More on BGP

1
More on BGP

• Check out the links on politics ICANN and net
  neutrality
• To read for next time
• Path selection big example
• Scaling of BGP

2
Decision Process

• Step 1 each border router
• Selects which of the paths advertised by its
  external neighbor can be imported
• May violate import policy
• May have routing loop
• Picks the best ones
• Optionally assign a LOCAL_PREF
• Sends its best paths to all its internal
  neighbors
• Step 2 All border routers compare all the paths
  learned through their neighbors and pick the
  single best path to each prefix

3
Path Selection Process

• For each of the paths to the same NLRI
• Ignore if next-hop is down
• Prefer Higher local preference
• If same Prefer locally originated route
• If same Prefer Shortest AS path
• If same Prefer Lowest origin (IGP better that
  EGP)
• If same Prefer Lowest MED
• If same Prefer eBGP over iBGP
• If same Prefer Path with lowest IGP cost to
  next-hop
• If same Prefer Lowest router id
• If same Prefer Lowest neighbor IP address
• Different vendors may have some slight
  variations!
• The RFC does not specify all the details (e.g.
  the preference for shorter AS paths

4
After path selection

• Export routes to external neighbors
• According to export policy
• Add the local AS in the AS_PATH
• Remove LOCAL PREF and optionally add
  MULTI_EXIT_DISC
• Originate routes inside the domain
• Through the appropriate IGP mechanisms
• OSPF has AS-external routes
• And can use the IGP cost to next-hop to select
  which one is better

5
OSPF external routes

• Certain routers in OSPF are ASBRs
• Inject external routes (type-3 LSAs)
• With themselves as the next hops
• OSPF routers compute next hops for these routes
  using the cost to reach the ASBR that originated
  them
• And install these in the RIB
• BUT clearly I can not inject all the routes in
  the Internet into OSPF, it will not scale
• We will see solutions to this later on

6
BGP implementation details

• Uses TCP to talk to neighbors
• Provides reliability
• Important since I have to exchange a lot of
  information
• Allows to send only changes and not all the
  information periodically like EGP
• But TCP has its own mechanisms for flow control
  and connectivity loss detection that may cause
  surprises

7
BGP messages

• Since TCP is a stream protocol I need a way to
  delimit packets
• BGP header
• Messages
• OPEN
• Used to negotiate parameters when establishing
  adjacency
• UPDATE
• Contains the reachability information
• NOTIFICATION
• Send when an error occurs
• KEEPALIVE
• To check the health of the adjacency

8
Paths

• BGP manages paths
• Path consists of
• Network Layer Reachability Information (NLRI) e.g
  12.50.45/24
• A sequence of PATH attributes that give info
  related to this destination
• PATH attributes
• Each have a Flags field
• Optional or well known (well known must be
  supported by all routers)
• Transitive or local (Transitive gets propagated,
  local not)
• Partial or not (partial applies only to part of
  the path)

9
Important path attributes

• ORIGIN (well known)
• Is this path learned from IGP, BGP or other
• AS_PATH
• The list of ASes (well known)
• NEXT_HOP
• Next hop to reach the prefix (well known)
• MULTI_EXIT_DISC (MED)
• Helps selection of paths (local, optional)
• LOCAL_PREF
• Helps selection of paths (well known)

10
Operation

• Establish a TCP connection with the neighbor
• Negotiate options and establish adjacency
• Send all the path info to the neighbor, receive
  its path info and store it
• Then only send updates/changes
• If an update is received that results in a
  different route process it and update neighbors
• Send keep-alives to monitor the health of the
  link
• Does not depend on TCP keep-alive mechanism

11
Prefix aggregation

• Need to be able to aggregate prefixes
• AS_PATH contains a number of AS_SET and/or
  AS_SEQUENCE
• AS_SET unordered set of ASes traversed
• AS_SEQUENCE ordered list of ASes
• When aggregating.
• Longest common prefix of the AS_SEQUENCE becomes
  the aggregated AS_SEQUENCE
• Union of all the rest becomes the new AS_SET
• Can repeat recursively
• EXAMPLE why it is good for the network

12
Scaling

• Route refresh
• Route dampening
• RR

13
Route Refresh

• Each time policy changes I need to find out
  about all the paths from the neighbor
• A path that was not valid before (and was
  discarded) may be valid under the new policy
• I could do a reset (I.e. re-establish adjacency
  with peer to relearn all the paths)
• Expensive and network disruptive
• Instead, initiate a route refresh
• Tell neighbor to resend me all its paths so I can
  re-evaluate policy
• Route Refresh capability, must be supported by
  both peer routers

14
Route Reflectors

• All BRs need to talk iBGP to all others
• Full mesh does not scale
• Use Route Reflectors (RR)
• Multiple RRs in an AS
• A RR has clients and may be connected (through
  iBGP sessions) to other RRs
• Much fewer iBGP sessions
• RR computes paths based on information it
  receives
• If path is learned from client, send it to all
  other clients and RRs
• If path is learned from other RR send it only to
  clients

15
More on RRs

• Essentially split the network into clusters
• One RR per cluster and multiple clients
• All RRs are fully meshed
• Use a cluster-id
• Need to have RR redundancy
• Not a good idea to have two RRs per cluster
• Each client talks to two RRs

16
Extensions to BGP over time

• Communities
• Logically group paths together for easier
  administration
• Can apply policies to a community and not to
  individual paths
• Simpler management, no need to add new policy
  rules for new paths
• Multi-protocol
• Carry different NLRIs
• Ipv4, ipv6, multicast
• VPNs (we will them later)
• Graceful restart (we will see it later)
• 4 bytes ASNs

17
Issues with BGP

• BGP dynamics are complex!
• We will see it through some papers
• Understand the structure of the internet and the
  properties of the topology
• How to measure it?
• Hot potato routing
• Multi-homing and provider selection
• Understand the convergence properties
• Role of policies in convergence
• Policy consistency
• Effect of the various mechanisms used

18
Convergence Dynamics

• 4 types of advertisements
• Withdraw, new, longer, shorter
• Withdraw and longer converge slower than new and
  shorter. Why?
• BGP explores alternate paths in case of withdraw
  and longer
• EXAMPLE
• Worse case can be n! steps
• In theory in fully connected graphs of course
• Also in BGP I can have a timer to collect
  multiple paths before I process them

19
Route Dampening

• Route flaps can be a pain
• A link that is flapping can cause ripples around
  the whole Internet
• Penalize flapping paths
• Keep track of updates per peer and per path
• If it flaps too much stop announcing it
• Parameters additive penalty, half-life, suppress
  threshold, reuse threshold
• Can be different on a per prefix/ per peer basis
  etc
• It was considered a good idea until 2000 or so
• Then people discovered that it was not working
  that well
• A single path withdrawal could appear like a flap
  as it propagates in the network
• Makes the path unavailable for long time (10s of
  minutes)