COMMS IRAD: Traffic Modeling

1
COMMS IRAD Traffic Modeling
2
IRAD Traffic Modeling Goals

• Assess current understanding of network traffic
  modeling
• Determine best approach for modeling Internet
  related traffic using OPNET
• Develop process for determination of model
  parameters from empirical data

3
Traffic Modeling Research Plan

• Data Collection and Analysis
• Literature Search
• Data Capture
• Tool Acquisition
• Analysis
• Custom Traffic Generator Implementation and Test
• Design
• Code
• Test

4
Data Collection Analysis Literature Search
Results

• Approximately 44 research papers obtained
  covering the following subjects
• Probabilistic Modeling (Poisson models, ARIMA,
  discrete Markovian processes, etc.)
• Self Similarity Models (Heavy tailed
  distributions, Fractal Gaussian Noise)
• Fractal Point Processes
• Multi-fractal Scaling Processes
• Network Traffic / User Service Traffic
  Characteristics

5
Literature Search ResultsAssessment of Industry

• Determined that traffic modeling has progressed
  over past 7 years as follows
• Poisson models modified with heavy tails (called
  M Pareto models) inadequate and poor results to
• Discrete-time Batch Markovian Arrival Processes
  better but still inadequate to
• Fractal Gaussian Noise better with self
  similarity feature but does not exhibit observed
  multifractal scaling to
• Fractal Point Processes (such as Fractal Binomial
  Noise Driven Process) which exhibit monofractal
  scaling to
• Multifractal based on Fractal Point Processes
  better with self affinity scaling feature to
• Multifractal conservative cascades and discrete
  wavelet transform synthesis
• All methods assume stationarity for the rate
  process even though observed network traffic
  follows diurnal cycles (not problematic as when
  modeling network stress via background traffic
  only the peak times need be considered)

6
Data Collection AnalysisReported Observations

• Traffic Characterization
• Possible sources of fractal behavior include
  heavy tail distributions of WEB file size, user
  idle times, FTP file size, transmit idle times
  of LAN host
• WAN Traffic
• Self similar (monofractal) at larger time scales
• Recent observations indicate multifractal scaling
  at short time scales in data traces
• Aggregate WAN is self similar if user initiated
  sessions arrive in a Poisson fashion w/ heavy
  tailed distribution having infinite variance
• Self similarity is independent of the network
• Network Impact
• When studying traffic over small time scales
• Local properties of WAN traffic are consistent
  with multifractals
• Multifractal scaling has little to do with the
  user session characteristics
• Multifractal scaling is the result of protocols
  and end to end congestion control methods
• The transition between multifractal and self
  similar scaling occurs at times scales of
  approximately the Round Trip Time (RTT) of
  packets in the network
• For small time scales the conservative cascade
  closely matches the way network mechanisms
  influence individual TCP connections

7
Literature Search ResultsMethodologies Adopted

• Most promising approaches identified are
• Superposition of Fractal Point Processes (Fractal
  Shot Noise Driven Poisson Process, Fractal
  Binomial Noise Driven Poisson Process, Fractal
  Renewal Process, and superposition of FRPs)
• Already incorporated into OPNET rel. 7 Modeling
  tools
• The FBNDP can model monofractal distributions
• The FSNDP can model multifractal distributions (3
  distinct scaling regions)
• Multifractal Wavelet Method (using conservative
  cascades)
• Intuitive mechanistic modeling technique, similar
  to that used to analyze turbulence
• Can be used to generate general multifractal
  scaling

8
Literature Search ResultsCurrent Areas of
Investigation

• Most pressing needs for improvement are
• Traffic research needs to mature from just
  numerical curve fitting to empirical data
  towards the ability to predict traffic statistics
  based on network configuration, protocols, and
  user session characteristics
• Math models can be generated but understanding of
  the precise causes and mechanisms for
  multifractal scaling is still rather heuristic
• The problem is exacerbated in that long term
  observations are hampered due to continual
  modifications in protocols, routing algorithms,
  configurations, etc. in the networks

9
Literature Search ResultsCurrent Areas of
Investigation

• Debatable Issues
• Which method more truly models self similarity
  aspects of network traffic both in terms of
  fitting the data and by associating cause with
  effects
• Some argue that flow driven FPP is representative
  of user sessions (flows) and naturally models the
  network traffic
• Others argue that probabilistic cascades with
  pdfs determined by multi- resolution
  approximation via orthogonal wavelet coefficients
  can more accurately model TCP traffic as it
  naturally models the network aspects mechanisms
  of traffic aggregation in reverse (here the
  aggregate data rate is successively divided by
  each stage in the cascade so as to result in
  multifractal time scaling)
• Another network modeling concern is the
  appropriate granularity level
• Start with individual user profiles and combine
  them to model aggregate OR
• Generate expected aggregate traffic based on
  combination of empirical data and mathematical
  models

10
Data Collection Analysis Data Capture

• Sources Identified for Modeling Internet Traffic
• Bellcore (Morristown Research Eng. Facility)
• LAN Traces for Aug. and Oct. 1989 were found and
  downloaded
• Each trace consists of first million packets
  starting at 1125 am and 1100 am for Aug. 29th
  89 and Oct. 5th 89 respectively
• National Laboratory for Applied Network Research
• Offers freely available high quality network
  traces (this has not been investigated under this
  IRAD as of yet)
• In-house collected traces
• TASC has network monitoring tools which can be
  used to collect traffic statistics
• Work has begun on collecting these traces

11
Data Collection AnalysisAnalysis

• Studies undertaken to complete analysis
• Fractal Mathematics
• Fractal Point Processes
• Wavelet Processing (multiresolution
  approximation)
• Data Analysis
• Matlab scripts and functions were written to
• Perform processing of raw trace data into rate
  process data
• Perform data reduction (statistical analysis at
  different time scales) of traffic rate data
• Mean, standard variance, Allan variance, Index of
  Dispersion Counts
• MLE (Least squares fitting) of reduced data to
  Fractal Binomial Noise Driven Poisson Process
  Model Parameters (Hurst Parameter, Fractal Onset
  Time Scale, mean packet rate, of processes,
  source activity ratio, and cutoff parameter)

12
Data Collection AnalysisAnalysis

• OPNET Investigations
• Use of Raw Packet Generators (RPG) added in
  release 7 of OPNET was completed
• Test project files created for analyzing
  effectiveness of RPG to modeling empirical data
• Confirmed correct operation by comparison with
  Bellcore August 89 trace data and OPNET
  generated project file

13
(No Transcript)
14
Bellcore Ethernet LAN Trace Data (Packet arrivals
in 16 sec. Intervals)
15
Measured IDC for Bellcore Trace vs Fitted IDC
using FBNDP Model
Note The red trace is fitted model data using
T00.0040 and ?0.6759
Model Curve Equation
16
Comparison for Bellcore pAug Trace(As Reported
by B. Ryu and S. Lowen)
Note These results compare favorably with our
results although it is not clear why the least
square fitted results are not identical.
17
OPNET Generated FBNDP (FMPP PowON-PowOFF) Model
Comments 1. Station 1 (Stn_1) transmits via the
hub to station 2 (stn_2) 2. Self similar traffic
generated by raw packet generator above MAC
sub-layer 3. Parameters selected correspond to
Bellcore pAug trace using std variance fitting
Process Model for Stations Node
Models
18
OPNET Generated Self SimilarTraffic
OPNET Generated FBNDP (PowON-PowOFF) using
Standard Variance Traffic is shown averaged over
T0.01, 0.1, 1.0, 5.0 sec intervals
19
Comparison of OPNET Generated FBNDP Trace
Bellcore Measured Trace
20
Work Planned for Next Quarter

• Custom Traffic Generator Implementation
• Design Identify and Delineate Multifractal
  Wavelet Method (MWM) Processing Steps
• Code Implement Conservative Cascade Multifractal
  Traffic Generator in C/C and Matlab using
  Wavelab
• Test Assess MWM modeling performance
• Code Develop OPNET Based MWM Traffic Generator