MATE: MPLS Adaptive Traffic Engineering Anwar Elwalid Cheng Jin Steven Low Indra Widjaja Bell Labs Michigan altech Fujitsu

1
MATEMPLS Adaptive Traffic EngineeringAnwar
Elwalid Cheng Jin Steven Low Indra WidjajaBell
Labs Michigan altech Fujitsu

• 2006

2
Talk Outline

• MPLS Traffic Engineering
• Overview of MATE
• Theoretical Results
• Simulation Results

3
Best of Both Worlds

• MPLS IP form a middle ground that combines the
  best of IP and the best of virtual circuit
  switching technologies
• ATM and Frame Relay cannot easily come to the
  middle so IP has!

4
Label Encapsulation

• MPLS between L2 and L3
• MPLS Encapsulation is specified over various
  media types. Top labels may use existing format,
  lower label(s) use a new shim label format.

5
Label Substitution

• Have a friend go to B ahead of you using one of
  the two routing techniques (hop-hop, source). At
  every road they reserve a lane just for you. At
  every intersection they post a big sign that says
  for a given lane which way to turn and what new
  lane to take.

6
MPLS Explicit Routing

• Multiple Label-Switched Paths (LSPs) between an
  ingress-egress pair can be efficiently established

7
The Need for Traffic Engineering

• No automatic load balancing among LSPs

8
Design Goals

• Distributed load-balancing algorithm
• Need no extra network support
• Minimal packet reordering required
• General framework for traffic engineering
• Internet Draft draft-widjaja-mpls-mate-02.txt

9
Two-State Adaptive Traffic Engineering
10
Functional Units in Ingress LSRs

• Probe packets are sent to estimate the relative
  one-way mean packet delay and packet loss rate
  along the LSP

11
Traffic Engineering Problem

• For each Ingress-Egress pair s
• Input
• Offered Load as
• Set of LSPs Ps (an LSP p)
• Output
• Vector of traffic splits ls lsp as

12
Problem Formulation

• Define a cost Cp , for an LSP p, as a function of
  link utilization lsp
• Each ingress-egress pair minimizes the sum of the
  cost function of each LSP subject to a feasible
  traffic split
• Min C(ls) Cp (lsp)
• s.t. lsp as, lsp gt 0

13
Understanding the Cost Function

• Not necessarily a perfect cost function
• Help steer network toward desirable operating
  point
• Allows systematic derivation and refinement of
  practical traffic engineering schemes

14
Solution Approach

• Optimality Criterion
• Optimal if paths with positive flow have minimum
  (and equal) cost derivatives
• Gradient Projection Algorithm
• Shift traffic from paths with highest derivatives
  to paths with lowest derivatives by a small
  amount each iteration

15
Asynchronous Environment

• Feedback delays (probe measurements)
• non-negligible
• different delays for LSPs
• time-varying
• Many ingress-egress routers shift traffic
• independently
• at different times
• likely with different frequencies

16
Convergence under AsynchronousConditions

• The algorithm will converge provided the cost
  function satisfies certain requirements
• Starting from any initial rate vector l(0), the
  limit point of the sequence l (t) is optimal,
  provided the step size is sufficiently small
• Bound on step size estimates the effect of
  asynchronism

17
Packet-level Discrete Event Simulator

• Entities Packets, Routers, Queues, and Links
• Simulated Functional Units
• Measurement and Analysis
• Traffic Engineering
• Assume traffic already filtered into bins
• Both Poisson and Long-range dependent traffic
  (DAR)

18
Experiment Setup
19
Aggregate Utilization on Shared Links
20
Packet Loss on Shared Links
21
Conclusion

• MPLS Adaptive Traffic Engineering
• an end-to-end solution without network support
• distributed load-balancing
• steer networks toward optimal operating point
  under asynchronous network conditions
• validated in simulation