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

[Phil Bedard]

We've looked at collapsing layers every once in awhile and do have some
smaller locations with a combined pair of P/PE devices.  The ability to
eliminate interconnects is attractive.  There are many more features we
use on the PE side which can introduce bugs into the core layer so the
ability to have transit traffic not be affected by those bugs is almost
necessary in our case.  Having a separate core also sets us up for some of
the devices coming in the future which may take the place of a full
featured MPLS router with something with a simplified control/forwarding
plane (stripped-down LSR) or a device with combined packet/optical which
can take the place of both our P nodes and DWDM transport nodes.  I'd like
to see more virtualization to be able to use the same chassis/line cards
as two separate routers and the ability to virtually connect them via an
internal fabric.  I know this is doable on the multi-chassis systems from
C/J but those systems are expensive and add additional complexity.

Right now the reason collapsing is a non-starter is we don't have the
density on any current MPLS router to use a combined P/PE in most
locations unless we go to multi-chassis configurations.

Phil 

On 10/20/11 12:08 PM, "Keegan Holley" <keegan.holley at sungard.com> wrote:

>2011/10/20 Pavel Lunin <plunin at senetsy.ru>
>
>> Folks, let me make a tiny semi-philosophical summarization on this :)
>>
>> Any layer of hierarchy in packet networks gives you another chance to
>> benefit from statistical multiplexing (trick, because of which packet
>> switching rules). The more flows from individual users/subscribers you
>> aggregate in a single link/box, the less dispersion of their
>>insensitivity
>> and the denser you pack the packets on a line (the more it's like TDM),
>> thus
>> the bigger the benefits.
>
>
>You could do the same by just adding more links in the same layer.  If you
>have a core layer but only two links to it that is still your bottle neck.
>Also MPLS doesn't do statistical multiplexing or load-balancing the same
>way
>TDM would.  Most flows will follow the same path from point A to point B
>so
>even though it's packet switched it has some of the constraints of circuit
>switched topologies.  Statistical multiplexing within a box or within the
>queues in a box is a different discussion.
>
>
>> In addition if your core serves users from
>> different time-zones, its efficiency will be even higher because of the
>> 'natural TDM' human biology creates.
>>
>
>lol I'm pretty sure people still use their lines at night.  Backup traffic
>has to go somewhere.
>
>>
>> However you won't earn anything if the streams are already well packed
>> (someone has already skimmed the cream off). As Mark has mentioned,
>>there
>> is
>> no point in selling switched vll/vpn/anything to customers who fully
>>fill
>> their last-miles of capacity close to that of your core-links. Well, at
>> least if the number of such customers is statistically significant.
>>This is
>> what L1 multiplexing is invented for.
>>
>
>This depends on your pricing.  Hopefully those customers that are allowed
>to
>saturate the PE links are paying enough to finance bandwidth/box upgrades.
>Hopefully, the financials work out so that profit is made if 100 customers
>fill the pipes or 10 or 2 for that matter.
>_______________________________________________
>cisco-nsp mailing list  cisco-nsp at puck.nether.net
>https://puck.nether.net/mailman/listinfo/cisco-nsp
>archive at http://puck.nether.net/pipermail/cisco-nsp/