PiTree: Practical Implementations of ABR Algorithms Using Decision - - PowerPoint PPT Presentation

▶
pitree practical implementations of abr
SMART_READER_LITE
LIVE PREVIEW

PiTree: Practical Implementations of ABR Algorithms Using Decision - - PowerPoint PPT Presentation

Aug 17, 2022 17 likes •192 views

PiTree: Practical Implementations of ABR Algorithms Using Decision Trees Paper # P5C-04 Zili Meng Jing Chen Yaning Guo Chen Sun Hongxin Hu Mingwei Xu Adaptive Bitrate (ABR) Algorithms Client Video server Quality Quality Bandwidth

slide-1
SLIDE 1

PiTree: Practical Implementations of ABR Algorithms Using Decision Trees

Paper # P5C-04 Zili Meng Jing Chen Yaning Guo Chen Sun Hongxin Hu Mingwei Xu

slide-2
SLIDE 2

Adaptive Bitrate (ABR) Algorithms

2

High

Quality Time

Medium Low

Video server Client Quality Time Bandwidth ABR Time

Partially borrowed from “Neural Adaptive Content-aware Internet Video Delivery” in USENIX NSDI 2018.

P5C-04 https://transys.io/pitree

slide-3
SLIDE 3

Adaptive Bitrate (ABR) Algorithms

3

Harmonic Mean (Rate) 2011 2012 2013 2014 2015 2016 2017 2018 2019 Neural Networks Deeper NNs Linear Programming Hidden Markov Model

Deep Learning

Piecewise Linear (Buffer)

Explicit Formula Complex Optimization

P5C-04 https://transys.io/pitree

slide-4
SLIDE 4

Drawbacks – Heavyweight

  • Large page size.

– Pensieve [SIGCOMM’17] increased HTML page size by 4x. – Page load time is increased by ~10s.

  • Long decision latency.

– RobustMPC [SIGCOMM’15] increased decision latency to seconds. – Decision latency > chunk length.

  • High operating expenses.

– Up to millions of concurrent viewers.

P5C-04 4

R = RobustMPC [SIGCOMM’15]: ILP P = Pensieve [SIGCOMM’17]: DNN H = HotDASH [ICNP’18]: 2xDNN

Server-side Implementation Client-side Implementation

https://transys.io/pitree

slide-5
SLIDE 5

Our Contribution: PiTree

  • Design & train ABR algorithms offline as usual.
  • Convert the model into a decision tree.
  • Deploy the decision tree online.

P5C-04 5

Offline training/design network traces decisions

Sophisticated ABR model (e.g., DNN, ILP)

Online deployment

https://transys.io/pitree

Decision tree

slide-6
SLIDE 6

Design Choice: Why Decision Tree?

  • Non-parametric and

expressive.

P5C-04 6

internal nodes leaf nodes 𝑇1 𝑇2 state space decision boundary low local approximation error

https://transys.io/pitree

slide-7
SLIDE 7

Design Choice: Why Decision Tree?

  • Non-parametric and

expressive.

  • Lightweight for video

players.

P5C-04 7

A decision tree with 100 leaf nodes: Page size increase <1% Decision latency <1ms laptop smartphone

https://transys.io/pitree

slide-8
SLIDE 8

Design Choice: Why Decision Tree?

  • Non-parametric and

expressive.

  • Lightweight for video

players.

  • Following the decision

logic of ABR algorithms.

P5C-04 8

Decision tree of BBA [SIGCOMM’14].

https://transys.io/pitree

slide-9
SLIDE 9

Design Challenge: Sequential Dependency

  • ABR Control is a sequential

decision-making process.

P5C-04 9

ABR Controller bandwidth logs bitrate Internet HTTP buffer occupancy Buffer request chunk

https://transys.io/pitree

slide-10
SLIDE 10

Design Challenge: Sequential Dependency

  • ABR Control is a sequential

decision-making process.

  • One wrong prediction may

drive the student off teacher’s trajectory.

P5C-04 10

prediction error unexperienced state space

teacher student

https://transys.io/pitree

slide-11
SLIDE 11

PiTree: Imitation Learning – Follow the Leader

P5C-04 11

decision tree (s,a) set virtual player deploy onto video clients correct actions states (in)correct actions teacher.predict() add into add into train retrain traffic traces videos

https://transys.io/pitree

slide-12
SLIDE 12

For design details and theoretical analysis, please refer to our paper.

slide-13
SLIDE 13

Evaluation – Quality of Experience (QoE) Ratio

13

P5C-04

  • Summary of experiments

– 3 QoE metrics. – 3 sets of bandwidth traces. – 3 ABR algorithms.

  • QoE ratio =

QoEPiTree QoEOriginal

  • Average QoE ratio > 97%.
  • Median QoE ratio > 98%.
  • Details in the paper.

0% 25% 50% 75% 100% 90% 95% 100% 105% 110% CDF QoE Ratio Pensieve HotDASH RobustMPC 98%

https://transys.io/pitree

slide-14
SLIDE 14

Evaluation – Page Size

14 P5C-04

0.6% 1.4% 3.1% 1.2% 4.6x 5.8x 500 1000 1500 2000 2500 Page Size (KB)

https://transys.io/pitree

14.5s @ 1Mbps Our experiments: <0.1s @ 1Mbps

RobustMPC [SIGCOMM’15]: ILP Pensieve [SIGCOMM’17]: DNN HotDASH [ICNP’18]: 2xDNN

slide-15
SLIDE 15

Evaluation – Latency

15 P5C-04 https://transys.io/pitree

PC Mobile 10s 1s 0.1s 10ms 1ms 0.1ms 10μs ~4s

Intel Core i7-8550 Qualcomm Snapdragon 821

1ms

RobustMPC [SIGCOMM’15]: ILP Pensieve [SIGCOMM’17]: DNN HotDASH [ICNP’18]: 2xDNN

slide-16
SLIDE 16

Takeaways

  • Current ABR algorithms are increasingly heavyweight.
  • PiTree convert complex ABR algorithms to decision trees to

deploy them in a lightweight way.

– Use imitation learning to address the action dependency.

  • PiTree can significantly reduce the algorithm overhead with

negligible QoE loss.

– Page size reduced by up to 5x, decision latency reduced by up to 1000x.

P5C-04

16

https://transys.io/pitree

slide-17
SLIDE 17

Thank you! Questions and comments?

Try your ABR algorithms with PiTree! https://transys.io/pitree zilim@ieee.org