Re: Infinity
From: Ingvar (ingvar_at_hexapodia.net)
Date: 01/10/05
- Next message: Cameron MacKinnon: "Re: Infinity"
- Previous message: Cameron MacKinnon: "Re: Smug scheme weenies?"
- In reply to: Rahul Jain: "Re: Infinity"
- Next in thread: Cameron MacKinnon: "Re: Infinity"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Date: 10 Jan 2005 10:13:34 +0000
Rahul Jain <rjain@nyct.net> writes:
> Ingvar <ingvar@hexapodia.net> writes:
>
> > Rahul Jain <rjain@nyct.net> writes:
> >
> >> And you would change all your math to use floating point (something
> >> which your fellow advocate does not understand at all) instead of the
> >> far faster integer math just because of this single need... and in one
> >> of the most speed-critical code paths in all of computing?
> >
> > In general, no. I'd probably use something that was sufficiently
> > outside normal expected values to represent "infinity" (if we're
> > using unsigned 32-bit arithmetic and expect normal path costs to be in
> > the range of 30-40, perhaps using 2^31 as an "infinity" is safe
> > enough). Thus, should teh rest of the routing mesh be sufficiently
> > brokwen that we reach a *real* path cost of multiple orders of
> > magnitude larger than normal, we'd start considering our
> > previously-uncosidered route due to the sheer cost of teh other paths.
>
> Hmm... that would be a rather strange situation indeed, and I'm not sure
> that you'd want to use a route that was explicitly asked to not be used
> in that case. There may be security/trust issues involved.
Backup routes would be a typical case. Though they're usually
implemented with interface watchdogs (interface A goes down, this
brings up interface B; as long as A is available, we don't want
*anything to go across B), so it may not apply.
> > And, these days, I am not actually sure FP math is much slower than
> > integer math.
>
> If 7 times is "not much", ok.
Oh? Fair enough, I thought the difference was smaller, these days.
> > The expensive thing is doing route lookups and that
> > doesn't involve running a routing protocol, it just involves looking
> > at a ready-prepared FIB (it's usually a merge of the results from
> > several routing protocols), so we're only doing routing calculations
> > in the "slow path" anyway.
>
> Hmm... but you'd need to compute routing costs via all routes for each
> different destination, right? Wouldn't it be different for different
> destinations and possibly for different ports for different
> destinations? (Say, if you wanted to keep, latency-sensitive
> communications to a specific network on a high-cost-per-byte link.)
It depends on the exact method and protocol used. For a link-state
protocol, each node will be doing a spanning-tree search to find the
next hop for all destinations. For a distance-vector protocol, you
only have to compare what you get from your immediate neighbours and
send on information based on the best route you have received.
> > Using an "infinity" value isn't about
> > speeding up the run-time, it's to simplify implementation.
>
> Wouldn't something like (throw 'compute-cost nil) be more sensible?
>
> (let ((best-cost t)
> (best-route nil))
> (dolist (route routes)
> (let ((cost (catch 'compute-cost
> (compute-cost dest route))))
> (when (and cost
> (or (not best-route) (< cost best-cost)))
> (setq best-cost cost
> best-route route)))))
>
> Not sure how that's particularly complex to write.
Now generalise that to a wire protocol... You'd need either
type-tagged values or designate a numeric value as your "nil".
> > The unfortunate protocol I was alluding to, where this horizon is on
> > the small side is (of course) RIP, where "infinity" is 15, making it
> > hard using path cost to affect route selection, even in not-very-large
> > networks.
>
> Fun. Good thing we double RIP in civilized societies. ;)
Aye, though the new RIP isn't (much) better than the old RIP (it has
explicit netmasks and this is a Good, though it still has an amazingly
narrow horizon).
//Ingvar
-- Coffee: Nectar of gods / Aesir-mead of sysadmins / Elixir of life
- Next message: Cameron MacKinnon: "Re: Infinity"
- Previous message: Cameron MacKinnon: "Re: Smug scheme weenies?"
- In reply to: Rahul Jain: "Re: Infinity"
- Next in thread: Cameron MacKinnon: "Re: Infinity"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Relevant Pages
|
|
