prof. k.j.blow, dr. marc eberhard and dr. scott fowler

44
Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler Adaptive Communications Networks Research Group Electronic Engineering Aston University Significance of Joint Density Plots in Markov Internet Traffic Modelling AHMED D. SHAIKH

Upload: kieve

Post on 13-Jan-2016

31 views

Category:

Documents


0 download

DESCRIPTION

Significance of Joint Density Plots in Markov Internet Traffic Modelling. AHMED D. SHAIKH. Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler Adaptive Communications Networks Research Group Electronic Engineering Aston University. Outline. Outline. Traffic Modelling Approaches. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Adaptive Communications Networks Research Group

Electronic Engineering

Aston University

Significance of Joint Density Plots in Markov Internet Traffic Modelling

AHMED D. SHAIKH

Page 2: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Outline

Page 3: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Outline

Page 4: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Traffic Modelling Approaches

Two types of Approaches:

• Black Box models• Internal structure is unknown. Opaque to user.

• Examples: HMM, MMPP, BMAP

• White Box models• Transparent structure. Has a physical meaning.

• Examples: Classic Markov Models, On-Off models

Page 5: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Markov Models – An Introduction

• Probabilistic models defining a stochastic process with finite number of states observing the Markov Property.

• Transitions occur with a fixed

transition rate Rij.

• States can model activities of traffic

sources on a network.

• Inter-Arrival times are exponentially

distributed.

• Packet level statistics obtained from Monte Carlo simulations are expressed in IPT (Inter-Packet times)

Page 6: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Outline

Page 7: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Simple Two State Markov Traffic Model

The sequence of packets will be ABABABABABABAB…..

Page 8: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Two state Model (Analytical analysis contd..)

• Two state models will have equal number of visits to each state.

So, V1 = V2 = 0.5

• Probability densities of time spent in each state:

• The Probability Density function of IPT for a two state model is:

Page 9: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Two state Markov Model (Numerical vs. Analytical results )

Page 10: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Two state Markov Model (Numerical vs. Analytical results –

Symmetric rates)

Page 11: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Higher order Statistics for Markov Models

• Higher Order Distributions

• Markov Models can also produce higher order statistics.

• Possible to study the sequence of IPTs and a variety of other unique features associated with the network traffic statistics.

• The Joint Density function for the two state Markov Model is given by:

Page 12: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Second Order Statistics – Joint Density (Results for Symmetric 2-

state model)

Page 13: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Higher Order Statistics – Joint Density

(Results for Asymmetric 2-state Model)

Page 14: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Outline

Page 15: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

N-state models with Poisson statistics

The general form equation for the IPT PDF of N-state Markov Models where every state is emitting packets is:

PDF (N-state) = V1 P1(t) + V2 P2(t) + V3 P3(t)……... + VN PN(t)

Page 16: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Outline

Page 17: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Two state Model with non-Poisson statistics

The sequence of packets is AAAAAAAAAAAA……

The PDF equation for the IPT is:

Page 18: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

PDF for the two state model with only one state emitting packets

Page 19: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Joint Density – 2 state model with one packet emitting state / source

Page 20: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

PDF for IPT for N-state Markov Models with only one state

emitting packets

The general form analytical equation of the PDF of IPT for Markov loop Models with only one state emitting packets is:

Page 21: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Use of Gamma Markov Models

Page 22: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Taking it further - A Gaussian Markov Model

• Now in the general equation of the Gamma distribution, we know that as N approaches infinity, the gamma distribution can be approximated by a normal or Gaussian distribution.

• Gives a normal distribution with mean

Variance

• Gaussian Distribution PDF.

Page 23: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Gaussian Markov Models

Page 24: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Outline

Page 25: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Modelling Real World Example – IP Traffic Measurement at UDP Port

15010 - VoIP

Page 26: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Fitting a Gaussian Markov Model

Gaussian Model(PDF) = V1* Gaussian(μ1,σ1) + V2 * Gaussian(μ2,σ2) + …+V6 * Gaussian(μ6,σ6)

Page 27: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Comparing the Joint Densities

Page 28: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Outline

Page 29: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Understanding Packet Sequences from Joint Density Results

Page 30: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

The significance of the Joint Density Plots

• Let us consider a 3 + 1 states Model where V1 = V2 = V3 = 1/3. (Markov Model A)

• Packet sequence can be ABBACACABCABBACAACA……….

Page 31: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

PDF and Joint Density – Markov Model A

Page 32: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Markov Model ‘B’

• Let us now consider a 3 state Loop Model where V1 = V2 = V3 = 1/3. (Markov Model B)

• Packet sequence must be ABCABCABCABCABCABC…..

Page 33: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

PDF and Joint Density – Markov Model B

Observation: Two different models have the same PDFs yet different Joint Densities. The Joint density Plots give more statistical details on Packet Sequences.

Page 34: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Outline

Page 35: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Understanding the curve of periodicity

Page 36: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Modelling Periodic Events with Markov Models

Page 37: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Small ∆ for Markov Models C and D - S∆ model

Page 38: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Large ∆ for Markov Models C and D - L∆ model

Page 39: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Multiple Periodicity

Page 40: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Use of S∆ and L∆ model sets to model measured results

Page 41: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Use of S∆ and L∆ model sets to model measured results

Page 42: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Outline

Page 43: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Summary and Conclusions• Summary:

• Observed first and second order statistics for N-state Markov Models with Poisson and Non-Poisson statistics and confirmed our anlaytical understanding of the models with simulated results.

• Established the significance of the Joint Density Plots and explored the use of simple Markov models to model unique features of Joint Density Traffic Statistics Results.

• Conclusions:

• The Joint Density Plot contains much more statistical information on the activities and nature of the traffic sources than the PDF.

• Modelling PDFs alone will result in reproducing first order statistics. Use of Joint Density Plots is Recommended to model source behaviour. Simple Markov Models can be used to model the unique features of Joint Densities.

Page 44: Prof. K.J.Blow, Dr. Marc Eberhard and Dr. Scott Fowler

Thank you!

Questions or comments?

The man himself:

Andrey Markov

(1856 - 1922)