RE: IETF54- Informal discussion on BC Model for DS-TE

["Ash, Gerald R (Jerry), ALASO" <gash@att.com>]

Francois,

> Russian Dolls model does not "mandate" the use of 
> preemption. It just uses it for what it has been specified 
> for (ie bounce off LSPs when needed).
> 
> My impression is that it is just not possible to simultanesouly :
>          -(i) ensure bandwidth sharing (ie no bandwidth wastage)
>          -(ii) ensure bandwidth isolation (ie a CT cannot 
>               have some of its bandwidth taken by another CT)
>          -(iii) refuse to use preemption
> 
> I believe SPs have requirements for (i) and (ii) and don't have a problem 
> with using preemption, which is an existing TE mechanism.

Not all SPs assume the use of preemption.  So the default BC model should not assume ('require') the use of preemption to operate efficiently.

> I believe the Russian Doll model is a very good way to achieve (i) and 
> (ii).  Yes it involves preemption to do so, but that is exactly why 
> preemption was specified for (ie let some LSPs which arrive later bounce 
> earlier LSPs).

A model like RD that only works efficiently if preemption is used is problematic.  Wai Sum Lai's I-D (http://search.ietf.org/internet-drafts/draft-wlai-tewg-bcmodel-00.txt) has also pointed out some of the other issues encountered by the RD model under overload.
 
> I don't think it would make any sense to sacrifice (i) or (ii) just for the 
> sake of not using preemption.

It is possible to achieve (i) and (ii) *without* using preemption (you don't have to 'sacrifice' anything).  For example, use of dynamic bandwidth reservation together with maximum allocation is one approach (such a 'BC model' has been used in large-scale applications for many years).  

My input to the informal discussion is that a lot more discussion (and analysis) is needed before settling on a default BC model.

Jerry Ash