On-Demand Media Streaming Over the Internet Mohamed M. Hefeeda, Bharat K. Bhargava Presented by Sam Distributed Computing Systems, 2003. FTDCS 2003. Proceedings.

On-Demand Media Streaming Over the Internet

Mohamed M. Hefeeda,Bharat K. Bhargava

Presented by Sam

Distributed Computing Systems, 2003. FTDCS 2003. Proceedings. The Ninth IEEE Workshop on Future Trends of , 2003

2

Agenda

Introduction Peer-to-peer model Peer-to-peer streaming protocol Architecture

Index-search algorithm Index-dispersion algorithm

Simulations Evaluation Conclusion

3

Introduction

Overall system capacity is the total number of concurrent clients (peers) the system can support

Mainly Limited by out bound network bandwidth

Others limitationsProcessing power, memory size, I/O speed

4

Peer to Peer (p2p) Model

Peer acts as min-server and client.As a client, it requests media files from the

systemThen the client stores segments of the media

files that it is already consumed for a specific period of time

It can provide these segments to other requesting peers in the system

5

P2p Model

Denote PP = {P1, P2, … PN} the set of all peers in the system

For every Pi , where 1 ≤ I ≤ N, Ri (in Kb/s) : max. rate peer Pi is willing the share with others

Gi (in bytes) : max. storage space the peer is willing to allocate to store segments of one or more media files

Ci : max. number of connections that can be opened to serve requesting peers

6

P2p Model

Media files The set of movies currently available in the system Bit rate = R kb/s Each divided into N segments, a minimum unit which a p

eer can cache. Multiple peers can stream to the requesting peer, sendin

g a rate less than R.

7

P2P Streaming Protocol

The requesting peer checks for the availability of the desired media file in the system

Stream segment by segment

The peer cache the segments

Three Phases

8

Phase IThe requesting peer checks for the availability of

the desired media file in the system

Store other peers information about

• IP address

• Available streaming rate

• Some reliability information from the peer’s history

128.0.20.16 100k b/s Low

9

Phase IThe requesting peer checks for the availability of

the desired media file in the system

Sort based on

• Proximity to client (i.e. hop count)

• Available streaming rate

• Average on-line time

128.0.20.16 100k b/s Low

10

Phase IThe requesting peer checks for the availability of

the desired media file in the system

• End with a verification step to make sure all segments are available.

• Otherwise, backoff and try later

11

Phase IIStream segment by segment

For every segment si, the protocol concurrently connects to all peers that are scheduled to provide pieces of that segment.

Define Pj as the set of peers supplying segment j (sj).

If a peer Px Є Pj has a rate Rx ≤ R,

it provides |sj|(Rx/R) bytes starting at wherever peer Px-1 ends

Since every peer supplies a different piece of the segment

and

All pieces of the segment will be downloaded by the end of time of ς, which is the time to stream the whole segment

jP

x jxj sRRs1

)/(

12

Phase IIIThe peer cache the segments

Runs the dispersion algorithm

Caching the right segments of the media files in right places

Discuss in later section

13

P2P Architecture

Two approaches for the model Index approach : a set of servers maintained

information about the currently participating peers

Overlay approach : does not assign special roles to any peer.

Index approach is presented.

14

Pros and Cons of Index Approach

Pros Easier and faster to deploy The load (eg. CPU, bandwidth, storage) is load More appropriate to commercial media provider,

for accounting and charging customersCons Alleviates the scalability Single point of failure arises

15

Index searching

To make the index server up-to-date If a peer gracefully shuts down

-> a daemon sending notification message to index server Request client checks the list of candidate peers by

pinging them. Client returns to index server the status in one message

16

Index Search Algorithm

a. Get all peers within the same clusterb. Ask those peers to send each segment

d. if those peers still can not provide the shortage, back offc. If peers within the cluster can not provide enough segments, ask the peers outside the cluster

17

Index Dispersion

Objective : Store enough copies of the media files in each cluster to serve all expected client requests from that cluster.

Client Py sends the request and also declare the willingness to cache up to Ny segments serving others with rate Ry in future

The algorithm should make sure that on average, same number of copies of each segment is cached.

18

Index dispersion algorithm

Consider 1 media file with N segments, rate R kb/s, and duration T hours. The index server maintains the followings

For each peer Px

Nx : # of segment cached by Px

Rx : the streaming rate for Nx

Ux : (0≤Ux≤1) fraction of time Px is online

For each cluster c

Lc : next segment to cache

qc : average request rate (per hour) within the cluster.

ac : average # of copies of the movie cached by peers in cluster

xP xxx

c uN

N

R

Ra

19

Index Dispersion Algorithm

Global

A, the average # of copies of the movie cached by all peers in the system

Q, average movie request rate in system

c caA

c cqQ

Q : global required capacity

(1/T)A : global available capacity

20

Index Dispersion Algorithm

a. Identify which cluster the peer residesb. Compute ac, qc, A, Qc. If the demand is larger than available capacity within the cluster

OR

If the global available capacity is too low compared with the global request rate

d. Allow the cache Ny segments in a cluster-wide round-robin fashion.

Py allows to cache Ny segments with rate R

21

Simulation

Parameters 13,000 nodes, hierarchical, Internet-like topology A seeding peer with a limited capacity introduces a

media file into the system. Other peers joins the system and request the file UDP and CBR If the peer does not find all segments with full rate, it

back off again after an exponentially increased waiting time. For a particular threshold, it is rejected.

22

Simulation results

Average Service Rate

23

Simulation results

Average Waiting time

24

Simulation results

Load on the seeding peer

25

Evaluation

Similar to our project The Index-based search, acting like directory

server, is trivial Not deployed the index-based dispersion

algorithm in simulations Low caching ratio seems useless in their system Multiple video files may need to delete the other

file cache.

26

Conclusion

Based on the P2P streaming protocol, two important of these steps are presented: Index search algorithm

locating peers with the required media files Index dispersion algorithm

disseminating media files into the system