Igor,
I obviously can't speak for all data center people as well, but I
completely agree with your below comments from my own perspective.
As far as pushback to BGP (which I personally like a lot) I could add
that non of my hosts can run BGP code therefor it is a non starter for
the deployment of end to end (physical host to physical host) over L3
virtualization.
Considering now public announcement of OVS being part of linux
distributions I would find it counter productive to propose
virtualization technology which is not compatible with most common
operating systems which resources (example: VMs) we are to really
virtualize.
I do realize that networking people think of network as a layer to
deliver services. This is just like transport folks were 100% convinced
in the past that their layer as critical plane to networking people.
However as we have observed in the past that transport and networking
people have merged now I think we need to face the requirement for
networking side and host side to merge as well to deliver competitive
services.
R.
On Thu, 19 Apr 2012, AshwoodsmithPeter wrote:
::
::>I am very much concerned about this. I know that this point is not
::>shared by all, but for the DC folk I've talked to (and there are
::>others I've talked to that say *exactly* the same thing), MPLS/BGP is
::>simply a non-starter.
::
:: While not a statistical sample by any means I've been on the receiving
:: end of some similar comments from customers but my feeling is that they
:: are not reacting to the data plane but more to a certain implementation
:: of the control plane / usability. That is after all what they see and
:: touch.
I obviously can't speak for all datacenter people, but I can tell you that
the reason for why MPLS is a non-starter for me inside the datacenter is
that my switches there simply don't support MPLS forwarding (and, neither
do my server NICs for that matter), and backwards compatibility with the
(modern) hardware that I have deployed today is a very hard requirement. I
also know for a fact that I'm not alone in that ;)
As far as pushback to BGP, I suspect that it comes from 3 things:
1) You are absolutely right - some of it is religious (but, not much you
can do about that)...
2) quite a bit of it is implementation related - an implementation of a
protocol for convergence across internet-scale, with all the hacks
and optimizations around that particular set of use-cases, may not be the
right thing to use inside the datacenter due to those convergence
properties..
3) (hopefully) a lot of it is coming from the fact that quite a few DC
operators believe that full transactional semantics that can be used by
some form of a guaranteed consistency model is absolutely mandatory for
this mapping system. Since as far as I know, that's not something BGP has,
then it is simply not the right tool for the job, and any number of
distributed database/state synchronization methods are a significantly
better fit, and should be used instead..
Hopefully that will provide some context for the pushback...
Thanks,
-igor
--------------------+----------------------+------------------
Igor Gashinsky | Network Architecture | Yahoo! Inc.
[email protected] | cell 917.807.2213 | Do You... Yahoo?
--------------------+----------------------+------------------
_______________________________________________
nvo3 mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/nvo3
_______________________________________________
nvo3 mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/nvo3
