RE : RE : Last call for draft-ietf-tewg-interarea-mpls-te-req-01.txt

["LE ROUX Jean-Louis FTRD/DAC/LAN" <jeanlouis.leroux@francetelecom.com>]

Hi Tony

Please see inline

Regards,

JL


>-----Message d'origine-----
>De : Tony Li [mailto:tony.li@tony.li] 
>Envoyé : jeudi 27 mai 2004 21:44
>À : LE ROUX Jean-Louis FTRD/DAC/LAN
>Cc : Adrian Farrel; te-wg@ops.ietf.org; ccamp@ops.ietf.org
>Objet : Re: RE : Last call for 
>draft-ietf-tewg-interarea-mpls-te-req-01.txt 
>
>
>>> As an example, I would like to be able to leak detailed topology 
>>> information from my IS-IS L2 database down into an L1 area. 
> I would 
>>> NOT want to leak bandwidth
>>> information.
>>
>> I don't see the gain, as regards inter-area TE,
>> in leaking detailed topology information from L2 into L1 without
>> leaking any bandwidth information...
>>
>
>
>The gain is that the bandwidth information is in continual flux, and 
>leaking
>that just sloshes L1.  Plus, by the time signaling gets there, the 
>information
>will likely be out of date.
>
>Combine this with the fact that the primary reason to use this 
>information
>is to ensure an optimal exit from the L1 area, and there doesn't seem 
>like
>a real need to worry about instantaneous capacity.

But an optimal exit, as regards cost, may not allow to find a feasible e2e path, if there is not enough capacity from
this exit to the destination.

>
>Again, what I'm suggesting is just an example of one overhead vs. 
>optimality tradeoff.
>
>
>>> I would want to leak an abstraction of another L1 area
>>> down into an
>>> L1 area, but I would definitely want that to be an
>>> abstraction, NOT the
>>> full
>>> database.
>>
>> Of course, else you come back to a single area network !
>> But IMHO such abstraction, to be useful, i.e. to allow computing an 
>> inter-area TE path,
>> would require to take into account a large set of constraint 
>> combinations,
>> and this would likely lead to major impact on IGP scalability.
>>
>
>
>You're making the assumption that the head end of the LSP is computing 
>the entire
>path.  
>Again, I'm advocating a somewhat different position (ala 
>Nimrod): the path
>is computed with more refinement as you get closer to the 
>destination.  
>The
>head end might be able to provide an explicit path through the 
>originating L1
>area, and based on its L2 topology information, it might select the 
>exit point
>from that area.  

Based on which information does the Head End select the ABR ?

>computing
>the remainder of the path, and the L2 router would have current 
>information
>about L2 and some topology information about the destination L1 area.  
>It
>might compute a path that transits L2 optimally and also selects the L2
>exit router.  The L2 exit then computes an optimal transit of the L1 
>destination.
>
>By chaining together 3 locally optimal paths, we will NOT 
>achieve global
>optimality, but we will get a good first approximation for a fraction
>of the cost.

Not so sure that such scheme will get a good first approximation for a fraction 
of the cost.
The ABR selected by the Head-End based on its L1 and L2 information may be the worst ABR, as regards the end-to-end path.
Actually you can definitely not control that. Further more such scheme cannot avoid cranckback risks, 
and it seems from your last comment that you want to avoid that...

Let's take the following example
A1, A2, A3, A4 are ABRs
Area0 topology is leaked into Area1 and Area2. 
All metrics are set to 10 except A1-A2 metric that is set to 1
Avaialble bandwidth is 100M on all links except on link A2->R5, A1->A3, and A2->A4 where available bw=50M

 Area1          Area0       Area2
  --------A1--------A2--------
R1         |         |        R5
  --------A3--------A4--------- 

We want to setup a TE-LSP from R1 to R5 with bandwidth 60M
R1 selects an ABR based on Area1 and Area0 topology info. With your scheme it may select A1
Then A1 will select A2, and A2 will find that there is no path to the destination,
Then Crankback on A1 => No path to destination.
Then Crankback on R1 that selects A3....


>
>
>> Not so sure, there are schemes that allow computing an optimal 
>> inter-area path without adding any byte in LSA/LSP...
>>
>
>
>Which would take us back down the crankback path.  No thank you.

I was refering to PCE approaches where the inter-area path is selected thanks to 
a recursive CSPF computation on ABRs. Such schemes allow computing an OPTIMAL path, without any IGP extension and without crankback...


Regards,

JL