mate: mpls adaptive traffic engineering anwar elwalid cheng jin steven low indra widjaja bell labs...
DESCRIPTION
Best of Both Worlds n MPLS + IP form a middle ground that combines the best of IP and the best of virtual circuit switching technologies n ATM and Frame Relay cannot easily come to the middle so IP has!TRANSCRIPT
MATE:MPLS Adaptive Traffic Engineering
Anwar Elwalid Cheng Jin Steven Low Indra WidjajaBell Labs Michigan altech Fujitsu
20062006
Talk Outline
MPLS Traffic EngineeringMPLS Traffic Engineering Overview of MATEOverview of MATE Theoretical ResultsTheoretical Results Simulation ResultsSimulation Results
Best of Both Worlds
MPLS + IP form a middle ground that combines MPLS + IP form a middle ground that combines the best of IP and the best of virtual circuit the best of IP and the best of virtual circuit switching technologiesswitching technologies
ATM and Frame Relay cannot easily come to the ATM and Frame Relay cannot easily come to the middle so IP has!middle so IP has!
Label Encapsulation
MPLS – between L2 and L3MPLS – between L2 and L3 MPLS Encapsulation is specified over various MPLS Encapsulation is specified over various
media types. Top labels may use existing media types. Top labels may use existing format, lower label(s) use a new “shim” label format, lower label(s) use a new “shim” label format.format.
Label Substitution Have a friend go to B Have a friend go to B ahead of youahead of you using one of using one of
the the twotwo routing techniques routing techniques (hop-hop, source).(hop-hop, source). At At every road they reserve a lane just for you. At every road they reserve a lane just for you. At every intersection they post a big sign that says every intersection they post a big sign that says for a given lane which way to turn and what new for a given lane which way to turn and what new lane to take.lane to take.
MPLS Explicit Routing
Multiple Label-Switched Paths (LSPs) between an Multiple Label-Switched Paths (LSPs) between an ingress-egress pair can be efficiently establishedingress-egress pair can be efficiently established
The Need for Traffic Engineering No automatic load balancing among LSPsNo automatic load balancing among LSPs
Design Goals Distributed load-balancing algorithmDistributed load-balancing algorithm Need no extra network supportNeed no extra network support Minimal packet reordering requiredMinimal packet reordering required General framework for traffic engineeringGeneral framework for traffic engineering
Internet Draft: draft-widjaja-mpls-mate-02.txtInternet Draft: draft-widjaja-mpls-mate-02.txt
Two-State Adaptive Traffic Engineering
Functional Units in Ingress LSRs
Probe Probe packets are sent to estimate the packets are sent to estimate the relative relative one-one-way mean packet way mean packet delaydelay and packet and packet loss loss rate along rate along the LSPthe LSP
Traffic Engineering Problem For each Ingress-Egress pair s:For each Ingress-Egress pair s: InputInput
Offered Load: Offered Load: aass
Set of LSPs: Set of LSPs: PPs s (an LSP (an LSP pp)) OutputOutput
Vector of traffic splits: Vector of traffic splits: ss spsp = = aass
Problem Formulation Define a cost CDefine a cost Cpp , for an LSP , for an LSP pp, as a function , as a function
of link utilization of link utilization llspsp
Each ingress-egress pair minimizes the sum Each ingress-egress pair minimizes the sum of the cost function of each LSP subject to a of the cost function of each LSP subject to a feasible traffic splitfeasible traffic split
Min C(Min C(ss) = C) = Cpp ( (spsp))
s.t. s.t. spsp = = aass, , spsp > > 00
Understanding the Cost Function Not necessarily a perfect cost functionNot necessarily a perfect cost function Help steer network toward desirable Help steer network toward desirable
operating pointoperating point Allows systematic derivation and refinement Allows systematic derivation and refinement
of practical traffic engineering schemesof practical traffic engineering schemes
Solution Approach Optimality CriterionOptimality Criterion
Optimal if paths with positive flow have Optimal if paths with positive flow have minimum (and equal) cost derivativesminimum (and equal) cost derivatives
Gradient Projection AlgorithmGradient Projection Algorithm Shift traffic from paths with highest Shift traffic from paths with highest
derivatives to paths with lowest derivatives to paths with lowest derivatives by a small amount each derivatives by a small amount each iterationiteration
Asynchronous Environment Feedback delays (probe measurements):Feedback delays (probe measurements):
non-negligiblenon-negligible different delays for LSPsdifferent delays for LSPs time-varyingtime-varying
Many ingress-egress routers shift trafficMany ingress-egress routers shift traffic independentlyindependently at different timesat different times likely with different frequencieslikely with different frequencies
Convergence under AsynchronousConditions
The algorithm will converge provided the cost The algorithm will converge provided the cost function satisfies certain requirementsfunction satisfies certain requirements
Starting from any initial rate vector Starting from any initial rate vector (0), the (0), the limit point of the sequence {limit point of the sequence { (t)} is optimal, (t)} is optimal, provided the step size is sufficiently smallprovided the step size is sufficiently small
Bound on step size estimates the effect of Bound on step size estimates the effect of asynchronismasynchronism
Packet-level Discrete Event Simulator
Entities: Packets, Routers, Queues, and Entities: Packets, Routers, Queues, and LinksLinks
Simulated Functional UnitsSimulated Functional Units Measurement and AnalysisMeasurement and Analysis Traffic EngineeringTraffic Engineering Assume traffic already filtered into binsAssume traffic already filtered into bins
Both Poisson and Long-range dependent Both Poisson and Long-range dependent traffic (DAR)traffic (DAR)
Experiment Setup
Aggregate Utilization on Shared Links
Packet Loss on Shared Links
Conclusion MPLS Adaptive Traffic EngineeringMPLS Adaptive Traffic Engineering
an end-to-end solution without network an end-to-end solution without network supportsupport
distributed load-balancingdistributed load-balancing steer networks toward “optimal” steer networks toward “optimal”
operating point under asynchronous operating point under asynchronous network conditionsnetwork conditions
validated in simulationvalidated in simulation