Multimedia Communications Spring 2006-07 Delay of Voice Traffic - - PowerPoint PPT Presentation

Delay of Voice Traffic Over IP Netw orks

Shahab Baqai LUMS

CS 584 / CMPE 584

Multimedia Communications

Spring 2006-07

2

λ μ Talk Spurt Silence

Characteristics and Requirements of Voice Traffic Voice characteristics

– Low rates (8Kb/s for G.729A, 64Kb/s for G.711) – Low variability resulting from Silence Suppression

Voice requirements: real-time communication

– Low delay: 200-300ms round-trip, 100-150ms one-way (Dmax) – Low jitter: for smooth playback – Low packet loss: at most 2%

3

10 10

1

10

2

10

• 6

10

• 5

10

• 4

10

• 3

10

• 2

10

• 1

Packet Formation Time Tf (ms) Maximum Tolerable Packet Loss Independent packet loss Correlated packet loss

Tolerable Packet Loss Versus Packet Formation Time

4

Case for Separation of Voice From Other Traffic Measurements on the Internet as well as simulations show that mixing voice and TCP data traffic leads to either

– Large delays for voice (100-500ms average delay) – Very low utilization of network resources (often less than 20%)

Voice must also be separated from bursty UDP traffic (e.g., VBR video)

– Voice delay affected if aggregate peak rate exceeds available bandwidth – Results in low bandwidth utilization

5

Case for Separation of Voice From Other Traffic

VBR video stream (730 Kb/s

average, 2.3Mb/s peak)

450 Kb/s aggregate voice

load

VBR Video Voice mixed with video Voice separated from video 1 0.1 1 10 100 10-5 10-4 10-3 10-2 10-1 Delay CCDF 7 Delay (ms)

1-4 video streams

10 12 12

10Mb/s, 7hops

6

Analysis of Voice Delay and Jitter Voice better served if separated from other traffic in the network

– Voice given its own circuit – Voice given priority over other traffic – Voice given its share of the bandwidth using WRR

In this context, study effect of

– Residual transmission time of lower priority traffic – Available bandwidth – Scheduling scheme – Packet formation time

7

max

) ( D D Q P T l T D

play path i i i i f

≤ + + + + + =

∑

∈

Encoder Packetization De-packetization

Sender Receiver

Playback Buffer Network Decoder

• Delay budget

End-to-end Delay Components

8

Encoding and Packetization

Analog Signal Encoder In Encoder Out Packetization Out

Frame f Lookahead l Processing pe

Encoding Delay Dencoder Encoding and Packetization Delay Dencoder+Dpack

9

End-to-end Delay Components

) negligible assumed be (can , ) ( ) ( ) 1 ( f p p p D P T Q p l kf p D P T Q f k p l f D

d e i d play i i i e i d play i i i e

≤ + + + + + + + = + + + + + − + + + =

∑ ∑

Tf = Packet formation time

play i i i i

D P T Q l kf D + + + + + =

∑

) (

Dnet = Network delay

10

End-to-end Delay Components

Constant,

function of:

• Encoding Scheme

(frame size, look- ahead)

• Packet size

Constant,

function of:

• Path in the network

(links speed, links propagation delay)

• Packet size

Source of Jitter, random, function of:

• Path in the network
• Traffic load and

characteristics

• Scheduling scheme
• Packet size

play path i i path i i i f

D Q P T l T D + + + + + =

∑ ∑

∈ ∈

) (

11

Playback Buffer Jitter unknown and random Thus, playback buffer needed

– Packets delayed by playback buffer delay Dplay – Insures continuous playback

Dplay depends whether sender and receiver clocks are

– Synchronized (e.g., GPS) – Not synchronized

12

Synchronized Clocks

Network

tsender Sender

Encoding and packetization De-packetization and decoding

Receiver time treceiver

Dplay treceiver-tsender Dmax max

) ( have must We D D t t

play sender receiver

≤ + −

13

Synchronized Clocks, Unknown Jitter

To accommodate highest amount of jitter (and minimize loss),

∑ ∑ ∑

− + + + − = − − =

i i i i i i f send rec play

Q P T l T D t t D D ) ( ) ( choose

max max

) ( then jitter) (no If

max max

∑ ∑ ∑

+ + + − = = =

i i i i f play play i i

P T l T D D D Q

∑ ∑

− ≤

i i play play i i

Q D D Q for buffered and received packet , If

max max

rejected packet , If

max play i i

D Q >

∑

max play

D

14

Synchronized Clock, Known Jitter If maximum bound on jitter known (by some measure), i.e.

max

) max( Q Q

i i =

∑

max max

then Q Dplay =

max play

D

max

) ( and Q P T l T D

i i i f

+ + + + =

∑

15

Non-Synchronized Clocks RTP timestamps packets However, no guarantee that sender and receiver are synchronized Hence, receiver does not know at what time the packet was sent

16

When first packet received at the destination, it must be delayed by Qmax in order to take into account the jitter ΣiQi

) ( max

max

∑

∈

= =

path i i path play

Q Q D

Receiver Sender Display Dplay= Qmax Qmax

... ... ... Play-out Buffer Delay

17

End-To-End Delay Requirements Toll quality real-time communication needed

– Round-trip delay must be in the range 200-300 ms – That is, D ≤ Dmax, where 100 ms ≤ Dmax ≤ 150 ms

Amount of jitter allowed 10-50 ms, function of

– Acceptable end-to-end delay Dmax – Formation time Tf – Propagation delay

• Hence, the one-way end-to-end delay becomes

⎩ ⎨ ⎧ − + + + + =

∑

∈

ed synchroniz and , ed synchroniz non and , 2 ) (

max max

D S Q D S Q P T l T D

path i i i f

18

Network Scenario

Assess queuing delay incurred by a voice stream traveling through a given number of hops Consider hops to be independent Delay observed depends on characteristics of interfering traffic

Target Source Target Receiver

19

Sources of Jitter

Queuing delay behind voice packets in the same queue

– Depends on the burstiness of the voice traffic pattern (Variability of traffic at the source introduced by Silence Suppression)

Residual transmission time of lower priority packets

– Increases the burstiness of the outgoing voice streams

20

Models for Voice Traffic Voice traffic alone on the network

– Voice traffic does not incur transmission time from lower priority packets – Traffic variability minimal – One hop: queuing delay ΣDi/D/1

Even though Silence Suppression reduces voice rate, effect on delay negligible

– H hops: convolution of one hop delays

21

Models for Voice Traffic (cont.)

Voice traffic incurs residual transmission time of lower priority packets

– Large traffic variability possible

Decreases with the number of hops Tail of inter-arrival time between packets bounded by exponential distribution

– One hop: delay percentile is sum of Queuing delay percentile M/D/1 Delay percentile from (uniformly distributed) residual transmission time – H hops: delay percentile is sum of Delay percentile from convolution of M/D/1 queuing delays Delay percentile from convolution of transmission times

22

Models Versus Simulation Results

Modeling Simulation Q CCDF, 4 streams, 5 hops, 384Kb/s 1 10 100 10-4 10-3 10-2 10-1 1 Queuing Delay Q (ms)

Voice Alone Residual transmission time

• f lower priority

packets incurred

23

Effect of Residual Transmission Time of Lower Priority Packets CCDF of Queuing Delay

T1, G.729A, Tf=30ms, 50% utilization

0.01 0.1 1 10 100 10-6 10-4 10-2 1 Queuing Delay (ms)

Voice Alone Residual transmission time of lower priority packets incurred

1 hop 2 hops 5 hops

24

0.1 1 10 10-6 10-4 10-2 1 Queuing Delay (ms)

CCDF of Queuing Delay

G.729A, Tf=30ms, 50% utilization 1 hop 10 hops

T3 T1

1 hop 5 hops

Effect of Bandwidth

25

Choice of Scheduling Scheme

CCDF of Queuing Delay

Weighted Round Robin (WRR) Versus Priority Queuing (PQ) 10-4 0.1 1 10 Queuing Delay (ms) 100 10-6 10-2 1 PQ WRR, 1.5Mb/s WRR, 10Mb/s

T3, G.729A, Tf=30ms, 5 hops, 1.35Mb/s Voice Load

26

Choice of Packet Formation Time

• r = rate of encoded bit
• stream
• R = rate of packetized bitstream

Incentive to use largest formation time possible (given Dmax, propagation and queuing delays)

8

f

T H r R + =

Total Header H=46-69Bytes

Data Link/MAC Encoded Voice RTP 12 Bytes UDP 8 Bytes IP 20 Bytes

R/r 0.5 1 1.5 2 2.5 3 3.5 4 20 40 60 80 100 G.729A G.711 Formation Time Tf

27

Effective Header Size, IEEE 802.11 Wireless Network

• Physical layer header must be transmitted at lowest speed

– Hence, effective header size increases with available bandwidth

Physical Layer No RUI H. Comp. RUI H. Comp. DSSS, 1Mb/s 97 Bytes 67 Bytes DSSS, 2Mb/s 120 Bytes 90 Bytes 802.11b, 5.5Mb/s 200.5 Bytes 170.5 Bytes 802.11b, 11Mb/s 327 Bytes 297 Bytes 802.11a, 6Mb/s 89 Bytes 59 Bytes 802.11a, 24Mb/s 134 Bytes 104 Bytes 802.11a, 54Mb/s 209 Bytes 174 Bytes

28

20 40 60 80 100 5 10 15 20

Formation Time Tf R/r

802.11b 802.11 T1/T3 G.723.1 G.729A G.711

Packetization Overhead

29

Choice of Packet Formation Time (2) In WANs, bandwidth largely available and propagation delays large

– Benefit limited

For transatlantic links, incentive for bandwidth saving large, but propagation delays large In local areas, call locality allows more significant gains if formation times larger than 30ms allowed (45% increase with T3 links)

– Significant only in the case of wireless LANs (where bandwidth is scarce)

Conclusion: formation time of 30ms (30Bytes) adequate for G.729A

– If Tf ≥ 30ms, potential in bandwidth saving limited – If Tf < 30ms, potential for delay reduction limited

30

Dynamic Packetization in Wireless Networks Choose packet size according to data rate and propagation delay

– Efficient use of the bandwidth – Allow up to ~70% increase in number of streams supported

Possible remedies against data rate variations

– Drop connections, Increase packet loss

Instead, increase packet size

– If needed, increase delay tolerance (e.g. from 100 to 150ms)

31

Analysis of Voice Delay and Jitter Summary

Giving high priority to voice traffic leads to adequate performance If enough bandwidth available for voice traffic

– Jitter negligible, can be ignored

If bandwidth limited

– Lack of preemption of lower priority packets leads to large jitter

Appropriate packet size

– 30 Bytes for G.729A, 10 Bytes for G.711 good compromise – Further optimization advantageous in local areas

In wireless networks, dynamic packetization useful

– Efficient use of bandwidth resources – Robustness against data rate variations