Dynamic Adaptive Streaming over HTTP Dataset

Stefan Lederer, Christopher Müller, and Christian Timmerer

ABSTRACT But leveraging existing and cheap HTTP infrastructures is not the
The delivery of audio-visual content over the Hypertext Transfer only reason for adaptive HTTP streaming. Due to the increasing
Protocol (HTTP) got lot of attention in recent years and with availability of mobile high speed Internet connections like
dynamic adaptive streaming over HTTP (DASH) a standard is WLAN/3G/4G and the huge smartphone boom in the recent years,
now available. Many papers cover this topic and present their mobile video streaming is becoming more and more important. As
research results, but unfortunately all of them use their own shown in Figure 1, the annual growth of mobile video will be
private dataset which –– in most cases –– is not publicly available. more than 100% over the next years until a total share of up to
Hence, it is difficult to compare, e.g., adaptation algorithms in an 66% of the total mobile traffic in 2014. However, basic
objective way due to the lack of a common dataset which shall be progressive HTTP download/streaming is not suitable for these
used as basis for such experiments. In this paper, we present our mobile environments with considerable high bandwidth
DASH dataset including our DASHEncoder, an open source fluctuations. The video stream has to adapt to the varying
DASH content generation tool. We also provide basic evaluations bandwidth capabilities in order to deliver the user a continuous
of the different segment lengths, the influence of HTTP server video stream without stalls at the best possible quality for the
settings, and, in this context, we show some of the advantages as moment, which is achieved by dynamic adaptive streaming over
well as problems of shorter segment lengths. HTTP [5][6].
In this context, the emerging ISO/IEC MPEG standard Dynamic
Categories and Subject Descriptors Adaptive Streaming over HTTP (DASH) [7] may get a lot of
D.5.1 [Multimedia Information System]: Video attention in the upcoming years and probably will become an
important as well as company-independent adaptive HTTP
General Terms streaming standard for all types of devices like PCs, Smartphones,
Algorithms, Measurement, Standardization, Documentation. Set-Top Boxes, etc. Hence, it is important to provide easy content
generation tools and test content for researchers and developers.
Keywords Unfortunately, there is currently no common test content
Dynamic Adaptive Streaming over HTTP, DASH, Dataset,
dedicated to DASH available, showing the basic possibilities of
Encoder, Content Generation Tool
the standard. In this paper we present our easy-to-use DASH
1. INTRODUCTION content generation tool called DASHEncoder, which enables the
Online video portals like YouTube or Netflix deploy progressive user to encode and multiplex DASH-compliant content in
download or adaptive video on demand systems and count segmented and continuous file structure with the associated Media
millions of users watching their content every day. Real-time Presentation Description (MPD). Furthermore, we provide a
entertainment produces already nearly 50% of the U.S. peak DASH dataset comprising full-length sequences of various genres
traffic nowadays, with almost 30% of the total U.S. downstream encoded at different representations in terms of bitrate, resolution,
traffic produced by Netflix [1]. Due to these enormous and quality.
infrastructure and bandwidth needs those costs became a huge The remainder of this paper is organized as follows. Section 2
problem for content service providers (CSP). describes related work. The content generation is described in
HTTP streaming and especially adaptive HTTP streaming got Section 3 and Section 4 provides details about the actual dataset.
much attention in the past few years. Proprietary systems like In Section 5 we use the dataset to evaluate the influence of the
Microsofts’’ Smooth Streaming [2], Adobes’’ HTTP Dynamic segment size and various Web server configurations to the
Streaming [3] or Apples’’ HTTP Live Streaming [4] provide streaming performance. The paper concludes with Section 6
similar approaches by leveraging existing content delivery including future work.
network (CDN) and proxy cache infrastructures which costs are
significant lower than dedicated streaming server like the Flash
2. Related Work
A detailed evaluation of the media formats of current adaptive
Media Server or other competing products.
streaming systems and a low overhead container format for
adaptive streaming is described in [8]. [9] analyzes the quality
Permission to make digital or hard copies of all or part of this work for improvements when using adaptive HTTP streaming over
personal or classroom use is granted without fee provided that copies are multiple heterogeneous network interfaces.
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
Microsoft Smooth Streaming is presented in [10] with two test
otherwise, or republish, to post on servers or to redistribute to lists, sequences of about 10 minutes of the open source movie Big
requires prior specific permission and/or a fee. Buck Bunny. They offer the clip in different resolutions up to
MMSys’’12, February 22-24, 2012, Chapel Hill, North Carolina, USA. 720p and 1080p respectively, with 8 quality levels from about 230
Copyright 2012 ACM 978-1-4503-1131-1/12/02...$10.00. kbit/s up to 3 Mbit/sec. The content generation for Microsoft

89
 Figure 2: DASHEncoder Content Generation Steps.
Figure 1: Mobile Data Volume Growth [5].
Smooth Streaming is done by the IIS Transform Manager, which # x264 Options
performs a fully automated transcoding of content to Microsoft #=========================
Smooth Streaming compatible fragmented MP4 files at different bitrate : 200|400|600|1000
statistics : stat.temp
quality levels.
gop : 48
Adobe offers test videos for their Dynamic HTTP Streaming profile : baseline
together with the standard installation of their Flash Media Server preset : slow
[11]. This includes sample encodings of approx. 2 minutes input : /home/stefan/foreman_cif.y4m
passes : 1
ranging from 150 kbit/s to 1500 kbit/s which comes together with
const-filesize : 0
the Flash Media Server. For content generation they also offer a # MP4Box Options
command line based packaging tool called File Packager which #=========================
performs the segmentation and packaging of the raw video/audio fragment-size : 2000
streams. The encoding of the videos has to be done separately, segment-size : 2000
e.g., using the Adobe Encoder or x264 as an open source rap-aligned
alternative [3]. segment-name : foreman
folder-prefix : foreman
Currently, the only available DASH content is the adapted 3GPP # MPD Options
Adaptive HTTP Streaming (AHS) test content of Ericsson Labs #=========================
which is available in[12]. They offer two clips, Mountain Bike mpd-name : foreman.mpd
(HVGA resolution) and Snowboard (VGA resolution) with a url-root : http://localhost/foreman_stats/
set-base-url
length of approx. 4 and 5 minutes respectively. This test content is
offered at different bitrates ranging from 200 kbit/s to 1200 kbit/s Figure 3: DASHEncoder Configuration File Example.
but it maintains the same resolution. However, this content uses platforms like YouTube, without any time consuming manual
segments with a length of 10 seconds which is a restriction when work of system administrators.
testing, e.g., high bandwidth fluctuations or adaption algorithms
like in [9]. The DASHEncoder is available open source at [14] with the aim
that other developers will join this project. The content generated
Finally, MP4Box from GPAC [13] supports the generation of with the DASHEncoder is compatible with [15] which can be
fragmented MP4 files and basic MPDs for separate used as a decoder and player respectively.
representations. At the time of writing of this paper, MP4Box
supports only one representation which is also used as the basis We developed the DASHEncoder in a very generic way which
for our DASHEncoder introduced in the next section. allows us to add new encoding and multiplexing tools with very
low effort and with the usage of the existing parameters and
3. Content Generation configuration files respectively. Even other video codecs and/or
For this dataset we developed a new content generation tool –– on container formats can be integrated in the same way. For example,
top of GPAC’’s MP4Box –– for DASH video on demand content in addition to the currently integrated AVC/MP4 tools presented
called DASHEncoder. Using our DASHEncoder the user does not in this paper, it is also possible to generate fragmented WebM
need to encode and multiplex each quality level of the final files which use the VP8 video encoder and a subset of the
DASH content separately. The DASHEncoder steps are depicted Matroska container format. Furthermore, it shall be also possible
in Figure 2. It generates the desired representations (quality/bitrate to integrate, e.g., the Adobe packaging tool for HTTP Dynamic
levels), fragmented MP4 files, and MPD file based on a given Streaming in DASHEncoder and use the same encoded AVC files
configuration file (e.g., as shown in Figure 3) or by command line for the generation of Adobe HTTP Dynamic Streaming compliant
parameters respectively. content.
Given the set of parameters the user has a wide range of For the current dataset we used our DASHEncoder with x264 [16]
possibilities for the content generation, including the variation of for the video encoding and MP4Box [13] for the multiplexing as
the segment size, bitrate, resolution, encoding settings, URL, etc., well as for the MPD generation on representation level. Based on
which is shown by the example of a DASHEncoder configuration these separate MPDs our software generates a combined MPD file
file in Figure 3. Due to this easy batch processing of multiple containing all representations. This MPD file fits to the generated
encodings to one final DASH source it is also possible to use this file structure of the DASHEncoder and the provided URL of the
application fully automatically, e.g., on user-driven social video configuration file so that one has no need to edit this file after the
content generation process.

90
 Table 1. Segment Length Quality Comparison of the Big Buck Table 3. Target Bitrate Levels for each Genre.
Bunny Sequence (300 kbit/sec., 480x360).
# Animation Sport Movie
Segment 1 sec. 2 sec. 4 sec. 6 sec. 10 sec. 15 sec. 1 50 kbit/s, 320x240 100 kbit/s, 320x240 50 kbit/s, 320x240
Length (24) (48) (96) (144) (240) (360) 2 100 kbit/s, 320x240 150 kbit/s, 320x240 100 kbit/s, 320x240
(GOP) 3 150 kbit/s, 320x240 200 kbit/s, 480x360 150 kbit/s, 320x240
PSNR 35,83 36,51 36,98 37,14 37,31 37,31 4 200 kbit/s, 480x360 250 kbit/s, 480x360 200 kbit/s, 480x360
5 250 kbit/s, 480x360 300 kbit/s, 480x360 250 kbit/s, 480x360
Table 2: Overview of our DASH dataset. 6 300 kbit/s, 480x360 400 kbit/s, 480x360 300 kbit/s, 480x360
7 400 kbit/s, 480x360 500 kbit/s, 854x480 400 kbit/s, 480x360
Name Source Quality Length Genre 8 500 kbit/s, 480x360 700 kbit/s, 854x480 500 kbit/s, 854x480
Big Buck Bunny [17] 1080p YUV 09:46 Animation 9 600 kbit/s, 854x480 900 kbit/s, 854x480 600 kbit/s, 854x480
Elephants Dream [18] 1080p YUV 10:54 Animation 10 700 kbit/s, 854x480 1,2 Mbit/s, 854x480 700 kbit/s, 854x480
Red Bull Playstreets 1080p 6 Mbit AVC 01:37:28 Sport 11 900 kbit/s,1280x720 1,5 Mbit/s,1280x720 900 kbit/s,1280x720
The Swiss Account [19] 1080p 6 Mbit AVC 57:34 Sport 12 1,2 Mbit/s,1280x720 2,0 Mbit/s,1280x720 1,2 Mbit/s,1280x720
Valkaama [20] 1080p 6 Mbit AVC 01:33:05 Movie 13 1,5 Mbit/s,1280x720 2,5 Mbit/s,1280x720 1,5 Mbit/s,1280x720
Of Forest and Men [21] SD 10:53 Movie 14 2,0 Mbit/s,1280x720 3,0 Mbit/s,1920x1080 2,0 Mbit/s,1920x1080
15 2,5 Mbit/s,1920x1080 4,0 Mbit/s,1920x1080 2,5 Mbit/s,1920x1080
In order to retrieve the best suitable bitrate vs. resolution 16 3,0 Mbit/s,1920x1080 5,0 Mbit/s,1920x1080 3,0 Mbit/s,1920x1080
configuration we analysed the PSNR values of the encoded 17 4,0 Mbit/s,1920x1080 6,0 Mbit/s,1920x1080 4,0 Mbit/s,1920x1080
content of several overlapping bitrate configurations. For this we 18 5,0 Mbit/s,1920x1080 5,0 Mbit/s,1920x1080
decoded the encoded videos with ffmpeg and upscaled the lower 19 6,0 Mbit/s,1920x1080 6,0 Mbit/s,1920x1080
resolution encodings with a bilinear scaling filter to the resolution 20 8,0 Mbit/s,1920x1080
of the source video. Based on this we calculated the PSNR value usually faces a tradeoff between high and low motion scene, fade-
of each encoding compared to the original source video. Using in and fade-out, and also quality changes. It is also not always
this metric we retrieved the best bitrates to change the resolution possible to achieve a continuous constant bitrate of an encoded
of the encoded content to provide a continuous quality increase video, especially at higher bitrates due to, e.g., low motion scenes.
through low to high bitrate levels. Therefore, full-length sequences are needed in order to provide
For the subsequent multiplexing and segmentation step it is more realistic results.
necessary to maintain fixed I-frame positions of the video which Our DASH dataset comprises the sequences as shown in Table 2
is achieved by restricting the GOP-size of the encoder to the and located at [22]. All animation and movie sequences are
desired segment size of the content. As a consequence, smaller licensed under Creative Commons-BY 2.0, so they are free to
segment sizes have a disadvantage because of their high number share and modify. Both sport sequences are free to use for
of I frames which need more bits in the overall bitstream. research purposes.
Furthermore, smaller GOP sizes also provide a lower encoding
performance and quality which is also shown in [1]. Table 1 4.2 Representations, Segment Lengths and
shows the PSNR values for different segment sizes and provides
evidence that this needs to be considered in the evaluation process
MPDs
for the segment sizes of DASH streaming systems. Table 3 shows the representation levels of our dataset. We
decided to distinguish the quality level configuration between
The time needed for the content generation of one representation types of video and to offer at least two sequences with the same
in 1080p resolution based on a YUV source video and using 2 configuration. This is useful because of the different scene
Xeon 2,33 GHz CPUs corresponds to the video playback time on characteristics (high motion in sport, low detail in Animation,
average, however lower resolution representations can be etc.) of the chosen video types. It is also important to mention that
generated significantly faster. In this process the time needed is the sport and movie genres have only a maximum quality of 6
mainly dependent on the x264 encoder and the settings used Mbit/sec. due to fact that our source video files are encoded at 6
therefore, the multiplexing and MPD generation in contrast is Mbit/sec. Another limitation arises at the animation videos due to
done in less than 5 % of the total time. the lack of achieving the target bitrate of the x264 encoder at
higher bitrates for this type of content. Therefore, the higher
bitrate encodings are finally encoded at a lower bitrate than
4. DASH Dataset desired (the MPDs entries represent the bitrate of the encoded
4.1 Introduction videos).
During our research on adaptive video streaming and DASH we
Additionally, the DASH dataset is encoded and multiplexed using
always had the problem that there is no DASH test content
different segment sizes inspired by commercial products ranging
available to perform objective comparisons of streaming
from 2 seconds (i.e., Microsoft Smooth Streaming [2]) to 10
algorithms and systems. In general there is hardly any DASH
seconds per fragment (i.e., Apple HTTP Streaming [4]) with some
content available at all, except the two test sequences from [12].
steps in between and at the lower end. Thus, we end up with the
Some papers used common test sequences like Foreman et al. for segments sizes of 1, 2, 4, 6, 10 and 15 seconds. The length of the
their video streaming tests, which are in general quite short for segments is a very important decision when providing content for
those kind of experiments, in the case of Foreman the length is adaptive streaming. Shorter segments may provide a larger
300 frames. One approach is to concatenate the same sequence overhead in term of requests, which is shown in the evaluation
over and over again in order to have a longer sequence for the section, and result in more quality switches, with the cost of a
evaluation. This approach has the disadvantage that the video lower encoding efficiency. On the other hand, longer segments are
characteristics are quite the same over the whole experiment not very handy in environments with high bandwidth fluctuations
which is often not the case in the real-world. In such cases one like mobile networks, because the client cannot adjust the video

91
 Figure 5: Network Emulation Environment.
and proxies, new applications on top of DASH, etc. The main
advantage of our dataset is that it now enables the objective
comparison of research results across different institutions
working in this field. The dataset and a short compendium of its
characteristics can be found in [22] where it is possible either to
stream the video directly from our Web server or download the
files via FTP.

5. Evaluation
5.1 Introduction
Together with the dataset we provide some basic evaluations of
the streaming characteristics of the different segment length
versions of the videos and their impact on the average bitrate of
the streamed video. Therefore, we evaluate the streaming
Figure 4: PSNR Value Example of different Representations. performance of each segment length version in terms of average
stream as fast as it would be needed in case of a significant bitrate and interpolate a graph to find the optimal segment length
bandwidth drop. for the given network scenario.
In this dataset we offer any content in a version with separate During our first tests we realized different Web server
segments and the corresponding MPD. Additionally, we also offer configurations on our hosting provider’’s sites. Some of them
an unsegmented version of the videos and the corresponding MPD allow persistent connections, as defined in HTTP 1.1 [23], and
for the movies of the animation genre, which allows for byte- some restrict the reuse of connections for multiple requests by
range requests. The provided MPDs of the dataset are based on closing the connection after the response. This is signaled by the
the 2nd Draft International Standard (DIS) of DASH [7] and are Connection: close field of the HTTP 1.1 response header. Other
also available in the version compatible with the current providers limit the Keep-Alive-Time to 1 second, which is quite
implementation of the DASH VLC Plugin [15]. short in comparison to the 15 second standard configuration of the
Apache Web server [24]. Due to this difference in todays’’
4.3 PSNR Statistics available Web server configurations we also decided to analyze
The average bitrate of the streamed video is a good metric for the impact thereof w.r.t. the streaming performance and segment
DASH clients and their stream switching algorithms. But the lengths.
bitrates do not necessarily correspond to the quality of the video,
especially because of the differences in encoding efficiency 5.2 Evaluation Setup
between high- and low-motion scenes. Usually the encoder As a first step, we built a test environment (shown in Figure 5 and
compensates those differences by adjusting the bitrate over the similar to [25]) emulating standard Internet connections to show
whole video to get the best overall quality. However, due to the the impact of the HTTP server configuration on DASH.
constant bitrate of DASH representations, which is more or less Therefore, we used a standard Apache Web server as source of our
constant all over the video (or at least the current DASH-Period), DASH streaming scenario which has a wide range of
this is not possible in our scenario. Thus, it can be useful to configuration possibilities regarding connections. The DASH
consider the PSNR as a metric for stream switching algorithms in client is a modified version of the DASH VLC plugin [15], with
certain cases. some minor changes regarding logging. Additionally, we used two
Therefore, we also provide PSNR values on frame level of the Big Ubuntu nodes which act as gateways to emulate the typical
Buck Bunny encodings. Those values are available for each frame network characteristics of a last mile Internet connection (e.g.
of each representation and each segment length version of this ADSL). One node –– network emulation –– emulates the Internet
video, which sums up in more than 1.7 million PSNR values. One service provider characteristics by adding a delay of 150 ms using
can find them together with the dataset in comma separated value NetEm (packet loss/packet reordering are currently not considered
(CSV) as well as Microsoft Excel files [22]. As an example, to keep this example evaluation simple). The other node –– traffic
Figure 4 shows the PSNR values of 4 representations of about 850 shaping –– simulates scenarios of fluctuating bandwidth
frames of the 2-second segment length version of our Big Buck conditions.
Bunny encoding.
5.3 System Parameters
4.4 Dataset Usage For our evaluations we used a standard installation of the Apache
Our dataset can be used for many kinds of evaluations in the Web server and enabled/disabled the KeepAlive On/Off setting in
context of MPEG-DASH like e.g. the comparison of different the httpd.conf to show the impact of persistent and non-persistent
stream switching algorithms, the caching performance of CDNs connections. As already mentioned we simulated the network by

92
 Figure 6: Non-/Persistent Connetions - 2 sec. segments. Figure 8: Segment length/Bandwidth Evaluation.
Segment sizes greater than 6 seconds are not influenced so much
by the connection settings of the Web server. Due to the lower
amount of segments and, thus, the smaller overhead produced by
reconnects, the difference between both configurations reduces by
2% for the 15-seconds segment length version of Big Buck Bunny
(cf. Figure 7). One can see the pretty equal media bitrate need of
both evaluations as well as the lower amount of requests shown
by squares and triangles in the plot. The requests are not shown in
Figure 6, due to the high frequency of the requests (i.e., every two
seconds) compared to one request every 15 seconds shown in
Figure 7.
Finally, we evaluated the optimal segment size of our network
configuration scenario for both cases, with and without persistent
connections. Therefore, we analyzed the performance of the 1, 2,
4, 6, 10 and 15 seconds segment length of Big Buck Bunny of our
dataset and interpolated a graph showing the performance of the
segment sizes in term average media bitrate. As shown in Figure 8
Figure 7: Non-/Persistent Connetions - 15 sec. segments. the optimal segment size for this network setting would be
adding a delay of 150 ms and introducing bandwidth fluctuations. between 2 and 3 seconds using persistent connections and
The bandwidth varies over time with a minimum of about 386 between 5 and 8 without using them. The bitrate of the optimal
kbit/sec. and a maximum of 1536 kbit/sec. for the downlink. For segment lengths of both configurations differs by about 50
our evaluation we used the Big Buck Bunny movie of our dataset, kbit/sec.
which is available in 6 different segment lengths of 1, 2, 4, 6, 10,
and 15 seconds, each with 21 representations. 6. Conclusion and Future Work
In this paper we provide, to the best of our knowledge, the first
5.4 Results DASH dataset comprising different genres, most of them in HD
Our evaluations showed a significant performance drop by quality. In comparison to other sources for video test content,
streaming from a Web server which does not allow persistent which mostly offer clips with only a few seconds of length, this
connection, especially for short segments lengths like 2 seconds or dataset is especially designed for DASH-based evaluations and
less. One can see this in Figure 6, which shows the comparison consists of videos between 10 and 90 minutes.
between the DASH streaming performance of the 2-second
Each video of this dataset is provided in different segment lengths
segment length version of Big Buck Bunny with and without
as well as in different representations ranging from 50 kbit/sec. up
persistent connections.
to several MBit/sec. One can use this dataset to compare DASH-
In our evaluation the streaming using persistent connections based streaming clients/systems in an objective way with a
corresponds with the findings from [25] and [26] that for TCP- common content basis. Additionally, we provide PSNR values for
based streaming one needs about the double bandwidth of the every frame of the Big Buck Bunny representations, which
media bitrate to achieve a sufficient performance. Streaming provide even more evaluation possibilities for researchers.
without persistent connections results in a significantly higher
bandwidth demand and the average media bitrate in our We also developed a convenient content generation tool called
evaluation of 2-seconds segment in Figure 6 is about 34% lower DASHEncoder, which we used for the encoding of this dataset.
than using persistent connections. The average bitrate in the worst This fully configurable and extensible DASH encoding tool is
scenario with 1-second segments drops by 46% in comparison to a offered as open source at [14].
Web server with persistent connections enabled as shown in
Figure 8.

93
Finally, we showed an evaluation of the optimal segment size for Multimedia systems (MMSys11), ACM, New York, NY,
a given network configuration based on our dataset. We also USA
showed the influence of enabled versus disabled persistent [10] Smooth Streaming Showcase,
connections of Web servers to the streaming performance of http://www.iis.net/media/experiencesmoothstreaming, (last
DASH. Therefore, we build a simulation environment for access: Dec. 2011).
performing tests to get the optimal segment size for those different
scenarios. [11] Flash Media Server,
http://www.adobe.com/products/flashmediaserver/, (last
Based on this dataset we plan to do research on stream switching access: Dec. 2011)
algorithms and DASH-based systems in general. We also plan to
update and extend our dataset in the future, like adding alternative [12] Ericsson AHS Content at ITEC/Alpen-Adria-Universität
encodings or audio tracks to the existing videos. Additionally, we Klagenfurt, http://www-itec.uni-klu.ac.at/dash/?page_id=6
will include versions with different sub-segment lengths and other (last access: Dec 2011)
media formats. [13] Le Feuvre, J., Concolato, C., Dufourd, J.-C., Bouqueau, R.,
Moissinac, J.-C. 2011. Experimenting with Multimedia
7. ACKNOWLEDGMENTS Advances using GPAC, Proceedings of ACM Conference on
This work was supported in part by the EC in the context of the Multimedia (ACM MM) 2011, ACM, New York, NY, USA
ALICANTE (FP7-ICT-248652) and SocialSensor (FP7-ICT-
[14] DASHEncoder at ITEC/Alpen-Adria-Universität Klagenfurt,
287975), and QUALINET (COST IC 1003) projects. Special
http://www-itec.uni-klu.ac.at/dash/?page_id=282 (last
thanks to the Red Bull Media House for providing us the Red Bull
access: Dec. 2011)
Playstreets video. They own the rights of the content but the usage
for scientific purposes is permitted. Special thanks also to the [15] Müller, C., Timmerer, C. 2011. A Test-Bed for the Dynamic
team of Valkaama and to the producer Tim Baumann for Adaptive Streaming over HTTP featuring Session Mobility,
providing us this content. In Proceedings of the second annual ACM conference on
Multimedia systems (MMSys11), ACM, New York, NY,
8. REFERENCES USA
[1] Schonfeld, E., Tech Crunch, 2011. Netflix Now The Largest [16] X264, http://www.videolan.org/developers/x264.html, (last
Single Source of Internet Traffic In North America (posted access: Dec. 2011)
on May 17th, 2011),
http://techcrunch.com/2011/05/17/netflix-largest-internet- [17] Big Buck Bunny Movie, http://www.bigbuckbunny.org (last
traffic/ (last access: Dec. 2011) access: Dec. 2011)
[2] Microsoft Smooth Streaming, [18] Elephants Dream Movie, http://www.elephantsdream.org,
http://www.iis.net/download/smoothstreaming (last access: (last access: Dec. 2011)
Aug. 2011). [19] The Swiss Account, http://lt11.com/2011/07/18/the-swiss-
[3] Adobe HTTP Dynamic Streaming, account-2 (last access: Dec. 2011)
http://help.adobe.com/en_US/HTTPStreaming/1.0/Using/ind [20] Valkaama Movie, http://www.valkaama.com, (last access:
ex.html, (last access: Dec. 2011) Dec. 2011)
[4] Pantos, R., May, W. 2010. HTTP Live Streaming, IETF [21] Of Forest and Men, http://www.offorestsandmen.org, (last
draft (Jun. 2010) http://tools.ietf.org/html/draft-pantos-http- access: Dec. 2011)
live-streaming-04 (last access: Dec. 2010).
[22] DASH Dataset at ITEC/Alpen-Adria-Universität Klagenfurt,
[5] Stockhammer, T. 2011, Dynamic Adaptive Streaming over http://www-itec.uni-klu.ac.at/dash/?page_id=207 (last
HTTP –– Design Principles and Standards, In Proceedings of access: Dec. 2011)
the second annual ACM conference on Multimedia systems
[23] Fielding, R. et al, RFC 2068 - Hypertext Transfer Protocol --
(MMSys11), ACM, New York, NY, USA
HTTP/1.1,
[6] Sodagar, I., The MPEG-DASH Standard for Multimedia http://www.w3.org/Protocols/rfc2616/rfc2616.html (last
Streaming Over the Internet, IEEE Multimedia, vol. 18, no. access: Dec. 2011).
4, pp. 62-67, Oct.-Dec. 2011
[24] Apache HTTP Server Project, http://httpd.apache.org/ (last
[7] ISO/IEC DIS 23009-1.2, Information technology —— access: Dec. 2011).
Dynamic adaptive streaming over HTTP (DASH) —— Part 1:
[25] Kuschnig, R., Kofler, I., Hellwagner, H. 2010. An
Media presentation description and segment formats.
Evaluation of TCP-based Rate-control Algorithms for
[8] Riiser, H., Halvorsen, P., Griwodz, C., Johansen, D. 2010. Adaptive Internet Streaming of H.264/SVC, Proceedings of
Low overhead container format for adaptive streaming. the First Annual ACM SIGMM Conference on Multimedia
Proceedings of the First Annual ACM SIGMM Conference Systems (MMSys), ACM, New York, NY, USA
on Multimedia Systems (MMSys), ACM, New York, NY,
[26] Wang,B., Kurose, J., Shenoy, P., and Towsley,D. 2008.
USA
Multimedia streaming via TCP: An analytic performance
[9] Evensen, K., Kaspar, D., Griwodz, C., Halvorsen, P. 2011. study.ACMTrans.Multimedia Comput. Commun. Appl. 4, 2,
Improving the Performance of Quality-Adaptive Video Article 16 (May 2008)
Streaming over Multiple Heterogeneous Access Networks, In
Proceedings of the second annual ACM conference on

94