An Analytical Model for Network Flow Analysis Ernesto Gomez, Yasha Karant, Keith Schubert Institute for Applied Supercomputing Department of Computer Science CSU San Bernardino The authors gratefully acknowledge the support of the NSF under award CISE 98-10708
Outline Networks and Flows History Statistical Mechanics Self-similar traffic Traffic creation and destruction Master Equation and traffic flow
One View of Network
Network Flows
Outline Networks and Flows History Statistical Mechanics Self-similar traffic Traffic creation and destruction Master Equation and traffic flow
Brief History Shannon-Hartley (classical channel capacity) C=B log2(1+SNR) Leland, Taqqu, Willinger, Wilson, Paxon, … Self-similar traffic Cao, Cleveland, Lin, Sun, Ramanan Poisson in limit
Stochastic vs. Analytic Stochastic best tools currently Opnet, NS Problems limiting cases Improving estimates Analytic (closed form equations) Handles problems of stochastic Insight into structure Fluid models Statistical Mechanics
Outline Networks and Flows History Statistical Mechanics Self-similar traffic Traffic creation and destruction Master Equation and traffic flow
Overview Large number of entities Bulk properties Equilibrium or non-equilibrium properties Time-dependence Conservation over ensemble averages Can handle classical and quantum flows
Density Matrix Formalism Each component Label by state n = node source and destination f = flow index c = flow characteristics t = time step
Density Matrix II Probability of a flow Element in Density Matrix is Averaged Properties
Outline Networks and Flows History Statistical Mechanics Self-similar traffic Traffic creation and destruction Master Equation and traffic flow
Poisson Distribution is mean Thin Tail
Problem with Poisson Burst Long-range dependence Extended period above the mean Variety of timescales Long-range dependence Poisson or Markovian arrivals Characteristic burst length Smoothed by averaging over time Real distribution is self-similar or multifractal Proven for Ethernet
Real versus Poisson
Pareto Distribution Shape parameter () Location parameter (k) Smaller means heavier tail Infinite varience when 2 ≥ Infinite mean when 1 ≥ Location parameter (k) t≥k
Pareto Distribution
Outline Networks and Flows History Statistical Mechanics Self-similar traffic Traffic creation and destruction Master Equation and traffic flow
Flow Origination Unicast Multicast One source One destination Many segments Multicast Many destinations
Multicast Possibilities
Outline Networks and Flows History Statistical Mechanics Self-similar traffic Traffic creation and destruction Master Equation and traffic flow
Probability in Density Matrix Tr = eHt (H is energy function) Tr= (1+t/tns)-1 Cauchy Boundary conditions hypersurface of flow space Ill behaved Gaussian quadrature, Monte Carlo, Pade Approximation
Unicast Flow Time
Future Directions More detailed network Bulk properties Online tool