RE: GMPLS Last Calls

[neil.2.harrison@bt.com]

Hi Dave.....I know exactly what you are saying and its something Andy Reid
and I have chewed over once or twice when considering a functional
architecture description of MPLS.

I was certainly not exhaustive nor precise in my description and
definitions....I was more interested in trying to emphasise the 'no free
lunch' point that making one aspect simpler usually makes some aspect more
complex (normal engineering trade-off).

I see 2 basic forms of MPLS LSPs:

1	One that is a close adjunct function of the IGP.....this is LDP
LSPs, and LDP is required since one cannot easily distribute labels
piggy-backed on a link-state routing protocol like OSPF (though you can with
BGP).

2	One that is not a close adjunct function of the IGP, and requires to
be signalled....this is RSVP-TE and CR-LDP LSPs.

The 'no free lunch' bit here means the former gives an operator more
headaches when considering fault-management and the definition of SLAs for
VPN services for example, whilst potentially getting some gains in
scalability......so that's the trade off I was inferring.

Its a bit like the polarised CO vs CNLS transfer mode view of the
world.....in reality technologies fall some-way between the 2 pure extremes,
and its knowing where the compromises should occur for a given (service)
case.....since neither alone can yield an optimum solution for all cases.
Indeed, one could say this is exactly why we have/need MPLS to augment IP.

Neil
-----Original Message-----
From: David Allan [mailto:dallan@nortelnetworks.com]
Sent: 07 June 2001 16:37
To: neil.2.harrison@bt.com; yakov@juniper.net; mpls@UU.NET;
ccamp@ops.ietf.org
Cc: eric.gray@sandburst.com
Subject: RE: GMPLS Last Calls


Neil, you wrote 
-       LDP induced LSPs do not really create a layer network.....since they

allow merging, and this in itself causes some hard QoS problems....like how 
does one define availability in a mpt-pt tree?  [Its the old issue of 'no 
free lunch'...if you push down a problem in 1 dimension (eg scalability), it

creates a problem somewhere else (eg measuing availability in VPNs say)]; 
-       Explicitly routed LSPs on the other hand do form true layer 
networks....since they need to be signalled they exist as true trail objects

in their own right. 
I'm not sure the existence of merging or QoS is necessarily a criteria for
excluding LDP "induced" LSPs as a layer. Intuitively it strikes me that for
the purposes of MPLS a layer can exist independently of which subset of
topological constructs in the architecture toolkit it subscribes to. ER-LSPs
overlaid on an LDP infrastructre using integrated routing to me would
constitute a useful definition of a "layer". It is only when I ruthlessly
exclude absolutely everything but the elemental  p2p undirectional trail
from the discussion that such a hard definition applies, unfortunatley much
of the value statements around MPLS disappear at roughly the same time.
cheers 
Dave