1 of 19

Adaptive Streaming Playback Statistics Dataset

Thiago Teixeira, Bo Zhang, Yuriy Reznik

Brightcove Inc.

2 of 19

Modern Era of Streaming

  • More than 50% of the Internet traffic
  • Heterogeneous devices
    • Desktops / Laptops
    • Mobiles
    • Tablets
    • Connected TVs and other embedded devices
  • Delivered by HTTP-based protocols such as HLS or DASH
  • Client adaptation can be influenced by multiple factors, including
    • network performance
    • player size
    • device decode capabilities
    • etc.

​

2

3 of 19

3

  • Video encoded into HLS and DASH formats
  • Encoding profiles are custom generated
  • Delivered by CDNs
  • Pulled by multiple devices

Delivery Platforms

4 of 19

Player Events Sent to Analytics

4

5 of 19

Some Things We Use Analytics for

  • Volume and Audience Measurement
    • Which content was watched
    • By how many viewers
  • User Engagement Measurement
    • How long the video was played
    • Distribution of playback times
  • QoE Measurement
    • What was the avg. video quality played
    • Distribution of start-up times
  • System Modeling and Optimizations
    • Adaptation logic
    • System performance and improvements
    • AI/ML

5

Typical uses of analytics

systems in practice

Possible uses of playback

statistics in research and engineering community

6 of 19

Objectives of This Work

  • Produce a streaming dataset suitable for use in the research community
    • Report metrics that are common across various existing analytics systems
    • Capture several real-world large-scale streaming playback events
    • With diverse types of networks, clients, delivery devices
    • With diverse types of content and encoding profiles

​

  • Intended Uses (not limited to)
    • Analysis of the behavior of streaming clients
    • Design of optimal ABR adaptation algorithms
    • Training of related AI/ML models
    • Analysis of the performance of adaptive streaming systems
    • Context-aware encoding optimizations
    • Multi-screen streaming delivery optimizations

6

7 of 19

Streaming Playback Datasets

7

8 of 19

Event Description

8

Available at https://github.com/brightcove/streaming-dataset

9 of 19

Parameters Reported

9

10 of 19

Event I Statistics

10

User devices Operating systems Networks Player resolutions

​

​

​

​

​

Encoding profile

​

rendition

video_codec

codec_profile

width

height

framerate

video_bitrate

audio_bitrate

total_bitrate

0

h.264

baseline

426

240

29.97

300

192

492

1

h.264

baseline

640

360

29.97

500

192

692

2

h.264

baseline

852

480

29.97

750

192

942

3

h.264

baseline

960

540

29.97

1000

192

1192

11 of 19

Event II Statistics

11

User devices Operating systems Networks Player resolutions

​

​

​

​

​

Encoding profile

​

​

rendition

video_codec

codec_profile

width

height

framerate

video_bitrate

audio_bitrate

total_bitrate

0

h.264

main

640

360

30

900

192

1092

1

h.264

main

960

540

30

1700

192

1892

2

h.264

high

1280

720

30

2400

192

2592

3

h.264

high

1280

720

30

4040

192

4232

12 of 19

Event III Statistics

12

User devices Operating systems Networks Player resolutions

​

​

​

​

​

Encoding profile

​

​

rendition

video_codec

codec_profile

width

height

framerate

video_bitrate

audio_bitrate

total_bitrate

0

h.264

baseline

360

202

29.97

141

32

173

1

h.264

baseline

480

270

29.97

300

64

364

2

h.264

main

640

360

29.97

770

128

898

3

h.264

main

1280

720

29.97

1750

192

1942

13 of 19

Event IV Statistics

13

User devices Operating systems Networks Player resolutions

​

​

​

​

​

Encoding profile

​

​

rendition

video_codec

codec_profile

width

height

framerate

video_bitrate

audio_bitrate

total_bitrate

0

h.264

baseline

480

270

23.976

450

128

578

1

h.264

baseline

640

360

23.976

800

128

928

2

h.264

main

768

432

23.976

1000

128

1128

3

h.264

main

1024

576

23.976

1500

128

1628

4

h.264

main

1280

720

23.976

2100

128

2228

14 of 19

Example Use Case

14

15 of 19

Analysis of Client Adaptation

Event IV

15

Stream

Profile

Resolution

Framerate

Bitrate

1

Baseline

480x270

23.976

450k

2

Baseline

640x360

23.976

800k

3

Main

768x432

23.976

1000k

4

Main

1024x576

23.976

1500k

5

Main

1280x720

23.976

2100k

Encoding profile

Usage of top bitrate rendition does not increase even when bandwidth is no longer a limit !!!

Observations:

  • Player sizes apparently impact selection of the streams! This happens not only with web players, but also with mobile apps.
  • Bandwidth-adaptation is no longer the only dimension of the problem!

16 of 19

Simple Client Model

16

17 of 19

How Does It Compare with the Real World

17

18 of 19

Conclusions

  • Streaming data from four large real-world events
  • One of many use cases provided
    • Building a streaming client model
  • Brings many opportunities for subsequent research
    • Using ML/AI for more sophisticated modeling of clients and systems
    • Training of such models
    • Optimal design of streaming clients
    • Analysis of the performance of adaptive streaming systems
    • Context-aware encoding optimizations
    • Multi-screen delivery optimizations

18

19 of 19

THANK YOU!

19