In this presentation we introduce a novel approach to model heterogenous
network traffic. The approach is based on a coarse discrete approximation
of the traffic. This allows to use linear Diophantine theory to analyze
the approximated model. The theory reduces the model to a certain system
of linear Diophantine equations, whose solutions are restricted with
nonnegative integer vectors. Basis of this system represents a
decomposition of the traffic onto the most significant combinations of
data types (sources). Each combination is characterized by a stable
linear proportion between volumes of its data types. To find the basis we
designed and implemented a new fast grammar--based algorithm that works
in pseudo-polynomial time.
We experimented with the model on real data traffic of the external
channel of the University of Petrozavodsk. Besides the model validation
and test, the main result of the experiment is a confirmation that linear
proportions between some data types can be considered as traffic
invariants, both on different time scales and different time intervals.
This makes the proposed decomposition to be a stable characteristic
property of traffic.
The model can be used: i) to estimate the current state of a network
link, its stability, structure, trends, required throughput, indication
on principal changes; ii) to be used as a base for an emulator of a
network link.