Source-linked AI summary
Network Characteristics of Video Streaming Traffic
Ashwin Rao, Yeon-Sup Lim, Chadi Barakat, Arnaud Legout, Don Towsley, Walid Dabbous
TL;DR
The paper addresses limited knowledge of the network characteristics of Netflix and YouTube traffic, which represent a large share of Internet traffic. It analyzes streaming across applications and containers, identifies three strategies, and models their aggregate effects. The results show that strategy depends on application and container, while higher-resolution streaming can smooth aggregate traffic but linearly increase its data rate.
Problem
Little is known about Netflix and YouTube streaming strategies, although rate-controlled traffic and unviewed downloads may affect network traffic.
Method
The paper analyzes Netflix and YouTube traffic across applications and containers, then derives a mathematical model of aggregate streaming traffic.
Results
Streaming strategy depends on application and container, and the study identifies three strategies ranging from bulk TCP transfer to non-ack clocked traffic.
Takeaways & Limitations
Higher-resolution streaming can produce smoother aggregate traffic while linearly increasing aggregate video-streaming data rate.
Takeaways & Limitations
The HTML5 accumulation-ratio distribution may be an artifact of measurement or media-player techniques for determining encoding rate.
Abstract
from arXiv · showhide
Video streaming represents a large fraction of Internet traffic. Surprisingly, little is known about the network characteristics of this traffic. In this paper, we study the network characteristics of the two most popular video streaming services, Netflix and YouTube. We show that the streaming strategies vary with the type of the application (Web browser or native mobile application), and the type of container (Silverlight, Flash, or HTML5) used for video streaming. In particular, we identify three different streaming strategies that produce traffic patterns from non-ack clocked ON-OFF cycles to bulk TCP transfer. We then present an analytical model to study the potential impact of these streaming strategies on the aggregate traffic and make recommendations accordingly.
1. INTRODUCTION
Video streaming is a major source of Internet traffic, yet the strategies used by Netflix and YouTube and their network effects are poorly understood. The paper analyzes these strategies across applications and containers, identifies three traffic patterns, and models their aggregate impact.
- 25-40% of all Internet traffic is attributed to video streaming, with Netflix and YouTube among its dominant sources.
- The paper analyzes how Web browsers, mobile applications, and Flash, HTML5, or Silverlight containers affect Netflix and YouTube traffic.
- Three streaming strategies produce traffic properties ranging from bulk TCP file transfer to non-ack clocked traffic.
- The identified strategies are characterized for Netflix and YouTube traffic and depend on the application and container combination.
- The paper derives a mathematical model of aggregate video-streaming traffic to support network dimensioning and traffic engineering.
2. VIDEO STREAMING BACKGROUND
Netflix and YouTube let users stream video through PC or mobile browsers and native mobile applications, using different containers. The study focuses on TCP connections carrying video content because they generate most streaming traffic.
- Users access Netflix and YouTube on PCs or mobile devices through Web browsers or native mobile applications.
- YouTube supports Flash and HTML5, with Flash default on PCs and HTML5 default in its native Android mobile application.
- Netflix uses Silverlight for Web-browser streaming and a native Netflix application on mobile devices.
- The analysis restricts attention to TCP connections transferring video content because they contribute the bulk of video-streaming traffic.
3. STREAMING STRATEGIES
Video downloads begin with buffering and continue in a steady state that regulates transfer relative to the video encoding rate. The paper distinguishes no, short, and long ON-OFF-cycle strategies by their transfer patterns and block sizes.
- Video streaming begins with buffering, then enters steady state; ON-OFF cycles limit the download rate during steady state.
- Playback starts once sufficient data is buffered, without waiting for the buffering phase to finish.
- The accumulation ratio is average steady-state download rate divided by video encoding rate; values at least one help prevent interruptions.
- Each steady-state cycle transfers one block at available bandwidth during an ON period, followed by an idle OFF period.
- No ON-OFF cycles transfer all data during buffering, resembling file transfer but potentially overwhelming the player and wasting data after interruptions.
- Short ON-OFF cycles transfer blocks smaller than 2.5 MB and use idle periods to maintain an accumulation ratio slightly above one.
- Long ON-OFF cycles transfer more than 2.5 MB during ON periods and resemble repeated buffering phases separated by long idle periods.
4. METHODOLOGY
The measurements combine YouTube and Netflix video datasets with packet capture across browsers, native mobile applications, and multiple network environments. The datasets vary by service, container, resolution, and device context.
- The study uses six datasets: four for YouTube and two for Netflix.
- YouTube datasets cover Flash, HD, HTML5, and native mobile-application videos, with 5,000 Flash and 2,000 HD videos.
- YouTube video encoding rates range from 0.2 Mbps to 1.5 Mbps for Flash and 0.2 Mbps to 4.8 Mbps for HD videos.
- Netflix datasets were sampled from 11,208 instantly watchable videos, selecting 200 for NetPC and 50 for NetMob.
- Measurements used Internet Explorer, Firefox, and Chrome on PCs, with Flash, Silverlight, or HTML5 support configured as applicable.
- TCP packets were captured while videos were streamed from four wired or Wi-Fi network locations in France and the United States.
- YouTube measurements covered all four locations, whereas Netflix measurements used only the Academic and Home networks.
5. MEASUREMENT RESULTS
Measurements classify YouTube and Netflix video traffic into three streaming strategies and examine how application and container choices shape those strategies.
- The analysis classifies YouTube and Netflix traffic according to three streaming strategies.The strategies span bulk TCP file transfer to non-ack clocked traffic.
- Flash streaming is not throttled by applications; any rate control is performed by YouTube servers, making the strategy application-independent.For HTML5, YouTube servers do not explicitly control transfer rate, while applications throttle it themselves.
- Netflix uses short ON-OFF cycles on PCs, short cycles on iPad, and long cycles on Android native applications.The PC result is independent of the web browser, whereas native mobile applications differ by platform.
- Traffic characterization uses buffering-phase download amount, block size, and accumulation ratio.Computing accumulation ratio requires the videos’ encoding rates.
- For HTML5 videos, the encoding rate is estimated from HTTP Content-Length because the WebM header contained an invalid frame-rate entry.
5.1 YouTube Streaming Strategies
YouTube uses short ON-OFF cycles, long ON-OFF cycles, or no ON-OFF cycles, depending on browser, application, container, encoding rate, and available bandwidth. These strategies differ in buffering, block sizes, and TCP behavior.
- Short ON-OFF cycles: Flash videos receive approximately 40 seconds of playback during buffering, while HTML5 on Internet Explorer typically downloads 10 MB to 15 MB.For a 1 Mbps video, 10 MB corresponds to 80 seconds of playback.
- Short ON-OFF cycles: HTML5 on Internet Explorer has a wide accumulation-ratio range, with mean 1.06 and median 1.04, partly attributed to encoding-rate estimation techniques.
- Short ON-OFF cycles: Flash and HTML5 on Internet Explorer use short ON-OFF cycles, with dominant block sizes of 64 kB and 256 kB, respectively.
- Long ON-OFF cycles: Chrome and Android produce long ON-OFF cycles by periodically pulling blocks larger than 2.5 MB, with OFF periods reaching approximately 60 to 80 seconds.The TCP receive window periodically becomes empty, and the congestion window is not reset after OFF periods.
- Long ON-OFF cycles: On iPad, streaming strategy depends on video encoding rate and end-to-end available bandwidth.One observed video used 37 successive TCP connections in 60 seconds, whereas another used one connection.
- No ON-OFF cycles: HTML5 on Firefox and Flash HD videos use no ON-OFF cycles because the whole video is downloaded during buffering.These sessions have no steady state phase when neither server nor client limits the transfer rate.
- ACK-clock discussion: Streaming blocks can bypass ACK-clock probing after idle periods, and the absence of an ACK clock can increase network loss rates.For Flash videos, an entire 64 kB block can be sent without probing available bandwidth.
5.2 Netflix Streaming Strategies
Netflix uses application-dependent streaming behavior, with buffering amounts, block sizes, and ON-OFF patterns differing between web browsers and native mobile applications.
- Netflix streaming shows short ON-OFF cycles for web browsers and the native iPad application, but long ON-OFF cycles in another viewing configuration.The supplied passage truncates the configuration associated with the long cycles.
- 50 MB is downloaded during buffering on PCs, compared with about 10 MB for the native iPad application.The buffering amount depends on the application used.
- Netflix buffering downloads multiple copies of the same video content at different encoding rates.This behavior is reported during the buffering phase.
- Netflix block sizes depend on the application, with large blocks in the native Android application producing long ON-OFF cycles.Short ON-OFF cycles mainly use blocks smaller than 2.5 MB.
- Netflix uses many TCP connections; ack-clocks appear for new connections carrying one block and for Android native streaming, but not when a connection sends multiple blocks.The authors do not speculate on why Netflix uses many connections.
5.3 Discussion
YouTube and Netflix streaming strategies vary with application and container, producing traffic behavior that can differ substantially from standard TCP and affect aggregate traffic.
- Flash streaming is application-independent, whereas HTML5 streaming varies across applications because YouTube servers do not limit HTML5 transfer rates.HTML5 therefore exhibits a wide range of buffering and steady-state patterns.
- Netflix uses the same strategy across web browsers but different strategies for Android and iOS native mobile applications.
- Three strategies range from no ON-OFF cycles, equivalent to TCP file transfer, to short and long ON-OFF cycles requiring application-layer rate control.The ON-OFF strategies restrict transfer rates below available bandwidth.
- Short and long ON-OFF cycles restrict application-layer transfer rates, so their traffic may not resemble a standard TCP flow and may lack ack-clocks.
- A migration between applications or containers can affect aggregate video streaming traffic, motivating the paper’s mathematical model.The paper specifically identifies migration from Flash to HTML5 and increased mobile-device use as possibilities.
6. MODEL FOR AGGREGATE VIDEO TRAFFIC
The paper models aggregate video-streaming traffic from session arrivals, video encoding rates, durations, and download strategies, including sessions interrupted before completion.
- Model assumptions: The model represents video sessions as a homogeneous Poisson process with rate λ, fixed encoding rate e_n, fixed duration L_n, and size S_n = e_nL_n.It also assumes overprovisioned end-to-end bandwidth relative to each video encoding rate.
- No interruptions: The model first analyzes no ON-OFF cycles, where each video downloads continuously at the available end-to-end bandwidth during an interval of duration D_n.
- No interruptions: Equations (3) and (4) provide the mean and variance of aggregate traffic for no ON-OFF streaming and can be used for network dimensioning.A proposed bandwidth rule is E[R(t)] + α√V_R, with α ≥ 1 controlling tolerable bandwidth violations.
- No interruptions: The mean and variance of aggregate traffic are independent of the streaming strategy when users do not interrupt downloads.This allows strategy selection for goals such as server load without overwhelming the network.
- No interruptions: Increasing video encoding rates increases aggregate traffic rates while making aggregate traffic smoother because variance grows linearly with encoding rate.
- Interruptions: For interruptions caused by lack of interest, the model quantifies unused bytes and bandwidth waste using buffering amount, accumulation ratio, viewing fraction, and video duration.Equation (7) gives a condition limiting unused bytes, while Equations (8) and (9) compute wasted bandwidth.
- Interruptions: With 40 seconds of buffering, accumulation ratio 1.25, and 20% watched, videos shorter than 53.3 seconds are downloaded before viewers see 20%.This example assumes fast buffering and uses YouTube Flash measurements.
7. RELATED WORK
Prior work examined video content, viewing patterns, caching, and rate limitations, while this paper focuses on network traffic strategies and aggregate-traffic modeling.
- Earlier studies characterized YouTube content popularity and viewing patterns, including caching approaches intended to improve experience or reduce network traffic.
- The paper distinguishes its focus on YouTube and Netflix network traffic from prior work on content characterization.
- Previous studies observed YouTube rate limitations and buffering differences but did not identify the three streaming strategies discussed here.Prior Netflix work observed rate limitation only during the steady-state phase.
- The authors claim to be the first to identify and characterize three YouTube and Netflix streaming strategies and derive a model of aggregate video traffic.
8. CONCLUSION
The paper characterizes Netflix and YouTube streaming strategies, models their aggregate traffic effects, and considers adaptation to user interruptions. It also identifies network loss rate as an unexamined aspect for future work.
- A shift from Flash to HTML5 or increased mobile-device use could significantly change network traffic characteristics across a large population.
- The model shows that higher-resolution streaming can smooth aggregate traffic while linearly increasing the aggregate data rate.
- Buffer amount and accumulation ratio can be adapted to account for interruptions caused by lack of user interest.
- The analysis does not consider how the three streaming strategies affect network loss rate.