Maria,
Not sure I follow (or fully agree with) your reasoning here.
I agree that there is at least a small scale advantage in separating
the control and
forwarding planes. However it is less than some might expect.
Take the case where a single control element is able to control 1,000
nodes. At
the same time, those 1,000 nodes are potentially at risk for a failure of that
one control
element. This can be addressed using redundant control elements, but this has
related
complexity costs that will impact both CAPEX and OPEX.
Management complexity should be pretty much the same, whether one does
configuration associated with 1,000 nodes in one, in ten, in a hundred, or in a
thousand
control elements.
In fact, management complexity (among other forms of complexity) can
become
quite "interesting" in a number of scenarios - mostly related to failures in
either control or
data-plane connectivity that produce irregularities and inconsistencies in real
vs. perceived
forwarding connectivity.
In terms of software complexity, a component of complexity relates to
the scale
in which the software operates. Many experienced folks see distributed
simplicity as a
bit better solution than concentrated complexity. The complexity associated
with keeping
control and data plane connectivity in alignment also has a substantial impact
on software
complexity.
In terms of having an "open protocol" between control and data plane
elements,
we have done a lot of this kind of work in the IETF. Do we have a clear
understanding of
the additional work that may be required?
--
Eric
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of
NAPIERALA, MARIA H
Sent: Monday, September 24, 2012 9:39 AM
To: Thomas Nadeau; Yakov Rekhter
Cc: Thomas Nadeau; [email protected]; Stiliadis, Dimitrios (Dimitri); Aldrin Isaac
Subject: Re: [nvo3] draft-drake-nvo3-evpn-control-plane
Tom,
decoupling PE control plane from the forwarding function has many advantages
but mainly it substantially increases operational scale - PE/control element is
able to control multiple (1000+) compute nodes spread across different servers
and other devices. The software complexity (e.g., managing policy functions,
gathering of operational information like stats, events, diagnostics, etc.) is
implemented in the control plane elements only. These reduce overall cost of a
data center deployment.
In addition, having an open protocol between a control plane and a forwarding
plane of a PE allows sending local forwarding rules to forwarding device(s).
XMPP is an open standard, light-weight, extendable (can carry various data
objects), and flexible protocol known to application environment.
Maria
> -----Original Message-----
> From: Thomas Nadeau [mailto:[email protected]]
> Sent: Thursday, September 20, 2012 7:23 PM
> To: NAPIERALA, MARIA H; Yakov Rekhter
> Cc: [email protected]; Stiliadis, Dimitrios (Dimitri); Aldrin Isaac;
> Thomas Nadeau
> Subject: Re: [nvo3] draft-drake-nvo3-evpn-control-plane
>
>
> Maria,
>
> The only issue that is being raised seems to be one of which control
> plane to run, not whether or not we need one. I think everyone agrees
> on that. In the way of BGP versus XMPP, perhaps you could elaborate
> why you think BGP is a bad choice?
>
> --Tom
>
>
> On 9/20/12 8:41 AM, "NAPIERALA, MARIA H" <[email protected]> wrote:
>
> >>
> >> Do you think that an NVE that implements only XMPP has no control
> plane
> >> at all ?
> >
> >It does not implement the (BGP) control plane of the overlay (e.g.,
> route
> >selection should be done on a controller and not on the NVE), and it
> >should not directly participate in any other routing protocols.
> >
> >Maria
> >_______________________________________________
> >nvo3 mailing list
> >[email protected]
> >https://www.ietf.org/mailman/listinfo/nvo3
_______________________________________________
nvo3 mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/nvo3
_______________________________________________
nvo3 mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/nvo3
