Traffic Characteristics

Data trafficVoice trafficvideo traffic

Delay Constraints

Multimedia Traffic

Integrated services

Circuit switching vs. packet switchingTelephone vs. data services

• Circuit switchingDedicated, fixedRealtimePay for distance and time

• Pakcet switchingShared, flexibleNon-realtimePay for amount of data

• QoS guaranteed switching/networkShared and dedicated, flexible and fixedNon-realtime and realtimePay for QoS ⇒⇛=> 인터넷 종량제

Analytic models of data application traffic characteristics

Application variable model parameters mean

telnet responder Log-normal X=4500, σx=7.2 31582

Duration (sec)

Log-normal X=240, σx=7.8 1979

nntp originator Log-normal X=211.5, σx=3 5295

smtp originator Log-normal X=210, σx=2.75 1708

ftp connection Log-normal X=3000, σx=16 140083

Burst size Pareto α=1, k=105.5 ∞

http File size Pareto α=1.1, k=105.5 3478505

Multimedia Traffic

Reference : Vern Paxson, “Empirically-derived analytic models of wide-area TCP connections”

Upper 5% of bursts contribute 90% of the data.

The Log-Normal Distribution

2

2

2

)(

2

1)( y

myy

y

y eyf

)2ln/(xdxdy

Multimedia Traffic

),(~ 2ymyNY

YX 2 XY 2log

mymx 2

YX

21

)(ln2

)/ln(2

ln2

1)(

y

mxx

mX

X x

xxf

X is log2-normal iff i.e.

Geometric mean of X

Geometric standard deviation of X

The Log-Normal Distribution

xm=3000 bytes,

σx=16

Mean=140083 bytes

Multimedia Traffic

Probability density function (pdf) Complementary cumulative density function (cdf)

100 102 104 105

)(xfX xX Pr

Log scale

The Pareto Distribution

pdf heavy-tailed distribution

CDF

Where h is a slow varying function at ∞

Multimedia Traffic

21,0,)/()( kxxkxF C

1)( xkxf X

20),(][)( xhxxXPa

)(lim)( XVarb

x

0,1)(

)(lim

xth

txht

The Pareto Distribution

α=1, k=105.5

Multimedia Traffic

Probability density function (pdf)

Complementary cumulative density function (cdf)

105 106 107

)(xfX xX Pr

Data Traffic Characteristics

Shaped by TCPSlow start and double window size at positive ACK

Loss of a packet restarts the process.

Not suitable to realtime continuous traffic

Self-similarity characteristics

Highly bursty by heavy-tailed distributions of data

Same degree of burstness observed at all time scale

Long range dependency

Slow decay in autocorrelation function

Experienced in both LANs and WANs

No statistical multiplexing gain

Incompatibility to real-time multimedia traffic

Multimedia Traffic

Voice traffic

Multimedia Traffic

Voice Traffic

CBR (Constant Bit Rate)Uncompressed 64 kbps

8000 samples/sec × 8bits/sample

Voice coding standardsMOS (Mean Opinion Score)

ITU-T Rec. Codec Data rate MOS Delay

G.711 PCM 64 kbps 4.4 0.75 ms

G.726 ADPCM 32 kbps 4.2 1 ms

G.728 LD-CELP 16 kbps 4.2 3~5 ms

G.729 CS-ACELP 8 kbps 4.2 10 ms

G.729a CS-ACELP 8 kbps 4.2 10 ms

G.723.1 MP-MLQ 6.3 kbps 3.98 30 ms

G.723.1 ACELP 5.3 kbps 3.5 30 ms

Speech Activity

One speaker talking : 64~73 %Both speakers talking : 3~7 %Both speakers silent : 33~20 %

Multimedia Traffic

SilenceAvg. 1.8 sec

TalkspurtAvg. 1.2 sec

VoIP System and Impairments

Multimedia Traffic

Impairments

Speech quality Delay impairments

Compression Packet Loss Interactivity(<150 ms)

Echo

Delay Jittertalk silence

Packetizer

silence

talkspurt talkspurt

Network

Sender

Depacketizer &Playout buffer

Encoder

Decoder &concealment

Receiver

Voice source

VoIP quality

Multimedia Traffic

Best(very satisfied)

High(satisfied)

Medium(some usersDissatisfied)

Low(many usersDissatisfied)

Poor(nearly alldissatisfied)

Mean Opinion Score(MOS)

Desirable

Acceptable

5.0

4.3

4.0

3.6

3.1

2.6

Speech TransmissionQuality ( user satisfaction )

Video traffic

Multimedia Traffic

Network convergence

Multimedia Traffic

Video people

Network people

Broadcasting,Storage

Circuit switching

Packet switching

Dedicated bitrateFixed rate (CBR)

Shared bitrateVariable rate (VBR)

Video traffic needs merits of both. Dedicated feature of circuit switching Shared feature of packet switching

Leaky bucket model!!

b

r

Bitrate?

MPEG video (Ex: SDTV with Ravg=5Mbps)GOP (500ms, 10Mbps) Frame (30ms, 30Mbps) Macroblock (0.1ms, 100Mbps)

Network (e.g. IETF)Instantaneous transmission speed of any packet is 100Mbps in a 100Mbps LAN Average < Guaranteed < Peak bitrateConstant bitrate(CBR), variable bitrate(VBR)

Multimedia Traffic

10 Mbps5 Mbps

1hr 2hr

VBR Video

SIN CITY

0123456789

1 119 237 355 473 591 709 827 945 1063 1181

Multimedia Traffic

MaxMeanMin Bitrate, R

Distortion

bad

good

Too good

rp

Bitrate

(Mbps)

Time (GOP)

VBR is optimal for compressed video.

CBR Transmission of VBR Video?

Naturally compressed video is VBR.VBR traffic is better suited to VBR channel.

Double Leaky Bucket

Peak bitrate (p, bp)

Guaranteed bitrate (r, p)

p

bp

rb

Double leaky buckets

Multimedia Traffic

output

input

b

r

M

p

Realtime traffic

non-realtime traffic

Resource Allocation for VBR

Resource = bandwidth + buffer (realtime) (non-realtime)

When realtime service needs excess bandwidth, non-realtime service packets are buffered.

Multimedia Traffic

Non realtime traffic

Realtime traffic

Resource Allocation in a Router

Double leaky bucket for VBR traffic

(peak, buffer1), (guaranteed, buffer2)

One shared egress channel

(peaks of realtime traffics) < (total BW)

(guaranteed of all traffics) < (total BW)

Dual buffer : for peak rate and guaranteed rate

(buffer1`s) < (Realtime Buffer Limit)

(buffer2`s) < (non-Realtime Buffer Limit)

Multimedia Traffic

Multiplexing rt-VBR`s and nrt`s

VBR : peak (p,pb), guaranteed (r, b)Rt-VBR : realtime e.g. VOD, VOIP, video phone etc.nrt : non-realtime e.g. Web, ftp, E-mail, etc.

Multimedia Traffic

KBS (p1,pb1)(r1,b1)FTP : nr1

MBC (p2,pb2)(r2,b2)Web : nr2

VOIP (p3,pb3)(r3,b3)E-mail : nr3

Brt, fast buffer

Bnrt, slow buffer

Rma

x

Instantaneous bitrateP=p1+p2+p3<Rmax

Average bitrateR+NR=r1+r2+r3 +nt1+nr2+nr3 <Rmax

Fast buffer PB=pb1+pb2+pb3<Brt

Slow buffer B=b1+b2+b3 <Bnrt

Realtime (P,PB)(R,B)

Non-realtime NR

Buffer management

Multimedia Traffic

Buffer

B

Data (bytes)OverflowBecause of excess feed

UnderflowLess feed, nothing to

decode

Available network

bandwidth I

Available network

bandwidth II

Maximum buffered data

Sum of di + B

Media dataSum of di

t (frame#)

Scheduling (VBR channel)

Multimedia Traffic

Buffer

B

Data (bytes)

Optimal schedule A(t)

Constant rate schedule (average)

Maximum buffered dataW(t)

Media dataD(t)

t (frame#)(a)

(b)

(d)

(e)

(c)δ

Statistical multiplexing of VBR video

As multiplexing more VBRs, peak-to-mean ratio↓Multiplexing VBRs similar to CBR

Multimedia Traffic

Averaged frame sizes of 1, 2, 4, 8 VBR video streams

Leaky bucket for VBR video

Max Buffer Size by rate

1.E+00

1.E+01

1.E+02

1.E+03

1.E+04

1.E+05

1.E+06

2.9 3.2 3.5 3.8 4.1 4.4 4.7

r [Mbps]

b [k

Byte

s]

Original

Merge 2

Merge 4

Merge 8

b

r

Leaky-bucket methodBuffer size (bucket depth) ‘b’ to guarantee no overflow at the drain rate ‘r’The more multiplexed, the smoother

r

b

Statistical Multiplexing of VBR Video Traffics

Multiplexed VBRs becomes CBR-like.Back to the vision of ATM in 1990s Cons : Self-similarity

(mean , deviation )For CBR,

For n VBR`s

30-40% saving in total bandwidthSatellite in which bandwidth is much costy.

0)(

)(

mean

deviation

nn

n

mean

deviation

)(

)(

Statistical Multiplexing Gain

28

CBR

n=10

0.5

1

10-1

10-2

10-3

m

n=1

1.2m 2m

%7.662.1

)2.12(

m

m

Statistical multiplexing

gain

CBR

n=10

n=1

m

Statistical multiplexing

Pdf and negative cdf

29

+ =

5Mbps 5Mbps 10Mbps

N(μ,σ)=N(5Mbps, 1Mbps)

N(10Mbps, 1.7Mbps)

N(5Mbps, 0.85Mbps)

N(5Mbps, 1Mbps)

5Mbps

normalize

backward cumulation

5Mbps

0.5

probability

1

xX Pr

0.5

Statistical multiplexing gain

Multimedia Traffic

Complementary cumulative density function of statistical multiplexed traffic

[log scale] Complementary cumulative density function of statistical multiplexed traffic

gain

Log scale

0.5

End-to-end delay components

Multimedia Traffic

playpathi

ipathi

iif DQPTlTD

)(

Constant, function of:

• Encoding Scheme(frame size, look –ahead)• l : Packet size = k*framesize

Constant, function of:

• Propagation delay = Distance / light speed• Link speed= Packet size / channel speed

Source of Jitter, random, function of :

• Path in the network• Traffic load and Characteristics• Scheduling scheme• Packet size

WaitPacketizationEncoding

time

LookAhead

Encoder delay

Propagation delay

queuing delay

Decoder/ display delay

End-to-end Delay Requirements

Hence, the one-way (end-to-end) delay becomes

Toll quality real-time communication neededRound-trip delay must be in the range 200-300msThat is, ,where

Amount of jitter allowed 10-50ms, function ofAcceptable end-to-end delay Dmax

Formation time Tf

Propagation delay

Multimedia Traffic

edsynchronizDandSQ

edsynchroniznonDandSQPTlTD

iiif

,max

max,2)(

maxDD msDms 150100 max