Vinay,
See
also my draft, draft-hummel-mpls-hierarchical-lsp-00.txt:
An
hierarchical LSP (H-LSP) shall concatenate normal LSPs (and/or hierarchical
LSPs of lower hierarchical level) just like an normal LSP
concatenates
physical links. However, it takes some changes in CR-LDP signalling as
to establish hierarchical LSPs:
Explicit Routing: the ingress of the
H-LSP provides the complete sequence of sub H-LSPs to be concatenated (=
change in ER-TLV).
Implicit Routing: the ingress (the
transit) node of the H-LSP needs to send the LSP-ID of the
first (the next) sub H-LSP to be concatenated (=new
TLV).
The messages (LABEL-REQUEST,
LABEL-MAPPING,...) have to be sent THRU "Control-Plane" sub H-LSPs (like thru
tunnels) whose endpoints comply with the
endpoints of the to be concatenated
"User Plane" sub H-LSPs.
The LSP-IDs of
the involved "Control-Plane" sub H-LSPs should be derivable from the LSP-IDs of
the involved "User-Plane" sub H-LSPs.
I stressed that, based
on a small but contiguous partial mesh of LSPs an effective full mesh can be
accomplished by building H-LSPs.
The N-square problem
would completely disappear. Costs, i.e. work for the H-LSPs would only happen
at the rim of the network. Indeed the network would
not
even know that the
H-LSPs exist, as all messages are tunnelled thru the "Control Plane" sub
H-LSPs.
I would appreciate if
more people supported this work on H-LSPs.
NH=> I tried reading
your ID some time ago and I could not make any sense of it....not sure if its
the content, the way its written or just that I am stupid. However, I
also asked a colleague's opinion and he had the same view.
However, I don't
understand what all the fuss is here.....AFAI can tell, this is not news
to trad network operators as we have been creating nested network hierarchies
for years, whose behaviour/manifestation is largely what you are concluding
towards the end of your mail above. Could I suggest a read of G.805 on
client/server relationships (vertical partitioning), and please don't overlook
the *commercial and technical* aspects of (horizontal) partitioning within a
single layer network, ie who owns what. One key observation here is that
unless you own all the layers to, and including, the duct layer network (yes,
its truly a network) you can say nothing with any confidence on
resilience......every layer network inherits (recursively) the immediately
lower layer network's connectivity, and this starts with the duct
connectivity. Corrollary:- Any model that assumes some all-seeing
routing instance across all layers is extremely naive (in a commercial sense)
because no operator is ever going to own all layers to the duct
everywhere.....and as soon as you lease you lose sight of the lower layers,
and I could never imagine any operator would be willing to advertise this sort
of info (even if it was technically feasible, which I doubt).
Conclusion=> so-called peer-models cannot scale beyond a single
wholly-owned-to-the-duct domain....which is somewhat restrictive if you think
about it.
The N^2/2 issue is always
overplayed, and it *never* goes away despite what some say.......if site X
wants to connect to site Y you must have connectivity in some form or other in
*some* dimension.
Further, the *topology* of
layer network N (whatever co technology) is determined by the link connections
in that layer (ie adjacent-node 1 hops)....in turn these link connections in
layer N are created by trails (= LSPs = usually many hops) in layer N+1 (ie
next lower layer network). This relationship recurses all the way to the
duct. Now if you make the trails in each layer network dynamic (ie
set-up on demand), then the topology of layer N cannot exist until the
topology of layer N+1 exists, which in turn cannot exist until the topology of
layer N+2 exists....etc to the duct. The key point here is that such
layered networks can only be stable and commercially viable *if* there is a
significant increase in the trail holding times as one approaches the
duct......and the duct network itself has the longest holding times. We
have done the maths on such models ages ago (can't share them unfortunately)
and offering SVC/BoD L1 services does not create a compelling business case
(understatement), which gets worse as you go lower and lower in the
hierarchy.....but I suspect you don't need me to tell you this, you just have
to look (i) at who is currently buying what/why at L1 (if anything) and
(ii) in our industry's press and note who is going
bust.
However, if you believe I
am overlooking something in the above Heinrich please enlighten
me.
regards,
Neil