Summary Route with Detailed Reachability George Swallow, Clarence Filsfils, Stefano Previdi [e-mail] [e-mail] [e-mail]

1
Summary Route with Detailed Reachability
George Swallow, Clarence Filsfils, Stefano
Previdiswallow_at_cisco.com cfilsfil_at_cisco.com
sprevidi_at_cisco.com
IS-IS WG - IETF 70
2
Agenda

• Motivation
• Solution Sketch
• Open Issues
• Encoding technique
• Inconsistent advertisements

3
Motivation

• Scalability and convergence
• IGP convergence
• BGP Convergence
• Example
• L3VPN over L2TPv3
• Other Uses
• PIM
• RSVP-TE mesh

4
Some Brief Notes on IGP Convergence

• Testbed
• CRS1
• 2500 ISIS prefixes
• Tier 1 ISP Topology
• Time measured by traffic loss

Prefix Number

• Time for ISIS LSP generation, SPF recalculation
  is very quick

• Substantial time is required for update of
  structures on linecards for individual prefixes

• Time shown is IP, for MPLS LFIB needs updating too

5
L3VPN over L2TPv3
10.10.0.2
10.10.0.3
10.10.2.3

• VPN packets are encapsulated in L2TPv3
• For many VPNs, multiple next-hops are carried in
  BGP using a Route Distinguisher (RD)
• Switch to new route occurs on BGP withdrawal or
  indication from ISIS that the next-hop is not
  reachable (aka BGP NH tracking)
• To scale IS-IS, operators would like to summarize
  PE loopbacks
• However summarizing hides detailed reachability,
  BGP convergence then depends on BGP withdrawal

6
Reachability and Routing

• Currently IS-IS makes no distinction between
  having a route and having reachability
• We call a route to an IP prefix IP reachability
• As we move toward sophisticated control planes
  and highly efficient forwarding planes, a tension
  has developed

7
Changing Economics and Priorities

• Early 90s
• Expensive memory
• Slow processors
• Applications tolerant of slow convergence
• Summarization considered good for dataplane and
  control plane
• Today
• Memory much cheaper
• Processors much faster
• Applications demand fast convergence
• For convergence fewer routes means faster
  convergence in the dataplane (primarily FIB
  update time)
• But BGP next-hop tracking needs the reachability
  information
• Needs of the control and data planes have
  diverged!

8
Separating Routing and Reachability

• New routing advertisement - SRDR
• Summarized route
• Detailed reachability
• Straw-man format
• Use the Extended IP Reachability TLV
• Add a sub-TLV
• Bit vector of reachable hosts
• Vector length 2(number of ignored bits)

9
Example

• Area 2 has 10.10.2.0/25 assigned as its address
  range
• The following addresses appear in ABR2s database
  for Area 2
• 10.10.2.1 - 10.10.2.27
• 10.10.2.46
• 10.10.2.74 - 10.10.2.87
• then the bit mask encoding would advertise a
  summary route to 10.10.2.0/25 with an associated
  128-bit mask like this
• 0 1 2 3
• 01234567890123456789012345678901
• --------------------------------
• 01111111111111111111111111110000
• 00000000000000100000000000000000
• 00000000001111111111111100000000
• 00000000000000000000000000000000

10
Bit-Vector Characteristics

• Limited to 1024 bits by TLV/sub-TLV encoding
• Fixed size
• Good for memory mgmt
• Good for LSP fragmentation issues
• Cannot exceed allowable sub-TLV size
• Not compact for sparse allocation
• Works well for IPv4 given the assumptions in the
  following case study

11
Bit-Vector Case Study

• Assume up to 20k routers in network
• Break this into 50 domains
• Average of 400 routers / domain
• Assume PE are numbered in blocks of /24 addresses
• Utilized 33 due to admin inefficiency
• Requires 5 /24 per domain 250 total
• Each /24 would need 32 bytes of bit-vector
• 8k bytes total

12
Detailed Reachabilty Encoding

• These assumptions should carry over to IPv6 if
  provides allocate loopbacks from /120 addresses
• Authors would like feedback on the assumptions
  from Service Providers

13
Open Issues

• Encoding
• Encoding scheme not cast in concrete
• One variation would be to have a new TLV instead
  of a sub-TLV of the Extended IP reachability TLV
  which would eliminate the /22 limitation on
  summarized addresses
• Inconsistent advertisements

14
Inconsistent Advertisements
10.10.1.1

• How do ABR1, ABR2 react to inconsistent
  advertisements from ABR3, AB4?
• How does PE1 react to inconsistent advertisements
  from ABR1, ABR2
• If no ECMP, then just relay the selected routes
  information

15
Inconsistent Advertisements

• Should only happen in two cases
• Race condition between L1L2 routers seeing a
  host/router come up or down
• Area partition
• Authors discussed several options
• Leaking /32 of addresses one L1L2 router
  advertises vs the other L1L2 routers
• Partition repair via L2 tunnel
• Analysis is not complete at this time