Scalable On-Demand Media Streaming with Packet Loss Recovery

A. Mahanti, D. L. Eager, (USask) M. K. Vernon, D S-Stukel (Wisc)

Presented by Cheng Huang

04.08.2004

2

Introduction

VOD is cool, but might also be expensive Naïve implementation: connection per user Fortunately, users favor a few “hot” movies. It is possible to have inexpensive approach to effectively

share resources (server, network, etc.) and serve lots of requests.

Target of the on-demand streaming scheme Scalable and reliable Fundamental trade-offs

Server bandwidth (network bandwidth), startup delay and client bandwidth

Client buffer requirement (might be of less concern)

3

Outline

Periodic Broadcast Schemes Bandwidth Skimming Scheme Conclusion

4

Outline

Periodic Broadcast Schemes Bandwidth Skimming Scheme Conclusion

5

Pyramid Broadcast

t i me

1 . . . . . .2 6543 7

1

b

2b

2b

2b

1 1 1 1 1 1 1 1 1 1 1 1 1

2 3 2 3 2 3 2 3 2 3 2 3 2 3

4 5 6 7 4 5 6 7 4 5 6 7 4 5

cl i entarri val

s1 s2 s3

start recei vi ng and pl ayi ng s1

start recei vi ng s2

start recei vi ng s3

vi deo sequence

serverbandwi dth

1

2

6

S. Viswanathan, T. Imielinski 1996

6

Skyscraper Broadcast

t i me

1 . . . . . .2 6543 7

1

b

b 11 1 1 1 1 1 1 1 1 1 1 1

cl i entarri val

s1 s2 s4

start recei vi ng and pl ayi ng s1

vi deo sequence

serverbandwi dth

1

1098 11 12 13 1514

s3 s5

2b 23 3 2 3 2 3 2 3 2 3 2 3 2

4b 45 54 5 4 5 4 5 4 5 4 5 4

6b 87 9 10 6 7 8 9 10 6 7 8 9 10

11b 1312 14 15 11 1312 14 15 11 1312 14 15

group 2 group 3

group 1

group 2

group 3

K. A. Hua, S. Shen 1997

7

Fibonacci Broadcast

t i me

1 . . . . . .2 6543 7

1

b

b 1 1 1 1 1 1 1 1 1 1 1 1 1

cl i entarri val

s1 s2 s4

start recei vi ng and pl ayi ng s1

vi deo sequence

serverbandwi dth

1

1098 11 12 13 1514

s3 s5

2b 2 33 2 3 2 3 2 3 2 3 2 3 2

4b 55 66 4 4 5 6 4 5 6 4 5 6

7b 98 10 11

b 1312 14 15 16 1817 19

16 1817 19

7 98 10 11 7 98 10 11 7 98 10 11

4 5 6 4 5

3 2 3 2 3

1 1 1 1 1

1312 14 15 16 1817 19 1312 14 15

8

Optimized Periodic Broadcast

t i me

1 . . .2 6543 7

1

b

b 1 1 1 1 1 1 1 1 1 1 1 1 1

cl i entarri val

s1 s2 s4

start recei vi ng s1

vi deo sequence

serverbandwi dth

1

1098 11 12 13 1514

s3 s5

2b 2 33 2 3 2 3 2 3 2 3 2 3 2

4b 55 66 4 4 5 6 4 5 6 4 5 6

7b 98 10 11

b 1312 14 15 16 1817 19

16 1817 19

7 98 10 11 7 98 10 11 7 98 10 11

4 5 6 4 5

3 2 3 2 3

1 1 1 1 1

1312 14 15 16 1817 19 1312 14 15

start pl ayi ng s1

A. Mahanti, D. L. Eager, M. K. Vernon, D. S.-Stukel 2001

9

Optimized Periodic Broadcast (cont.) General segment length

2

22

12

1

k

k

ll

ll

kkk

10

Generalized OPB

Given the client reception bandwidth as constraint b: maximum client reception bandwidth r: bandwidth of each stream s: number of streams client listen to concurrently

b = r*s

sk

sk

l

lr

l

r

lk

skjj

k

jj

k

1

1

1

1

1

11

Performance of OPB

12

Optimized Periodic Broadcast

t i me

1 . . .2 6543 7

1

b

b 1 1 1 1 1 1 1 1 1 1 1 1 1

cl i entarri val

s1 s2 s4

start recei vi ng s1

vi deo sequence

serverbandwi dth

1

1098 11 12 13 1514

s3 s5

2b 2 33 2 3 2 3 2 3 2 3 2 3 2

4b 55 66 4 4 5 6 4 5 6 4 5 6

7b 98 10 11

b 1312 14 15 16 1817 19

16 1817 19

7 98 10 11 7 98 10 11 7 98 10 11

4 5 6 4 5

3 2 3 2 3

1 1 1 1 1

1312 14 15 16 1817 19 1312 14 15

start pl ayi ng s1

13

Optimized Periodic Broadcast

t i me

1 . . .2 6543 7

1

b

b 1 1 1 1 1 1 1 1 1 1 1 1 1

cl i entarri val

s1 s2 s4

start recei vi ng s1

vi deo sequence

serverbandwi dth

1

1098 11 12 13 1514

s3 s5

2b 2 33 2 3 2 3 2 3 2 3 2 3 2

4b 55 66 4 4 5 6 4 5 6 4 5 6

7b 98 10 11

b 1312 14 15 16 1817 19

16 1817 19

7 98 10 11 7 98 10 11 7 98 10 11

4 5 6 4 5

3 2 3 2 3

1 1 1 1 1

1312 14 15 16 1817 19 1312 14 15

start pl ayi ng s1

pari ty

l oss

14

Loss Recovery Strategy

Digital Fountain approach Use Error Correction Codes to generate n

symbols from k original symbols (n >> k) Symbol: frame, packet, etc.

Recover the k original symbols from any k out of n coded symbols

1 52 63 4 1 2 3

a eb fc d g h i

ori gi nalsequence

codedsequence

start recei vi ng

5 64

j k l

15

Generalized RPB

Reliable Periodic Broadcast (RPB) Maximum tolerable loss rate p

need L/(1-p) packets in order to receive L packets Segment length (a = 1/(1-p))

sk

sk

l

lr

la

r

la k

skjj

k

jj

k

1

1

1

1

1

16

Performance of RPB

17

Outline

Periodic Broadcast Schemes Bandwidth Skimming Scheme Conclusion

18

Bandwidth Skimming Scheme

19

Bandwidth Skimming Scheme (cont.)

t i me

a1

cl i entarri val

mul t i cast starts

b1c1

substream asubstream bsubstream c

a2b2c2

a3b3c3

a4b4c4

a1b1c1

a2b2c2

a3b3c3

a5b5c5

substream asubstream bsubstream c

20

Reliable Bandwidth Skimming Pros

Bandwidth decreases as the client request rate decreases

VCR functionality – fast forward Protocol

Primary stream is transmitted at rate 1 Secondary stream is transmitted at rate p/(1-p) Merging happens between primary streams, as w

ell as secondary streams Real system in Univ. of Wisconsin

21

Reliable Bandwidth Skimming (cont.)

22

Conclusions

Propose two reliable on-demand streaming schemes RPB RBS

Schemes investigate the fundamental trade-offs of on-demand streaming service