28
1 On a Unified Architecture for Video-on-Demand Services Jack Y. B. Lee IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 4, NO. 1, MARCH 2002

On a Unified Architecture for Video-on-Demand Services

  • Upload
    norman

  • View
    29

  • Download
    3

Embed Size (px)

DESCRIPTION

On a Unified Architecture for Video-on-Demand Services. Jack Y. B. Lee IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 4, NO. 1, MARCH 2002. Outline. Introduction UVoD Architecture Performance Modeling Numerical Results Simulation Results Interactive Controls Conclusions. Introduction. - PowerPoint PPT Presentation

Citation preview

1

On a Unified Architecture forVideo-on-Demand Services

Jack Y. B. Lee

IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 4, NO. 1, MARCH 2002

2

OutlineIntroductionUVoD ArchitecturePerformance ModelingNumerical ResultsSimulation ResultsInteractive ControlsConclusions

3

Introductiontrue-VoD (TVoD)

Service quality is maximized

near-VoD (NVoD)System cost is minimized

unified VoD (UVoD)Cost-performance tradeoff

4

UVoD Architecture (1)

ttm

5

UVoD Architecture (2)

6

UVoD Architecture (3)

7

Admit-via-UnicastArrives at time ttm-1 < t < (tm – δ)

After (t – tm-1)

8

Recourse reduction over TVoDAdmit-via-Multicast

As multicast channels is fixed, Admit-via-Multicast users will not result in additional loadIncreasing the admission threshold δ then more user will be admitted to the multicast channels

Admit-via-UnicastSince 0 < (t – tm-1) < (T – δ) ≪ L, unicast channels are occupied for a much shorter duration compared to TVoD

9

Performance ModelingLatency (average waiting time)

Admit-via-MulticastAdmit-via-Unicast

Admission ThresholdChannel Partitioning

10

Waiting Times (1)Admit-via-Multicast

wM (δ) = δ / 2

Admit-via-UnicastArrival process

λu = ( 1 – δ / TR ) λ

Service timeUniform distribution between 0 < s < TR – δ

Approximation by Allen and Cunneen for G/G/m queue

11

Waiting Times (2)

Traffic intensity

Coefficient of variation

Average service time

Server utilization

Erlang-C function

12

Admission Threshold

}0,))()((|min{ xTxwxwx RUM

13

Channel PartitioningFind the optimum number of multicast channel such that the resultant latency is minimizedTheorem 1: The optimal proportion of available channels to multicast that minimizes the load at the unicast channels is given by

N

M

NLM

LNN opt

M

2

14

Numerical ResultsCorresponding Latency Formula

NVoD

The latency is constant at 360(900)s for 10(20) movies

TVoD

MNL

WNVoD

2

2)1(

),( L

N

uNEW C

TVoD

15

Admission Threshold verus Queueing Delay

16

Channel Partition versus Latency

17

Latency Comparison With TVoD and NVoD

18

System Capacity and Scalability (1)

%100}0,|max{

}0,|max{

uW

uWG

TVoD

UVoD

λ arrival rate in customers/s

u latency constraint in seconds

WUVoD latency fo UVoD

WTVoD latency fo TVoD

19

System Capacity and Scalability (2)

0.1 0.2

20

System Capacity and Scalability (3)

21

Simulation ResultsEnvironments

Simulation program is developed in C++ using CNCL version 1.10Run 31 days

Model ValidationAdmission Rescheduling

22

Model Validation (1)

23

Model Validation (2)

24

Admission Rescheduling (1)

When Admission Rescheduling?For heavy system loads, a user by Admit-via-Unicast may waiting exceed the time to the next multicast of the requestd movie

25

Admission Rescheduling (2)

26

Interactive Controls (1)Using Unicast Channels

Break current multicast video stream then restart at some pointTreat interactive controls as new-video requests starting at the middle of a movieCould increase waiting for both and interactive requests

27

Interactive Controls (2)Channel Hopping

Client has a buffer large enough to cache TR s

User pause at a movie time Tp

Case1: If resume before buffer overflow, nothing need to be done

Case2:Once buffer is full, stop bufferingLater resume immediately and determine the nearest multicast channel at movie time Tm ≤ Tp

28

ConclusionsThis paper propose and analyzes an architecture that unifies the existing TVoD and NVoDThrough admission-threshold and channel partitioning can achieve cost-performance tradeoffResults show large performance gain