Tehnology (IC- BNMT), ot., pp.. [] J. Famaey et. al, On the merits of SVC-based HTTP Adaptive Streaming, in in Pro. IFIP/IEEE International Symposium on Integrated Network Management (IM ),, pp.. [] R.Huysegems,B.D.Vleeshauwer, T.Wu,and W.V.Leekwijk, SVC- Based HTTP Adaptive Streaming, Bell Labs Tehnial Journal, vol., no., pp.,. [] S. Lederer, C. Müller, and C. Timmerer, Dynami adaptive streaming over http dataset, in Proeedings of the rd Multimedia Systems Conferene, ser. MMSys. New York, NY, USA: ACM,, pp.. [] S. de la Fuente et. al., idash: Improved Dynami Adaptive Streaming over HTTP Using Salable Video Coding, in in Pro. of the Seond Annual ACM Conferene on Multimedia Systems, ser. MMSys. New York, NY, USA: ACM,, pp.. [] Travis Andelin et. al., Quality seletion for dynami adaptive streaming over http with salable video oding, in Proeedings of the rd Multimedia Systems Conferene, ser. MMSys. New York, NY, USA: ACM,, pp.. [] C. Liu, I. Bouazizi, M. M. Hannuksela, and M. Gabbouj, Rate adaptation for dynami adaptive streaming over http in ontent distribution network, Image Commun., vol., no., pp., Apr.. [] Niels Bouten et. al, Minimizing the impat of delay on live SVC-based HTTP adaptive streaming servies, in Pro. IFIP/IEEE International Symposium on Integrated Network Management (IM ),, pp.. [] [Online]. Available: http://trae.eas.asu.edu/yuv/ [] H. Shwarz, M. Wien, and J. Vieron, Joint salable video model jsvm- software, Output Doument from Joint Video Team,. [] [Online]. Available: http://www.lart.org/