Rebera: Real-time Bandwidth Estimation and Rate Adaptation...

Rebera: Real-time Bandwidth Estimation andRate Adaptation for Video Calls

over Cellular Networks

E. Kurdoglu, Y. Liu, Y. Wang 1

Y. Shi, C. Gu, J. Lyu 2

1NYU Tandon School of Engineering

2Tencent Co. Ltd.

May 11, 2016

Kurdoglu, Liu, Wang, Shi, Gu, Lyu NYU Tandon School of Engineering, Tencent Co. Ltd.

Rebera: Real-time Bandwidth Estimation and Rate Adaptation for Video Calls over Cellular Networks

Outline

1 Introduction and Problem Setting

2 Rebera ComponentsMeasuring the Current BandwidthPredicting the Future BandwidthDetermining the Sending Rate (Bit-budget)Encoder & Video Stream StructureDynamic Frame Selection

3 Experiment Results



Video calls: very low-delay, interactive applications

1 Need to adapt transmission rate wrt network speed:Tx rate > Bandwidth ⇒ intolerable delaysTx rate < Bandwidth ⇒ under-utilized link

Accurate measurement and forecast of bandwidth is hard!

2 Need to adapt video encoding rate wrt to network speed

Accurate low-delay video rate control is hard!

3 Need resiliency against late or lost packets

Cellular links can be highly volatile: link speeds, delays...



A Video Call Scenario

ISP1 Internet ISP2

S D

BS BD

Challenge: measuring and predicting the e2e bandwidth

Assume: cellular links are the only bottlenecks

Model: cellular links as isolated queues (no cross-traffic)

uplink buffer in the device, downlink buffer in the BS



Outline






Rebera Overview

PacketTransmission

H.264/AVCHierP

CapacityPrediction

PacketReception

CapacityMeasurement

Decode &Display

sender

receiver

We periodically insert I-frames

Intra-period: time between two successive I-frames (≈1 sec)

Rate control performed at the beginning of each intra-period



Measuring the Current Bandwidth

Observation: We can measure bandwidth by sending packet trainsand measuring packet inter-arrival times at the receiver.

…

a2 a3 ap

z2

z3

zp

time⇒ mn ,

z2 + · · ·+ zpa2 + · · ·+ ap

,Zn

An

Key idea: Use the packets belonging to a frame as a packet train.

1 Estimate avg. ABW in thelast intra-period

2 Feed it back to sender

3 Slide the window

ZK

AK

Z1

A1

now

sliding window



Predicting the Future Bandwidth

Observation: Bandwidth measurement values are correlated.

Key idea: Forecast future bandwidth from past values with onlinelinear adaptive filtering (e.g. RLS)

time(sec)200 250 300 350 400 450 500

kbps

0

500

1000

1500MeasurementPrediction (RLS)

ck+1 =M−1∑i=0

wi (k)ck−i

M number of filter tapsw(k) filter tap vector of length Mck measured capacity



Determining the Sending Rate (≡ Bit-budget)

Observation: Forecast errors are inevitable.

Key idea: Estimate the (almost) worst-case scenario to find the(almost) safest max sending rate.

rk sending rateεk exact-to-predicted ratio (= ck

ck)

qk backlogged bits

Pr(rk+1 > ck+1) ≤ δ ⇒ Pr

(εk+1 <

rk+1

ck+1

)≤ δ

r safek+1 = ck+1 × δ-quantile of ε

Bit-budget for intra-period k + 1

bk+1 = r safek+1 × T − qk



Encoder & Video Stream Structure

Observation: Resiliency against packet losses is essential.

Key idea: Use hierarchical-P temporal layering.

GoP GoP

I0

P1

P2

P3

P4

P5 P7

I8P6

time

Intra-period1 Enables UEP X

2 Complexity & overheadkept minimum X

3 DFS does not causejitter. X



Dynamic Frame Selection for Hierarchical-P Video

Observation: When video rate > target bitrate, we can discardhigher layer frames and maintain a stable frame rate.

Requirement: Must make the “send/discard” decision on the fly.

Decision for frame P1?

Cannot wait for {P2,P4,P6}

GoP GoP

I0

P1

P2

P3

P4

P5 P7

I8P6

time

Intra-period

Sequential decision makingalgorithm considers:

1 Qualities of layers

2 Decoding dependencies

3 Smoothness of frameintervals



Outline






Our Testbed

NYU Network

S D

CS CD

eth0eth1

eth0eth1

Implemented in C++ (platform: Linux)Encoder: x264, modified to generate hierarchical-P w/ 3 TLsDecoder: FFmpegCellular links emulated by CellSim

FaceTime used as a benchmark

Performance metrics: ABW utilization, 95-percentile packet &frame delays



Evaluation under Cellular Bandwidth Traces

0 200 400 600 800 1000

Mbps

0

0.5 Tr1

0 200 400 600 800 1000

Mbps

0

0.5 Tr2

0 200 400 600 800 1000

Mbps

0

1 Tr3

0 200 400 600 800 1000

Mbps

012 Tr4

0 200 400 600 800 1000

Mbps

0

2Tr5

time (sec)0 200 400 600 800 1000

Mbps

0

2Tr6




ABW utilization percentage10 20 30 40 50 60 70

95

-pe

rce

ntile

fra

me

de

lays (

mse

c)

0

1000

2000

3000

4000

5000

6000

A-BA-C B-A

B-CC-A

C-B

att-tmobile

att-verizon3g

att-verizon4g

tmobile-att

tmobile-verizon3g

tmobile-verizon4g

verizon3g-att

verizon3g-tmobile

verizon4g-att

verizon4g-tmobile

A-BA-CB-A

B-C

C-A

C-B

att-tmobile

att-verizon3g att-verizon4g

tmobile-att

tmobile-verizon3g

tmobile-verizon4g

verizon3g-att

verizon3g-tmobile

verizon4g-att

verizon4g-tmobile

Double-link scenarios

FaceTimeRebera

Rebera avg. util.: 55%, avg. 95-percentile frame delay: 865 msecFaceTime avg. util.: 32%, avg. 95-percentile frame delay: 2.8 sec




ABW utilization percentage10 20 30 40 50 60 70

me

an

fra

me

de

lays (

mse

c)

0

200

400

600

800

1000

1200

1400

1600

1800

A-B

A-C

B-A

B-C C-AC-B

att-tmobile

att-verizon3g

att-verizon4g

tmobile-att

tmobile-verizon3g

tmobile-verizon4g

verizon3g-att

verizon3g-tmobile

verizon4g-att

verizon4g-tmobile

A-BA-C

B-A B-C

C-A

C-B

att-tmobileatt-verizon3g

att-verizon4gtmobile-att

tmobile-verizon3g

tmobile-verizon4g

verizon3g-att

verizon3g-tmobile

verizon4g-attverizon4g-tmobile

Double-link scenarios

FaceTimeRebera

Rebera avg. util.: 55%, avg. mean frame delay: 295 msecFaceTime avg. util.: 32%, avg. mean frame delay: 620 msec




sec0 100 200 300 400 500 600 700 800

kbps

0

500

1000

1500

2000verizon3g-att

CapacityFaceTimeRebera

sec0 100 200 300 400 500 600 700 800 900 1000

kbps

0

500

1000

1500

2000verizon3g-tmobile


sec0 100 200 300 400 500 600 700 800

kbps

0

500

1000

1500

2000att-verizon4g


sec0 100 200 300 400 500 600 700 800

kbps

0

500

1000

1500

2000tmobile-att




Conclusions

Proposed a real-time bandwidth estimation for video calls→ can be integrated with existing CC mechanisms

Sequential linear prediction allows forecasting→ can be further improved: ML, etc...

Hierarchical-P structure enables easy rate control and someresiliency

Proposed method integrated into last version of WeChat



Thanks!

Find the code and reproduce our work at

http://vision.poly.edu/

Q&A



Rebera: Real-time Bandwidth Estimation and Rate Adaptation...

Documents

Transcript of Rebera: Real-time Bandwidth Estimation and Rate Adaptation...