Reconstructing proprietary video streaming algorithms Maximilian - PowerPoint PPT Presentation

Reconstructing proprietary video streaming algorithms Maximilian Grüner , Melissa Licciardello, Ankit Singla Photo: ETH Zürich / Gian Marco Castelberg

Roadmap 2

Roadmap •Introduction 2

Roadmap •Introduction •Reconstruction 2

Roadmap •Introduction •Reconstruction •Experiments 2

Roadmap •Introduction •Reconstruction •Experiments •Limitations and future work 2

Roadmap •Introduction •Reconstruction •Experiments •Limitations and future work •Takeaways 2

Streaming in Industry Share of global internet traffic % Categories [Sandvine 2019] 3

Streaming in Industry Share of global internet traffic % Categories [Sandvine 2019] 3

Streaming in Industry 23.8 % Various, 23.1 % Netflix, 12.7 % YouTube, 4.2 % Twitch… Share of global internet traffic % Categories [Sandvine 2019] 3

Streaming in academia # Google Scholar citations Year 4

Streaming in academia # Google Scholar citations Year 4

Applications of reconstruction 5

Applications of reconstruction 5

Applications of reconstruction 5

Applications of reconstruction 5

6

6

6

6

6

Introduction into ABRs Played [Video from YouTube] Resolution Throughput 7

Introduction into ABRs Played [Video from YouTube] Resolution Throughput 7

Introduction into ABRs Played [Video from YouTube] Resolution Throughput 7

Introduction into ABRs Played [Video from YouTube] Resolution Throughput 7

Introduction into ABRs Played [Video from YouTube] Resolution Throughput 7

Introduction into ABRs Played [Video from YouTube] Resolution Throughput 7

Simple ABR in code def choose_next_quality(throughput_history): throughput_estimate = throughput_history[-1] if throughput_estimate == low: next_quality = ⬇⬇ else : ## throughput_estimate == high next_quality = ⬆⬆ return next_quality 8

Simple ABR in code def choose_next_quality(throughput_history): throughput_estimate = throughput_history[-1] if throughput_estimate == low: next_quality = ⬇⬇ else : ## throughput_estimate == high next_quality = ⬆⬆ return next_quality 8

Simple ABR in code def choose_next_quality(throughput_history): throughput_estimate = throughput_history[-1] if throughput_estimate == low: next_quality = ⬇⬇ else : ## throughput_estimate == high next_quality = ⬆⬆ return next_quality 8

Simple ABR in code def choose_next_quality(throughput_history): throughput_estimate = throughput_history[-1] if throughput_estimate == low: next_quality = ⬇⬇ else : ## throughput_estimate == high next_quality = ⬆⬆ return next_quality 8

Simple ABR in code def choose_next_quality(throughput_history): throughput_estimate = throughput_history[-1] if throughput_estimate == low: next_quality = ⬇⬇ else : ## throughput_estimate == high next_quality = ⬆⬆ return next_quality 8

Simple ABR as a tree throughput_estimate == low False True next_quality = ⬇⬇ next_quality = ⬆⬆ 9

(Slightly more) complex ABR in code def choose_next_quality(throughput_history): throughput_estimate = np.mean(throughput_history[-2:]) if throughput_estimate == low: next_quality = ⬇⬇ else : ## throughput_estimate == high next_quality = ⬆⬆ return next_quality 10

(Slightly more) complex ABR in code def choose_next_quality(throughput_history): throughput_estimate = np.mean(throughput_history[-2:]) if throughput_estimate == low: next_quality = ⬇⬇ else : ## throughput_estimate == high next_quality = ⬆⬆ return next_quality 10

Why do we need primitives 11

Why do we need primitives 11

Why do we need primitives 2 Rules 11

Why do we need primitives 5 Rules 11

Why do we need primitives 25 Rules 11

Why do we need primitives 100 Rules 11

Complex ABR in code def choose_next_quality(throughput_history, buffer_size): throughput_estimate = np.mean(throughput_history[-2:]) if throughput_estimate == low: download_time = chunk_size_high / throughput_estimate if download_time < buffer_size: next_quality = ⬇⬇ else : next_quality = ⬆⬆ else : ## throughput_estimate == high next_quality = ⬆⬆ return next_quality 12

Complex ABR in code def choose_next_quality(throughput_history, buffer_size): throughput_estimate = np.mean(throughput_history[-2:]) if throughput_estimate == low: download_time = chunk_size_high / throughput_estimate if download_time < buffer_size: next_quality = ⬇⬇ else : next_quality = ⬆⬆ else : ## throughput_estimate == high next_quality = ⬆⬆ return next_quality 12

Complex ABR in code def choose_next_quality(throughput_history, buffer_size): throughput_estimate = np.mean(throughput_history[-2:]) if throughput_estimate == low: download_time = chunk_size_high / throughput_estimate if download_time < buffer_size: next_quality = ⬇⬇ else : next_quality = ⬆⬆ else : ## throughput_estimate == high next_quality = ⬆⬆ return next_quality 12

Number of primitives 13

Number of primitives Features 13

Number of primitives Features X Lookback 13

Number of primitives Features X Lookback X Actions 13

Number of primitives Features X Lookback X Actions ~ 1300 Primitives 13

14

14

Services of interest Provider Description Alexa Rank 15

Services of interest Provider Description Alexa Rank YouTube Broad coverage 2, Global 15

Services of interest Provider Description Alexa Rank YouTube Broad coverage 2, Global ZDF German Public Service 47, Germany 15

Services of interest Provider Description Alexa Rank YouTube Broad coverage 2, Global ZDF German Public Service 47, Germany Pornhub Pornographic video sharing website 46, Global 15

Services of interest Provider Description Alexa Rank Provider Description Alexa Rank YouTube Broad coverage 2, Global YouTube Broad coverage 2, Global ZDF German Public Service 47, Germany ZDF German Public Service 47, Germany Pornhub Pornographic video sharing website 46, Global Pornhub Pornographic video sharing website 46, Global Arte French-German,cultural 270,France Fandom Gaming, pop-culture 91, Global SRF Swiss Public Service 45, Switzerland TubiTV Movies and series of all genres 1330, USA Twitch Live and VoD streaming service, gaming 39, Global Vimeo Artistic content 188, Global XVideos Pornographic video sharing website 67, Global 15

Inference features 16

Inference features Providers 16

Inference features Providers X Net-traces 16

Inference features Providers X Net-traces X Videos 16

Inference features Providers X Net-traces X Videos X Length 16

Inference features Providers X Net-traces X Videos X Length ~ 9 hours of training/testing streaming time 16

17

17

Why is agreement not sufficient 18

Why is agreement not sufficient Played [Video from YouTube] Resolution Throughput 18

Why is agreement not sufficient Played [Video from YouTube] Resolution Throughput 18

Why is agreement not sufficient Played [Video from YouTube] Resolution Throughput 18

Why is agreement not sufficient Played [Video from YouTube] Resolution Throughput 18

Why is agreement not sufficient Played [Video from YouTube] Resolution Throughput 18

Agreement 19

Agreement [Wikipedia] 19

Agreement [Wikipedia] 19

Agreement [Wikipedia] 19

Agreement [Wikipedia] 19

Agreement [Wikipedia] 19

Agreement [Wikipedia] 19

Agreement [Wikipedia] 19

Agreement F 1 = 2 ⋅ precision ⋅ recall precision + recall [Wikipedia] 19

Similarity 20

Similarity [Scikit-learn] 20

Similarity [Scikit-learn] 20

Similarity [Scikit-learn] 20

Similarity [Scikit-learn] 20

Similarity Score Similarity = 𝒬 ( Decision from reference distribution ) [Scikit-learn] 20

General metric analysis 21

General metric analysis 21

General metric analysis Best 21

General metric analysis Best Worst 21

General metric analysis Best Worst 21

General metric analysis Best Worst 21

General metric analysis Best Worst 21

General metric analysis Best Worst 21

General metric analysis Best Worst 21

22

22

22

22

22

Perceptual metric analysis 23