Re: Aggregating channels...

[Joe Abley <jabley@automagic.org>]

On Sat, Dec 15, 2001 at 09:39:32PM +0100, Henning Brauer wrote:
> On Fri, Dec 14, 2001 at 10:34:47PM +0100, Måns Nilsson wrote:
> > PI space is IMNSHO a requirement if you want to multihome. 
> 
> No, it is possible with PA, too, but requires even more seldom seen clue at
> the uplinks.

   ,--- A
  ME
   `--- B

ME is numbered using a PA netblock delegated from A, for which A
advertises a supernet. If ME advertises that prefix to both A and B,
and A is suppressing covered routes under its supernets in its own
advertisements, then the result will be that the path through A will
act as a backup to the path from the world to ME back through B.

Outbound traffic can be routed through either A or B at ME's
discretion.

In order to balance (share) world->ME traffic through the connections
to both A and B it is necessary for A to leak the covered prefix
coresponding to ME's A-derived PA netblock as well as A's supernet,
so the world sees two paths for the same-length prefix. This is the
additional clue requirement that Henning mentioned.

Some ISPs are intolerant of long-prefix PA blocks being advertised
through other ISPs, and even prohibit such behaviour in their terms
and conditions. The solution in that case is either to acquire PI
space (all three RIRs have protopolicies permitting small PI
delegations to organisations for the purpose of multihoming) or to
number out of PA space delegated from an ISP who is not intolerant
of CIDR-abuse.


Joe