Adaptive streaming playback statistics set

source ：Brightcove be the speaker ：Thiago Teixeira, Yuriy Reznik translate ： Zhong Hongcheng This speech first introduces the characteristics of streaming media . The analysis of streaming media playback data is helpful to the client logic 、 Optimization of streaming media system , therefore , The author provides a streaming media playing data set that can be used for research , Session information containing four real events 、 Client information 、 Streaming information 、 Broadcast information, network information, etc .

Catalog

• Introduce
  • ABR flow ： Main principles
• Data sets
• Examples of using datasets
  • Behavior modeling of streaming media client
• Conclusion

Introduce

The characteristics of modern streaming media ：

• Video content is distributed to Multiple devices , Include ：
  • PC、 laptop
  • mobile phone
  • The tablet
  • TV
• video Embedded in the web , perhaps Play full screen
• Use be based on HTTP Adaptive streaming media protocol distribution , for example HLS and DASH
• Client's Adaptive logic Mainly affected by the following factors ：
  • network bandwidth
  • Player window size
  • Device decoding performance

ABR flow ： Main principles

chart 1: Typical architecture of streaming media client

The figure above shows the typical architecture of the streaming media client , The key elements are ：

• Buffer： The buffer stores the loaded media segment , Generally, you can cache 10-30 Second content
• Bandwidth estimation： Bandwidth prediction , Used to calculate and predict the available network bandwidth
• Adaptation engine： Adaptive engine , Based on available network bandwidth 、 Bit rate ladder 、 Select a version based on buffer loading and other factors

chart 2: Examples of streaming media sessions

In a streaming session , May experience bandwidth fluctuations , This causes the adaptive switching event to occur , Streaming media has the following characteristics ：

• Will not play immediately , You need to fill the buffer first
• When it's playing , Adaptive handoff may occur due to bandwidth changes
• The client may pause to wait for the buffer because the buffer is exhausted
• The buffered fragments are sometimes discarded during client switching
• When the user aborts playback , The client also discards the buffered fragment

chart 3: Player analysis system logic

Typical player States and events are collected by an analysis system , The analysis system is a plug-in built into the client . The analysis system will collect the playback logs of the client , Collect key performance indicators such as ： Watching Events 、 Loaded data 、 Selected stream version, etc .

Collecting these analytical data can be used in many ways ：

• Audience analysis
  • What was viewed
  • How many viewers are watching
• User data analysis
  • What is the average viewing time of the video
  • What is the distribution of playback times
• QoE analysis
  • Average distribution quality analysis of the system
  • Start time distribution
• System modeling and optimization
  • Optimization of adaptive logic
  • Optimization of bit rate ladder
  • Design of distribution system

Data sets

Dataset in GitHub Open source ：https://github.com/brightcove/streaming-dataset

• Contains more than 5000 Million player records
• Contains four real-world streaming events
• Contains a variety of distribution devices 、 Network status and coding settings
• With the development of streaming media technology , More data may be added later , Include ：
  • new codec
  • Low latency distribution
  • UHD/HDR distribution

chart 4: Data set collection event Overview

The information collected includes ： Session information 、 Client information 、 Streaming information 、 Play information and network information , The details are as follows: .

chart 5: Data set collection information details

chart 6 The event is given 1 A brief summary of ：

chart 6: event 1 data

Examples of using datasets

Behavior modeling of streaming media client

By analyzing the data , We found that the player resolution has a great influence on the choice of convection , Bandwidth adaptation is not the only problem . This inspired us to improve the player modeling . The model that combines bandwidth adaptation and player resolution adaptation is more accurate .

chart 7: Client modeling vs. real values

Conclusion

The streaming media playback data set provided can be used for research , In the follow-up study , It can be used for ：

• Use machine learning to better model streaming media clients and systems
• The best design of streaming media client
• Performance analysis of adaptive streaming media system
• Content aware coding optimization
• Multi screen distribution optimization

Attach speech video ：

http://mpvideo.qpic.cn/0b2efqaaiaaa3qags6p6sfqvalgdaqwaabaa.f10002.mp4?dis_k=d46c785c30818b626a10caf0a2c86384&dis_t=1645153183&vid=wxv_2256065279388762125&format_id=10002&support_redirect=0&mmversion=false