1 Queuing Theory vs. System I dentification Network Configurations - - PDF document

SLIDE 1

1

Measurement Measurement-

• Based Modeling of

Based Modeling of I nternet Round I nternet Round-

• Trip Time Dynamics

Trip Time Dynamics using System I dentification using System I dentification

Hiroyuki Hiroyuki Ohsaki Ohsaki Department of Information Science and Department of Information Science and Technology, Osaka University, Japan Technology, Osaka University, Japan

• osaki@ist.osaka
• osaki@ist.osaka-
• u.ac.jp

u.ac.jp http://www. http://www.anarg anarg. .jp jp/ /

Research Background Research Background

If characteristic of If characteristic of packet transmission delay packet transmission delay is is known known… …

Can improve Can improve QoS QoS (Quality of Services) (Quality of Services) Can realize efficient congestion control Can realize efficient congestion control

Dynamics of packet transmission delay Dynamics of packet transmission delay

How packet transmission delay is affected by How packet transmission delay is affected by packet packet transmission rate transmission rate sender receiver

Research Objectives Research Objectives

Model dynamics of packet transmission delay Model dynamics of packet transmission delay

View network as a View network as a black box black box Measure input and output to network Measure input and output to network Estimate model parameters by Estimate model parameters by system identification system identification

Validate model accuracy Validate model accuracy

Using two validation methods Using two validation methods Comparison with simulation (time domain) Comparison with simulation (time domain) Comparison with spectral analysis result (frequency domain) Comparison with spectral analysis result (frequency domain)

receiver sender

Black Black-

• Box Modeling

Box Modeling

Definition of input and output Definition of input and output packet transmission rate round-trip time

input

• utput

input

• utput

black box

ARX ARX（ （ Auto Auto-

• Regressive

Regressive eXogenous eXogenous） ） Model Model

ARX model

u(k) y(k) e(k)

input (transmission rate)

• utput

(round-trip time) noise (effect of other traffic, etc)

) ( ) ( ..... ) 1 ( ) ( ..... ) 1 ( ) (

1 1

k e n k u b k u b n k y a k y a k y

b n a n

b a

+

• +

+

• +
• =

Determined by system identification

Modeling Procedure Modeling Procedure

• 1. Measure input and output data
• 2. Determine ARX model parameters from

measured input and output data

• 3. Validate accuracy of ARX model obtained

model parameters system identification ARX model input data

• utput data

SLIDE 2

2

Queuing Theory vs. System I dentification Queuing Theory vs. System I dentification

? ? good good model accuracy model accuracy not required not required required required model information model information cannot model cannot model can model can model statistical statistical characteristic characteristic yes yes no no control theory control theory applicability applicability can model can model cannot model cannot model dynamics dynamics black box black box white box white box modeling approach modeling approach system identification system identification queuing theory queuing theory

Network Configurations Network Configurations

Build model in three network configurations Build model in three network configurations

Complexity of network topology Complexity of network topology Effect of Effect of noise noise (e.g., background traffic) (e.g., background traffic) Location of Location of bottleneck link bottleneck link

N1 N1: LAN : LAN

Simple network topology Simple network topology

N2 N2: WAN : WAN

Complex network topology Complex network topology Access link is Access link is bottleneck bottleneck

N3 N3: WAN : WAN

Complex network topology Complex network topology Access link is Access link is not bottleneck not bottleneck

Network N1 Network N1

100Mbps

2000 2150 120 40 80 slot Mbps

Input (packet transmission rate)

0.41 0.44 0.42 0.43 2000 2150 slot

Output (Round-Trip Time)

ms

sampling interval = 0.9ms

FTP server sender FTP client receiver

Network N2 Network N2

sender

Internet

access link

bottleneck

2000 2150 slot Mbps

Input (packet transmission rate)

2000 2150 slot

Output (Round-Trip Time)

ms 0.08 0.02 0.04 0.06 150 500

sampling interval = 125ms

receiver

ISP

Network N3 Network N3

sender

Internet bottleneck?

receiver

2000 2150 slot Mbps

Input (packet transmission rate)

2000 2150 slot

Output (Round-Trip Time)

ms 20 5 15 10 20 80 40 60

sampling interval = 6ms

Validation using Simulation Validation using Simulation (Comparison in Time Domain) (Comparison in Time Domain)

model output input data 2

Compare output data (not used for model creation) with simulation output

• utput data 2

ARX model

input data 1

• utput data 1

SLIDE 3

3

Validation using Board Diagram Validation using Board Diagram (Comparison in Frequency Domain) (Comparison in Frequency Domain)

unknown

) sin( t ω

gain phase

A

ϕ

ARX model

• btain frequency characteristic

from ARX model

input data

• utput data

estimate frequency characteristic using spectral analysis Compare frequency characteristics of ARX model and one estimated using spectral analysis

) sin( ϕ ω + t A

Network N1: Simulation Network N1: Simulation

2300 2290 2280 2270 2260 2250 0.41 0.415 0.42 measured output model output slot Round-Trip Time

100Mbps FTP server sender FTP client receiver

Network N1: Board Diagram Network N1: Board Diagram

10

• 1

10 10

1

10

• 2

10

• 1

10 10

1

10

• 2

ARX model spectral analysis

phase (degree) amplitude

10

• 6

10

• 2

10

• 4

200 100 frequency (rad/s)

Network N2: Simulation Network N2: Simulation

2300 2290 2280 2270 2260 2250 measured output model output slot Round-Trip Time 150 450 300 250 200 350 400

sender

Internet

bottleneck receiver

Network N2: Board Diagram Network N2: Board Diagram

10

• 1

10 10

1

10

• 2

10

• 1

10 10

1

10

• 2

ARX model spectral analysis

phase (degree) amplitude

10

2

10

4

10

3

400 200 frequency (rad/s)

Network N3: Simulation Network N3: Simulation

2300 2290 2280 2270 2260 2250 measured output model output slot Round-Trip Time 25 20 30 35 40

sender

Internet

bottleneck? receiver

SLIDE 4

4

Network N3: Board Diagram Network N3: Board Diagram

10

• 1

10 10

1

10

• 2

10

• 1

10 10

1

10

• 2

ARX model spectral analysis

phase (degree) amplitude

10

2

10

2

10

1000 500 frequency (rad/s)

Conclusion Conclusion

Model dynamics of Model dynamics of packet transmission delay packet transmission delay

View network as a View network as a black box black box Measure input and output data Measure input and output data Determine model parameters by Determine model parameters by system identification system identification

Validate model accuracy Validate model accuracy

Simulation (comparison in time domain) Simulation (comparison in time domain) Bode diagram (comparison in frequency domain) Bode diagram (comparison in frequency domain)

Show effectiveness of black Show effectiveness of black-

• box modeling

box modeling

When network is When network is not so noisy not so noisy

Future Works Future Works

More accurate modeling of packet transmission delay More accurate modeling of packet transmission delay

Model structures Model structures

Parametric model Parametric model Non Non-

• parametric model

parametric model

Linearity Linearity

Linear model Linear model Non Non-

• linear model

linear model

Noise assumption Noise assumption

white noise white noise colored noise colored noise

Design a delay Design a delay-

• based congestion control using control

based congestion control using control theory theory More info: http://www. More info: http://www.anarg anarg. .jp jp/ /