![Page 1: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/1.jpg)
Online learning of video-streaming algorithms on live users
Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi Zhu, Sadjad Fouladi,James Hong, Keyi Zhang, Philip Levis)
Assistant Professor of Computer ScienceAssistant Professor of Electrical Engineering (by courtesy)Stanford University
![Page 2: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/2.jpg)
Outline
I Networked systems present unique challengesfor machine learning.
I Continual learning in situ mitigates modelmismatch and dataset shift.
I Our experience with continual learning forvideo streaming.
![Page 3: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/3.jpg)
Background
I Networking algorithms have many tunable parameters.
I Natural setting for machine learning, but:
Networked systems present uniquechallenges for machine learning.
![Page 4: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/4.jpg)
Example: Sprout (NSDI 2013) in publication
Sprout, NSDI 2013 (figure 7)
![Page 5: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/5.jpg)
Sprout in real life in America
Stanford Pantheon result (July 31, 2018, T-Mobile in California),https://pantheon.stanford.edu/result/3455/
![Page 6: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/6.jpg)
Sprout in real life in India
Stanford Pantheon result (August 1, 2018, Airtel in New Delhi),https://pantheon.stanford.edu/result/3474/
![Page 7: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/7.jpg)
Example: Remy in simulation (SIGCOMM 2013)
Better
![Page 8: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/8.jpg)
Remy in real life
RemyCC
Stanford Pantheon result (Nov. 3, 2018, Google Cloud Iowa to London),https://pantheon.stanford.edu/result/4678/
![Page 9: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/9.jpg)
Example: Vivace (NSDI 2018) in publication
Vivace, NSDI 2018 (figure 7)
![Page 10: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/10.jpg)
Vivace in real life
Stanford Pantheon result (August 1, 2018, AWS Brazil-HostDime Colombia),https://pantheon.stanford.edu/result/3470/
![Page 11: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/11.jpg)
Example: Pensieve (SIGCOMM 2017) in publication
Pensieve, SIGCOMM 2017 (figure 11)
![Page 12: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/12.jpg)
Pensieve in reproduction
Stanford CS244 student project,https://reproducingnetworkresearch.files.wordpress.com/2018/07/recreating pensieve.pdf
![Page 13: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/13.jpg)
Example: BBR (ACM Queue 2016) in publication
“BBR converges toward a fair share of the bottleneck bandwidth whether competingwith other BBR flows or with loss-based congestion control. [...] Unmanaged routerbuffers exceeding several BDPs, however, cause long-lived loss-based competitors tobloat the queue and grab more than their fair share.”
https://queue.acm.org/detail.cfm?id=3022184
![Page 14: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/14.jpg)
BBR in independent evaluation
Cubic vs BBR over a 12ms RTT 10G circuit
Geoff Huston, TCP and BBR, RIPE 76 (May 2018)https://ripe76.ripe.net/presentations/10-2018-05-15-bbr.pdf
![Page 15: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/15.jpg)
Example: spam filtering
SpamAssassin (spam filtering engine):
I Anybody can propose a spam-filtering algorithm.
I Central party learns best weights based on predictive power ofeach algorithm.
I Weights are then deployed in the field.
![Page 16: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/16.jpg)
Deployment: spam filtering
I 2007: a rule is added
I Rule: “Does the year match 200x ?”
I Catches a lot of spam.
I Extremely low false-positive rate!
I (but. . . big surprise on 1/1/2010)
![Page 17: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/17.jpg)
Example: Google Flu Trends
Proposal (2008): train a model to predict flu incidence fromhistorical search engine queries. Then deploy the model to predictflu in advance of the government.
![Page 18: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/18.jpg)
Google Flu Trends plot as of today
(http://www.google.org/flutrends/about/how.html)
Keith Winstein [email protected] CSAIL
The 2012–2013 Divergence of Google Flu Trends
![Page 19: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/19.jpg)
Most of plot is training data
(http://www.google.org/flutrends/about/how.html)
Keith Winstein [email protected] CSAIL
The 2012–2013 Divergence of Google Flu Trends
![Page 20: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/20.jpg)
Large divergence (3.7×) in New England (HHS region 1)
0
2%
4%
6%
8%
10%
12%
14%
2010 2011 2012 2013
Out
patie
nt v
isits
for
influ
enz
a-lik
e ill
ness
CDC
Keith Winstein [email protected] CSAIL
The 2012–2013 Divergence of Google Flu Trends
![Page 21: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/21.jpg)
Substantial divergence (+72%) in France
0
0.2%
0.4%
0.6%
0.8%
1%
1.2%
2010 2011 2012 2013
Out
patie
nt v
isits
for
influ
enz
a-lik
e ill
ness
Sentinelles
Keith Winstein [email protected] CSAIL
The 2012–2013 Divergence of Google Flu Trends
![Page 22: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/22.jpg)
Example: predicting unemployment claims from Twitter (9/2015)
Twitter, Big Data, and Job Numbers
People turn to Twitter to see what’s trending in the news. But LSA Economics ProfessorMatthew Shapiro has found a new way to harvest employment information from tweets andhashtags faster and more accurately than the government’s official reports.
LSA Economics Professor Matthew Shapiro is one of the rare people who can claim to be getting work done when looking atsocial media. He’s not following sports scores or breaking news; he’s tracking the nation’s labor market.
“When we started,” explains Shapiro, “we had no idea if we could track job loss with tweets, but over a two-year period, we’veseen the social media index perform quite well. In 2011-2012, for example, our index leveled off just like the official numbers. Wecaptured big job-loss fluctuations around Hurricane Sandy, and around the government shutdown in October 2013.”
There were times when Shapiro’s numbers matched the reports, and there were timeswhen they didn’t. When they differed, Shapiro’s numbers were more accurate than thegovernment’s.When the state of California got new computers, for example, there were delays in processing unemployment claims.Government data reflected the slowdown in processing applications, but social media captured a more accurate picture of whatwas happening in the labor market.
“Our series was stable,” says Shapiro, the Lawrence R. Klein Collegiate Professor of Economics, “so our numbers were, in someways, a better indicator.”
By the NumbersMany important economic indicators, including the Bureau of Labor Statistics unemployment rate and the University of MichiganConsumer Sentiment Index, are collected based on surveys. But it’s hard and costly to get enough people to answer a survey inorder to get a representative sample. It’s hard to verify whether the people responding to the survey are speaking for just
by Susan Hutton
![Page 23: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/23.jpg)
Predicting unemployment claims from Twitter (at time of publication)
![Page 24: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/24.jpg)
Predicting unemployment claims from Twitter (post-publication)
![Page 25: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/25.jpg)
Outline
I Networked systems present unique challengesfor machine learning.
I Continual learning in situ mitigates modelmismatch and dataset shift.
I Our experience with continual learning forvideo streaming.
![Page 26: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/26.jpg)
Unique challenges for ML in networking
I Model mismatch: training environment (simulator, emulator, ortestbed) does not match deployment environment (Internet).
I Dataset shift: drift in deployment environment over time
I Each participant has only partial information
I Failure may not be evident to any individual
I Adversaries, extreme heavy tails, . . .
![Page 27: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/27.jpg)
Proposed solution: continual learning in place
I “Deploy-first” approach
I Retrain frequently
I Compare continually against held-back “sane baseline”
![Page 28: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/28.jpg)
Outline
I Networked systems present unique challengesfor machine learning.
I Continual learning in situ mitigates modelmismatch and dataset shift.
I Our experience with continual learning forvideo streaming.
![Page 29: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/29.jpg)
Puffer: continual learning for Internet video streaming
I We created a live TV-streaming website with six channels
I Opened to public in late 2018
I Streaming about 1,000 hours of video per day
I Randomize users to different algorithms
I Goal: continual learning for video streaming
![Page 30: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/30.jpg)
Algorithms that affect video streaming
I Congestion control: when to send each packet
I Throughput prediction: how fast can server send in near future?
I Adaptive bitrate: what version of each “chunk” to send
![Page 31: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/31.jpg)
Demo
![Page 32: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/32.jpg)
Google ad for “tv streaming”
![Page 33: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/33.jpg)
Reddit ad
![Page 34: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/34.jpg)
Reddit (continued)
![Page 35: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/35.jpg)
Puffer algorithm
I Continual learning algorithm for video streaming
I Neural network predicts “how long would each chunk take?”I given internal TCP statistics (weakly cross-layer)I given size of proposed chunkI producing probability distribution (vs. point estimate)
I Control strategy: probabilistic model-predictive control
![Page 36: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/36.jpg)
Data Aggregation
Transmission Time Predictor MPC Controller
“OurTube”Video Server
bitrateselection
stateupdate
updatemodel
da
ily t
rain
ing
mo
de
l-ba
se
d c
on
trol
![Page 37: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/37.jpg)
Results
I Data from January 19–28, 2019
I 8,131 hours of video streamed
I 3,719 unique users
I 2 congestion-control schemes (BBR and Cubic), 7 ABR schemes
I Daily retraining for Puffer/BBR and Puffer/Cubic
![Page 38: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/38.jpg)
Stalls vs. SSIM (BBR)
Better
QoE
![Page 39: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/39.jpg)
Results (BBR)
Algorithm Time stalled Mean SSIM SSIM variation(over BBR congestion control) (lower is better) (higher is better) (mean, lower is better)
Puffer 0.06% 17.03 dB 0.53 dBBuffer-based 0.34% 16.58 dB 0.79 dBMPC 0.81% 16.32 dB 0.54 dBRobustMPC 0.29% 15.74 dB 0.63 dBPensieve 0.52% 16.19 dB 0.88 dBNon-continual Puffer 0.31% 16.49 dB 0.60 dBEmulation-trained Puffer 0.76% 15.26 dB 1.03 dB
![Page 40: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/40.jpg)
Stalls vs. SSIM (BBR)
Better
QoE
![Page 41: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/41.jpg)
Stalls vs. SSIM (BBR)
Better
QoE
![Page 42: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/42.jpg)
Stalls vs. SSIM (BBR, < 6 Mbps only)
Better
QoE
![Page 43: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/43.jpg)
Results in emulation don’t match real world
Better
QoE
![Page 44: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/44.jpg)
People on Puffer watched 2× as long
Algorithm Stream duration (mean ± std. err.)(any congestion control) (higher may suggest less frustrated users)
Puffer 20.3 ± 1.3 minutesBuffer-based 12.9 ± 0.8 minutesMPC 9.5 ± 0.5 minute sRobustMPC 10.0 ± 0.5 minutesPensieve 10.0 ± 0.5 minutesNon-continual Puffer 12.5 ± 0.7 minutesEmulation-trained Puffer 7.9 ± 1.3 minute s
Multivariate regression predicts each % of stall is assocated with 6 minutes less watch time.Each dB of SSIM predicts 2 minutes more watch time (R2 = 0.36).
![Page 45: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/45.jpg)
Contributors to transmission-time accuracy
TTP ablation Prediction error rate(lower is better)
No History + No cwnd inflight 19.8%No History + No delivery rate 17.9%No History + No RTT 24.0%No History 17.1%Full TTP 13.7%Harmonic Mean 17.9%
![Page 46: Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi ...€¦ · Online learning of video-streaming algorithms on live users Keith Winstein (with Francis Yan, Hudson Ayers, Chenzhi](https://reader033.vdocuments.mx/reader033/viewer/2022042919/5f5ff35f644de930170c6ca4/html5/thumbnails/46.jpg)
Takeaways
I Much work remains in calibrating training environments to the real world.
I We need better emulators.
I Continual learning in-situ appears to be a powerful tool for ML in networkedsystems.
I We are opening Puffer to the research community for others to develop and deploycongestion control and ABR algorithms on real traffic.
Keith Winstein, [email protected]