[c-nsp] Downsides of combining P and PE functions into a single box

[Keegan Holley]

2011/10/19 Jared Mauch <jared at puck.nether.net>

> On Wed, Oct 19, 2011 at 12:58:41PM -0400, Keegan Holley wrote:
> > The real question:
> > >
> > > Are you selling customer links that are near to or equal to the size of
> > > your core links(s).
> > >
> >
> > Why would anyone do this on purpose and not upgrade the core?  I
> understand
> > over-subscription but having your edge links the same speed as your core
> is
> > just asking for trouble.
>
>         Because the core link sizes are the same as the edge sizes.
>
>        This means you have to create duplicate links to haul it through
> more
> routers vs less.
>
>
> > > Anyone doing 10GE edge or looking at 100GE for customer-facing handoffs
> can
> > > save significant amounts of money by doing P/PE.  While there are
> tradeoffs,
> > > not having the cumulative cost of a packet being A+B+C and perhaps can
> be
> > > localized to a single device has value.  I'm surprised that Rolland
> doesn't
> > > see this as an optimization as it would be something the Arbor
> equipment
> > > could help you optimize.
> > >
> >
> > Not sure how you save money by buying extra routers.  That's a pretty
> > aggressive discount structure.
>
>         I'm saying buy less routers.
>
>        If your customer is talking to a peer, place them on the
> same device.  Don't have a 'peering edge' vs 'customer edge'.
>
>        It may make sense to terminate your 'core' links on the same
> device as well.  It may not.  This all depends.  The problem here is how
> people think about the network.  "There must be a core", or "you must
> transit a P device".
>

Oh... I think we were saying the same thing here.  It really depends on the
requirements of each individual network.

>
> > > While some may see these cost savings as inelegant, the idea of a core
> will
> > > continue to come under these pressures.  Keep in mind the fraction of a
> > > chassis you must allocate for these edge <-> core links and core <->
> core
> > > links.  These have real world costs.  There's a reason everyone didn't
> go
> > > out there and load-up on OC768 hardware and just stuck with N*10G.  The
> > > finances don't work out.
> > >
> >
> > Cards are cheaper than entire routers in most cases especially at N*10
> and
> > 40G speeds.  Assuming you want chassis based, with redundant control
> planes
> > and whatever the vendor uses for fabirc blades.  I'm not saying everyone
> > should throw their core P routers into a dumpster, but I don't see how
> > having them saves money.  You also have to add the cost of service
> contacts,
> > power, fingers and eyes to keep them running, etc..  I think people who
> need
> > separate cores should have them.  However, I don't see how P routers save
> > money or reduce complexity.
>
>         They don't in many cases.  I think you misunderstood my comments.
>

+1 on the misunderstanding.  My apologies.  I should be working anyway :)

>
>        - Jared
>
> --
> Jared Mauch  | pgp key available via finger from jared at puck.nether.net
> clue++;      | http://puck.nether.net/~jared/  My statements are only
> mine.
>
>