Slides - PAM 2007

LiTGen, a lightweight traffic
generator: application to mail
and P2P wireless traffic
Chloé Rolland*, Julien Ridoux+ and Bruno Baynat*
* Laboratoire LIP6 – CNRS
Université Pierre et Marie Curie – Paris 6
+
ARC Special Research Center for Ultra-Broadband
Communications (CUBIN),
The University of Melbourne
Generating IP traffic with accurate
timescales properties
Web
1

General framework: multiple applications

LiTGen, a lightweight traffic generator
–
Semantically meaningful structure
–
Does not rely on a network and/or TCP emulator
–
Fast computation

Measurement based validation

Application to mail and P2P wireless traffic
Mail
P2P
Etc.
LiTGen’s underlying model
2

Focus on the download path

Do not consider up/down interactions

Focus on TCP traffic

Approach
–
Application oriented & User oriented
–
Semantically meaningful hierarchical model
LiTGen’s underlying model
IS
SESSION
SESSION
TIS
NSESSION
OBJECT
OBJECT
OBJECT
IAOBJ
NOBJ
IAOBJ
PACKETS:
3
IAPKT
Basic vs. Extended LiTGen


4
Basic LiTGen
•
Renewal processes
•
Successive random variables (R.V.) i.i.d.
•
No dependency between different R.V.
Extended LiTGen
•
Renewal processes
•
Dependency introduced, the average packets
inter-arrival depends on the objects size: IApkt =
f(Nobj)
Calibration by inspection of the
wireless trace

Wireless trace: US ISP wireless network
Mail traffic to user 1
Download Application filter Mail
traffic
src port select.
User filter
traffic
Mail traffic to user 2
Mail traffic to user i
Mail traffic to user i
5
Objects id.
Objects
Session id.
Sessions
Validation methodology

Wavelet analysis of the packets arrival
times series (LDE)
Energy spectrum
comparison
Captured trace
Synthetic trace
6
?
Comparison of different kinds of
traffics spectra (1/2)
Web + Mail + P2P traffic
7
Comparison of different kinds of
traffics spectra (2/2)
Mail traffic
8
P2P traffic
Further validation:
semi-experiments (SE)


Does LiTGen reproduces the traffic
internal structure?
Semi-experiments
Manipulation of internal parameters
– Impact of the manipulation: importance of the
parameters modified ?
–
9
Example of SE:
P-Uni

Uniformly distributes packets arrival times within
each object

Examine impact of in-objects packets burstiness
1. Impact ?
Captured trace
P-Uni
2. Similar reaction ?
Synthetic trace
10
P-Uni
SE results: mail traffic
Captured trace
11
Synthetic trace
SE results: P2P traffic
Captured trace
12
Synthetic trace
Traffic sensitivity with regards to
the distributions


13
Random Variables (R.V.) distributions?
–
Heavy-tailed distributions important?
–
Source of correlation in traffic?
Investigation of each R.V. separately
–
Replace individually the empirical distribution of the
studied R.V. by a memoryless distribution
–
Model the other R.V. by the empirical distributions
–
Impact on the spectra?
–
Conclusion on the importance of the R.V. distribution
Mail traffic sensitivity
Insensitive distributions
14
Sensitive distributions
P2P traffic sensitivity
Insensitive distributions
15
Sensitive distributions
Conclusion

Extended LiTGen reproduces accurately
the traffic scaling properties

Investigation of the impact of the R.V.
distributions
The in-objects organization is crucial
– Heavy-tailed distribution
correlation
– Give insights for the development of
accurate traffic models
–
16
Future works

Dependency introduced in Extended LiTGen

Realistic performance prediction?
–
Burstiness: strong implications on queuing &
performance
–
Compare the performance of a model fed by
•
The captured traffic
•
The synthetic traffic from LiTGen
•
17
Simpler renewal processes
Thank you !
18
Trace originating on the Sprint
access network
19