Multiple Sender Distributed Video Streaming

Thinh Nguyen, Avideh Zakhor

appears on “IEEE Transactions On Multimedia, vol. 6, no. 2, April, 2004”

Outline

Introduction System Overview Rate Allocation Algorithm (RAA) Packet Partition Algorithm (PPA) Performance Evaluations Conclusions

Introduction

Existing CDNs, like Akamai Receivers only receive streaming content from one se

nder only throughput of a single sender may not meet the strea

ming requirement

Multiple streaming senders Effectively provide the require throughput by aggregat

ing content from multiple sources Combats unpredictable congestion in Internet

System Overview

Each sender estimates and sends its RTT to receiver Receiver uses the estimated RTTs and its own estimates of senders

’ loss rates to calculate the optimal sending rate for each sender (RAA) and send control packets to senders

Upon receiving the control packet, senders run a distributed PPA to determine which packets to send

System Overview

Assumptions Bandwidth bottleneck is not at the last hop Amount of FEC in a streaming session is fixed CBR videos replicated in all senders Average packet loss rate over long term is

independent of the instantaneous sending rate, i.e. packet loss only caused by cross traffic

Packet loss between two routes are independent

Rate Allocation Algorithm (RAA)

Goal To determine how to split the total rate S between M s

enders in order to minimize the probability of packet loss

Intuitions Splitting routes reduce bursty loss FEC work better in non-busty environments Simultaneous packet lost on all routes is rare and, thu

s, allow FEC to recover packets on other routes

Rate Allocation Algorithm (RAA)

Model packet loss as two states discrete Markov chain with transition probability matrix

gg gb

bg bb

P P

P P

Good Bad

Pgg

PgbPbb

Pbg

Rate Allocation Algorithm (RAA) Minimize

0 11 0

, , , , , ,A BN N j

A Bj N K i

C K N N P A i N P B j i N

, ,A BA B A B

N NN N N B B

subject to

j out of NA + NB packets are lost

more than N – K packets are lost

Rate Allocation Algorithm (RAA)

The search complexity is The authors believe that

“…from the implementation point of view, having more than then connections results in too large of an overhead, and makes the coordination of the senders too difficult for distributed streaming to be feasible in practice.”

Not a big fan of our architecture

1MN

Packet Partition Algorithm (PPA) Goal

To determine which packets should be sent by which senders in order to prevent duplicate packets and to minimize the startup delay

In traditional P2P systems like Kazza Sender sends continuous block of data of length prop

ortional to its sending rate As sending rate of individual sender is smaller than th

e playback rate, it incurs additional startup delay

Packet Partition Algorithm (PPA) PPA minimizes startup delay by ensuring

interleaved packets arrival from multiple senders

Control packets from receiver indicateSending rate of each senderEstimated delay from each senderSync, the starting sequence no. that all

senders use in PPA

Packet Partition Algorithm (PPA) Estimated time difference between arrival

and playback time of kth packet

' ' ', ,, 2

k k j k kA j k P k n j D j

Packet Partition Algorithm (PPA) All senders compute , for each pac

ket k, for all senders The sender that maximizes is assi

gned to send kth packet, in order to maximize the probability of meeting playback deadline

' ,kA j k

' ,kA j k

Packet Partition Algorithm (PPA) Consideration of Sync, k’

Received packets are in order as much as possible

Aggregate bit rate should remain constant during rate transition

Packet Partition Algorithm (PPA) k’ = minj( k’’(j) )

k’’(j) is the estimated seq. no. for the latest packet sender j sent

k’’(j) = k*(j) + 2D(j)S k*(j) is the seq. no. of the last packet received from se

nder j

Sending rates have to be temporarily increase after receiving the control packets

Performance Evaluations

Irrecoverable loss probability for various of FEC levels as a function of average bad times of route B, using optimal partition fortwo senders for (a) scenario X and (b) scenario Y.

Bandwidth of each route > 800kbps, packet size = 500 bytes

Performance Evaluations

Optimal partition for various FEC levels as a function of average bad times of route B for (a) scenario X and (b) scenario Y.NA denotes the number of packets per 30 sent on route A.

NA – the optimal number of packets out of 30 that should be sent on route A

More packets should be sent on route A if bad time of B increase The stronger the FEC, the more the packets should be sent on

bad routes

Irrecoverable loss probability ratio of sending all packets using one sender to that of using two senders as a functionof average bad times of route B for scenarios (a) X and (b) Y.

Performance Evaluations

Irrecoverable probably greatly improve if optimal rate allocation is used

Optimal RAA is more effective with stronger FEC

Performance Evaluations

(a) Irrecoverable probability as the function of different partitions of sending rate between two senders.(b) Irrecoverable loss probability ratios between sending all packets on route A over 50–50 and optimal rate splits.

Stronger FEC curves are flatter at the minimum

Performance Evaluations

Internet experimentpacket loss is artificially induced according to t

he Markov chain model Average good time = 1s, bad time = 0.02s Packet loss probability: good = 0, bad = 1

RS(60, 46)Packet size = 500 bytesRequired sending rate = 200 pkts/s (800kbps)

Internet simulations showing the number of lost packet per FEC block of 60 packets versus packet sequence for (a) streaming fromBelgium alone; (b) streaming from Belgium and Sweden; (c) throughputs of two senders; (d) variations in order of the received packets

Bad time change to 0.04s

Adjust sending rate

Performance Evaluations

Actual Internet experimentPlanetLab sites in Sweden and Hong KongPacket size = 500 bytesSending rate = 220 pkts/sUse hidden Markov model inference algorithm

1 on traces of packets to estimate network parameters

1M. Jordan and C. Bishop, “An Introduction to Graphical Models,”

Performance Evaluations

Irrecoverable loss reduces up to 10.2 times Average packet loss rate

Sweden = 1.3%, Hong Kong = 1.8% Sweden’s loss pattern is more bursty Lower average packet loss rate is not a good indicator

for sending packets at higher rate

Conclusion

A receiver-driven protocol for simultaneous video streaming from multiple senders Increase tolerance to packet loss

RAA determines the optimal sending in present of FEC protection

PPA ensures no senders send the same packet and minimize startup delay

Final Thought

When should a receiver decide to change the senders’ rate?

The complexity of RAA is very high, and thus not scalable