fec and pseudo-arq for receiver-driven layered multicast

FEC and Pseudo-ARQfor Receiver-drivenLayered Multicast

Philip A. Chou, Alex Mohr,Albert Wang, Sanjeev Mehrotra

(Microsoft, University of Washington, Stanford)

Problem

• Find ways to multicast real-time audio and video efficiently and robustly over the near-term Internet– Best-effort

• packet loss, jitter, out-of-sequence

– Heterogeneous• transmission rate, loss rate, delay, jitter, receiver capacity

– Dynamic

• Influence design of network services in longer term

Outline

• Multicast

• Receiver-driven Layered Multicast

• FEC and Pseudo-ARQ

Multicast

• Mechanism for broadcasting in packet networks: any receiver able to receive broadcast packetsby “tuning in” to broadcast address

• Our focus: IP networks

• Protocols:– How to dynamically maintain a multicast routing tree

• IGMP, DVMRP/PIM/CBT/…

– How to forward packets along the tree• IP Multicast

Packet Forwardingin a Multicast Routing Tree

S

RR

RR R

1 2

S,G | data :S,G - 1,2 :

multicast routing table

packet S = Source AddressG = Multicast Group Address

Grafting / Joining / Subscribing

S

RR

RR R RR

1 2 Join S,G

:S,G - 1,2 :

multicast routing table

Join S,G

:S,G - 1 :

multicast routing table

1

Pruning / Leaving / Unsubscribing

S

RR

RR RR

multicast routing table

R

Leave S,G

:S,G - 1 :

1

Leave S,G

:S,G- 1, 2 :

multicast routing table

1

Observations

• Bandwidth is conserved.

• Sender has no knowledge of receivers(likewise any interior router).

• Scalable: As new receivers join, no new resources (bandwidth, computation) are needed at the sender, and resources needed at any individual router are bounded.

• Joining and Leaving are a form of feedback.

Receiver-driven Layered Multicast (S. McCanne ‘96)

Base layer ... to multicast group G0

Enh. layer 1 ... to multicast group G1

Enh. layer 2 ... to multicast group G2

time

• Code source in layers (base, enh1, enh2, …)

• Send each layer to different multicast group

• Receivers subscribe to as many layers as desired

S0t S0t+1 S0t+2 S0t+3 S0t+4 S0t+5

S1t S1t+1 S1t+2 S1t+3 S1t+4 S1t+5

S2t S2t+1 S2t+2 S2t+3 S2t+4 S2t+5

Dynamic Joining/Leaving

S

RR

RRR

• Receivers subscribe and unsubscribe according to instantaneous capacity

transmission rate

loss

pro

b

1

0

The Case for FEC

S

RR

RRR

• Problem is that ambient congestion may always be present

transmission rate

loss

pro

b

1

0

FEC forReceiver-driven Layered Multicast• Amount of redundancy must be receiver-

driven to deal with heterogeneity, dynamics• Generate FEC in layers, using a systematic

rate-compatible code (for each source layer)• Send each layer to different multicast group• At each receiver, subscribe to number of

source layers, number of parity layers for each source layer, to optimize received quality

Generation of Parity Packets• Block each source layer into K packets per block

• Apply systematic Reed-Solomon type code to packets (bytewise) to produce N-K parity packets

Source layer ... to multicast group Gx

Parity layer 1 ... to multicast group Gx.1

Parity layer 2 ... to multicast group Gx.2

Parity layer 3 ... to multicast group Gx.3

StK StK+1 StK+2 S(t+1)K

Pt+1,1Pt,1

Pt+1,2Pt,2

Pt+1,3Pt,3

S(t+1)K+1 S(t+1)K+2

K

N-K

Optimization at Each Receiver

• Measure transmission rate and loss probability

• For this transmission rate and loss probability, minimize expected distortion (over number of source layers, and number of parity layers for each source layer)

• Subscribe to optimal number of source layers, and optimal number of parity layers for each source layer

Expected Distortion and Rate

)(

)()()1(

)1()1()1()1(

8

8

818

211087818

108717818

ii i

ii i

i ii

RpSD

pSD

pDDDppOpDpDppD

pDpppDppDD

i iRR

layer

D0

D1

…

D8

D2

1 2 30 8

D3

…

1

00

1 32 Ri=Ni /K

pi=p(Ri )

raw loss prob

Lagrangian Optimization

8

10

,,)(minmin

81 iiii

RRRRRpSDRD

i

pii p

DS

0 1 32 R=N /K

S1

S2

S3…

Unequal Error Protection

1 2 3 4 5 6 7 80

0.5

1

1.5

2

2.5

3

3.5

4

Layer

Red

unda

ncy

Optimal allocation for 20% packet loss, K=8, W=1

The Case for ARQ

• FEC does not achieve capacity C = 1p• ARQ makes optimal use of forward channel

• Adapts to loss (erasure) probability p

• Used extensively for data transmission (e.g., TCP) and media transmission (e.g., VOD)

• Must avoid “repeat-request implosion”

• Observe: for broadcast of real-time media, delay (number of repeat-requests) is bounded

Pseudo-ARQ forReceiver-driven Layered Multicast• Sender transmits delayed versions of source

• Receivers “request repeats” by subscribing and unsubscribing to delayed versions

Source stream ... to multicast group G

Delayed stream 1 ... to multicast group G1

Delayed stream 2 ... to multicast group G2

time

St St+1 St+2 St+3 St+4 St+5

StSt-1

St-1St-2 St

St+1

St+1

St+2 St+3

St+3

St+4

St+2

StK StK+1 StK+2 S(t+1)K

Pt+1,1Pt,1

Pt,2Pt-1,2

Pt,3Pt-1,3

S(t+1)K+1 S(t+1)K+2

Hybrid FEC & Pseudo-ARQ• Sender delays parity packets• “Repeated” info is actually parity info

Source layer ... to multicast group Gx

Parity layer 1 ... to multicast group Gx.1

Parity layer 2 (delayed 1 block)

... to multicast group Gx.2

Parity layer 3 (delayed 1 block)

... to multicast group Gx.3

K=3W=2 waves (decision epochs)

• Delay is blocklength times number of waves

Markov Decision Process

• Finite-state stochastic process in whicha decision can be made at each step to influence the transition probabilities,in order to maximize an expected reward

• Sequence of decision rules is a policy

• Optimal policy, which minimizes expected cost (D+R) of paths through state space, can be found by dynamic programming

State Space for FEC/Pseudo-ARQ

# pktsto req

# pktsrecvd

# pktsrecvd

03

4

# pktsto req

# pktsto req

# pktsto req

0

1

2

3

4

# pktsrecvd

# pktsrecvd

# pktsrecvd

# pktsrecvd

# pktsrecvd

# pktsrecvd

012

0

1

2

210

0

1

2

3

4

0

101

2

01

01

2

010 1

2

Caveats

• No experiments so far, only analysis– no actual network transmissions,

no source coding, no channel coding– source is modeled as D=222R

– channel is modeled as iid 20% packet loss,no packet jitter or resequencing

– assumed no join or leave latencies

• Little analysis of aggregate behavior

Analyzed Systems

1 - No error protection (original RLM)

2 - Equal error protection, redundancy determined by sender; K=8; N=11,14,17,20

3 - Unequal error protection, determined by receiver; K=8; Ni=Ni* (RLM w/FEC)

4 - FEC and Pseudo-ARQ; W=1,2,4,8, KW=8

Results of Analysis

0 2 4 6 8 10 12 14 16 18 200

5

10

15

20

25

30

35

40

45

50

Packet Transmission Rate (layers x avg redundancy)

Sig

nal t

o R

econ

stru

ctio

n N

oise

rat

io (

dB)

Recovery of an idealized source with 20% packet loss rate

(N,K)=(8,8)

(N,K)=(11,8)

(N,K)=(14,8)

(N,K)=(17,8)

(N,K)=(20,8)

EEPW=1(UEP)

W=2W=4

W=8

1

2

34

4

3

2

1

FEC/P-ARQ

FEC (receiver)

FEC (sender)

original RLM

Issues for Further Investigation

• Estimation of channel condition

• Accommodation for join/leave latency

• Clustering of channels

• Study of aggregate (social) behavior

• Implementation

Summary

• FEC & Pseudo-ARQ allow error controlwhen flow control alone is insufficient

• Layered (rate-compatible) FECallows receiver-driven approach

• Pseudo-ARQ is more effective than FECand uses only standard multicast machinery

• Hybrid FEC/Pseudo-ARQ enhances scalability• Many issues remain for further investigation