iDASH: improved Dynamic Adaptive Streaming over HTTP …web.cs.wpi.edu/~claypool/mmsys-2011/Day3-5_idash.pdfiDASH – Yago Sánchez MMSYS-2011 SVC can be efficiently encoded to get

iDASH: improved Dynamic Adaptive Streaming over HTTP using Scalable Video Coding

Yago Sánchez, Thomas Schierl, Cornelius Hellge, Thomas Wiegand - Fraunhofer HHI, Germany

Dohy Hong - N2N Soft, France Danny De Vleeschauwer, Werner Van Leekwijck, Bell Labs - Alcatel Lucent, Belgium

Yannick Lelouedec - Orange Labs FT, France

iDASH – Yago Sánchez MMSYS-2011

Motivation

HTTP Streaming

SVC

Caching

Results

Conclusion

OUTLINE


Service providers have resorted to use HTTP/TCP for Multimedia Delivery

 Relying on RTP/UDP may result in traversal problems with NATs and Firewalls

 HTTP/TCP allows re-useing existing HTTP cache infrastructures

Different type of users => poor cache performance

 Different equipment capabilities

 Different connectivity characteristics

Resources at the network limited  Capacity on the links shared

 Storage in the HTTP cache not enough

MOTIVATION


HTTP caches placed in the network reducing the load on the server and transit link

Video split in smaller segments/ chunks and transported over HTTP

Different representation of the video to allow the users to request the one that matches at best their capabilities

HTTP STREAMING - INTRODUCTION

Internet

server

R (kbps) = 768...

0 kbps

768 kbps

1536 kbps

768 kbps

768 kbps

768 kbps

768 kbps

768 kbps

transit link


Multiple Representations Video on Demand (MR-VoD) is less effective for caches performance than Single Representation Video on Demand (SR-VoD)

HTTP STREAMING – EFFECT OF MULTIPLE REPRESENTATIONS

00,10,20,30,40,50,60,70,80,91

0 1000 2000 3000 4000

cache-‐hit-‐ratio

cache capacity (C) [media units]

SR-‐VoDMR-‐VoD


With MR-VoD more data has to be sent from the server (though the transit link) than with SR-VoD

HTTP STREAMING – EFFECT OF MULTIPLE REPRESENTATIONS

0

50

100

150

200

250

0 1000 2000 3000 4000

Tran

sit link

traffic [M

bps]


SVC-‐VoDMR-‐VoD


Proposed solution:

 USE SCALABLE VIDEO CODING (SVC) TO IMPROVE CACHE PERFORMANCE

HTTP STREAMING


This works focuses on SNR Scalability

Representations are created from sub-streams of the whole SVC stream

Different representations mapped to Operation Points (OP)

 OPs based on complete Layers => If many it may increase the SVC overhead

 OPs also based on smaller parts of layers.

SVC (SCALABLE VIDEO CODING)


Different representations/Operation Points (OP)


SVC (SCALABLE VIDEO CODING) - EXAMPLE

GOP Border GOP Border

T0

T1

T2

T3




SVC (SCALABLE VIDEO CODING)


T0

T1

T2

T3

OP 0






T0

T1

T2

T3

OP 1






T0

T1

T2

T3

OP 2






T0

T1

T2

T3

OP 3






T0

T1

T2

T3

OP 5






T0

T1

T2

T3

OP 7


Caches store data supposed to be asked in the future to relieve the server from having to send directly the data

There are many different cache replacement algorithms that improve the cache performance based on some especial criteria/metric

 LRU: Least Recently used

 LFU: Least Frequently used

 CC: Chunk-based Caching

 …

CACHING - I


Caches store data supposed to be asked in the future to relieve the server from having to send directly the data

There are many different cache replacement algorithms that improve the cache performance based on some especial criteria/metric

 LRU: Least Recently used

 LFU: Least Frequently used

 CC: Chunk-based Caching

 …

CACHING - I

... ... Certain video file

Other files

Chunk #n Chunk #n+1 & #n+2

Chunk #n+3

Chunk #n+4 & #n+5


Cache performance can be easily improved by using SVC

 Cache storage better used than with MR-VoD

 Uneven distribution of request = more pronounced

CACHING - II


CACHING - CAPACITY USAGE COMPARISON

MR-VoD

SVC-VoD

File 1File 2File 3File 4File 5File 6File 7

File 1File 2File 3File 4File 5File 6File 7

Removed files from cache

rep1rep2

rep3

rep1rep2rep3


CACHING – EXAMPLE MR-VoD

Internet

wired wired

server

Q1Q2

Q3

Q2Q3

R (kbps) = 256 512 7681536 Kbps

1280 Kbps 768 Kbps

wireless

Q1

wireless

Q2

512 Kbps

256 Kbps

512 Kbps

768 Kbps


CACHING – EXAMPLE MR-VoD

Internet

wired wired

server

Q1Q2

Q3

Q2Q3

R (kbps) = 256 512 7680 Kbps

256 Kbps 768 Kbps

wireless

Q1

wireless

Q2

Q1

wireless

Q3

768 Kbps256 Kbps


CACHING – EXAMPLE SVC-VoD

Internet

wired wired

server

Q1Q2Q3

Q2Q3

R (kbps)844550256

844 Kbps

550 Kbps844 Kbps

wireless

Q1

wireless

Q2

844 Kbps

550 Kbps

256 Kbps

550 Kbps


CACHING – EXAMPLE SVC-VoD

Internet

wired wired

server

Q1Q2Q3

Q2Q3

R (kbps)844550256

0 Kbps

294 Kbps0 Kbps

wireless

Q1

wireless

Q2

Q1

wireless

Q3

844 Kbps256 Kbps


Statistics for users requests extracted from a deployed system

 Measured during a month

 More than 5000 files among which the users can choose

 3400 requests per day on average

 This real system is SR-VoD

Cross traffic in access links

 2 scenarios considered

- Heavy cross traffic in Access link - Light cross traffic in Access link

SIMULATIONS – SIMULATED SCENARIO


4 representations for each video to allow adaptation

SVC overhead of 10% compared to AVC

Rate Distribution of the representations

SIMULATIONS – AVAILABLE REPRESENTATIONS

Codec Rep. 1 Rep.2 Rep.3 Rep. 4

AVC 500 kbps 1000 kbps 1500 kbps 2000 kbps

SVC 500 kbps 1066 kbps 1633 kbps 2200 kbps


Needed throughput in transit link and capacity in caches for all reps

AVC = 5000 kbps per video = (video of 90 min) 3375 MB

SVC = 2200 kbps per video = (video of 90 min) 1485 MB

Rate Distribution of the representations

SIMULATIONS – AVAILABLE REPRESENTATIONS

Codec Rep. 1 Rep.2 Rep.3 Rep. 4

AVC 500 kbps 1000 kbps 1500 kbps 2000 kbps

SVC 500 kbps 1066 kbps 1633 kbps 2200 kbps


Four state Markov-chain simulated

Cross traffic characterization:

 Mean state sojourn time

 Average percentage of time in each state

CROSS TRAFFIC

1 2 3 4

p11

p12

p21

p22

p23

p32

p33

p34

p43

p44

[ ] ( ) ( )ii

iitii

tii p

ppttE i

i−

=−+=∑∞

= 111**1

0

ππ =Ppi *:

P=[pij]


Transition matrix

Mean state sojourn time

E[ti] = 40 min (for segment of 10 sec length)

Average percentage of time in each state 25% of the time in each state

HEAVY CROSS TRAFFIC

⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

=

996.0004.000004.0992.0004.000004.0992.0004.000004.0996.0

P


LIGHT CROSS TRAFFIC

Transition matrix

Mean state sojourn time

E[ti]=(approx.){2min, 2min, 10min, 40min}(for segment of 10 sec length)

Average percentage of time in each state p={9.1%, 9.5%, 19.1%, 62.3%}

⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

=

996.0004.000013.0985.0002.000004.09.0096.0001.09.0

P


Comparison of cache efficiency with MR-VoD and SVC-VoD

RESULTS-HEAVY CROSS TRAFFIC

Cache capacity (media units)

LRU CC

MR-VoD SVC-VoD MR-VoD SVC-VoD

500 30.9 % 45.6 % 42.9 % 56.6 %

1000 42.1 % 58.2 % 52.0 % 64.5 %

2000 54.6% 69.0% 61.5% 72.0%


00,10,20,30,40,50,60,70,80,91

0 1000 2000 3000 4000



MR-‐VoDlayer1layer3SVC-‐VoDlayer2layer4


RESULTS-HEAVY CROSS TRAFFIC


00,10,20,30,40,50,60,70,80,91

0 1000 2000 3000 4000



MR-‐VoDSVC-‐VoDlayer1layer2layer3layer4


RESULTS-LIGHT CROSS TRAFFIC


SVC can be efficiently encoded to get a high enough number of representations

HTTP-based VoD service can be easily improved by only using SVC

SVC-VoD results in a higher cache-hit-ratio and consequently in a lower traffic transmitted across the transit link between the server and the cache

Further work:

 Study the enhanced adaptability, faster response times with SVC

 Advantages of using SVC in DASH for Live Streaming

CONCLUSIONS


THANKS FOR YOUR ATTENTION!!!

Acknowledgments:

 The research leading to these results has received funding from the European Union's Seventh Framework Programme ([FP7/2007-2013] ) under grant agreement n° 248775

Documents

iDASH: improved Dynamic Adaptive Streaming over HTTP …web.cs.wpi.edu/~claypool/mmsys-2011/Day3-5_idash.pdfiDASH – Yago Sánchez MMSYS-2011 SVC can be efficiently encoded to get