Re: draft-ashwood-ccamp-gmpls-constraints

["Igor Bryskin" <ibryskin@movaz.com>]

Don,

You can find a detailed description of the Virtual Link mode in the Layer 1
VPN WG documents.

In brief, in this mode a domain is represented to the outside routing domain
not as a single node (as in case of the Virtual Node mode), but as a set of
PEs interconnected by virtual (more correctly, abstract) links. The Virtual
Link mode has some serious advantages compared to the Virtual Node mode.
Here is some of them:

1. In the Virtual Node mode there is some synchronization required between
PEs: in order for the outside routing domain to perceive the hidden domain
as a single node all PEs need either to generate exactly the same
advertisings ( specifically, they need to agree on the Virtual Node Router
ID, learn about every other PE and the links interconnecting the PE with the
outside routing domain, etc.) or identify the outside routing domain
segments interconnected exclusively by the means of the hidden domain and
elect for each of them a single PE that would generate the Virtual Node
advertisings. Neither of these approaches is trivial to implement. On the
contrary, PEs in the Virtual Link mode advertise information into the
outside routing domain completely independently.

2. In order to advertise a matrix of acceptable input-output link
combinations a PE must periodically solve ALL PEs -TO-ALL PEs constraint
based path computation problem. It is far more difficult problem to solve
compared to a single PE -TO-ALL PEs constraint based path computation (which
is as complex as a single source - single destination path computation)
required in the Virtual Link mode;

3. The constraints used during the computation of the input-output link
matrix is not advertised and not available for the external path computer,
which diminishes the value of the matrix advertising. In other words, even
when the external path computer uses the matrix as a constraint, there is
still a significant blocking probability of the LSP setup using the computed
path because there is no guarantee that the sets of the "external" and
"internal" path computation constraints match. There is no such problem in
the Virtual Link mode where the internal path computation constraints could
be advertised as abstract TE link attributes and hence could be considered
explicitly by the external path computer

4. The matrix of input-output link combinations does not provide information
about the cost of a particular input-output binding across the hidden
domain. This means that suboptimal path selection is quite possible. On the
contrary, each abstract TE link advertising has a TE metric sub-TLV;

5. It is not trivial to use the matrix of input-output link combinations as
a constraint, and some modifications of the external path computation engine
algorithms are required. There is no such a requirement for the Virtual Link
mode.

The major disadvantage of the Virtual Link mode, of course, is scalability:
the number of the abstract links grows proportionally to the square of
number of PEs. Because of that the Virtual Node mode could be the only
choice to hide a domain with large number of PEs.

Hope this helps.
Igor


----- Original Message ----- 
From: "Don Fedyk" <dwfedyk@nortel.com>
To: <ibryskin@movaz.com>
Cc: <ccamp@ops.ietf.org>
Sent: Thursday, August 04, 2005 6:03 AM
Subject: RE: draft-ashwood-ccamp-gmpls-constraints


Igor

Thanks for the feedback I would like to work with you to capture the
Virtual Link Mode into the draft.

Regards,
Don

> -----Original Message-----
> From: ibryskin@movaz.com [mailto:ibryskin@movaz.com]
> Sent: Thursday, August 04, 2005 5:32 AM
>
> Hi,
>
> I believe this is a very useful draft. The described blocking
> problem (a limited ability of a node to cross-connect
> resources on input and output links wrt a particular LSP)
> exists not only in the Virtual Node scenario: there could be
> "real" network elements experiencing the problem (perhaps,
> because of the hardware limitations). Hence there is a need
> for a routing controller to be capable to advertise a map of
> acceptable (or
> unacceptable) input-output link combinations, and for a path
> computer to account for such a constraint (which is not trivial).
>
> I also would suggest the authors to consider the Virtual Link
> mode, that is, representing the domain to the outside world
> as a bunch of PEs interconnected by abstract (virtual) links.
> This approach may require more advertisements compared to the
> Virtual Node mode; however, it does relieve external path
> computers from handling the interface maps, plus it gives the
> idea about the cost and attributes of feasible paths across
> the domain.
>
> Igor