CS2P: Improving Video Bitrate Selection and Adaptation with - - PowerPoint PPT Presentation

• Y. Sun, F. Lin, N. Wang @
• X. Yin, J. Jiang, V. Sekar, B. Sinopoli

@

• T. Liu @

CS2P: Improving Video Bitrate Selection and Adaptation with Data-Driven Throughput Prediction

Bitrate adaptation is key for QoE

2

• DASH = Dynamic Adaptive Streaming over HTTP
• Entail new QoE metrics, e.g., low buffering, high video quality
• Need intelligent bitrate control and adaptation

Prior work: Accurate throughput prediction can help!

Accurate throughput prediction à Better initial bitrate selection

3

— Fixed bitrate — Adaptive bitrate

Accurate throughput prediction à Better midstream adaptation

— Replicate the analysis by Yin et al. at SIGCOMM2015[1]

p𝑂𝑝𝑠𝑛𝑏𝑚𝑗𝑨𝑓𝑒 𝑅𝑝𝐹 =

/01234 567 89:6;:1<034 =>?@ABC

4

[1] X. Yin, et al. “A Control-Theoretic Approach for Dynamic Adaptive Video Streaming over HTTP”. ACM SIGCOMM, 2015. [2] T.

• Y. Huang, et al. “A Buffer-Based Approach to Rate Adaptation: Evidence from a Large Video Streaming Service”. ACM

SIGCOMM, 2014.

Open questions on predictability!

5

— Our understanding of throughput variability and

predictability is quite limited.

— What types of prediction algorithms to use?

—In the context of video bitrate adaptation

— Prior approaches: 30%+ of predictions with error ≥0.2

Our work and contributions

A large-scale analysis, providing data-driven insights for predicting the throughput accurately. Design of CS2P (Cross-Session Stateful Predictor): Improving bitrate selection and adaptation via throughput modeling. A practical implementation of CS2P and the demonstration of improvements in video QoE.

6

Outline

— Motivation

è Data-driven Observations

— CS2P Approach — Evaluation

7

— From operational platform of iQIYI.

— iQIYI is a leading online video

content provider in China.

— 20M+ sessions, 8 days in Sep. 2015,

— Each session records avg.

throughput per 6-second epoch.

Dataset description

8

Feature Coverage Client IP 3.2M Client ISP 87 Client AS 161 Province 33 City 736 Server 18

Observation 1: Significant variability within a session.

9

50% of sessions with n-stddev ≥ 30% 20% of sessions with n-stddev ≥ 50%

Observation 2: Stateful/persistent characteristics.

An example session.

10

Throughput variation across two consecutive epochs with a particular IP/16 prefix.

Observation 3: Similar session à Similar throughput

Throughput at different session clusters with particular IP/8 prefixes.

11

Observation 4: Complex relationship between session feature throughput

Combinations of multiple features often have a much greater impact than the individual feature.

12

The impact of the same feature on different sessions could be variable.

Outline

— Motivation — Data-driven Observations

è CS2P Approach

— Evaluation

13

Observation Idea

Many sessions exhibit stateful characteristics in the evolution of the throughput. CS2P learns Hidden-Markov Models (HMM) to capture the states and state transitions. Sessions sharing similar critical characteristics tend to exhibit similar throughput patterns. CS2P groups similar sessions sharing the same critical feature values and uses Cross-Session prediction methodology. The relationship between session features and throughput are quite complex. CS2P learns a separate model for each similar session clusters instead of using a global model.

14

Prediction Engine Step 1: Session Clustering Step 2: Model Training

Video Server Video Player Clients

1.Initial Throughput 2.Prediction Model Step 3: Throughput Prediction and Bitrate Selection

Throughput Measurements

Workflow of CS2P

15

Session clustering-finding critical features

16

All the sessions for training A given subset of session features Session under prediction Sessions matching selected features with Predict the throughput of with these filtered sessions Try another session feature subset Repeat these procedures to find the critical feature set, which yields the most accurate throughput prediction of

Throughput prediction with HMM

17

Model Train: A Hidden-Markov Model per cluster via EM algorithm (offline).

State 1 Gaussian (0.43,0.052) Mbps State 2 Gaussian(2.41,1.492 ) Mbps State 3 Gaussian(1.20,0.102 ) Mbps 0.972 0.876 0.970 0.055 0.012 0.016 0.020 0.069 0.010

Throughput prediction and bitrate selection (online).

Throughput 𝑋𝑢ϵ 𝑺 Hidden State 𝑌𝑢ϵ 𝝍 Xt-1 Xt Wt-1 Wt

Outline

— Motivation — Data-driven Observations — CS2P Approach

è Evaluation

18

Trace-driven simulation setup

19

iQIYI throughput trace:

• Non-overlapping traces
• f training and testing

Algorithms to compare:

• 1. History-based predictor:
• Last Sample, Harmonic-

Mean, Auto Regression

• 2. ML-based predictor:
• SVR, Gradient Boosting

Regression trees

• 3. CFA[1]

Bitrate selection method: Video source:

• “Envivio” from dash.js test

website

• Encoded in H.264/MPEG-4

in 5 bitrate levels

• State-of-art: MPC[2]

[1] J. Jiang, et al. “CFA: A Practical Prediction System for Video QoE Optimization”. In Proc. of USENIX NSDI, 2016. [2] X. Yin, et al.“A Control-Theoretic Approach for Dynamic Adaptive Video Streaming over HTTP”. In Proc. of ACM SIGCOMM, 2015.

20

l Midstream Epoch

pReduce median

error by 50%

l Multi-epoch Ahead

p9% prediction

error for 10 epoch ahead

p50% improvement

Throughput Prediction Accuracy

Takeaway

Midstream Throughput Reduce by 50%

Video QoE

21

— 𝑂𝑝𝑠𝑛𝑏𝑚𝑗𝑨𝑓𝑒 𝑅𝑝𝐹 =

/01234 567 89:6;:1<034 =>?@ABC

— QoE[1] is a linear combination of avg. video quality,

quality variation, total rebuffer time and startup delay.

[1] X. Yin, et al. “A Control-Theoretic Approach for Dynamic Adaptive Video Streaming over HTTP”. In Proc. of ACM SIGCOMM, 2015.

Midstream epoch

8% 5% 19%

22

Metrics

• vs. HM+MPC
• vs. BB
• Avg. Bitrate

10.9% 9.3%

Good Ratio

2.5% 17.6%

Bitrate Variability -2.3%

5.6%

Startup Delay

0.4%

• 3.0%

Overall QoE

3.2% 14.0% Takeaway:

• 1. CS2P improves most of the QoE metrics,

except longer startup delay than BB and higher bitrate variability than HM.

• 2. The overall QoE improvement of CS2P is

3.2% to HM and 14% to BB.

Pilot deployment: multi-city test

Conclusions

23

l Good prediction è Better bitrate selection & adaptation è

Improved video QoE

l Key insights on throughput variability

pEvolution of intra-session throughput exhibits stateful

characteristics.

pSimilar sessions have similar throughput structures.

l CS2P: Cross-session HMM-based approach

lOutperform prior predictors by 50% in midstream prediction error. lAchieve 3.2% improvement to HM and 14% to BB in video QoE.