[c-nsp] Conditional advertise-map

[Heath Jones]

Gert,

You've got me thinking about this whole non-determinism thing and something
just entered my brain.. The debates here and on other threads seem to be
based on the notion that aggregating prefixes in tcam becomes
non-deterministic due to not being able to guarantee future # of tcam
prefixes if the topology changes.

Does that mean that people think routing is completely deterministic now,
and that we can guarantee noone will go and advertise a shitload of /24's?
Or, does it mean that people think its more likely for a topology change to
affect the aggregation in a tcam because the egress interface will change?

It seems probability comes into play here, but my gut feeling is that very
few networks would run into any real issues with tcam aggregation.


Perhaps that is enough from me on the subject - I don't want to spam the
thread if people aren't interested in this stuff..
Please do let me know your thoughts either on here or directly!
Cheers
Heath



On 15 September 2010 18:43, Gert Doering <gert at greenie.muc.de> wrote:

> Hi,
>
> On Wed, Sep 15, 2010 at 06:38:26PM +0100, Heath Jones wrote:
> > I thought you'd say that...
> >
> > There is absolutely *NO* reason why an additional entry needs to occupy
> > space in TCAM in this case.
> > If you have 2 contiguous prefixes that can be directly aggregated to a
> > single prefix on the next subnet boundary, and that share the same
> > next-stage-treatment (egress, queueing etc), they can take up only 1
> spot.
> > If a router isn't preprocessing routing information before it goes to the
> > forwarding plane, then its not efficiently using the expensive
> resources..
>
> Well, the theory agrees with you.  Unfortunately, routers out there in
> the market (and remember, this is *cisco*-nsp) don't do this.
>
> I have asked for it for a number of times, but it's a tricky problem
> (like: you have a /16 that encompasses 256 /24, the RIB barely fits
> into the FIB space, and then the /16 goes away -> 253 extra FIB entries
> needed -> non-deterministic boom) *and* Cisco has no commercial interest
> in spending lots of engineering effort to the goal of selling *less*
> expensive hardware...
>
> gert
>
> --
> USENET is *not* the non-clickable part of WWW!
>                                                           //
> www.muc.de/~gert/
> Gert Doering - Munich, Germany
> gert at greenie.muc.de
> fax: +49-89-35655025
> gert at net.informatik.tu-muenchen.de
>