The MPLS WG Archive

[Curtis Villamizar <curtis@workhorse.fictitious.org>]

In message <FFFC48AEAA5F7447929F4F0D93FCC12D02B9280C@zcard031.ca.nortel.com>, "
David Allan" writes:
> Curtis:
> 
> Unfortunately we're overlapping. IMHO what has been discussed is the
> snooping and that implementations hash portions of or all of the stack. I'm
> primarily interested in seeing the specific protocol permutations pinned
> down instead of the blanket assume all permutations exist. In particular
> w.r.t the label stack as I am aware of implementations that hash the bottom,
> the top and would be very useful to know how much of the middle. 
> 
> If people are paranoid about publishing, perhaps feeding the information
> through Loa anonymously would be a solution. However in the absence of
> information other that what has appeared on the list, I'd consider the MUST
> support ECMP requirement difficult to meet and would like to have somthing
> to point to. I'd like to be able to know if the solution space has been
> fully explored and the requirement has been met.
> 
> cheers
> Dave


Dave,

Even if everyone described their *current* ECMP implementation in
detail, not everyone would like to be constrained forever to a
particular method of ECMP just to satisfy *ITU* OAM.

If ITU OAM can't work in the presence of ECMP without aprior knowledge
of how the ECMP split is accomplished this remains an issue of
applicability.  ITU OAM is not applicable in that case.

Curtis



> > Dave,
> > 
> > We've described all the ECMP methods in use either explicitly 
> > in this email thread or through reference to RFC2991 and 
> > RFC2992.  At least one implementation will split based on the 
> > bandwidth configured for an RSVP/TE LSP rather than an equal 
> > split.  Otherwise very few details have not been discussed.
> > 
> > Some implementations use only labels.  Some implementations 
> > look for an IP header *if configured to do so* and hash the 
> > IP src/dst, falling back to labels if the header is not IP.  
> > Most implementations (probably all) limit label stack depth 
> > before using hashed labels. Whether the bottom label of the 
> > whole stack is hashed will varry.
> > 
> > You are not going to get the actual hash function itself and 
> > the method of seeding it.  You also won't get a list of 
> > companies and which feature is released in what version and 
> > which line cards support what (that would be a competitive 
> > analysis that is not needed for this sort of technical discussion).
> > 
> > What *relevant* detail are you missing?
> > 
> > Curtis
> > 
>