Multiple Sender Distributed Video

Streaming

Thinh Nguyen and Avideh Zakhor

Presented by

Huang Wendong

IEEE TRANSACTIONS ON MULTIMEDIA 2004

Introduction Multiple sender distributed video streaming

Multiple senders and single receiver Independent routes from senders to the receiver Achieve higher throughput, increase tolerance to packet

loss and delay due to network congestion

Three main components in the proposed work

Rate allocation algorithm Packet partition algorithm Receiver-driven transport control

System Overview

Sender 1

Sender 2

Receiver

Data

Data

Control ACK

Control ACK

Packet Partition algorithm

Packet Partition algorithm

EstimateRTT

EstimateRTT

Rate Allocation algorithm

Estimate available

bandwidth

Estimate loss rate

Receiver-driven Transport Control

Receiver Calculate the optimal sending rate using RAA based

on the RTTs and estimated loss rates Monitor variations of each sender and readjust rate

distributions among senders accordingly Sender

Estimate and send its round-trip time to the receiver Run distributed PPA upon receiving the control packets

Control packet format Synchronization sequence number + RTT+ sending

rates

Rate Allocation Algorithm Objective

Minimize the irrecoverable loss under fixed FEC Be fair to existing TCP traffic

Setup of RAA Packet loss model: Two-state discrete Markov chain FEC scheme: RS(n, k), n and k are fixed for the entire

streaming process Bitstream model: single description and single layer

bitstream

Rate Allocation Algorithm(2) Mathematical model

Subject to

P(m,i,N) denotes the probability that i packets are lost out of N packets sent by m

is the interval between the successive transmitted FEC blocks in seconds

BA NN

KNj

j

iBABA NijBPNiAPNNKC

1 0

),,(),,(),,(

BB

AA

BA BN

BN

NNN

,,

Rate Allocation Algorithm(3)

},{ },{

),(),,(bgi bgj

mmij

mim NkNkmP

},{

))](1()1,()()1,1([),(bgl

lossm

mlj

mil

lossm

mlj

mil

mij jPpnkjPpnknk

)( 1MNO

Implementation Search over all possible values of and For M senders, the exhaustive search has

complexity of

AN BN

Packet Partition Algorithm Objectives

Run in distributed fashion All packets are sent by one and only one sender To minimize the startup delay

Mathematical model

Where the playback time of k-th packet with respect to

the estimated arrival time of the k-th packet sent by j

To maximize the time difference between arrival and playback time of the k-th packet

)(2)(,, jDjn kkj

( )kP k ),( kjAk , ,[ ( ) 2 ( )]j k kn j D j

( )kP k kT

Packet Partition Algorithm(2)

Distributed calculation Each sender effectively keeps track of all the values

of for all senders No need to synchronize all the sender’s clocks to a

global time

),( kjAk

10 14 13 12 11 15 18 17

18 16 14 12 10

17 15 13 11

Packet playback time

Sender 1

Sender 2

16

Sync

Packet Partition Algorithm(3)

The choice of the synchronization sequence number in the control packet

The lag among the senders needs to be as little as possible

The aggregate bit rate for all senders remains constant

where is the estimated sequence number for the latest packet sender j has just sent, before receiving the control packet.

min ( )jk k j

( )k j

Experiments Results Numerical Characteristics

Two common scenarios X and Y are considered FEC configuration: RS(30,27), RS(30,25), RS(30,23) irrecoverable loss probability change significantly

among various schemes Optimal packet partition scheme As the average bad time of route B increases, more

packets are sent by route A for all three levels of FEC protection;

At the same average bad time of route B, the number of packets sent on route A decreases with increased FEC level

Experiments Results(2) Single sender vs. multiple senders Irrecoverable packet loss is significantly reduced in

multiple sender scheme; Optimal RAA is more effective with stronger FEC

schemes Sensitivity Analysis of Optimal Sending

Rate Sensitivity of loss probability to deviations form

optimal sending rates In case of strong FEC protection,a slight variation in

the sending rate around the optimal value results in a smaller change in irrecoverable probability

Experiments Results(3) Internet Experimental Results Show

Multiple sender scheme can effectively combat burst packet loss In most cases, PPA provides interleaved packets with small lags

Conclusions The proposed work can effectively reduce

overall packet loss rate

Further work Adopt adaptive FEC schemes Employ scalable bitstream to accommodate the

additional data