The MPLS WG Archive

[Curtis Villamizar <curtis@avici.com>]

In message <393445FE.BD4D9876@marconi.com>, David Charlap writes:
> Curtis Villamizar wrote:
> > 
> > I was one of those people.  Doing the IP destination lookup is no
> > longer a limiting factor for modern routers.
> 
> Depends on the speed.  If you up the number and type of interfaces so
> that your fabric has to forward billions or trillions of packets per
> second, you will once again find differences in how the two forwarding
> engines (IP destination vs. label) perform.

The forwarding decision needs to be made independently on each
forwarding card if you want to build a scalable router.  This way the
amount of parallelism grows as the number of interfaces grows.  If the
IP forwarding decision is part of the fabric, that's simply a design
mistake if very large scalability is a design goal.

> Regardless of how fast the chips run, an IP best-match lookup is an
> inherently more complicated operation than the integer-matching lookup
> required by something that forwards based on label (or VC or DLC).

If you can do it in a few clock cycles in an microprogammed engine,
and its faster than the packet arrival rate it doesn't matter.  As
Greg Mirsky already pointed out, its been done in routers (from
multiple sources).

> -- David

Curtis