Source-linked AI summary
On Crowdsourced Interactive Live Streaming: A Twitch.TV-Based Measurement Study
Cong Zhang, Jiangchuan Liu
TL;DR
Crowdsourced live streaming depends on nonprofessional sources with limited, unstable resources, while live interaction adds timing demands. Using a two-month Twitch access-data crawl and captured local traffic, the study finds that viewing patterns reflect events and broadcaster sources, and that viewer delays and upload dynamics affect streaming experience.
Problem
Twitch-like platforms rely on nonprofessional sources with limited, unstable resources, while viewer interaction makes high-quality streaming and timely switching challenging.
Source: §I · p. 1Method
The study examines Twitch using a two-month REST-API crawl of access data and captured traffic from local broadcasters and viewers.
Source: §I · p. 1, §II · pp. 2–3Results
Measurements found that viewing patterns reflect events and broadcaster sources, viewer-side delay affects interaction, broadcast latency is much longer than messaging latency, and dynamic upload capacity affects smoothness.
Source: §I · pp. 1–2Takeaways & Limitations
With limited bandwidth, Twitch’s CBR-based adaptation makes it difficult to provide both good smoothness and low latency.
Source: §V · p. 7As a preliminary study, it considers limited scale and interactions, and similar large-scale interactive events had yet to appear beyond the Twitch Plays Pokemon pilot.
Source: §V · p. 6
Abstract
from arXiv · showhide
Empowered by today's rich tools for media generation and collaborative production, the multimedia service paradigm is shifting from the conventional single source, to multi-source, to many sources, and now toward {\em crowdsource}. Such crowdsourced live streaming platforms as Twitch.tv allow general users to broadcast their content to massive viewers, thereby greatly expanding the content and user bases. The resources available for these non-professional broadcasters however are limited and unstable, which potentially impair the streaming quality and viewers' experience. The diverse live interactions among the broadcasters and viewers can further aggravate the problem. In this paper, we present an initial investigation on the modern crowdsourced live streaming systems. Taking Twitch as a representative, we outline their inside architecture using both crawled data and captured traffic of local broadcasters/viewers. Closely examining the access data collected in a two-month period, we reveal that the view patterns are determined by both events and broadcasters' sources. Our measurements explore the unique source- and event-driven views, showing that the current delay strategy on the viewer's side substantially impacts the viewers' interactive experience, and there is significant disparity between the long broadcast latency and the short live messaging latency. On the broadcaster's side, the dynamic uploading capacity is a critical challenge, which noticeably affects the smoothness of live streaming for viewers.
I. INTRODUCTION
Crowdsourced live streaming expands participation through diverse sources, but nonprofessional broadcasters’ limited resources and viewers’ real-time interactions make high-quality, timely streaming challenging. A Twitch study examines these systems and finds that viewing patterns reflect both events and broadcasters, while viewer-side delays and dynamic upload capacity affect the experience.
I. INTRODUCTION
- Crowdsourced production broadens media creation from many sources to diverse, scalable participation, supported by devices that capture audio and video anywhere.Twitch lets general users broadcast to massive audiences, and its viewers grew from 20 million to 45 million while unique broadcasters tripled to 900 thousand in three years.Source: §I · p. 1
- Twitch-like platforms bridge viewers to nonprofessional sources with limited, unstable resources, making high-quality streaming and timely interaction challenging.Viewers can choose a preferred perspective and interact with broadcasters through real-time chat.Source: §I · p. 1
- Viewing patterns reflect both live events and broadcaster sources, while viewer-side delay strategy substantially affects interactive experience.The measurements also report a significant disparity between long broadcast latency and short live-messaging latency.Source: §I · pp. 1–2
- Dynamic broadcaster upload capacity is a critical challenge that noticeably affects viewers’ streaming smoothness.The study also finds a significant disparity between long broadcast latency and short live-messaging latency.Source: §I · p. 2
II. INSIDE THE TWITCH ARCHITECTURE
The authors combine a two-month Twitch REST-API crawl with captured traffic from local broadcasters and viewers to examine the platform’s architecture. Their account describes a heterogeneous source base, server-assisted video distribution, and separate live messaging infrastructure.
II. INSIDE THE TWITCH ARCHITECTURE
- The two-month REST-API crawl covered 2,923 active broadcasters and 105,117 performances, attracting over 17.8 million viewers.It collected stream counts, views, and stream metadata every ten minutes, and tracked broadcasters’ past-broadcast views and durations daily.Source: §II · p. 2
- About 76.7% of broadcasters used PCs/laptops, about 15.1% used PS4, and about 9.2% used Xbox One; OBS accounted for about 59% of PC/laptop streaming software.Multiple sources can participate in one event, including game players and a commentator.Source: §II · p. 2
- Twitch uses RTMP streaming servers across 14 regions, then HLS and a CDN to deliver streams, with load-balancing servers providing viewers’ playlists.Premium broadcasters can enable online transcoding for viewer-selected qualities; the authors report that 50% of sessions exceed 150 minutes.Source: §II · p. 2
- Interactive messaging servers route viewers’ live messages to broadcasters and other viewers; the servers are in North America and use IRC to deliver messages worldwide.The paper reports reasonably low message latency.Source: §II · p. 2
- The local measurement setup used three source-end PCs and five viewers to monitor traffic, with broadcaster video encoded locally in H.264 and sent to Twitch through RTMP.A network switch simulated changing upload bandwidth, and devices were synchronized to an NTP server for latency measurements.Source: §II · p. 3, Figure 2 · p. 3
III. VIEW STATISTICS AND PATTERNS
The section examines Twitch view characteristics, including stream volume and patterns of popularity and duration. In October 2014, concurrent streams peaked above 12,000, and game streams attracted more than one million views daily.
III. VIEW STATISTICS AND PATTERNS
- The analysis investigates view characteristics across live content, including source-driven and event-driven views.Source: §III · p. 3
- Figures 3 and 4 present streams ranked by views and the distribution of live-streaming duration.Source: Figure 3 · p. 3, Figure 4 · p. 3
- Concurrent streams peaked above 12,000, and game streams attracted more than one million views daily.Most concurrent streams were online gaming broadcasts.Source: §III · p. 3
A. Popularity and Duration
Twitch viewership is highly concentrated among a small share of broadcasters, while stream popularity and duration vary substantially. The observed durations and view distributions complicate resource planning for broadcasts.
A. Popularity and Duration
- For more than 105 thousand streams, the views-versus-rank plot has a long tail and does not fit a classical Zipf distribution.Weibull and Gamma distributions fit the rising portion of the plot.Source: A. Popularity and Duration · p. 4
- Top 0.5% of broadcasters contribute more than 70% of total views in general, and top 0.4% account for more than 90% in some extreme cases.The view distribution is more skewed than in YouTube and PPLive.Source: A. Popularity and Duration · p. 4
- About 30% of live contents last around 60–120 minutes, while another 30% run more than four hours.Exact broadcast duration depends on viewer interest and interaction and is hard to predict in advance.Source: A. Popularity and Duration · p. 4
B. Event- and Source-Driven Views
Twitch view patterns vary with both time and live events and broadcasters. The measured views can shift abruptly, and the timing of event-driven bursts and broadcast terminations is difficult to predict.
B. Event- and Source-Driven Views
- Concurrent views follow daily patterns, with more game-stream viewing during the day and evening than at midnight.The highest view count in the observed week occurred around midnight on October 4, then quickly declined.Source: B. Event- and Source-Driven Views · p. 4
- Game-category traces show sudden view rises and drops when competitions start or terminate, while the timing of bursts is difficult to predict.These patterns are described for League of Legends and Defense of the Ancients 2 broadcasts.Source: B. Event- and Source-Driven Views · pp. 4–5
- Crowdsourced live-stream views are more dynamic and unpredictable than conventional video services, and are both event- and source-driven.The paper notes that platform awareness of source and viewer status does not by itself ensure persistently good user experience.Source: B. Event- and Source-Driven Views · p. 5
IV. MESSAGING AND VIEW LATENCY
The viewer-side latency analysis focuses on representative devices and settings. It examines live messaging, broadcast, and switching latency.
IV. MESSAGING AND VIEW LATENCY
- The analysis examines live messaging latency, broadcast latency, and switching latency.Source: §IV · p. 5
- The representative viewers include a wired PC, a wireless PC, and three mobile tablet viewers.Source: §IV · p. 5
- The latency measurements use viewers with typical device and network settings.Source: §IV · p. 5
A. Live Messaging Latency
Live-message latency depends on network conditions and device type, but is generally low enough to support responsive interaction. The messages also carry participant-type and streaming-quality information to a Twitch statistics server.
A. Live Messaging Latency
- Live-message latency depends on both network conditions and device type.The measurements capture send and receive timestamps from five devices and three sources.Source: A. Live Messaging Latency · p. 5
- Desktop devices had nearly identical send-to-receive latencies, while mobile devices had lower receiving than sending latency.Source: A. Live Messaging Latency · p. 5
- Live-message latency was generally ≤400ms, enabling responsive interaction among viewers and broadcasters.Source: A. Live Messaging Latency · p. 5
- Live messages also sent participant type and streaming quality to Twitch’s statistics server, mp.twitch.tv.Source: A. Live Messaging Latency · p. 5
B. Broadcast Latency
Broadcast latency is the lag between a live event at its source and viewers watching it, and long delays affect interactivity. Measurements identify device-dependent delays, with Twitch processing and delivery strategy accounting for much of the latency.
B. Broadcast Latency
- Broadcast latency measures the time lag between a live event at its source and viewers watching the stream; long delays affect interactivity.Source: B. Broadcast Latency · p. 5
- Browser-based PCs had about 12 seconds of broadcast latency, versus about 21 seconds for client-based mobile devices across streaming rates.Source: B. Broadcast Latency · p. 5
- Twitch’s processing takes about 10s for all devices, while mobile devices receive three extra chunks, adding 12 seconds even on an ideal network.For desktop devices, Twitch converts RTMP streams to HTTP Live Streaming chunks of four seconds each.Source: B. Broadcast Latency · p. 5
- Mobile 3G incurred additional network delay that increased with streaming rate, while mobile WiFi latency remained almost constant.The extra network latency was significant only at very high streaming rates.Source: B. Broadcast Latency · p. 5
C. Source Switching Latency
Source-switching latency varies with bandwidth, network, and device type. Higher download bandwidth lowers switching latency in wired and wireless networks, but the relationship is not proportional.
C. Source Switching Latency
- Switching took 4 seconds on a high-speed wired network and 5.5 seconds on a low-speed wireless network.Source: C. Source Switching Latency · p. 6
- Higher download bandwidth enabled lower switching latency in both wired and wireless networks, but latency did not fall proportionally with bandwidth.Source: C. Source Switching Latency · p. 6
- Mobile-network devices generally switched faster than wired-network devices despite having much lower bandwidth, indicating device-dependent Twitch strategies.Source: C. Source Switching Latency · p. 6
D. Impacts of Broadcaster’s sources
Reducing broadcaster upload capacity increased dropped frames and viewer-side broadcast latency. When capacity recovered, Twitch reduced the delay only slightly, and results were consistent across additional broadcaster and viewer devices.
D. Impacts of Broadcaster’s sources
- Dropped frames consistently increased as broadcaster upload bandwidth decreased, while viewer-side broadcast latency rose stepwise.The latency had notable increases at 900 and 1200 seconds; Twitch attempted to maintain stable latency but could not guarantee smooth streaming.Source: D. Impacts of Broadcaster’s sources · p. 6
- The upload-capacity test used five limits from No Limit to 512 Kb/s, each held for 300 seconds, before returning to No Limit at 1500 seconds.The original encoding setting remained 4000 Kb/s.Source: D. Impacts of Broadcaster’s sources · p. 6
- The same test on two additional broadcaster devices and other viewer devices produced results consistent with Figure 7.Source: D. Impacts of Broadcaster’s sources · p. 6
- After upload capacity recovered at 1500 seconds, Twitch decreased broadcast delay only slightly despite the broadcaster again being able to provide stable streaming.The results indicate sensitivity to frequent changes in broadcaster network capacity, including for responsive interactions.Source: D. Impacts of Broadcaster’s sources · p. 6
V. CONCLUSION AND FURTHER DISCUSSION
The Twitch case study finds that viewing patterns reflect both events and broadcasters’ sources, while viewer-side delay and broadcasters’ dynamic upload capacity affect interaction and stream smoothness. The authors discuss limited study scope and potential ways to improve streaming quality.
V. CONCLUSION AND FURTHER DISCUSSION
- The two-month Twitch measurement found that viewing patterns reflect both events and broadcasters’ sources, while viewer-side delay and dynamic upload capacity affect interaction and viewer smoothness.The paper also reports a disparity between long broadcast latency and short live-messaging latency.Source: §V · p. 6
- The study is preliminary, with limited scale and interactions; similar-scale interactive events beyond the 1.6-million-player, 55-million-viewer Pokémon pilot had yet to appear.The paper links dissatisfaction with interaction in that pilot largely to latency disparity between live messages and broadcast content.Source: §V · p. 6
- Including massive sources in server-client optimization is challenging because their behavior is highly dynamic, although historical broadcaster and viewer activity may support advance behavior prediction.Joint server-client optimization is described as commonly used in state-of-the-art streaming services [9].Source: §V · p. 6
- Analyzing historical broadcaster and viewer activity may let providers predict behavior in advance and accordingly improve streaming quality.Source: §V · pp. 6–7
- With limited bandwidth, Twitch’s mainly CBR-based adaptation makes good smoothness and low latency difficult to offer together; proposed responses include adaptive viewer quality selection and crowd uploading.The source-side strategy under development seeks to leverage aggregated bandwidth from many sources for faster uploading; live messaging and social connections may also help.Source: §V · p. 7