Reducing startup time in MP4 on - demand video streaming services with movie atom caching

lication, MAC
are likely to play by the user. Such files can be daemon, will monitor the publishing area for new
obtained by using a recommendation system that files and automatically split each new file into
belong to content management system of a video two files containing moov atom and mdat atom
site, for example. respectively. As examplified in figure 5, v1.atom
 Although MAC algorithm is very simple, and v1.mdat are the two files splitted from the
implementing it into good working software original file v1.mp4. Those files will be moved
applications needs following requirements: to ”public area” and become ready for streaming.
 When a MAC client is commanded to view a
 • First of all, MAC should be extensions MP4 video, it will check for an apropriate atom
 or add-on components of existing solutions file in its local video cache. If such a file is
 for streaming MP4 files. Development found, it will be used as the atom header for
 of a new streaming server software is parsing multimedia data retrieved from server. In
 considered a bad design choice because of case MAC client has to download atom file from
 high implementation costs. server, the successfully downloaded atom file will
 also stored into the local video cache.
 • The streaming service for MAC protocol At server side, existing HTTP-based solutions
 should be compatible with non-MAC for streaming atom and mdat files of MP4
 clients. Again, developing a completely new movies are used. Specifically, files in public
 streaming server software is not a good idea area and publishing area are stream by typical
 because of potential incompatibility with webserver like nginx [10]. This can be
 existing clients. done simply by configuring those locations as
 accessible resources to the webserver and the web
 • Finally, with similar reasons as above, ones application hosted by the webserver. Additionaly,
 should not develop a new video player for a simple web application developed in PHP is
 MAC. Insteads, plugins should be developed deployed on the webserver for handling both
 to integrate with existing video players. MAC non-MAC (normal HTTP) MP4 video
38 X. T. Hoang, T. T. Nguyen / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 1 (2016) 33–41
 4. Performance Evaluation
 In this section, we present our setups for
 comparison startup delays between MAC and
 other protocols including progressive download,
 and HLS. In our experiments, the following
 elements are deployed:
 • A server machine on which: an FTP service
 (vsftpd) is running for allowing MP4 file
 uploads. An instance nginx webserver
 with supports for PHP 5 over fastCGI
 [11] are launched for providing HTTP
 streaming capabilities. And MAC daemon
 application are running for MAC protocol
 file preprocessing. All thosecomponents are
 run on a virtual 32-bit CentOS 6.x machine.
 • A client machine on which: required
 Fig. 4: Atom Caching/prefetching algorithm for startup
 delay reduction. player and its dependencies are installed.
 Those software components include ffmpeg
 library and MAC protocol plugin. All of
 them run on a Ubuntu 14.04 LTS 32-bit
 virtual machine.
 • And a virtual network setup by GNS3
 simulator [12]. In our setup, the virtual
 network contains only a virtual Cisco 2691
 router and two virtual links connecting the
 router with the server and client machines
 above. Such a simple virtual network is
 sufficient for us to create different values
 of end-to-end bandwidth between the client
 Fig. 5: Movie Atom Caching implementation. machine and the server machine.
 Our experiments for measuring startup delay
streaming. The web application will conceal are conducted as follows: Startup delays are
atom and mdat files under reference to original measured on 3 sample videos with the same
MP4 files and provide a layer for providing settings of bitrate, frame-rate, I-frame rate,
compatibility with both streaming protocols. and other parameters related to audio streams.
 Our MAC client is implemented using FFmpeg Table 2 summarizes those parameters. The
[6]. Specifically, we developed a protocol plugin three sample movies have durations, respectively,
for FFmpeg. Our plugin simply interfaces with 30, 60, and 90 minutes. The sample videos
streaming service and local video cache for are used for streaming with MAC, HLS, and
handling atom retrieval, caching and combining progressive download protocol under end-to-end
with mdat data from server. Thanks to high bandwidths of 512 Kbps and 2 Mbps. We believe
portability of FFmpeg, our MAC implementation that 2 Mbps is the typical access bandwidth
can be built for both Windows and of ADSL subscribers on the internet currently.
Linux platforms. And 512 Kbps can be used as a representative
 X. T. Hoang, T. T. Nguyen / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 1 (2016) 33–41 39
 Table 2: Parameters of sample MP4 files large as 60 seconds), MAC results in much
 smaller streaming failure rates in all cases.
 Specifically, streaming failure rates are less
 Parameter Value
 than 10%, for k = 0, and less than 1%
 Video frame rate 30 frames/s
 for k > 0.
 Video Frame size 1280 x 720 pixels
 I-Frame rate 1 I-frame per 50 frames • Startup delay of MAC is slidely smaller than
 Audio bitrate 192 Kbps HLS. The reason that HLS has low startup
 Audio sample rate 48 kHz delay is as follows. HLS is a combination of
 Number of audio channels 2 channels smaller segments. Each small segment will
 require only small header atom, thus incurs
 relatively small delay to start. Although
 reencoding the original file into multiple
residual bandwidth of such internet subscribers
 smaller files help in reducing startup delay,
during normal working conditions. Ones can
 it incurs more pre-processing cost. Total
say that 512 Kbps is a representative value
 data of all segment files are higher than the
for typical network condition, while 2 Mbps is
 original file. And data rate of the video
the representative one for very good network
 streaming will also higher.
conditions. For each sample video and a value
of the above bandwidths, 5 runs are performed
and the final result is obtained by averaging over 5. Related Work
the 5 results. Since we are interested only in how
startup delay can be reduced thanks to caching In the past, most streaming solutions used
moov atom, we let the the MAC client has moov streaming protocols such as RTP as a multimedia
atom the requested movie in all those runs. For transport protocol and RTSP as session control
the first time of requesting a MP4 video, where protocol. Today, popular video streaming
its moov atom is not cached yet, startup delay of services are exclusively based on HTTP. Major
MAC client actually is identical to that of HTTP advantages of HTTP-based streaming includes
progressive download. simplicity in deployment, firewall-friendly
 Figure 6 and figure 7 show comparisons traffics, and already available at almost all client
between MAC and progressive download in terms platforms. HTTP progressive download [2]
of startup delay under typical access bandwidth of best illustrates those advantages. A content
512 Kbps and large access bandwidth of 2Mbps. provider who wants to provide video streaming
Using results from those figures and models simply host MP4 files on its website, and then
described in 2 (equation 9, 10), we can conclude clients can enjoy view video streams only with
the followings: web browsers.
 For MP4 streaming over HTTP, it is crucial to
 • If waiting time threshold T0 that a user download moov atom of MP4 files in the first
 can wait is not higher than 20 seconds, place. To achieve that, a couple of methods
 progressive download is not acceptable to exist. Firstly, software tools, such as FFmpeg
 users in most of the cases. Meanwhile, MAC [6] and MP4 FastStart [13] can be used to
 provides relatively good user experience relocate moov atom to the begining of the file.
 with streaming failure rates are less than When movie files are progressively downloaded
 40%, 20%, 10%, 5% for patience factor k by clients, naturally, moov atoms are downloaded
 equals 0, 1, 2, 3 respectively. first and video playback can be started right away.
 Another method for downloading moov atoms
 • If some users are sufficiently patient to wait in the first place is used by some video players,
 until moov atoms are retrieved (T0 is as e.g. VLC and iOS’s video player. Following this
40 X. T. Hoang, T. T. Nguyen / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 1 (2016) 33–41
 Fig. 6: Startup delay comparisons (512 Kbps access Fig. 7: Startup delay comparisons (2.0 Mbps access
 bandwidth). bandwidth).
method, a client opens a http connection with a and DASH, a more advance file format called
standard HTTP Range request [14] to download fragmented MP4 is used for multimedia segments
the mp4 file from the begining of the file. As and XML-based format is used to packaging
long as it detects that those first bytes are not segments together. Since segments are small
moov atom, another HTTP Range request is sent video chunks with about dozens of seconds in
to server for requesting moov atom at the end of durations, segment atom headers are small in
the file. size and can be retrieved shortly. Startup delay
 Although the two methods improve HTTP is thus not as high as progressive download.
progressive download for MP4 streaming, they However, those solutions are not as popular as
do not solve the long startup delay problem MP4 format because of their newly introduced
as pointed out in this paper. By caching standards and less adopted in existing hardware
and prefetching moov atoms, MAC can far and software. Also, a large number of video
more improve user experience over progressive chunks could lead to several disadvantages
download while keeping simplicity and ease of including difficult in managing multimedia assets
deployment for the system. It is especially true and higher input/output operations on harddisks.
for streaming of full-length movies.
 Apple’s HTTP Live Streaming (or HLS),
Microsoft Smooth Streaming (MSS) [15], and
Dynamic Adaptive Streaming over HTTP (or
DASH) [16] are video streaming procotols that In comparisons with fragmentation approach
allows delivery of video content over HTTP. (like HLS, MSS, and DASH), MAC provides
They can also provide advanced features like a easy-to-use yet efficient improvement to an
adaptive bitrate streaming [17]. In all those already popular scheme. It does not require
solutions, the original media is splitted into reencode video into multi-file media assets, thus,
segments. Each segment can be seen as a small it produces less complexity and overheads. The
file with atom headers and data. Additional disadvantage is that, it is relatively limited
”packaging formats” are introduced to combine in providing advanced feature like streaming
all segments into a playlist to form a complete with adaptive bitrate. However, we believe
video content. In case of Apple HLS, multimedia that MAC can be used in combinations with
segments have MPEG-2 transport stream (TS) fragmented streaming solutions, e.g., caching and
files, and packaging format is simply a text prefetching atoms of segments, to improve further
file whose extension is ”.m3u8”. In MSS streaming experience.
 X. T. Hoang, T. T. Nguyen / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 1 (2016) 33–41 41
6. Conclusions [4] F. Dobrian, V. Sekar, A. Awan, I. Stoica, D. Joseph,
 A. Ganjam, J. Zhan, H. Zhang, Understanding the
 In this paper, we have shown that in streaming Impact of Video Quality on User Engagement, ACM
schemes where MP4 files are used as multimedia SIGCOMM Computer Communication Review 41 (4)
assets, time for downloading moov atoms, a (2011) 362–373.
 [5] X. Liu, F. Dobrian, H. Milner, J. Jiang, V. Sekar,
data segment in MP4 files which is important I. Stoica, H. Zhang, A Case for a Coordinated
for decoding process, is noticeable. This Internet Video Control Plane, in: Proceedings
amount of time leads to long startup delays of the ACM SIGCOMM 2012 Conference on
and poor quality of experience. User behavior Applications, Technologies, Architectures, and
 Protocols for Computer Communication, SIGCOMM
models were presented to show how startup ’12, ACM, New York, NY, USA, 2012, pp. 359–370.
delay affects the streaming system. Analysis on doi:10.1145/2342356.2342431.
the models reveals that reducing startup delay [6] F. Bellard, M. Niedermayer, et al., Ffmpeg,
can greately improve user quality of experience 
 [7] V. Organization, VLC Media Player,
and keep users using the system service. In  (2006).
order to cut down startup delay, a caching [8] G. Lawton, LAMP Lights Enterprise Development
mechanism, called MAC, was proposed and Efforts, Computer 38 (9) (2005) 0018–20.
implemented. Our performance evaluations [9] R. Pantos, W. May, HTTP Live Streaming draft-
 pantos-http-live-streaming-05, Published by the
showed that MAC improves startup delay Internet Engineering Task Force (IETF).
significantly in MP4 file streaming applications [10] W. Reese, Nginx: the High-Performance Web Server
and have competitive startup delay with modern and Reverse Proxy, Linux Journal 2008 (173) (2008)
fragmented streaming schemes with very little 2.
 [11] M. R. Brown, FastCGI: A High-performance Gateway
complexity and overhead. Interface, in: Fifth International World Wide Web
 Conference, Vol. 6, 1996.
Acknowledgments [12] Y. WANG, J. WANG, Use gns3 to Simulate
 Network Laboratory, Computer Programming Skills
 This work was supported by the project & Maintenance 12 (2010) 046.
CN.14.02 funded by VNU University of [13] MP4 FastStart, 
 Accessed: 2015-12-04.
Engineering and Technology. [14] R. Fielding, Lafon, Y., Ed., and J. Reschke, Ed.,”
 Hypertext Transfer Protocol (HTTP/1.1): Range
References Requests, Tech. rep., RFC 7233, June (2014).
 [15] A. Zambelli, IIS smooth Streaming Technical
 [1]N.F arber,¨ S. Dohla,¨ J. Issing, Adaptive Progressive Overview, Microsoft Corporation 3.
 Download Based on the MPEG-4 File Format, Journal [16] T. Stockhammer, Dynamic Adaptive Streaming Over
 of Zhejiang University SCIENCE A 7 (1) (2006) 106– HTTP–: Standards and Design Principles, in:
 111. Proceedings of the Second Annual ACM Conference
 [2] P. Gill, M. Arlitt, Z. Li, A. Mahanti, Youtube on Multimedia Systems, ACM, 2011, pp. 133–144.
 Traffic Characterization: A View From the Edge, in: [17] Melnyk, Miguel A and Stavrakos, Nicholas J and
 Proceedings of the 7th ACM SIGCOMM Conference Penner, Andrew and Tidemann, Jeremy and Breg,
 on Internet Measurement, ACM, 2007, pp. 15–28. Fabian, Adaptive Bitrate Management for Streaming
 [3] Information Technology – Coding of Audio-visual Media Over Packet Networks (2011).
 Objects – Part 14: MP4 File Format (2003).