Adaptive Delay Aware Error Control for Internet telephony Catherine Boutremans Jean-Yves Le Boudec IP Telephony Workshop’2001 Institute for computer Communication.

Adaptive Delay Aware Error Control for Internet

telephonyCatherine BoutremansJean-Yves Le Boudec

IP Telephony Workshop’2001

Institute for computer Communication and Applications

phone: + 41 21 693.5258; fax: +41 21 693.6610

[email protected]

http://icawww.epfl.ch

Framework• Real-time audio over Best Effort networks suffers

from varying packet loss rates, delays and available bandwidth.

• Forward Error Correction (FEC) is an efficient way to recover from packet losses but :– bandwidth requirement

– end-to-end delay

• Adaptive rate/error control was proposed but it does not try to optimize the end-to-end delay

Motivation

• Above a certain threshold (around 150ms): the end-to-end delay is annoying

• New differentiated services, such as Alternative Best Effort offer applications the trade-off between receiving lower end-to-end delay (and higher loss rates) or more overall throughput.

important to take the end-to-end delay into account in the adaptive control scheme

Adaptive Delay aware error control

• Problem definition:

develop an error control scheme for audio which is delay aware (namely, which chooses the FEC according to its impact on the end-to-end delay) for Best Effort Internet services such as:

1) Flat 2) ABE

• Out of scope:validation of perceptual models

Outline• A.Our Joint rate/error/delay control scheme

– 1. Error recovery taken from media specific FEC – 2. An RTCP-based Rate control scheme– 3. A new Delay control scheme – 4. Our Global optimization problem

• B.Simulation examples– 1. Flat Network: what do we improve?– 2. ABE: Is it worth to be green?

Dr

Rr

Dr

Rr

1.How media specific FEC Works

Qmax

Rmax

Qmax

Rmax

Audio Frame 3Audio Frame 4

primary/source encoding

Dp

Rp

Dp

Rprate

Qua

lity

... ... packet stream345

• Start with high rate audio encoding

234

• copy audio, compress, and delay

secondary/redundancy encoding35 4 reconstruction

• if packet lost, recover from redundancy

1.Efficiency of FEC mainly depends on packet loss

process• Loss process of audio packets in the Internet

can be modeled as low order Markov chain

• In this work: Gilbert Model

10

p

1-q1-p

q

qp

pPLR

qp

qp n

1)1(

)losses econsecutivProb(n

2.Our RTCP-based Rate Control scheme

• Audio streams have to share bandwidth fairly with TCP connections TCP-Friendly

• Equation-Based TCP Friendly rate control (Padhye):

• Rate controlled via packet size (constant packet intervals)

• Use RTCP for Feedback• EWMA filter of PLR in order to smooth the rate

)321()8

33(

32

_

2plrplrplr

tplr

RTT

sizepcktR

RTO

TCP

2. Our Rate control achieves fairness

0 50 100 1500

0.2

0.4

0.6

0.8

1

1.2

1.433.33% TFRC vs 66.66% Sack TCP, 5Mb/s, RED Queuing

Number of connections (TCP + TFRC)

No

rma

lize

d th

rou

gh

pu

t

TCP FriendTCP

Our scheme shares bandwidth fairly withTCP connections

3.Quality is function of rate and delay

• Audio quality is function of reconstructed rate (SNR,MOS,...)

• Audio quality is function of end-to-end delay: decreases rapidly above 150ms

0 50 100 150 200 250 300 350 4000.75

0.8

0.85

0.9

0.95

1

delay (ms)

utili

ty

0 1 2 3 4 5 6 7

x 104

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

encoding rate (bps)

util

ity

0 50 100 150 200 250 300 350 4000.65

0.7

0.75

0.8

0.85

0.9

0.95

1

delay (ms)

util

ity

0 1 2 3 4 5 6 7

x 104

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

encoding rate (bps)

util

ity

0 100 200 300 4000.3

0.4

0.5

0.6

0.7

0.8

0.9

1

delay (ms)

util

ity

3.Quality is function of rate and delay

• Consider the user perceived quality (utility) as a function of the reconstructed rate (R) and the overall end-to-end delay (D)

• We use several utility curves as we don’t know which one is the best.

4.Our Joint rate/error/delay control

• A source with flexibility to encode audio at rate• Unreliable network characterized by

– a loss distribution Gilbert Model : r.v.– a delay distribution – an available bandwidth TCP-Friendly rate constraint

• A utility function of delay and reconstructed rate

kbpsx 64,0

),( DRf

Consider:

1,0iX

4.Our Joint rate/error/delay control

Constrained optimization problem:

Define:• K = of copies of audio segment sent over the network• ,the delay spent on FEC• the r.v. the set of copies that make it across the

network KiXi i ,...,1,0

valpckt_interKK )1(

),(max

Ωx,K i

Knetworkii1,..,K1,...,K)(inbmax1

DxfΡ(Ω)maximize

Under TCPoverhead

K

ii RRx

1

Ki0xi ,...,1, MAXtionreconstrucafter PLRPLR

• General solution is derived using Lagrange Multipliers for small values of K and via numerical method SQP for K4

• The solution has the following properties:– – if (p+q)1 it pays to offset and to put more quality into the

end packets– if (p+q)1 better not to offset and

4.Our Joint rate/error/delay control

Solution:

2,...,Kixx i ,1

Kxxx ...21

B.Simulation examples• Single bottleneck

• Topology with small and long flows

• Bottleneck BW = 15Mbps, 5Mbps

• variable number of connections

d1 d1 d2d2

Example 1: when delay is important

0 20 40 60 80 100 1200.3

0.4

0.5

0.6

0.7

0.8

0.9

1Utility 3 ,5Mbits/s,d

prop=100ms

Number of flows

Util

ity

K = 1 K = 2 K = 3 K = 4 K = 5 Delay-Aware

0 20 40 60 80 100 1200.1

0.12

0.14

0.16

0.18

0.2

0.2299% percentile e2e delay ,5Mbits/s,d

prop=100ms

Number of flows

De

lay

(s)

K = 1 K = 2 K = 3 K = 4 K = 5 Delay-Aware

0 20 40 60 80 100 1200

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

Packet loss rate,5Mbits/s,dprop

=100ms

Number of flows0 20 40 60 80 100 120

0

50

100

150

200

250

300

350

400

450TCP-Friendly Rate ,5Mbits/s,d

prop=100ms

Ra

te (

Kb

its/s

)

Number of flows

Example 2: when delay is not important

0 20 40 60 80 100 1200.55

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

Utility 1 ,5Mbits/s,dprop

=100ms

Number of flows

Util

ity

K = 1 K = 2 K = 3 K = 4 K = 5 Delay-Aware

0 20 40 60 80 100 1200.1

0.12

0.14

0.16

0.18

0.2

0.2299% percentile e2e delay ,5Mbits/s,d

prop=100ms

Number of flows

De

lay

(s)

K = 1 K = 2 K = 3 K = 4 K = 5 Delay-Aware

0 20 40 60 80 100 1200

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

Packet loss rate,5Mbits/s,dprop

=100ms

Number of flows0 20 40 60 80 100 120

0

50

100

150

200

250

300

350

400

450TCP-Friendly Rate ,5Mbits/s,d

prop=100ms

Ra

te (

Kb

its/s

)

Number of flows

0 20 40 60 80 100 1200.65

0.7

0.75

0.8

0.85

0.9

0.95

1Utility 2 ,5Mbits/s,d

prop=100ms

Number of flows

Util

ity

K = 1 K = 2 K = 3 K = 4 K = 5 Delay-Aware

0 20 40 60 80 100 1200.1

0.12

0.14

0.16

0.18

0.2

0.2299% percentile e2e delay ,5Mbits/s,d

prop=100ms

Number of flows

De

lay

(s)

K = 1 K = 2 K = 3 K = 4 K = 5 Delay-Aware

Example 3: tradeoff between delay and audio distortion

0 20 40 60 80 100 1200

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

Packet loss rate,5Mbits/s,dprop

=100ms

Number of flows0 20 40 60 80 100 120

0

50

100

150

200

250

300

350

400

450TCP-Friendly Rate ,5Mbits/s,d

prop=100ms

Ra

te (

Kb

its/s

)

Number of flows

1. FLAT: Conclusion

• The Delay aware scheme increases utility by avoiding that the source waste delay on the FEC when it is not really needed.

2.ABE: Alternative Best Effort

• ABE is a novel service for IP networks which offers the choice between receiving a lower end-to-end delay or more overall throughput.

• Packets are marked either green or blue.• Green packets receive a low, bounded queuing

delay but they receive more losses during bouts of congestion.

• Blue packets receive more throughput (and less losses) but also more delay jitter.

Is it worth being green?

• Green packets receive a lower delay but they experience more losses (and hence, receive less throughput).

• Losses are repaired using FEC but FEC increases the e2e delay and the BW requirement.

• With small RTT, difference is minor.

0 50 1000.05

0.1

0.15

0.2

0.25

Packet loss rate

0 50 100

0.9

0.95

1Utility

Load in network

Blue Green

0 50 1000.04

0.06

0.08

0.1

0.12

0.1499% percentile e2e delay

Load in network

De

lay (

s)

0 50 1000.5

1

1.5

2

2.5x 10

5 TCP-Friendly Rate

Ra

te (

b/s

)

• Higher RTT (2 bottlenecks): green is better if load not too high.

10 20 30 400.1

0.15

0.2

0.25

0.3Packet loss rate

10 20 30 400.5

0.6

0.7

0.8

0.9

1Utility

Load in network

Blue Green

10 20 30 400.05

0.1

0.15

0.2

0.2599% percentile e2e delay

Load in network

De

lay

(s

)

10 20 30 402

4

6

8

10

12x 10

4 TCP-Friendly Rate

Ra

te (

b/s

)

• Higher RTT (1 bottleneck): green always better

0 50 1000

0.05

0.1

0.15

0.2Packet loss rate

0 50 1000.75

0.8

0.85

0.9

0.95

1Utility

Load in network

Blue Green

0 50 1000.1

0.12

0.14

0.16

0.18

0.299% percentile e2e delay

Load in network

De

lay

(s

)

0 50 1000.5

1

1.5

2

2.5x 10

5 TCP-Friendly Rate

Ra

te (

b/s

)

2. ABE: Conclusions

• It is worth accepting to receive less bandwidth (and more losses) except in trivial cases where– the RTT is small anyway– the network is badly congested

• Need for adaptive Color choosing algorithms

Conclusions

• We proposed an adaptive Delay aware error control scheme.

• We showed it could prevent a source from wasting delay on FEC when not necessary.

• It helped us to figure out that it was worth trading throughput for delay.