Predictive Buffering for Streaming Video in 3G

Networks

Varun SINGH, Jörg OTT, Igor Curcio

Problem Statement

•  22% of mobile broadband in the US is YouTube [MobileTrends, 2011]. –  [note] http://m.youtube.com uses RTSP instead

of HTTP based progressive download.

•  Problems with 3G connectivity: – Mobility (fading, interference, cell loading,

handovers) è affects available throughput

Streaming

•  If there is loss of connectivity – Pre-buffer, typically, ~5 to 10s

•  May be insufficient in the case of 3G

– Rate-switching usually happens after disruption is detected

ATT Coverage map

http://www.wireless.att.com/coverageviewer/

Architecture

Streaming*

Server*

Network*Congestion*Map*Service*

RTP

Throughput Updates Look-ahead Request

Available Throughput

RTCP

Streaming*Client*

Signaling

Alternative: RTSP SPEED parameter

Streaming Server! Streaming Client! Network Coverage Map Server!Client startup…!

[ASYNC] Acquiring Satellite…![ASYNC] Current Position known!Media Session Initiated!

RTP Media! Database Lookup!Finds locations around requested location. !If (coverage hole is found) !then complete hole Info is sent!else only neighborhood info is sent!

RTP continues!(…)!

RTP Session Started!SRs and RRs will be scheduled according to RFC3550+RFC4585!

If (Coverage hole is found)!then Schedule Predictive Rate Control! !

Estimated Capacity Vector

(ECV)!Varying Transmission Rate or Rate Switch!

RTP Media!

Stream switched to lower rate!

RTP Media!

Stream switched back to normal rate!

Database Updates!Throughput of the Location is updated by using moving average of the new value with older values!

Look-ahead

•  Known travel route – Client can calculate maximum size of pre-buffer

for the whole trip.

•  Area look ahead – Client can only calculate optimum buffer for the

known outages

– Client subscribe to locations with poor connectivity for updates

Early Scheduling

Late Scheduling

Client operation

t (in sec)!

BW (i

n kb

ps)!

Playback Rate!

Time To Outage (Ttto)!

Channel Capacity!

Receiver Rate!4. Data pre-buffered due to predictive feedback!

Initial !pre-buffering!

Time Duration of Outage (Tgap)!

2.Congestion hole detected! 3. Client schedules

variable transmission rate or switches to lower rate!

1. Look-ahead for holes!

5. Client resets to normal media rate or transmission rate!

Calculate impact of coverage hole

function(speed, location)

Grouping of throughput values

•  Calculate STDEV for each region [1,3]

•  Calculate AVERAGE/median for each region [2]

•  Definition of region –  Areas with similar

throughput

This is out-of-scope of the paper

Use K-means square algorithm Grouping based on throughput

Protocol Design and implementation

•  HTTP between Coverage Map Server and client –  REpresentational State Transfer (REST) APIs –  JSON encoded responses

•  How to throttle the rate –  Dictionary of {time, throughput} –  RTSP Speed parameter

•  Gstreamer using x264 and JRTPLib •  PostgreSQL, C++

HelSinki Bus Routes

Second half of 2010 ~400,000 updates •  40-50 bus rides •  Walking around

the city/campus

Throughput Around the University Area

Bandwidth along a travel route

0

500

1000

1500

2000

2500

3000

60.187902,24.828374

60.186355,24.826504

60.174801,24.800772

60.175371,24.70583

60.175371,24.70583

60.163694,24.79771

60.161182,24.760993

60.161997,24.748103

60.165134,24.732384

60.162382,24.801506

60.185611,24.825956

60.162604,24.774443

60.159163,24.785112

60.1827,24.795065

60.186201,24.823197

thro

ughp

ut (k

bps)

locations (lat,lon)

Intermediate Quartile Range

Performance Analysis

0 500

1000 1500 2000

0 50 100 150 200 250Thro

ughp

ut (k

bps)

a) No Adaptation

3G LinkReceiver Rate

0 500

1000 1500 2000

0 50 100 150 200 250Thro

ughp

ut (k

bps)

time (s)

c) Rate Switching

0 500

1000 1500 2000

0 50 100 150 200 250

b) Omniscient

0 500

1000 1500 2000

0 50 100 150 200 250time (s)

d) Late Scheduling

Video Quality

•  Average PSNR over multiple runs

•  No adaptation: 27.5

•  Omniscient: 43.12

•  Rate-switching: 42.75

•  Late-switching: 48.43

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1

CC

DF

NCMSratio

OmniscientLate Scheduling

Early Scheduling

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1

CC

DF

NCMSratio

At t=1h

At t=8h

NCMSratio= Response/Actual

Conclusions

•  We find that the information provided by coverage map service is suitable for – Predictive rate-switching

– Predictive pre-buffering

•  Future Work: Integrate with a DASH system