Network characteristics of a
2 likes986 views
Network Characteristics of Video Streaming Traffic in YouTube 1. The study analyzes the network characteristics of video streaming on YouTube by examining the differences between PC and mobile streaming strategies and user behavior. 2. It finds that PC streaming uses a single TCP connection while mobile uses multiple connections to cope with storage constraints, and that streaming strategies differ based on application, container, and device used. 3. It also observes that users generally do not change video resolutions, watch only a portion of videos before abandoning them, and prefer short videos, with mobile streams wasting more bandwidth than PC streams due to inefficient buffer handling.
![Understanding user behavior
[2] Alessandro Finamore, Marco Mellia, Maurizio M. Munafò, Ruben Torres, and Sanjay G. Rao
YouTube everywhere: impact of device and infrastructure synergies on user experience.
In Proceedings of the 2011 ACM SIGCOMM conference on Internet measurement conference (IMC
'11).](https://image.slidesharecdn.com/termpaper-presentation-130504171847-phpapp01/85/Network-characteristics-of-a-19-320.jpg)
![Overall waste of bandwidth
Overall the wasted amount of data during peak hours
⢠for PC-player, 39%
⢠for Mobile-player, 47%
160Mb/s of YouTube
@ peak hours
67Mb/s of
traffic is wasted!
Bitrate[Mb/s]](https://image.slidesharecdn.com/termpaper-presentation-130504171847-phpapp01/85/Network-characteristics-of-a-32-320.jpg)
Network characteristics of a
- 1. Network Characteristics of Video Streaming TrafďŹc in YouTube - Deepti Ghadiyaram
- 2. References 1. Ashwin Rao, Arnaud Legout, Yeon-sup Lim, Don Towsley, Chadi Barakat, and Walid Dabbous. Network characteristics of video streaming traffic. In Proceedings of the Seventh COnference on emerging Networking EXperiments and Technologies (CoNEXT '11) 2. Alessandro Finamore, Marco Mellia, Maurizio M. Munafò, Ruben Torres, and Sanjay G. Rao. YouTube everywhere: impact of device and infrastructure synergies on user experience. In Proceedings of the 2011 ACM SIGCOMM conference on Internet measurement conference (IMC '11).
- 3. Motivation
- 4. ⢠One of the most popular and bandwidth intensive video streaming services of todayâs internet. ⢠Accounts for 20-30% of the internet traffic. ⢠Over 4 billion hours of video watched each month. ⢠72 hours of video uploaded every minute. ⢠25% of global YouTube views come from mobile devices. ⢠About 1 billion views on YouTube mobile per day. Source: http://www.youtube.com/yt/press/statistics.html ISP1 ISP2 YouTube CDN Data Center
- 5. Containers Desktop Browsers Native Mobile Applications What are the Network Characteristics of Video Streaming Traffic?
- 6. ⢠What exactly happens during video streaming? ⢠Strategies to stream videos ⢠How differently are PC and mobile requests handled? ⢠How do different strategies impact the network performance? ⢠How differently do users watch videos on YouTube? ⢠Resolution switch. ⢠Average duration of the content Specifically,
- 7. How are streaming requests handled?
- 8. What are the similarities/differences in the evolution of the download? PC Mobile
- 9. PC-player Download Mobile-player Download Temporal evolution of the HTTP messages exchanged between the client and the YouTube servers
- 10. Download scheme comparison ⢠PC player ⢠Download controlled by the video server ⢠1 video session = 1 TCP connection ⢠Mobile player ⢠Download controlled by the client ⢠1 video session = N TCP connections ⢠Reason for this behavior in mobile player ⢠To cope with tighter constraints in terms of storage availability for mobile devices. ⢠Usually, only a fraction of the complete video is watched, thus this strategy saves network bandwidth.
- 11. Generic Behavior of Video Streaming
- 12. Data ⢠YouTube videos ⢠Flash, HTML5, and HD (Flash) ⢠Mobile ⢠Locations ⢠France ⢠Academic (Wired; Wi-fi for mobile) ⢠Residential (Wi-fi) ⢠USA ⢠Academic (Wired; Wi-fi for mobile) ⢠Residential (Wired) Dataset Size or source YouFlash 5000 from searching on browser YouHD 2000 from searching on browser YouHtml 2500 from YouFlash + 500 from YouHD YouMob Videos from searching on native application on iPad
- 14. Download Amount Time Block Size Off Generic behavior of video streaming Slope â end-to-end available bandwidth
- 15. Advantages ⢠Buffering phase ⢠Ensures that the player has a sufficient amount of data to compensate for the variance in the end-to-end available bandwidth. ⢠Reduced transfer rate during steady state ⢠Ensures that the amount of video content does not overwhelm the video player while keeping the amount of buffered data during the buffering phase constant or increasing. ⢠Enables to increase the number of videos that can be streamed in parallel. ⢠Important for mobile devices which may not be able to store the entire video. Download Amount Time
- 16. Short On Off Cycles No On Off Cycles Long On Off Cycles OFF OFF Observation : Streaming strategies vastly differ
- 17. Streaming Strategies: summary Service YouTube Container Flash HD (Flash) HTML5 IE 9 Short No Short Firefox Short No No Chrome Short No Long iOS (native) - - Based on encoding rate Android (native) - - Long
- 18. Strengths: 1. Migration from one application to another, or from one container to another, can impact the aggregate video streaming traffic. ⢠migration from Flash to HTML5 or ⢠an increase in the usage of mobile devices 2. These observations help the community to be mindful of the impact any drastic changes are going to have on the streaming traffic. Extension: To actually simulate the situations where such a sudden migration and analyze the impact this might have on the network congestion. Drawback: The impact of these streaming strategies on the network loss rate has also been not studied in this paper.
- 19. Understanding user behavior [2] Alessandro Finamore, Marco Mellia, Maurizio M. Munafò, Ruben Torres, and Sanjay G. Rao YouTube everywhere: impact of device and infrastructure synergies on user experience. In Proceedings of the 2011 ACM SIGCOMM conference on Internet measurement conference (IMC '11).
- 20. Data Collection Tool ⢠Tstat* 1. TCP statistics ⢠Per-connection statistics (#bytes, #pkts, ...) 2. Deep Packet Inspection to inspect the HTTP messages ⢠Classify the type of content and device ⢠Identify the âcontrolâ messages ⢠Per-video statistics (video duration, resolution, codec, ...) (*) http://tstat.polito.it
- 21. Dataset ⢠Week-long collections starting on February 25th 2011 ⢠5 vantage points in Europe and US ⢠Both Residential ISPs and Campus networks ⢠Mobile-player access YouTube via WiFi ⢠No 3G/4G in our data sets. ⢠Datasets accounted for 16TB of traffic and more than 9 million videos
- 22. Understanding user behavior ⢠Which type of content different users retrieve? ⢠Does it change using different devices?
- 23. People with very different cultural bias using very different terminals and with different Internet access bandwidth produce and consume the same type of content: short videos which can be quickly watched from YouTube. Comparing video durations
- 24. ⢠Several formats supported ⢠Hidden to the user ⢠Mainly used FLV and MP4 default PC-player default Mobile-player Available Video Formats
- 25. Understanding user behavior ⢠Do users change the video resolution during a session? ⢠How much of the video is actually played? ⢠âŚ. and how much is downloaded?
- 26. Percentage of resolution switch ⢠Users stick to the default playback parameters! ⢠Why so? (according to the authors) ⢠Users are not aware of this possibility- Meh! ⢠It is âdifficultâ to change resolution ⢠Inertia ⢠Low-to-High is the most common resolution switch.
- 27. Fraction of video downloaded Ρ Fraction of video downloaded Downloaded bytes Full video bytes = = Download only a portion of the video Download more than the entire video ??
- 28. Fraction of video downloaded Download only a portion of the video ⢠80% of the sessions were aborted by Users before completion. Reasons: ⢠Possible mismatch between video expectation and actual content. ⢠Bad quality of experience! Influence on the network: ⢠>20% of aborted sessions downloaded more than 5 times what could be played! ⢠This is due to aggressive buffering policies at the player
- 29. Ρ Fraction of video downloaded Downloaded bytes Full video bytes = = Download more than the entire video
- 30. Why is there so much wastage of bandwidth for mobile? video Playout buffer Initial condition video Each chunk of video is delivered in a separate flow (HTTP Range) Until the delivery of β bytes the playback do not starts 1 β β 2 video 1 β 2 3 4 5 6 1 2 sent 3 4 5 sentplayed If there isnât enough space in the buffer ďą Data already sent are wasted ďą Need to retransmit the data Playout buffer Playout buffer Recall: No guarantee to store the content on the file system One possible cause: not optimized control of the playout buffer Explains Ρ > 1
- 31. ⢠The client keeps downloading content ignoring that the buffer is full ⢠No correct handling of flow control ⢠Possible bug in the player framework? ⢠But, didnât we earlier that data requests in mobile devices are controlled by the client? Mobile-player Download
- 32. Overall waste of bandwidth Overall the wasted amount of data during peak hours ⢠for PC-player, 39% ⢠for Mobile-player, 47% 160Mb/s of YouTube @ peak hours 67Mb/s of traffic is wasted! Bitrate[Mb/s]
- 33. ⢠Strength ⢠Helps in identifying the amount of bandwidth wastage that is happening giving way for researchers to think about how to reduce the same by implementing better streaming policies. ⢠Weakness ⢠Certain assumptions made in this paper are not well-founded. ⢠The reason for an early abort of a video is user's lack of interest. ⢠Several factors such as bad video quality, continuous rebuffering events, continuous switching of bit rates etc. ⢠Extension ⢠A more crucial study would be to analyse the behaviour of mobile video streaming on cellular networks (3G/4G) and understand the network characteristics of the same.
- 34. Exploiting the observations ⢠Users usually do not switch resolutions during a session. ⢠Users do not watch the complete video. ⢠Users prefer watching short videos to a great extent. ⢠Users abort a video if the quality of experience is bad.
- 35. Conclusions ⢠Streaming strategies depend on the application, container and the device being used to stream videos. ⢠Interesting behavior patterns in users can be observed and exploited in designing better streaming strategies.
- 36. Thanks!
Editor's Notes
- #10: Briefly, on a regular PC, the client downloads the web page describing the video which contain a combination of text and other objects needed to display the page. A generate204 request is sent to the video server that is supposed to serve the video. This starts the video prefetch which has two main goals: first, it forces the client to perform the DNS resolution of the video server hostname. Second, it forces the client to open a TCP connection towards the video server. Note that the video server replies with a â204 No Contentâ response, as implied by the command name, and no video content is downloaded so far. At this point, the Flash player sends a HTTP video playback request to get the video. Because of server congestion or lack of content, the server can redirect the client to other servers. In this case, the video server replies with a HTTP â302 Foundâ response which specifies the hostname of another video server to contact. The player then resolves the hostname, and sends a new video playback request. This process can repeat until a valid video server is found. The final video server of the chain replies with the usual HTTP â200 OKâ response, which initiates the stream of video data to the client. Mobile devices use a different protocol as seen in Figure 1 with no prefetch message and with the video content being downloaded in âchunksâ, with every chunk requested on a separate TCP connection. The video server then replies with a â206 Partial Contentâ response.
- #13: To understand the streaming strategies of the different devices and media containers, four datasets of YouTube videos were gathered in [2] namely YouFlash, YouHD, YouHtml, and YouMob depending on the player used to play those videos. The YouFlash and YouHD datasets respectively contain randomly selected 5000 Flash videos and 2000 HD videos. The YouHtml dataset contains 2500 videos from the YouFlash dataset and 500 videos from the YouHD dataset; these videos can be played using the HTML5 player. For the YouMob dataset, videos were searched using the native YouTube application on an iPad.
- #17: The streaming strategy of no ON-OFF cycles is that neither the server nor the client limit the rate of data transfer and the whole video is downloaded during the buffering phase with no steady state phase. This streaming strategy is observed for HTML5 videos on Firefox, and for Flash HD videos.It was observed that OFF periods were in the order of 60 seconds and that the TCP receive window periodically became empty thus showing Chrome throttles the data transfer rate by periodically pulling large blocks of data resulting in long ON-OFF cycles.
- #34: For example, it brings out the fact that the selected content delivery mechanism on mobile player can introduce significant waste, in some cases leading to the download of twice as much content as the video size itself â an important aspect that certainly deserves more attention when the network resources are scarce