IP Network Traffic Engineering

1
IP Network Traffic Engineering

• Albert Greenberg
• Internet and Networking Systems Research Lab
• ATT Labs - Research Florham Park, NJ

See http//www.research.att.com/jrex/papers/iee
enet00.ps (to appear in IEEE Network Magazine,
special issue on Internet Traffic Engineering,
March 2000). Joint work with Anja Feldmann,
Carsten Lund, Nick Reingold and Jennifer Rexford.
2
IP Network Traffic Engineering

• Goal? In operational IP networks, improve
  performance and make more efficient use of
  network resources, by better matching the
  resources with traffic demands
• How? By integrating
• traffic measurement
• network modeling
• selection and configuration of network management
  and control mechanisms.
• Time Scale? Tens of minutes, hours, days,
• Applications?
• Troubleshooting performance problems.
• Why is this link congested?
• Incremental load balancing
• How to tune intradomain (OSPF, IS-IS) routing
  weights, or interdomain (BGP) import policies?
• Capacity planning and optimization
• How to estimate facilities cost from forecasted
  demands and optimal design?
• Focus of this talk ISP backbone networks
• (See framework draft of new IETF, Traffic
  Engineering Working Group)

3
Traffic Engineering in IP Networks

• Topology
• Connectivity and capacity of routers and links
• Demands
• Expected load between points in the network
• Routing
• Tunable rules for selecting a path for each
  traffic flow
• Performance objective
• Balanced load, low latency, service level
  agreements,
• Question Given the topology and traffic demands
  in an IP network, how do you decide which routes
  to use?

4
Short Answer?

• The desired detailed, up to date, network-wide
  views of the topology are unavailable
• The prevailing traffic demands are unknown
• The network doesnt adapt path selection to the
  load
• The static routes arent necessarily optimized to
  the traffic
• These challenges arise because IP networks are
• Decentralized
• Self-configuring
• Connectionless
• Operating in loose confederation with peers
• Attributes that contributed to success and
  dominance of IP

5
Example Congested Link

• Detecting that a link is congested
• Utilization statistics every five minutes from
  SNMP
• Active probes suffer degraded performance
• Customers complain
• Reasons why the link might be congested
• Increase in demand between some set of
  source-destination pairs
• Failed router/link in our network causes change
  in our routes
• Failure or policy change in another ISP changes
  traffic flow
• How to determine why the link is congested?
• How to relieve the congestion on the link?

6
Long Answer!

• Derive topology from network configuration
  information
• Compute traffic demands from edge measurements
• Model path selection achieved by IP routing
  protocols
• Build a query and visualization environment for
  what-if analysis

7
Toolkit Architecture
Analysis/Visualization
Routing Model
Important to separate models from methods and
data used to populate models
Info Model
Configuration
Measurements
8
Configuration

• Information
• Backbone topology, link capacities, and router
  locations
• Layer 2 and layer 3 links (e.g., ATM PVCs)
• Intra-domain and inter-domain routing (e.g., OSPF
  weights)
• Customer location and IP addresses external IP
  addresses
• Administrative policies, conventions
• Construct
• Unified views of the network topology, and of
  customer and peer reachability
• Main sources router configuration files,
  forwarding tables

9
Measurements

• Performance statistics
• Impact of traffic demands on the network
• delay, loss, throughput from active probes
  between edge systems
• Utilization, loss statistics from passive
  monitoring of links, nodes
• Mapping of statistics onto the network topology
• Traffic Demands
• An accurate view of the demands themselves is
  extremely useful for effective traffic
  engineering
• A large fraction of the traffic is interdomain,
  and a large number of customers are multihomed
• Model traffic demands as loads from an edge
  interface to a set of candidate edge interfaces

10
Information Model

• Abstraction of IP networks
• Different views
• router complexes, router, physical (layer 2),
  abstract (for routing)
• Objects representing
• routers, links, and traffic demands
• Methods for manipulating objects
• finding and selection of objects
• linkage of objects, e.g., router complexes to
  routers
• statistics histogram, tables, etc.
• Salient features
• Captures important global network properties
• Supports routing simulation (e.g., change of OSPF
  weights)
• Trade off between accuracy and simplicity of model

11
Visualization of Link Utilization and Delay in
Backbone
Utilization (from passive measurement) link
color (high to low) Delay (from active probes)
link width (high to low)
12
Routing Model

• Capture selection of shortest paths to/from
  (multihomed) customers and peers splitting of
  traffic across multiple shortest paths
  multiplexing of layer 3 links over layer 2 trunks

13
Routing Model (continued)

• Intradomain (OSPF) routing emulator
• Extract backbone topology and link weights
• Compute all shortest paths (Dijkstras algorithm)
• Split load evenly along all shortest paths
• Emulates Cisco-style use of multiple routes

0.25
0.25
0.5
1.0
0.25
1.0
0.25
0.5
0.5
0.5
14
Visualization of Traffic Flow in Backbone
Color/size of node proportional to traffic to
this router (high to low) Color/size of link
proportional to traffic carried (high to low)
15
Systems

• Configuration
• construction of network topology layer 2, 3
  connectivity, capacity, OSPF weights, customer
  and peer IP addresses, router locations
• Measurements
• Performance (active delay, loss, throughput
  passive link and node utilization)
• Traffic demands
• Information model
• physical level, IP level, router-complex level,
  abstract level
• router attributes, link attributes
• Routing model
• shortest-path routing, OSPF tie-break,
  multi-homing, interdomain routing
• bookkeeping to accumulate traffic load on each
  link
• Visualization/analysis environment
• querying to subselect links and nodes
  histograms what-if capabilities
• coloring and sizing to illustrate link and node
  statistics

16
Key Ideas

• data (configuration, routing, measurement)
  models (topology, demands, routing) analysis
• Generate accurate global views of the network,
  and provide mechanisms to infer network-wide
  implications of changes in traffic, configuration
  and control
• Architectureseparate systems for measurement,
  models, methods to populate models, analysis
• Can and must evolve with change to underlying
  infrastructure and network architecture
• Interfaces for modules above
• E.g., design and optimization (e.g., Bernard
  Fortz and Mikkel Thorup, "Internet Traffic
  Engineering by Optimizing OSPF Weights," Proc.
  IEEE INFOCOM, March 2000. http//www.ieee-infocom.
  org/2000/papers/165.ps)
• (informed) provisioning and reconfiguration
• Closing the loop
• Improving performance and making more efficient
  use of network resources