[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [RAM] Incremental Deployment of LISP
At 07:19 PM 3/30/2007 +0200, Eliot Lear wrote:
Hi Fred,
You seem to be assuming that the mapping will be driven by
the first packet out the door; I'm not assuming that at all.
For inter-site communications, there will always be an initial
FQDN-to-locator resolution before a session starts, and the
mapping can be done then before any packets are sent.
Sorry, Yes. I had my head in LISP, where really there is no end host
interaction. I think that's a good thing, in as much as it becomes
deployable faster due to a smaller set of processors being touched. But
in a context, say, like HIP, you would want to do that exact resolution
for a HIT.
Eliot
This is beginning to hint at something that I have wondered about
for a while:
It seems to me that there could be cases where a router (perhaps
in a private network, or a CE router, or perhaps even a PE router)
receives some packets that already use a PA address, and some
that are using a PI address and that need to be encapsulated.
For example this might occur because deployment of LISP is not
precisely equally rapid in each private network, or is not equally
rapid in each site within a larger private network.
How is a router supposed to know which packets are using an
address that is available in the global BGP top-level routing table
(such as a PA address), and which packets are using an address
that needs to be mapped?
For those routers that are in the default-free core of the Internet
this might be easy: If you have a route, then use it. Else map.
Of course these routers are probably large enough that the ratio
between the speed of the data plane and the speed of the
control plane is so large (perhaps in many cases a factor of
10,000 or greater) that you don't want them receiving any
control packets specific to LISP, or at least any one control
packet that they do receive better correspond to at least
10,000 data packets, on average. To me this implies that you
want to push the mapping function outside of the default free
core of the Internet.
However, for many routers they will have a default route, and
will not have the full global BGP routing table. I am thinking that
the routers where you might want to do LISP will mostly (or
perhaps entirely?) have a default route for most remote
locations. Thus, if a packet has a destination address that
matches the default route it *might* be a PA address and
*might* be a PI address.
Is the thought to partition the address space? The only other
solution that I have thought of is to have strictly controlled
topology (so that there is a clear boundary where the mapping
is done).
Thanks, Ross
_______________________________________________
RAM mailing list
RAM at iab.org
https://www1.ietf.org/mailman/listinfo/ram