MATE: MPLS Adaptive Traffic Engineering

1
MATE MPLS Adaptive Traffic Engineering
Elwalid, A Jin, C Low, S Widjaja, I. INFOCOM
2001, pp. 1300-1309. vol.3

• Gyu Myoung Lee
• Broadband Network Laboratory
• Tel) [phone]

2
Term Project Schedule

• Objective
• Using IP-centric control plane (e.g., GMPLS)
• To provide QoS in optical network
• optimal network utilization and load balancing
  using multipath routing
• aggregating the resources of multiple paths and
  reducing blocking probabilities
• Key technology
• Traffic engineering and QoS routing
• Multipath routing
• GMPLS control technology in IP over optical
  network
• Target
• Survey paper related with the existing multipath
  routing mechanism
• Apply multipath routing algorithm to supporting
  QoS using GMPLS control plane in IP over optical
  network

3
Presentation Schedule

• 9. July (TUE)
• Study of Multipath Routing for QoS Provisioning
• Ref) network control and management for next
  generation internet
• 25. July (THU)
• An adaptive flow-level load control scheme for
  multipath forwarding
• 30. July (TUE)
• Dynamic constrained multipath routing for MPLS
  networks
• 1. August (THU)
• A constrained multipath traffic engineering
  scheme for MPLS networks
• 7. August (WED)
• Fault tolerance and load balancing in QoS
  provisioning with multipath MPLS paths
• 8. August (THU)
• MATE Multipath Adaptive Traffic Engineering
• 14. August (WED)
• Performance analysis of burst level bandwidth
  allocation using multipath routing reservation

4
Contents

• Abstract
• Introduction
• Motivation
• Overview
• MATE implementation techniques
• Experimental methodology
• Simulation results
• Conclusions
• Appendix MATE stability

5
Abstract

• Multipath adaptive traffic engineering (MATE)
• Targeted for switched networks such as MPLS
  networks
• Main goal
• Avoid network congestion by adaptively balancing
  the load among multiple paths based on
  measurement and analysis of path congestion
• Adopts a minimalist approach in that intermediate
  nodes are not required to perform traffic
  engineering or measurements besides normal packet
  forwarding
• Does not impose any particular scheduling, buffer
  management, or a priori traffic characterization
  on the nodes

6
Motivation - 1

• Internet traffic engineering is emerging as key
  tool for satisfy customers demands for fast,
  reliable, and differentiated service
• Traffic engineering maximizing operational
  network efficiency while meeting certain
  constraints
• QoS routing meet certain QoS constraints for a
  given source-destination traffic flow
• The emergence of MPLS
• Provides basic mechanisms for facilitating
  traffic engineering
• Explicit routing allows a particular packet
  stream to follow a pre-determined path
• A packet is forwarded along the LSP by swapping
  labels
• Support of explicit routing in MPLS does not
  entail additional packet header overhead

7
Motivation - 2

• Several algorithms has been proposed to add
  traffic engineering capability in traditional
  datagram network using shortest path
• Load sharing cannot be accomplished among paths
  of different costs ? Explicit LSPs and flexible
  traffic assignment can solve this problem
• Traffic/policy constrains are not taken into
  account ? Constraint base routing
• Modification of link metrics to re-adjust traffic
  mapping tend to have network-wide effects causing
  undesirable and unanticipated traffic shift ?
  LSPs are pinned down
• Traffic demands must be predictable and known a
  priori ? Proposed in this paper

8
Motivation - 3

• MPLS traffic engineering mechanisms
• Time dependent mechanism
• Historical information based on seasonal
  variations in traffic is used to pre-program LSP
  layout (changes on a relatively long time scale)
  and traffic assignment
• May not be able to prevent significant imbalance
  in loading and congestion
• State dependent mechanism
• Used to deal with adaptive traffic assignment to
  the established LSP according to the current
  state of the network (based on utilization,
  packet delay, packet loss, etc)
• In this paper
• LSP layout has been determined using a long-term
  traffic matrix
• Focus on load balancing short-term traffic
  fluctuations among multiple LSPs between an
  ingress node and an egress node

9
Overview - 1

• The features of MATE (state-dependent TE)
• Distributed adaptive load-balancing algorithm
• End-to-end control between ingress and egress
  nodes
• No new hardware or protocol equipment in
  intermediate nodes
• No knowledge of traffic demand is required
• No assumption on the scheduling or buffer
  management schemes at a node
• Optimization decision based on path congestion
  measure
• Minimal packet reordering
• No clock synchronization between two nodes

10
Overview - 2

• Assumptions
• Several explicit LSPs have been established using
  CR-LDP or RSVP-TE or configured manually
• No two IE pair use same LSP, even though some of
  their LSPs may share links
• The goal of in the ingress node
• Distribute the traffic across the LSPs
• The loads are balanced and congestion is thus
  minimized
• The traffic to be balanced by the ingress node is
  the aggregated flow (traffic trunk) that shares
  the same destination

11
Overview - 3

• MATE functions in an ingress node

12
Overview - 4

• Functions
• Filtering and Distribution
• facilitate traffic shifting among the LSPs in a
  way that reduce the possibility of having packets
  arrive at out of order
• Traffic engineering
• Decides on when and how to shift traffic among
  the LSPs
• Monitoring phase detect an appreciable and
  persistent changes
• Load balancing phase try to equalize the
  congestion among the LSPs
• Measurement and Analysis
• Obtain one-way LSP statistics such as delay and
  loss using probe packets
• Estimator of LSP statistics from the probes may
  be obtain using bootstrap resampling techniques

13
MATE implementation techniques

• Traffic filtering and distribution
• Two-stage traffic distribution
• Traffic measurement and analysis
• Probe packet packet delay and packet loss
• Bootstrap

14
Traffic filtering and distribution

• Two-stage traffic distribution
• First Stage the engineered traffic can be
  filtered and distributed into the N bins in
  numbers of ways
• Per-packet basis without filtering
• Filter the traffic on a per-flow basis (based on
  ltsource IP address, source port, destination IP
  address, destination port, IP protocolgt tuple)
• Filter the traffic by using a hash function on IP
  fields (based on CRC)
• Second Stage maps each bin to the corresponding
  LSP

15
Traffic measurement and analysis - 1

• Measurement process
• Only the ingress and egress node are required to
  participated in the measurement process
• For the purpose of balancing the load among LSP,
  the available bandwidth appears to be a desirable
  metric to measure ? difficult in MATE
• Packet delay is can be a reliably metric
• The probe packet is time-stamped at ingress node
  at time T1 and recorded at the egress node at
  time T2
• Total packet delay (i.e., queueing time,
  propagation time, and processing time) - T2-T1
  Td
• A group of packet delay ET2-T1Td
• The reliability of the estimator can be evaluated
  by bootstrapping
• The value Td is not required when only the
  marginal delay is needed
• Clock synchronization is not necessary

Clock difference
16
Traffic measurement and analysis - 2

• Packet loss probability is another metric that
  can be estimated by a group of probe packet
• Estimated by encoding a sequence number in the
  probe packet to notify the egress node how many
  probe packets have been transmitted by the
  ingress node, and another field in the probe
  packet to indicate how many probe packets have
  been received by the egress node
• When a probe packet returns, the ingress node is
  able to estimate the one-way packet loss
  probability

17
Traffic measurement and analysis - 3

• Bootstrap
• The bootstrap is a powerful technique for
  assessing the accuracy of a parameter estimator
  in situation where conventional technique are not
  valid
• Most other techniques assumes that the size of
  the available set of sample space value is
  sufficiently large( e.g Central limit theorem
  asymptotic results)
• In MATE, we can use bootstrap to obtain reliable
  estimates of the congestion measures of the mean
  delay and cell loss rate from a given set of of
  measurement obtained via the probe packets
• By selecting a desirable confidence interval, we
  get a dynamic way of specifying the number of
  observation

18
Traffic measurement and analysis - 4

• A basic procedure for computing a confidence
  interval

19
Experimental methodology - 1

• Objective of simulation
• Within a network that has multiple LSPs between
  some ingress and egress nodes, traffic
  distribution under the MATE algorithm is stable
  and load balancing is achieved
• Networking Environment
• Traffic conditions vary due to changes in
  network load (link utilization)
• Traffic conditions vary due to correlation and
  dependencies
• Two types of traffic
• Engineered traffic needs to be balanced
• Cross traffic background traffic that we have
  no control over
• Traffic model
• Poisson shot-range dependencies
• DAP(p) process (discrete autoregressive process
  of order p) large degree of dependencies (our
  experiments, p 10)

20
Experimental methodology - 2

• Two network topology
• A single ingress-egress pair connected by
  multiple LSPs
• Multiple ingress-egress pairs where some links
  are shared among the LSPs from different pairs
• A considerable interaction between the pairs

21
Experimental methodology - 3

• Two implementation
• A small random delay is introduced before the
  algorithm moves from the monitoring phase to the
  traffic engineering phase upon detection of
  change in traffic conditions
• This damping mechanism reduces synchronization
  among multiple ingress nodes
• a coordination among the ingress nodes so that
  only one ingress node at a time enters the
  traffic engineering phase
• This obviously requires a special coordination
  protocol

22
Simulation results - 1
Unbalance situation one heavily congested LSP
and five lightly loaded LSPs Successfully reduce
the engineered traffic from the overload link and
distribute them to the underutilized links
The loss rate on the first LSP dropped from 40
to a value that is too small to observe
23
Simulation results - 2
24
Simulation results - 3

• Engineered traffic
• Travel from the ingress node to the egress node
• The probe traffic required in the each phase of
  the algorithm is around 0.5 of the engineered
  traffic
• Cross traffic
• Enters through the intermediate nodes and exit at
  the egress nodes
• In order to balance traffic, the algorithms must
  shift traffic

25
Simulation results - 4
26
Simulation results - 5
27
Simulation results - 6
28
Conclusions - 1

• Our focus on this paper
• Apply adaptive traffic engineering to utilize
  network resource more efficiently and minimize
  congestion
• MATE
• Distributed adaptive load balancing algorithms
• No new H/W or protocol requirement in
  intermediate node
• No knowledge of traffic demand is required
• No clock synchronization between two nodes

29
Conclusions - 2

• Simulation results
• MATE can effectively remove traffic imbalances
  among that may occur among multiple LSPs
• High packet loss rates can be significantly
  reduced by properly shifting some traffic to less
  loaded LSPs
• Our algorithms have good stability and
  convergence properties
• Future work
• Consider more realistic networking environment
• Examine the impact of MATE on the application
  level

30
Appendix - MATE stability

• Present analytical model of MATE and prove
  stability and optimality
• Model

S1,2,3,s A set of IE pair Psa set of LSPs
(Ps? 2L) L1,2,3,l a set of unidirectional
links xsp amount of input traffic on LSP
p?Ps rstotal input traffic xs(xsp, p?Ps) rate
vector of s x(xsp, p?Ps,s ?S) the vector of all
rates
31
Optimization - 1

• The flow on a link l has a rate that is sum of
  source rates on all LSPs that traverse link l

• Objective(Convex cost-linear constraints)

Total cost
32
Optimization - 2

• Derivate of the objective functions with respect
  to xsp

Cl(xl) the first derivate length of link
l ?C/?Xsp the (first derivate) length of LSP p

• Therom1(Result from Kuhn-Tucker theorem)
• the rate vector x is optimal if and only if,
  for each pair s , all LSPs have minimum(and
  equal) first derivate lengths

33
Asynchronous algorithm - 1

• Gradient Projection Algorithms
• ? step size, xs(t) (xsp(t), p?Ps) ss rate
  vector at time t
• ?Cs(t)(?C/?Xsp(x(t)), p?Ps )the vector of first
  derivate length of LSP p
• Not Realistic
• All updates are synchronized
• Zero feedback delay

34
Asynchronous algorithm - 2

• The new rates calculated by the IE pairs may be
  reflected in the link flows after certain delays
  the flow rate available at link l time t and is
  an weighted average of past sources rates xsp(t)

35
Asynchronous algorithm - 3

• The weight alsp(t,t) depend on (l,s,p,t)
• Latest data only
• Latest average
• IE pair s estimates the first derivate length of
  LSP by asynchronously collection a certain number
  of measurements(using probe packets)

36
Q A - Discussion

• Traffic engineering using multipath in
  GMPLS-based optical network
• Objective
• Optical resource optimization (load balancing)
• Reduce blocking probability
• Fault tolerance
• Control plane (Routing Signaling)
• Traffic aggregation and classification
  (flow-level control)
• Constrained-based explicit route setup (consider
  multipath)
• What kind of constraints ?
• Maximum hop count constraint
• Maximum path count constraint
• Node/link affinity constraint
• Path selection policy