Traffic Engineering Beyond MPLS

Traffic Engineering Beyond MPLS
Apricot 2004 Tutorial

February 24, 2004
Kuala Lumpur, Malaysia
Arman Maghbouleh John Evans

Cariden Technologies, Inc. Cisco Systems, Inc.
arman @ cariden.com joevans @ cisco.com
(c) cariden technologies, cisco systems

TE Beyond MPLS Tutorial Apricot 2004 1
Carrier IP Backbone Engineering Models
Simple Dynamic Controlled
• Emphasis on • Emphasis on • Emphasis on

Scalability Smart Network Asset Utilization
• Low Overhead • Service-Aware • Optimize Offline
Protocols Protocols – Static Explicit
– Pure IP – MPLS CSPF MPLS/ATM PVC
– No CoS – Diffserv/–TE
– 50% Upgrade
Simple++
• Pure IP for scalability

• Capacity Planning/TE for QoS (CoS for insurance)
• Metric-Based Offline TE for Control

Goals
• Investigate Assumptions Behind Models
– Dynamic
• Internet traffic is highly variable and bursty.
– Simple
• Capital expenditures not significant.
– Controlled
• Shortest path first protocols do not provide enough
levers of control.
– Simple++
• Smart Network Engineering vs. Smart Networks
• Demonstrate Simple++

Summary
• Traffic Characteristics
– Long term is smooth and predictable
– Uncorrelated microbursts
– High utilization with little delay at high capacities
– Little need for dynamic routing or queue management
• Simple++
– Traffic Matrix (Measure, or Estimate)
– Capacity plan based on failure simulation
– TE without Layer 2 Overlay
• Computer-Aided Metric-Based TE ≈ as Efficient of
Theoretical Optimum (though more scalable)
• Multiple Routes to High Availability
– Fast Reroute
– Fast Convergence

MPLS TE Aspects
• Covered Here
– Efficient Use of Assets
– QoS
– Fast Reroute
• Not Covered Here
(less backbone relevance)
– Admission Control
– Route Pinning

What is Covered
Core IP / MPLS Network
Low Loss/Latency/Jitter High

Availability
Diffserv IP Traffic NSF/

Engineering FRR SSO
Fast IGP BGP

Convergence
Ad Hoc IGP Metric-
Based TE Security
MPLS TE

Agenda
I. Traffic Characterization
II. Traffic Matrices
III. TE Introduction
IV. Metric-Based TE
V. Convergence

Traffic Characterization
• Long Term
(minutes +)
II. Traffic Matrices • Short Term
(milliseconds)
IV. Metric-Based TE
V. Convergence

Traffic Characterization
• Long-Term
100%
– Measured Traffic micro-bursts
• E.g. P95 (day/week)
– Accommodate failure
and growth failure & growth
• Short-Term
– Critical scale for queuing
– Determine over-
measured traffic
provisioning factor that
will prevent queue
buildup against micro-
bursts
0%
24 hours
High- vs. Low-Bandwidth Demands
Cleveland -> Denver Washington D.C. -> Copenhagen

Mean=64Kbps, Max=380Kbps Mean=106Mbps, Max=152Mbps
P95=201Kbps, Std. dev.=66Kbps P95=144Mbps, Std. dev=30Mbps

Variance vs. Bandwidth
• Around 8000 demands

between core routers 1 Mbps
• Relative variance
decreases with increasing
bandwidth [5]
• High-bandwidth demands
seem well-behaved
• 97% of traffic is carried by
the demands larger than 1
Mbps
(20% of the demands!)

Long Term Traffic Summary
• Most traffic carried by (relatively) few big
demands
• Big aggregated demands are well-behaved
(predictable) during the course of a day and
across days
• Little motivation for dynamically changing
routing during the course of a day

Short-term Traffic Characterization
• Investigate burstiness within 5-min intervals
• Critical timescale for queuing, like 1ms or 5ms
• Analyze statistical properties
• Only at specific locations
– Complex setup
– A lot of data

Fiber Tap (Gigabit Ethernet)
Tap
Analyzer

Raw Results
30 sec of data, 1ms scale
• Mean = 950 Mbps

• Max. = 2033 Mbps
• Min. = 509 Mbps
• 95-percentile: 1183 Mbps

• 5-percentile: 737 Mbps
• (around 250 packets per

1ms interval)

Traffic Distribution
Histogram (1ms scale)
• Fits normal probability
distribution very well
(Std. dev. = 138 Mbps)
• No Heavy-Tails
• Suggests small
overprovisioning factor

Autocorrelation, Lag Plot (1ms scale)
• Scatterplot for consecutive

samples
• Are periods of high usage
followed by other periods of
high usage?
• Autocorrelation at 1ms
is 0.13 (=uncorrelated)

Traffic: Summary
• Long Term Traffic Patterns
– Smooth for big (relevant) flows
– Predictable Trends
– Less motivation for dynamic routing
• Millisecond Time Scale
– Uncorrelated
– Not Self-Similar Long-term well-behaved traffic
– Less headroom required for QoS as circuit capacity
increases

Theoretical Models
• M/M/1 • Self-Similar
• Markovian • Traffic is bursty at many or all

– Poisson-process timescales
– Infinite number of sources
• “Scale-invariant burstiness (i.e. self-
similarity) introduces new complexities
• “Circuits can be operated at over 99% into optimization of network performance
utilization, with delay and jitter well and makes the task of providing QoS
below 1ms” [2] [3] together with achieving high utilization
difficult” [4]
• (Various reports: 20%, 35%, …)

Empirical Simulation
• Feed multiplexed sampled traffic data into FIFO queue
• Measure amount of traffic that violates the delay bound
Example: 92% Utilization

Sampled Traffic
126 Mbps
FIFO Queue
Sampled Traffic Fixed Service Rate
206 Mbps 572 Mbps 622 Mbps
Sampled Traffic
Monitor Queuing Delay
240 Mbps

Queuing Simulation: Results
+ 622 Mbps
+ 1000 Mbps

Queuing Simulation Results
• 1 Gbps (Gigabit Ethernet)

– 1-2 ms delay bound for 999 out of 1000 packets
(99.9-percentile):
• 90%-95% maximum utilization
• 622 Mbps (STM-4c/OC-12c)

– 1-2 ms delay bound for 999 out of 1000 packets
(99.9-percentile):
• 85%-90% maximum utilization

Theory vs. Simulation (1Gbps)
- M/M/1 Model
+ Simulation

Multi-hop Queueing
1 hop 2 hops
Avg: 0.23 ms Avg: 0.46 ms
P99.9: 2.02 ms P99.9: 2.68 ms

Multi-hop Queueing (1-8 hops)

Queueing: Summary
• Queueing Simulation:
– 622Mbps, 1Gbps (backbone) links
• overprovisioning percentage in the order of 10% is
required to bound delay/jitter to less than 1-2 ms
– Lower speeds (≤155Mpbs)
• overprovisioning factor is significant,
– Higher speeds (2.5G/10G)
• overprovisioning factor becomes very small
• P99.9 multi-hop delay/jitter is not additive

Role of Backbone CoS
• Insurance for Issues Beyond Planning
– Denial of Service Attacks
– Catastrophic Failure
(e.g., earthquake, terrorist attack)
• Traffic Separation Under Massive Load

– Coarse-grained service types
– ATM-style queue management not necessary with
high speed links
• (See example in the demo section)

COS Example

Worst-Case Failure per Class
Business
Internet
Voice

Traffic Characterization Summary
• Long Term Traffic Patterns
– Smooth for big (relevant) flows
– Predictable Trends
• Millisecond Time Scale
– Uncorrelated
– Not Self-Similar
• High Utilization, Little Delay
on High Speed Backbone Links
• QoS via Capacity Planning
– CoS insurance for failure of capacity planning/TE

Traffic Matrices
II. Traffic Matrices • Measurement Methods

• Estimation Methods
IV. Metric-Based TE
V. Convergence

Core traffic matrix
• Options
– Full mesh of TE tunnels and Interface MIB
– NetFlow BGP Next Hop TOS Aggregation
– NetFlow MPLS Aware
– MPLS LSR MIB
– BGP Policy Accounting
– Interface MIB and Estimation

Core traffic matrix
• Full mesh of TE tunnels and Interface MIB
– Tunnel interface stats provide bandwidth usage
between all entry and exit points on core
– Data collected via SNMP from headend Router
– Requires full mesh of TE tunnels
– No support for per-CoS routing into tunnels yet

Core traffic matrix
• NetFlow
– MPLS aware Netflow
• Provides flow statistics per MPLS and IP packets
• FEC implicitly maps to BGP next hop / egress PE
– NetFlow BGP Next Hop TOS Aggregation
• v9 includes accounting based upon BGP next hop
NetFlow
• MPLS LSR MIB
– MPLS-LSR-MIB mirrors the Label Forwarding
Information Base (LFIB)
– FEC implicitly maps to BGP next hop / egress PE

Core traffic matrix
• BGP Policy Accounting
– Allows accounting for IP traffic differentially by
assigning counters based on:
• BGP community-list (included extended)
• AS number
• AS-path
• destination IP address
• For more details on above methods see:
– Benoit Claise, Traffic Matrix: State of the Art of Cisco
Platforms, Intimate 2003 Workshop in Paris, June
2003, http://www.employees.org/~bclaise/

Demand Estimation
• Problem:
– Estimate point-to-point demands from measured link
loads
• Network Tomography
– Y. Vardi, 1996
– Similar to: Seismology, MRI scan, etc.
• Underdetermined system:
– N nodes in the network
– O(N) links utilizations (known)
– O(N2) demands (unknown)

Example
A B
6 Mbps
D C
y: link utilizations
A: routing matrix
x: point-to-point demands
Solve: y = Ax -> In this example: 6 = AB + AC

Example
Solve: y = Ax -> In this example: 6 = AB + AC

Additional information
6 Mbps E.g. Gravity Model (every
source sends the same percentage
as all other sources of it's total
traffic to a certain destination)
AB
Example: Total traffic sourced
at Site A is 50Mbps.
Site B sinks 2% of total
network traffic, C sinks 8%.
0
0 AC 6 Mbps AB = 1 Mbps and AC = 4 Mbps
Final Estimate: AB = 1.5 Mbps and AC = 4.5 Mbps

Real Network: Estimated Demands
Cariden
Demand
Deduction
Tool
GBLX
Network

Estimated Link Utilizations!
Cariden
Demand
Deduction
Tool
GBLX
Network

AT&T Labs Procedure
• NANOG 29: “How to Compute Accurate Traffic

Matrices for Your Network in Seconds”
– Implemented on AT&T IP backbone (AS 7018)
– Hourly traffic matrices for > 1 year (in secs)
– Used in reliability analysis, capacity planning, TE

Demand Estimation Results
• Individual demands:
– Can be inaccurate.
• Estimated worst-case link utilizations:
– Accurate!
• Explanation:
– Multiple demands on the same path
indistinguishable, but their sum is known
– If these demands fail-over to the same alternative
path, the resulting link utilizations will be correct

Traffic Matrix Summary
• Existing Options
– MPLS
– Netflow
• New Options
– Netflow BGP Next Hop Aggregation
– Estimation Based on Link Utilization
• Individual Demand Estimation can be
inaccurate
• Estimated Link Utilizations very Accurate

TE Introduction
• Objectives
• Payback
IV. Metric-Based TE • Limitations
• Relation to
Network Design
V. Convergence

IGP Traffic Engineering
• Manipulate Internal Routing
– SPF Metrics (OSPF/IS-IS Metrics/Costs/Weights)
– Explicit Routes
• Minimize Maximum Utilization

– Normal (Non-Failure) Conditions
– Single-Element Failure Conditions (typical)
+ Latency, Policy Constraints
• Given
– Topology
– Source-Destination Traffic Matrix

Strategic versus Tactical
• Strategic TE (focus of this presentation)
– Aimed at $ Savings
– Medium Term Engineering/Planning Process
– Configure in Anticipation of Failures, Traffic Changes
• Resilient Metrics, or
• Primary and Secondary Disjoint Paths, or
• Dynamic Tunnels, or …
• Tactical TE
– Aimed at Fixing Problems
– Short Term Operational/Engineering Process
– Configure in Response to Failures, Traffic Changes

Strategic TE Payback
Without TE With TE
• Real Example
– Delay 6 OC-192 Circuits for a year
(17 circuits under 50% upgrade policy)
– Capital + Operational Savings ≈ $1M/OC-192/year

TE Limitations
• Cannot Create Capacity
– Bottlenecks need capacity not TE
• Limited by Topology
– E.g., V-O-V topologies allow no Strategic TE
Only two directions in each “V” or “O” region
One taken under normal, other under failure
No routing choice for minimizing failure utilization

TE versus Design Diagnostic
• Proxy for Optimal $/bit Calculation
• Calculate Maximum Link Utilization
Current Routing Multicommodity
Flow
No Failure A C
Worst-Case
B D
Failure
• C/D ≈ 1/2 -> Design Limits Efficiency

C/D ≈ 3/4 -> Efficient Design
• A»C or B»D -> Inefficient Routing

A≈C or B≈D -> Efficient Routing

Metric-Based TE
IV. Metric-Based TE
• Case Study
• Performance Evaluation
V. Convergence • Comparison to MPLS TE

Case Study
• Proposed OC-192
U.S. Backbone
• Connect Existing
Regional
Networks
• Anonymized
(by permission)
• Live Demo
(Some Stills)

Plot Legend
• Squares ~ Sites (PoPs)
• Routers in Detail Pane
(not shown here)
• Lines ~ Physical Links
– Thickness ~ Speed
– Color ~ Utilization
• Yellow ≥ 50%
• Red ≥ 100%
• Arrows ~ Routes
– Solid ~ Normal
– Dashed ~ Under Failure
• X ~ Failure Location

Traffic Overview
• Major Sinks in
the Northeast
• Major Sources
in CHI, BOS,
WAS, SF
• Congestion
Even with
No Failure

Manual Attempt at Metric TE
• Shift Traffic
from Congested
North
• Under Failure
traffic shifted
back North

Worst Case Failure View
• Enumerate
Failures
• Display Worst
Case Utilization
per Link
• Links may be
under Different
Failure Scenarios
• Central Ring+
Northeast Require
Upgrade

Cariden Metric TE
• Change 16
metrics
• Remove
congestion
– Normal
(121% -> 72%)
– Worst case
link failure
(131% -> 86%)

New Routing Visualization
• ECMP in
congested
region
• Shift traffic to
outer circuits
• Share backup
capacity: outer
circuits fail into
central ones

Metric-Based TE Evaluation
• See 100
NANOG 27 90
(theoretically optimal max utilization)/max utilization

APRICOT ‘04 80
• Study on Real
70
Networks
60
50
• Single Set of 40
Metrics 30
Achieve 20
80-95% of 10
Theoretical 0
Network A Network B Network C Network D Network E Network F US WAN
Best across Delay Based Metrics Optimized Metrics
Demo
Optimized Explicit (Primary + Secondary)
Failures

MPLS TE
• MPLS Traffic Engineering gives us an “explicit” routing

capability (a.k.a. “source routing”) at Layer 3
– Lets you use paths other than IGP shortest path
– Allows unequal-cost load sharing
• MPLS TE label switched paths (termed “traffic
engineering tunnels”) are used to steer traffic through
the network

MPLS TE Components – Refresher
• Resource / policy information distribution
• Constraint based path computation
• RSVP for tunnel signaling
• Link admission control
• LSP establishment
• TE tunnel control and maintenance
• Assign traffic to tunnels
TE Beyond MPLS Tutorial Apricot 2004

MPLS TE Components (1)
R1
R4
R7
R8
R2 R3
R5 R6
• Resource / policy information distribution

– OSPF / IS-IS extensions are used to advertise “unreserved
capacity” and administrative attributes per link

R1
R4
R7
R8
R2 R3
R5 R6
• Constraint based path computation

– Constraints (required bandwidth and policy) are specified
for a TE “tunnel”
– Constraint based routing – PCALC on head-end routers
calculates best path that satisfies constraints based upon
the received topology and policy information
• prune unsuitable links from the topology and pick shortest
path on the remaining topology
R1
R4
PAT
H R7
PA
H
PAT
HT
PAT
H R8
R2 R3
R5 R6
• RSVP for Tunnel Signaling

– Output of constraint based routing is an explicit route used
by RSVP (with extensions) for tunnel signaling
• ERO = R1->R3->R4->R7->R8

R1
Admission R4
Control PAT Admission
H R7 Control
PA
H
PAT
HT
PAT
Admission H R8
Control
R2 R3
R5 R6
• Link admission control

– At each hop – determines if resources are available
• If Admission Control fails, send PathError
• May tear down (existing) TE LSPs with a lower priority
• Triggers IGP information distribution when resource
thresholds are crossed

R1
R4
PAT
H R7
PA
H
PAT
TH
RE
RES PAT
V V H R8
RES
SV
Use label 30 RES

V
R2 R3 Use label 4 Use label 12
POP
R5 R6
• LSP Establishment
– RESV confirms bandwidth reservation and distributes
labels
• downstream on demand label allocation
– MPLS used for forwarding – overcomes issues of IP
destination based forwarding
R1
R4
PAT
H
PA H
R7
T PAT
RE
H RES PAT
V V H R8
RES
SV
RES
V
R2 R3
R5 R6
• TE tunnel control and maintenance

– Periodic RSVP PATH/RESV messages maintain tunnels

R1
R4
R7
R8
R2 R3
R5 R6
• Assign traffic to tunnels

– Head-end routers assign traffic to tunnels using:
• Static routing, Autoroute or PBR

MPLS TE Components: Minimum Config
(config-if)# mpls traffic-eng tunnels
(config-if)# ip rsvp bandwidth 150000 150000
(config)# router ospf 1
R1 (config-router)# mpls traffic-eng area 0
R4
R7
R8
R2 R3
R5 R6
(config)# interface tunnel 1
(config-if)# ip unnumbered Loopback0
(config-if)# tunnel destination 24.1.1.1
(config-if)# tunnel mode mpls traffic-eng
(config-if)# tunnel mpls traffic-eng priority 0 0
(config-if)# tunnel mpls traffic-eng path-option 1 dynamic
(config-if)# tunnel mpls traffic-eng autoroute announce

MPLS TE Deployment Strategies
MPLS TE
Ad hoc:
Systematic: Few TE tunnels set up to
All traffic transported move a subset of traffic
using TE tunnels away from congested links
Hierarchical
Full Core
or Regional Tunnels paths
mesh mesh
mesh typically static and
determined offline
Can be static (offline) or
dynamic (online)

Systematic Deployment: Full Mesh
• Requires n * (n-1) tunnels, where n = # of head-ends

• Reality check: largest TE network today has ~100 head-
ends
! ~9,900 tunnels in total
! max 99 tunnels per head-end
! max ~1,500 tunnels per link
• Provisioning burden may be eased with AutoTunnel
Mesh

Systematic Deployment: Core Mesh
• Reduces number of tunnels required

• Can be susceptible to “traffic-sloshing”

Traffic “sloshing”
Tunnel #1
1 1
1 A C E 1
1 1 1 Y
X
1 1
B 1 D 2 F
Tunnel #2
• In normal case:
– For traffic from X ! Y, router X IGP will see best path via
router A
– Tunnel #1 will be sized for X ! Y demand
– If bandwidth is available on all links, Tunnel from A to E
will follow path A ! C ! E

1 1
1 A C E 1
1 1 1 Y
X Tunnel #1
1 1
B 1 D 2 F
Tunnel #2
• In failure of link A-C:

– For traffic from X ! Y, router X IGP will now see best path
via router B
– However, if bandwidth is available, tunnel from A to E will
be re-established over path A ! B ! D ! C ! E
– Tunnel #2 will not be sized for X ! Y demand
– Bandwidth may be set aside on link A ! B for traffic which
is now taking different path

1 1
1 A C E 1
1 1 1 Y
X Tunnel #1
1 1
B 1 D 2 F
Tunnel #2
• Forwarding adjacency could be used to overcome traffic

sloshing
– Normally, a tunnel only influences the FIB of its head-end
• other nodes do not see it
– With Forwarding Adjacency the head-end advertises the
tunnel in its IGP LSP
• Tunnel #1 could always be made preferable over tunnel #2
for traffic from X ! Y

Hierarchical or Regional Mesh

Ad hoc Deployment
OC12
OC48
• Explicit path configured on head-end for each tunnel to

offload traffic from congested links
• Can be useful when faced with:
– Unexpected traffic demands
– Long bandwidth lead-times

MPLS TE deployment considerations
• Systematic (strategic) or ad hoc (tactical)
deployment
• Statically (explicit) or dynamically established
tunnels
– If dynamic – must specify bandwidths for tunnels
• Otherwise defaults to IGP shortest path
– Dynamic tunnels introduce indeterminism
• Can be addressed with explicit tunnels or prioritisation
scheme – higher priority for larger tunnels
• Tunnel sizing and how often to re-optimise?

Tunnel Sizing
• Tunnel sizing is key …
– Needless congestion if actual load exceeds expected
max (even by a little bit)
– Needless tunnel rejection if reservation > actual
• Enough capacity for actual but not for the tunnel
reservation
• Traffic reverts to SPF, which is presumably set for
latency not for traffic distribution
• … as is the relationship of tunnel bandwidth to
QoS
– Actual heuristic will depend upon dynamicism of
tunnel sizing

Tunnel Sizing
• Static (offline) Sizing
– Statically set reservation to percentile of expected
max load (e.g. P95)
– Periodically readjust – not in real time

Tunnel Sizing
• Dynamic (online) Sizing: autobandwidth
– Router automatically adjusts reservation (up or
down) potentially in near real time based on traffic
observed in previous time slot:
1. Monitor the 5 min average counter (as in show
interface command)
2. keep track of the largest 5 min average over a
configurable interval
3. re-adjusting the tunnel bandwidth based upon the
largest 5 min average for that interval
4. After the interval has expired, the largest 5 min
average is cleared (set to 0)
– Tunnel churn if autobandwidth periodicity high
• Tunnels de-establish and establish needlessly during
the day as links fill up
– Tunnel bandwidth not persistent

Pipes, Hoses, and Tunnels
Pipe Services Hose Services
• Point-to-point commodity • Point-to-multipoint

– Defined ICR and ECR commodity
between two specified – Defined ICR and ECR to
points cloud
• TE bandwidth based • TE bandwidth based
upon sold ICR / ECR upon monitored load
• Less Risk of Traffic- • More Risk of Traffic-
Tunnel Size Mismatch Tunnel Size Mismatch
•Always OK to use Offline Explicit or Metric-Based TE

TE Summary
• Strategic TE important to resilience and cost
savings
• Computer-Aided Metric-Based TE is a new
option
• MPLS TE has many deployment considerations
• Metric-Based TE close to theoretical optimum,
even under failure conditions

Convergence
IV. Metric-Based TE
V. Convergence • Fast SPF Convergence

• Fast Reroute

Options for IP Traffic engineering
Core IP / MPLS Network
Low Loss/Latency/Jitter High

Availability
Diffserv IP Traffic NSF/

Engineering FRR SSO
Fast IGP BGP

Convergence
Ad Hoc IGP Metric-
Based TE Security
MPLS TE

IGP fast convergence
• Historical IGP convergence ~ O(10-30s)
– Focus was on stability rather than fast convergence
• Optimisations to IGPs enable reduction in
convergence to <1s for first 500 prefixes in a
well designed backbone
– with no compromise on network stability or
scalability
– where POS links are used - slower for non-POS
• Allows higher availability of service to be
offered across all classes of traffic
• For more details see conference session on
“Fast IGP Convergence”, Wednesday 25
February 16:00-16:30

IGP Fast Convergence
• IGP convergence time depends upon a number
of factors
– Propagation delay – distance from failure detecting
node
– Flooding delay – number of hops from failure
detecting node to rerouting node
– Number of nodes in the network
– Number of prefixes
– Position of prefixes in terms of order of processing
• Hence IGP convergence time is not
deterministic
– Difficult to define a maximum bound for loss of
connectivity

MPLS TE Fast Reroute (FRR)
• If …
– recovery around failures is needed in few 100s of ms
– or time to reroute around a failure needs to be more
deterministic
• Then …
– MPLS TE fast reroute is required
• MPLS TE FRR is faster and more deterministic
than IGP convergence

MPLS TE FRR link/node protection
• FRR uses local detection and protection at the
point of failure
– Use POS for rapid detection
– Fast local protection at the point of failure: in ms
– No dependency on propagation, flooding etc
– Uses a pre-established back-up tunnel to protect all
appropriate tunnels on a link
• Uses nested LSPs (stack of labels) – original LSP
nested within link protection LSP
– Switching entries pre-calculated before failure

MPLS TE FRR link protection
• How to protect Tunnel1
against the failure of the
PE1
red link? P1 Tunnel1 P2 PE2
– LSP restoration will take

2.2.2.2
a few seconds
• Using Fast Re-Route
(FRR) link protection can
ensure restoration in PE3 P3 P4 PE4
<<1s

Resilience Strategy: two pronged
approach
• FRR allows for temporary protection of TE
LSPs affected by a link/node failure, while
their head-end is reoptimizing
– Local detection and protection at POF
• Uses a back-up tunnel to protect all appropriate
tunnels on a link
– Uses nested LSPs (stack of labels) – original LSP nested
within link protection LSP
• Fast—O (100 milliseconds)
• May be sub-optimal
– Path restoration
• Repair made at the head-end
• An optimized long term repair
• Slower—O (seconds)

FRR Refresher (1)
• Tunnel1 is configured
as fast reroutable on PE1 P1 Tunnel 1 P2 PE2
headend (PE1)
–Session_Attribute’s 2.2.2.2
Flag = 0x01 in the
path message
PE3 P3 P4 PE4
(config)# interface Tunnel1

(config-if)# description VOIP_TUNNEL
(config-if)# tunnel mpls traffic-eng bandwidth sub-pool 10000
(config-if)# tunnel mpls traffic-eng path-option 1 dynamic
(config-if)# tunnel mpls traffic-eng fast-reroute

FRR Refresher (2): Configuration
• Explicitly routed back-up

PE1 PE2
P1 Tunnel1
P2 Tunnel99 is configured on P1
to P2 via P4
2.2.2.2
• No “tunnel mpls traffic-eng
Tunnel99 autoroute announce” !
–The back-up tunnel MUST
PE3 P3 P4 PE4 only be used when a failure
occurs
(config)# interface Tunnel99
(config-if)# tunnel mpls traffic-eng bandwidth 10000
(config-if)# tunnel mpls traffic-eng path-option 1 explicit name tu99
(config-if)# exit
(config-cfg-ip-expl-path)# ip explicit-path name tu99 enable
(config-cfg-ip-expl-path)# next-address 10.0.14.4 ![P4]
(config-cfg-ip-expl-path)# next-address 10.0.42.2 ![P2]

FRR Refresher (3): Configuration
• On P1 configure Tunnel99
to backup valid tunnels
PE1
on P1-P2 link P1 Tunnel1 P2 PE2
2.2.2.2
Tunnel99
PE3 P3 P4 PE4
(config)# interface POS2/0

(config-if)# description Link to P2
(config-if)# ip address 10.0.12.2 255.255.255.252
(config-if)# mpls traffic-eng tunnels
(config-if)# ip rsvp bandwidth 150000 150000 sub-pool 30000
(config-if)# mpls traffic-eng backup-path Tunnel99
(config-if)# pos ais-shut

FRR Refresher (3): before failure
20.20.20.20 27
IP Packet PE1 P2 PE2
P1
Tunnel1
20.20.20.20 20.20.20.20
2.2.2.2
Tunnel99
PE3 P3 P4 PE4
PE1# sh tag for 20.20.20.20

Local Outgoing Prefix Bytes tag Outgoing Next Hop
tag tag or VC or Tunnel Id switched interface
28 27 1.1.1.1/32 0 TU1 point2point

FRR Refresher (4): before failure
20.20.20.20 27 20.20.20.20
IP Packet PE1 20.20.20.20 10
P2 PE2
P1
20.20.20.20 Tunnel1 20.20.20.20
2.2.2.2
Tunnel99
PE3 P3 P4 PE4
P1# sh tag for ...

Local Outgoing Prefix Bytes tag Outgoing Next hop
tag tag or VC or Tunnel Id switched interface
27 10 [T] 1.1.1.1/32 0 POS2/0 point2point
[T] Forwarding through a TSP tunnel.

FRR Refresher (5): after failure
20.20.20.20 27 20.20.20.20
IP Packet PE1 P1 P2 PE2
20.20.20.20
Tunnel1 20.20.20.20
2.2.2.2
20.20.20.20 10 51
Tunnel99
20.20.20.20 10
PE3 P3 P4 PE4
t1. P1-P2 link fails

t2. Data plane: P1 will immediately swap 27 <-> 10 (as before) and
pushes 51 (done for all protected LSPs)
t3. Control Plane registers a link-down event. RSVP PATH_ERR
message sent
t4. P4 will do PHP
t5. P2 receives an identical labelled packet as before
– Global label allocation
MPLS TE FRR
• Rapid local protection
1. Link Failure Notification
– PoS alarm detection in <10ms
2. RP updates LFIB
• Replace a swap by a swap-push
3. LFIB change notified to the linecards
• 1 message covers all the entries that need
modification
4. LFIB rewrite
• In parallel – distributed on all the linecards

FRR – why do it?
• For telephony users:
– If the connectivity is lost for >150ms, a glitch may
be perceived
• 150ms equates to at least 2 lost samples for 50ms
packetisation interval
– If the loss of connectivity lasts for several seconds,
the phone call may be dropped
• Hence FRR required where very tight SLAs are
required
– Allows highest availability of service to be offered for
VoIP class

MPLS TE FRR – deployment scenarios
MPLS TE FRR
Systematic: Ad hoc:
Deployed to provide Deployed only to protect
complete protection key components whose
for the failure of every failures will have a severe
link and/or node impact on services

• Full mesh of TE tunnels
is not needed for systematic
approach PE1 PE2
P1 P2
• Can instead use next-hop
(NH) tunnels on every link 2.2.2.2
– Single hop tunnel on every
link in each direction
– Run autoroute on every P3 P4 PE4
PE3
tunnel
– As tunnels are 1 hop, due to
penultimate hop popping, in normal operation:
• no labels are imposed
• packets are not label switched
• traffic follows the IGP shortest path

• Allows FRR to be used for
link protection without
needing a TE full mesh PE1 PE2
P1 P2
– Recovery time becomes
a function of number of 2.2.2.2
LSPs / prefixes
• Can similarly use next-
next-hop (NNH) tunnels
PE3 P3 P4 PE4
to protect every node
• Allows decisions on need
for TE and FRR to be
independent

MPLS TE FRR – bandwidth protection
• Backup tunnels can
be configured with
non-zero or zero L3’s view
bandwidth R2
• Zero bandwidth
R4
backup tunnels
R1
provide more
efficient use of
resources
R3
Unlikely two failures
– Assuming single will occur at the same
element failures time!

MPLS TE FRR – bandwidth protection
• With zero bandwidth tunnels some local
congestion might occur during rerouting
– Conflict between resource efficiency and tight SLA
guarantees
• Use Diffserv to mitigate this short-term congestion
• Use LSP reoptimization to handle the long-term
congestion
• Simulation/modelling tools may be useful to
figure out more optimal configurations under
different link/node failure scenarios

Convergence Summary
• Number of technologies to increase core
convergence and hence core network
availability
– IGP fast convergence
• Where recovery in < ~1s is acceptable
– MPLS TE FRR
• Where faster recovery or more determinism is required
• Could adopt a hybrid approach
– MPLS TE FRR – to protect key resources or services
such as VoIP
– Fast IGP convergence – for everything else

Summary
• Traffic Characteristics
– Long term is smooth and predictable
– Uncorrelated microbursts
– High utilization with little delay at high capacities
– Little need for dynamic routing or queue management
• Simple++
– Traffic Matrix (Measure, or Estimate)
– Capacity plan based on failure simulation
– TE without Layer 2 Overlay
• Computer-Aided Metric-Based TE ≈ as Efficient of
Theoretical Optimum (though more scalable)
• Multiple Routes to High Availability
– Fast Reroute
– Fast Convergence

Traffic Engineering References
• B. Fortz, J. Rexford, and M. Thorup, “Traffic Engineering With Traditional IP
Routing Protocols” in IEEE Communications Magazine, October 2002.
• D. Lorenz, A. Ordi, D. Raz, and Y. Shavitt, “How good can IP routing be?”,
DIMACS Technical Report 2001-17, May 2001.
• Cariden “IGP Traffic Engineering Case Study”, Cariden Technologies, Inc.,
October 2002.
• B. Fortz and M. Thorup, “Internet traffic engineering by optimizing OSPF
weights” in Proceedings of IEEE INFOCOM, March 2000.
• B. Fortz and M. Thorup, “Optimizing OSPF/IS-IS weights in a changing world”
IEEE Journal on Selected Areas in Communications, volume 20, pp. 756-767,
May 2002.
• L. S. Buriol, M. G. C. Resende, C. C. Ribeiro, and M. Thorup, “A memetic
algorithm for OSPF routing” in Proceedings of the 6th INFORMS Telecom, pp.
187188, 2002.
• M. Ericsson, M. Resende, and P. Pardalos, “A genetic algorithm for the weight
setting problem in OSPF routing” J. Combinatorial Optimization, volume 6, no.
3, pp. 299-333, 2002.
• W. Ben Ameur, N. Michel, E. Gourdin et B. Liau. Routing strategies for IP
networks. Telektronikk, 2/3, pp 145-158, 2001.

Traffic Characterization References
[1] Steve Casner, Cengiz Alaettinoglu and Chia-Chee Kuan,
A Fine-Grained View of High-Performance Networking, NANOG 22
http://www.nanog.org/mtg-0105/casner.html
[2] Chris Liljenstolpe, Design Issues in Next Generation Carrier Networks,
MPLS 2001 Conference
[3] Peter Lothberg, A View of the Future: The IP-Only Internet,
NANOG 22, http://www.nanog.org/mtg-0105/lothberg.html
[4] Zafer Sahinoglu and Sirin Tekinay, On Multimedia Networks:
Self-Similar Traffic and Network Performance, IEEE Communications
Magazine, January 1999
[5] Robert Morris and Dong Lin, Variance of Aggregated WebTraffic, IEEE
INFOCOM 2000, Tel Aviv, March 2000, pages 360-366.
[6] Anna Charny, Jean-Yves Le Boudec, Delay bounds in networks with aggregate scheduling,
April 14 2001.
[7] Thomas Bonald, et al, Statistical Guarantees for Streaming Flows Using Expedited
Forwarding, INFOCOM 2001.
[8] Roberts Traffic Theory and the Internet, IEEE Communications Magazine, January 2001.
[9] Jin Cao, William S. Cleveland, Don X. Sun, A Statistical Model for Allocating Bandwidth to
Best-Effort Internet Traffic, to appear in Statistical Science, 2004
[10] Chuck Fraleigh, Fouad Tobagi, Christophe Diot, Provisioning IP Backbone Networks to
Support Latency Sensitive Traffic, Proc. IEEE INFOCOM 2003, April 2003
[11] Cao, J., W.S. Cleveland, D. Lin, D.X. Sun,Internet Traffic Tends Towards Poisson and
Independent as the Load Increases. In Nonlinear Estimation and Classification, New York,
Springer-Verlag, 2002

Traffic Engineering Beyond MPLS

Uploaded by

Copyright:

Available Formats

Traffic Engineering Beyond MPLS

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Traffic Engineering Beyond MPLS

Uploaded by

Copyright:

Available Formats

Traffic Engineering Beyond MPLS

Apricot 2004 Tutorial

Arman Maghbouleh John Evans

(c) cariden technologies, cisco systems

• Emphasis on • Emphasis on • Emphasis on

• Pure IP for scalability

TE Beyond MPLS Tutorial Apricot 2004 2

TE Beyond MPLS Tutorial Apricot 2004 3

TE Beyond MPLS Tutorial Apricot 2004 4

TE Beyond MPLS Tutorial Apricot 2004 5

Core IP / MPLS Network

Low Loss/Latency/Jitter High

Diffserv IP Traffic NSF/

Fast IGP BGP

TE Beyond MPLS Tutorial Apricot 2004 6

II. Traffic Matrices

TE Beyond MPLS Tutorial Apricot 2004 7

TE Beyond MPLS Tutorial Apricot 2004 8

Cleveland -> Denver Washington D.C. -> Copenhagen

TE Beyond MPLS Tutorial Apricot 2004 10

• Around 8000 demands

TE Beyond MPLS Tutorial Apricot 2004 11

TE Beyond MPLS Tutorial Apricot 2004 12

TE Beyond MPLS Tutorial Apricot 2004 13

TE Beyond MPLS Tutorial Apricot 2004 14

• Mean = 950 Mbps

• 95-percentile: 1183 Mbps

• (around 250 packets per

TE Beyond MPLS Tutorial Apricot 2004 15

TE Beyond MPLS Tutorial Apricot 2004 16

• Scatterplot for consecutive

TE Beyond MPLS Tutorial Apricot 2004 17

TE Beyond MPLS Tutorial Apricot 2004 18

• Markovian • Traffic is bursty at many or all

TE Beyond MPLS Tutorial Apricot 2004 19

Example: 92% Utilization

TE Beyond MPLS Tutorial Apricot 2004 20

TE Beyond MPLS Tutorial Apricot 2004 21

• 1 Gbps (Gigabit Ethernet)

• 622 Mbps (STM-4c/OC-12c)

TE Beyond MPLS Tutorial Apricot 2004 22

TE Beyond MPLS Tutorial Apricot 2004 23

TE Beyond MPLS Tutorial Apricot 2004 24

TE Beyond MPLS Tutorial Apricot 2004 25

• P99.9 multi-hop delay/jitter is not additive

TE Beyond MPLS Tutorial Apricot 2004 26

• Traffic Separation Under Massive Load

• (See example in the demo section)

TE Beyond MPLS Tutorial Apricot 2004 27

TE Beyond MPLS Tutorial Apricot 2004 28

TE Beyond MPLS Tutorial Apricot 2004 29

TE Beyond MPLS Tutorial Apricot 2004 30

II. Traffic Matrices • Measurement Methods

TE Beyond MPLS Tutorial Apricot 2004 31

TE Beyond MPLS Tutorial Apricot 2004 32

TE Beyond MPLS Tutorial Apricot 2004 33

TE Beyond MPLS Tutorial Apricot 2004 34

TE Beyond MPLS Tutorial Apricot 2004 35

TE Beyond MPLS Tutorial Apricot 2004 36