End-to-End Analysis of Distributed Video-on-Demand Systems Padmavathi Mundur, Robert Simon, and Arun K. Sood IEEE Transactions on Multimedia, February.

End-to-End Analysis of Distributed Video-on-Demand Systems

Padmavathi Mundur, Robert Simon, and Arun K. SoodIEEE Transactions on Multimedia, February 2004

Outline

Motivation Hierarchical VoD architecture Analytical model Evaluation methodology and results Conclusion

Motivation

In a real environment, if a video requires R mbps transmission rate, allocate R mbps bandwidth is not accurate enough

From network view, analyze the bandwidth required for videos

Hierarchical VoD architecture

Data flow at server

Double buffer technique RSVP

Disk scheduling and double buffer scheme in the server

3541672

1. RAID-5storage

2. SCAN EDF

scheduling

(RSVP)

Token bucket + WFQ

Traffic regulator at server (1/2) Leaky bucket

Control average rate

sendpkts

packets wait

packetsto

network

r pkts/sec

Traffic regulator at server (2/2) Token Bucket

Control average rate Control input burst size

removetoken

packets wait

packetsto

network

r tokens/sec

buckets holds up to b tokens

Weighted fair queuing (WFQ) at server Provide different priority to different packets

Combine token bucket & WFQ

Token bucket scheme controls the average output rate

WFQ allocates different resource to different users

Token bucket + WFQ provide delay upper bound

Receiver B

Receiver A

Sender

Session (Ipa,PID,Port)

path (2)

Merge point

Session (Ipa,PID,Port)

Session (Ipa,PID,Port)

IGMP (1)

IGMP(1)

Resv(3)

Resv (3)

Path message

Resv message

IGMP message

DataPacket (4)

Resource reservation protocol (RSVP) along the routing path

Review the whole data flow

RAID 5 storage

SCAN EDF scheduling

Double buffer technique Token bucket +

WFQRSVP

Admission control scenario

Remote cluster

Remote cluster

Localcluster

Localdistribution

network

Networkconnections

requestDisk admission control

Check available bandwidth

Analysis – admission control

Server disk

, if accept the request

Network

p

d

R

Rn nnc 1

jrp ARR

overall disk bandwidth

client playback rate

bandwidth available on jth link

reserved rateclient playback rate

Analytical model – use delay bound to calculate reserved bandwidth

WFQ + Token bucket

rJ

bJ

wJ

r1

b1

w1

……

J

j j

rr

A

MD

MJBR

1

maxmax

)1(

J

j jrr

r

A

M

R

MJ

R

BD

1

maxmax

)1(

maxMM : max packet size for the flow

: MTU

rB : retrieval block size

Performance evaluation – request handling policy Redirect:

A blocked request at one resource is simply redirected to other resources

Split-based Sharing the loads to other resources

Simulation setup – environment

Remote cluster1

Remote cluster2

Localcluster

Network1 Network2

requests

•Servers in local cluster: 5

•Storage capacity per local server: 500 GB

•Disk transfer rate at local server: 1.2 Gbps

•Hops to remote cluster1: 3

•Hops to remote cluster2: 6

•Max. Transmission Unit: 1500 Bytes

•Maximum packet size: 1500 Bytes

•Network bandwidth: 2488 Mbps

•End-to-end delay 300 ms

•Size of video collection 150

•Size of videos in GBytes: 2.46 to 4.8

•Service time in hours: 0.68 to 2.03

•Video popularity: according to Zipf distribution

•Request arrival interval: adopt Poisson distribution

Simulation setup – request handling policies Redirect

Redirect order: LC RC1 RC2 Split

Split50-60: 50% are served in LC, 60% of the remains are served in RC1, the rests are in RC2

Split-redirect Split first, also contains redirect policy

Simulation setup – scenarios

Replicated video collection (RVC) All videos are available on local or remote servers

Distributed video collection (DVC) Only a partial set of videos is available on the

local cluster, the requests for non-available parts are served by remote clusters

Simulation results – compare performance of request handling policies in RVC Purpose: test the performance of the VoD system us

ing different request handling policies Redirect policy performs better than the other two p

olicies

Simulation results – difficulties with split-based policies in RVC The lines are

crossed over in the previous figures (Ex: split-50-60 and split-60-60)

It is difficult to pick an efficient split for a given workload

Split-60-60 performs better at low load

Split-50-60 performs better at heavy load

Simulation results – performance at each resources for split policies in RVC Use individual resource

performance to help explain the crossover and divergence behavior

Simulation results – efficient split policy in RVC

Split requests proportional to their resource

It may difficult to know remote clusters since they may be dynamically shared with other user populations

Simulation results – varying the number of videos on local server in DVC

local storage size

local video number

100 30

200 76

300 135

400 147

500 150

Distribute the available storage capacity at the local cluster to videos in proportion to their popularity

Redirect policy only Class1: top 20% popular, class2: 20~60% popular, class3: last 40%

popular

Conclusion and distribution

Develop a method to analyze distributed VoD systems

Use an extensive simulation to the distributed VoD architecture and evaluate several request handling policies