Performance Modeling of MPLS Networks with Link Failures and Repairs

1
Performance Modeling of MPLS Networks with Link
Failures and Repairs

• Dénes Papp, Tien Van Do, Xuan Mai T. Truong
• COST 285 8th MCM meeting
• Sep. 21-22, 2006
• Pisa, Italy

2
Outlines

• Background goals
• Analytical model
• Solution of the model
• Validation with simulation
• Results
• Further research

3
Load balancing in MPLS networks

• Multipath routing in MPLS
• Load balancing mechanism in case of LSP failures
  and repairs
• Repair strategies

4
Goals

• Propose a model which
• calculates network performance parameters
• could take into account LSP failures and repairs
• offers an easy way to implement new repair
  strategies

5
The proposed analytical model

• Mapping LSPs to servers
• c heterogeneous servers in parallel
• GE-distributed service time parameters
• L queuing capacity

6
The proposed analytical model

• Using a generalized Markovian node model, called
  HetSigma (MM?Kk1CPPk/GE/c/L G-queue)
• Arrival process (X)
• Repair and failure can described by a Markovian
  process (Z)
• The system process Y has the generator matrix as
  the Kronecker sum of Qx and Qz.
• steady state probabilities with the same
  methodology developed for HetSigma model.

7
Tools for solving the analytical model

• Mathematica (Wolfram)
• Probabilistic Symbolic Model Checker, PRISM
  (University of Birmingham)
• Own system description language
• It can handle large number of states (105 - 106)

8
Validation with simulation

• Sample network, queue length vs. arrival
  intensity
• Differences between analytical and simulation
  model
• Assume that only one link can fail along an LSP
• Calculation of repair intensity

9
Numerical results

• Number of waiting customers and packet loss
  probability at different link repair intensity

10
Further research

• Investigating new repair strategy based on
  priority
• Refinement of determination of priority order
• Finding an appropriate function for weighting
  failure/repair rates and LSP bandwidth component.