Endtoend Multihoming draftohtae2emultihoming00.txt

1
End-to-end Multihoming-00.txt

• Masataka Ohta
• mohta_at_necom830.hpcl.titech.ac.jp

2
THE PROBLEM

• Full Routing Table Too Large
• because of random IPv4 address allocation
• to be solved with IPv6
• because of Multihoming
• of TLAs should be controlled
• below what?
• below 1,000 seems to be easy
• below 100 is not difficult, hopefully

3
TLI
SLI
Subscribers
Typical Scenario of IPv6 Multihoming
4
1
1
1
1
1
TLI
3
3
2
2
2
SLI
3
8
2
4
5
Subscribers
Number of Prefixes with E2E Multihoming
5
Multihoming

• Typical IPv4 multihoming
• Advertise an address range through multiple (not
  necessarily 2) routes
• Explosion of
• of routing table entries
• of ASes

6
Rest of the Internet
ISP A
ISP B
Multihomed Site
Singly Homed Site
H
7
Rest of the Internet
131.112.0.0/16? 131.113.0.0/16
131.112.0.0/16
ISP A
ISP B
131.113.0.0/16
131.112.0.0/16
131.112.0.0/16
Multihomed Site
Singly Homed Site
H
Multihoming by (Intelligent) Routing
8
Rest of the Internet
131.112.0.0/15
131.112.0.0/16
ISP A
ISP B
131.113.0.0/16
131.112.0.0/16
131.112.0.0/16
Multihomed Site
Singly Homed Site
H
Multihoming by (Intelligent) Routing
9
End to EndMultihoming (1)

• A host has multiple addresses
• Application or transport tries all the
  destination addresses
• Each address range can be aggregated
• No routing table entry explosion
• No AS number explosion

10
End to EndMultihoming (2)

• When to Try Alternative Addresses?
• Application/Transport dependent
• Controlled by the intelligence of end systems
• TCP will have a default timeout period
• Which Address Should be Tried Next?
• Routing table can give hints
• Lack of a routing table entry means lack of
  reachability to a host
• Existence of an aggregated routing table entry
  does not mean reachability to the host
• Metric information in routing table can help too

11
Rest of the Internet
131.0.0.0/8
133.0.0.0/8
ISP A
ISP B
131.113.0.0/16
131.112.0.0/16
133.112.0.0/16
Multihomed Site
Singly Homed Site
H
133.112.32.132, 131.112.32.132
End to End Multihoming
12
Because It IS End to End

• No change to router functionality
• MUST change API on hosts
• Or the hosts are singly homed
• Wrong to assume
• Intelligent routers help dumb nodes
• A host can and is recommended to have a default
  free global (but now small) routing table
• A real dumb host is dumb and singly homed

13
Do Intelligent End SystemsRequire Standard IGP?

• No.
• Standard protocol to distribute (but not compute)
  routing table to hosts is required
• RIPv2 (with metric 15) seems to be good enough.
  Or RA?
• BGP routers generate metric from policy based
  preference

14
Source Address Selection?

• Wrong topic for multihoming
• Both sources select destination addresses
• At the destination, reverse forward DNS lookup
  of source address gives all the address of the
  source
• The destination selects an appropriate
  destination address for reply
• No source address selection by source meaningful
• Or, protocols may be modified to carry them

15
TCP API/Protocol Changes

• Source have no reachability information of SYNACK
  reply addresses at the destination
• Destination should select the address
• Multiple PCB entries for a connection
• How to give the multiple addresses of the source
  to the destination?
• DNS?
• Let SYN carry all the addresses?

16
DNS Changes

• DNS can not give addresses for DNS reply
• Clients should choose source address reachable
  from the name server
• First, choose randomly
• Should (re)try with other source addresses
• A lot of delay
• Or, change protocol?
• Query carries all the addresses of the source

17
End to End Multihomingand DNS/SMTP

• DNS and SMTP servers already deploy E2E
  multihoming
• NS/MX servers may have multiple A records
• If a server has multiple addresses
• All the addresses are tried
• It is of course as the most important,
  required-to-be-rubust applications of the Internet

18
88

• DNS reverse lookup by lower 8 bytes only
• Hosts are identified by lower 8 bytes (IID)
• A compact DNS name carried by all the packets of
  all the protocols
• Makes modification to application/transport
  protocols for E2E multihoming easier
• Not Mike Odells GSE one (violate E2E)
• Teraoka san will present his version (85)