MapDistributionExplicit Routing Architectures

1
Map-Distribution/Explicit RoutingArchitectures

• J. Noel Chiappa
• Midnight sun routing Workshop
• Luleå, Sweden
• 17-19 June, 2002

2
Overview

• Architectural fundamentals
• Classes of routing architecture
• Reason why a different class of routing
  architecture from the one in use today is
  necessary

3
Architectural Fundamentals

• Terminology and Concepts
• Names
• Objects
• Routing Fundamentals
• Maps
• Abstraction hierarchies
• Attributes
• Routing Tables
• What is a Routing Architecture?

4
Terminology

• I am far from thinking that nomenclature is a
  remedy for every defect in art and sciencestill
  I cannot but feel that confusion of terms
  generally springs from, and always leads to,
  confusion of ideas.
• John Louis Petit, Architectural Studies in
  France, 1854

5
Names and Objects

• Must separate
• name for a thing
• object, i.e. the thing in itselfto which those
  names refer
• Must separate
• Form of the name for a thing
• Generic concept of a name for a thing

6
Namespaces

• 1 or more names of types of names per object
• Single namespace used to refer to different
  objects
• Namespaces may or may not have structure

7
Objects

• Intrinsic to discussion of routing architecture
• Network
• Route or path
• Node
• Interface
• Not intrinsic to discussion of routing
  architecture
• Host
• Router
• Switch

8
Names

• Address
• Structured
• Topologically
• Minimizes destination tracked
• Allows quick location on a map
• Refer to either
• Interface
• Topology aggregate
• One namespace 2 referent objectsis a source of
  potential confusion

9
Names contd

• DNS names
• Structured for human usability
• Structure facilitates
• Distribution
• Lookup

10
Maps

• Representations of the fundamental underlying
  topology with which routing has to deal
• In one simple model, terminology of graph theory
  is used
• Nodes open-ended list of attributes
• Arcs uni-directional link which connect nodes
  and have no attributes

11
Maps contd

• Node can represent some part of the physical
  network
• As large as an ISP
• As small as an interface
• Contiguous portion of a graph representing entire
  physical network can be represented as
• Single node
• Set of nodes
• Either of which is an abstraction for that part
  of the network
• Abstractions can be recursive

12
Abstraction hierarchies

• Name for the naming structure that creates a
  nested set of abstractions for parts of a network
• Consists of naming boundaries around parts of
  network
• E.g. the familiar hierarchical addressing
  structure
• Creates hierarchical addresses
• Interfaces
• Topological aggregates

13
Example Physical Topology
14
Example Abstraction Hierarchy

• For any single Physical Topology there can be
  many Abstraction Hierarchies
• Physical Topology and Abstraction Hierarchy are
  different things

X
Y
A
B
15
Attributes

• Nodes characterized by list of optional
  attributese.g.
• Bandwidth
• Delay
• Delay variance
• Error Rate
• Cost
• Allowed Users
• Nodes have required inherent attributes
• Connectivity of node

16
Routing Tables

• A database used to make decisions about where to
  send traffic
• Often a linear array
• Indexed by destination address
• Provides pieces of information
• Outbound interface
• Next entity to send packet to
• Other information used to calculate paths

17
Example Routing Table
18
What is aRouting and Addressing Architecture?

• Addressing architecture includes
• System of naming
• Networks
• Interfaces
• Topology aggregates
• Routing Architecture
• A way of exchanging information as to where named
  things are
• A way of computing and selecting path between
  sets of communicating entities
• A way of causing traffic to take these paths

19
What is aRouting ArchitectureNot?

• It is not simply a protocol, i.e.
• A set of packet formats
• Instructions on how to process items
• It considers the entire question of how the
  network organizes the handling of user traffic

20
Non-Example of Routing Architecture

• BGP
• iBGP
• eBGP
• MP-BGP
• OSPF
• IS-IS
• RIP

21
Example of Routing Architecture

• Entire collection of BGP variants and all IGPs,
  including specifications on their interaction.
• Nimrod
• Islay

22
Classes of routing architecture

• Currently 2 fundamentally different ways of
  classifying routing architectures.
• Path Selection
• Data Types

23
Class Path Selection

• Fully distributed
• Commonly known as Hop by Hop (HbH)
• Each node along path makes completely independent
  decision of path data is to follow
• Explicit
• One node selects entire path
• When source chooses path, commonly known as
  Source routing
• In more general case it is not necessary that
  source include entire route in packet, therefore
  term Explicit (E) is preferred.

24
Spectrum of Path Selection Alternatives

• Spectrum between E and HbH includes
• Hierarchical Explicit One node picks entire path
• But in terms of higher level abstractions, paths
  across those abstractions selected by other
  entities.
• Hierarchical Distributed Exit router based
  routing
• At a given level of abstraction no node more
  responsible for selecting path then any other
  node
• Node at lower level cannot override higher
  layers choice of next high-layer abstraction to
  transit

25
Comparison ofPath Selection Mechanisms

• Hop-by-hop architectures have the following
  disadvantages
• They effectively require global consistency in
  the routing databases to prevent routing loops.
• Explicit architectures have the following
  advantages
• They trivially allow the testing and deployment
  of new path-selection algorithms.
• They can allow the users much more control over
  the path of their traffic.
• They are more naturally immune to loops.
• They can be made much more robust.
• They can be secured against almost all attacks.
• Some optimization problems cannot be solved by
  local optimization algorithms, which means no HbH
  architecture can handle them.

26
Class Data Type

• Routing architecture depends on kind of data
  passed between nodes.
• Destination Vector (DV)
• Node passes sets of (designation, information)
  n-tuples to immediate neighbors
• Information used to build routing tables
• Map distribution (MP)
• Nodes distribute information about immediate
  surrounding through the network
• Information is used to construct topological
  maps.

27
More on DV

• Selection of path from a given destination is
  usually evenly distributed
• For any level of abstraction no node has more
  significant responsibility in selecting the path
• In the course of running the algorithm to select
  the path, intermediate results in that
  computation are passed between independent nodes

28
More on MP

• Selection can be localized or on an a HbH basis,
  but the actual computation is not fully
  distributed
• No intermediate results passed around network
• Computation can be delayed until the path is
  needed
• Computation can be partial
• Path selection algorithm can inspect map directly
  to select path
• Alternatively a classical routing table can be
  built.

29
Comparison of Data Exchanges

• DV architectures have the following advantages
• They require less computing and storage overhead
  than MD architectures.
• MD architectures have the following advantages
• They are much more practical to secure against
  many attacks.
• They can react more quickly to changes (in
  topology, etc.).
• Their stabilization time is not only shorter on
  average, but also much more bounded.
• They require less bandwidth to react to
  significant changes in network topology.
• They allow the incremental deployment of new
  attributes.
• Constraint-based routing (which includes most QOS
  routing) is only practical with MD architectures.
  
• Highly secure routing is only practical with MD
  architectures.

30
Combining Path Selection with Data Type
DV
MD
HbH
Path Selection
E
Data Type
31
QoS Routing and Traffic Engineering?

• QoS Routing means
• selecting paths on multiple constraints,
• using constraints related to service levels (e.g.
  bandwidth, delay), not just on a single simple
  metric like the number of hops.
• Traffic Engineering means
• allocating paths for traffic aggregates in such a
  way that an optimal use is made of network
  resources.

32
Routing Architectures for QoS

• QoS routing cannot be done with classic HbH
  architectures
• Some constraints (e.g. overall delay) cannot be
  calculated in a hop-by-hop fashion.
• Unless path selection decisions are tightly
  coordinated (i.e. no partially-deployed
  attributes), loops can result.
• QoS routing cannot be done with DV architectures
  
• Each combinations of constraints requires a
  separate routing table, leading to combinatorial
  explosion.
• MD-E architectures are the only ones suitable for
  QoS.

33
Load-Based Routing

• QoS can also imply load-based routing, where
  traffic patterns change in response to offered
  load. The increased dynamicity of LBR make it
  more susceptible to oscillatory behaviour
• Distributed computations of DV architectures are
  harder to stabilize.
• Generally, with Explicit architectures, paths are
  fixed, again increasing resistance to oscillatory
  behaviour.
• Overall, MD-E architectures are more suitable for
  LBR.

34
Routing Architectures for TE

• TE is an optimization process, and getting good
  optimization with local traffic placing
  algorithms can be difficult global algorithms
  can do better. There are several different
  aspects of local
• Local knowledge only (i.e. routing tables, as
  opposed to a complete map).
• Local control only (i.e. the ability to select
  the next hop only, not the complete path).
• MD-E architectures, which are capable of being
  more global along both axes, are thus much better
  suited to TE.
• Simply routing each aggregate independently may
  not produce a pattern of traffic flows which the
  available network resources can support.
• For such cases, a centralized allocation
  algorithm, not a distributed one, is needed.
• Such a centralized algorithm must have both
  complete knowledge (i.e. MD) and complete control
  (i.e. Explicit).

35
The Current Internet Routing Architecture

• The current Internet routing architecture has the
  following key characteristics
• Destination Vector type
• Hop-by-hop path selection
• Do not be misled by the deployment of IGP's which
  use MD architectures, such as OSPF and IS-IS, in
  limited areas of the network.
• The overall operation of the current Internet
  routing architecture is DV
• The data passed between AS's consists of routing
  tables, not maps.
• The path across a number of AS's is selected
  piecemeal, not in a unitary fashion. (I.e. if
  downstream ISP Y has two paths to destination
  site D, upstream ISP X can't select which of
  those two it wants to use all it can do is give
  the traffic to Y, which picks which one it will
  use.)

36
Problems With The Current Routing Architecture

• The overall DV-HbH model does not allow a lot of
  desirable features
• User control of paths
• QoS Routing
• Load-Based Routing
• Traffic Engineering
• Not very robust.
• Secured only by extensive configuration, which
  limits the flexibility, especially in response to
  failure.
• Poor tools for doing abstraction, and controlling
  it.
• The simplistic EGP/IGP split.

37
Recent Routing Work

• The Internet engineering community has been
  experimenting with MD/E architectures for some
  time.
• They have not gotten wide-spread deployment, for
  two likely reasons
• They are very different from existing routing
  protocols, particularly the DV-based system which
  people are most familiar with, and with the
  incredible pressure to provide service, it was
  easier to stay with what was familiar, and known
  to work.
• The MD/E systems which were done are more
  complex, because they provide greater
  capabilities, but there was apparently not enough
  need for the capabilities they provide.
• Recently, work in the MPLS community has started
  to explore MD/E architectures (e.g. QOSPF
  together with an LSP setup protocol).
• The work is poorly integrated with the
  internetwork layer.