March-04 1

Proxy Cache Management for Fine-Grained Scalable Video Streaming

Jiangchuan LiuThe Chinese University of Hong Kong

Xiaowen Chu and Jianliang XuBaptist University of Hong Kong

Infocom’04, March 2003

March-04 2

Outline

Introduction and Motivations Problem Settings and Solutions Performance Evaluation and Comparison Conclusion and Future Work

March-04 3

Proxy

Proxy

Video Server

Backbone Network

Video Caching

Proxy caching

saving video objects at proxies close to clients

Temporal locality Geographical locality

March-04 4

Unique Features

Video objects vs. Web objects

High data rate, yet adaptive Long playback duration

► Various interactions: ־ random access ־ early termination

► Huge volume ־ one-hour MPEG-1, about 675 MB

March-04 5

Partial Caching Interval caching (Dan96, Tewari98…)

Prefix caching (Sen99, Jin02…)

Segment caching (Wu01, Miao02, Chae,Chen03…)

1 4 5 6 7 8 9 3 4 5 6 7 8 9

r1r2 r1r2

1 2

2 3

March-04 6

Common Assumptions

Continuous playback No interactions or premature termination Homogeneous segment access rate

Homogeneous clients Identical access bandwidth

Time partitioning only Non-adaptive caching Non-scalable single-rate video

March-04 7

Our Assumptions and Objectives Assumptions

User interactivity: early terminations Heterogeneous client access bandwidths

Vehicle Fine-Grained Scalable (FGS) video

Objectives Rate adaptive caching and streaming Understanding the benefits (?) of FGS caching

־ Vs. Replication (Hartanto02), Transcoding (Tang02)

March-04 8

Related Work and Differences

Video staging (Zhang00)

Quality adaptation (Yu00, Rejaie00) Adaptation per user basis Replacement Prefetching

Blocking performance (Kangasharju02) Admission control Homogeneous access rate

Caching for VCR-operations (Fahmi01)

cut-off rate

March-04 9

Outline

Introduction and Motivations Problem Settings and Solutions Performance Evaluation and Comparison Conclusion and Future Work

March-04 10

System Model and Operations

FGSstream

(bi)

ClientRequest

(ci)

ProxyRequest

UncachedPortionFilter

VideoRepository

Video Server Video Proxy

Monitor

Fil

ter/

Ass

embl

er

ProxyManager

BackboneNetwork

Cache

t i mebase l ayer

Enh l ayer

rate

March-04 11

Model Parameters

L Length of the video, t sL L L= +

tL Length of the prefix

sL Length of the suffix

ETp Probability of early terminations

H Proxy cache size for the video

baser Base layer rate of the video

l Client request rate for the video

M Number of client classes

ip Probability that a client is in class i , 0ip >

ic Client bandwidth of class i

ib Streaming rate for a class i client ( base i ir b c£ £ )

ia Utility of a client in class i , /i i ib ca =

March-04 12

Problem(1): Caching Strategy

Explore the rate adaptability of FGS in caching Problem:

Given cache size and client utility level, which portion of an FGS video should be cached

Objective Min transmission cost

Difficulty: Heterogeneous bandwidth demands Non-uniform segment access rate A 2-D space: time and rate

־ Greedy is not optimal

March-04 13

Problem(1): Solution

2-segment Exhaustive search

Multi-segment Access probability – segment/rate

March-04 14

Problem (2): Utility Assignment

Explore the rate adaptability of FGS in both caching and streaming

Problem Given cache size and backbone bandwidth limits, jointly decide

the caching strategy and utility assignment of each client. Objective

Max expected client utility

max

, 1

MH i ii

U ph l a=

= å

,

. .st

/ 1base i ir c a£ £ , 1,2,...,i M=

,

1 1i i i ic ca a- - £ , 2,3,...,i M=

ˆHB Bh£

March-04 15

Problem(2): Iterative Solution

Difficulty Utility assignment

optimal caching strategy (problem 1) Caching strategy

optimal utility assignment

(dynamic programming) Iterative optimization

2-Segment – exhaustive search on cache partition

March-04 16

Optimization for Multiple Objects

Heterogeneity of objects Access rate Access bandwidth Client distribution

Cache partitioning Backbone Bandwidth partitioning

2-D Knapsack Pseudo-polynomial partitioning algorithm

March-04 17

Outline

Introduction and Motivations Problem Settings and Solutions Performance Evaluation and Comparison Conclusion and Future Work

March-04 18

Sample Configuration

Client Multiple classes Uniform, skewed

Utility function Linear

2-segments: early termination Probability = 0.3 A conservative configuration !

March-04 19

Backbone Bandwidth Reduction

0

10

20

30

40

50

60

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Normalized Cache Size

Bac

kbon

e B

andw

idth

Red

uctio

n (%

)

Compared to MaxLen

Compared to MaxRate

MaxLen: length first MaxRate: rate first

March-04 20

Joint Optimal Caching & Utility Assignment

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Normalized Backbone Bandwidth

Exp

ecte

d U

tility

CacheSize = 0.1

CacheSize = 0.3

CacheSize = 0.5

March-04 21

Results for Multiple Videos

0

3

6

9

12

15

0 0.2 0.4 0.6 0.8 1

Skew Factor

Util

ity Im

pro

vem

en

t (%

) x

Joint OptimalBand OptimalCache Optimal

Baseline uniform cache partition + proportional bandwidth partition

March-04 22

Scalable Video or Replicated Video?

Backbone bandwidth reduction with FGS

0

10

20

30

40

50

60

70

80

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Normalized Cache Size

Ba

ckb

on

e B

an

dw

idth

Re

du

ctio

n (

%)

UniformS-narrowS-wide

Optimal caching for replicated video 1D knapsack

March-04 23

Scalable Video or Transcoding ?

1

10

100

1000

10000

FGS

Filter

FGS

Assembler

SFD

Filter

FD-DC

Filter

RQ

Filter

Co

mp

uta

tion

Tim

e x

x (

μs/

fra

me

)

Given a frame interval of 30 ms, our PC can support about 300 concurrent filter/assembler operations

March-04 24

Outline

Introduction and Motivations System Description Problem Settings and Solutions Performance Evaluation and Comparison Conclusion and Future Work

March-04 25

Conclusion FGS-based proxy caching Key problems

Optimal caching strategy Optimal utility assignment Optimization for multiple videos

Performance Evaluation Backbone bandwidth reduction Utility improvement Comparision

־ FGS caching vs. Replication caching

־ FGS filtering vs. Transcoding

March-04 26

Future Work

Utility functions Tradeoff: accuracy/speed

Multi-segments Fastforward, backward

Practical issues Error control Synchronization Signaling

March-04 27

Thanks

Q & A

March-04 28

Scalable Video or Replicated Video?

0

10

20

30

40

50

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Normalized Backbone Bandwidth

Util

ity I

mpr

ovem

ent

(%)

'

Cache Size = 0.1

Cache Size = 0.3

Cache Size = 0.5

Utility improvement with FGS