[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: LC Comments on "considerations" scaling analysis



Hi Eric, working group,

Eric Rosen:
I will just  reiterate my request for  Appendix A to be removed  and for any
recommendations depending upon it to be removed.

Indeed, your email only consists in reiterating the arguments of your previous email on the subject. (Ok, not only, I admit: you also use bigger words to sound more convincing.)

What I see is that:
- your claims that about the analysis in Appendix A are abusive (you still did not produce any quote of something that would be incorrect, and did not address my reply to your claims of incorrectness) - the scalability-related observations of this analysis are used elsewhere in the document among other arguments, not alone, and lead to recommendations that are well balanced between theoretical aspects and practical aspects ("[recommendation] that implementations support both the BGP-based and the full per MPVN PIM peering solutions for PE-PE transmission of customer multicast routing until further operational experience is gained with both solutions")

So, I don't see a rationale to follow your request.

I have a few more precise comments inlined below (I'll take the time of trying to address your different comment, even if I honestly find them quite abusive, and even if experience tells me that you are unlikely to reply).

-Thomas




Eric Rosen:
If you are going to recommend one scheme over another based on a scalability
analysis, then the analysis must  take into account all the scaling factors.
You do not have the option  of saying "I didn't consider memory utilization,
that's out of scope",
(I don't get why you say this, memory utilization due to mVPN routing state maintenance *is* accounted for in the analysis in Appendix A)

or "I didn't consider sparse mode, that's out of scope
(even though it is an important factor in virtualyl every deployment)".
Please let me remind you that SSM is a subset of PIM Sparse Mode, so the scenario through which a scalability comparison is made in Appendix A is certainly "sparse mode" ("sparse", if I may insist, as opposed to "dense"). The subset of sparse-mode being looked at in this scenarios (SPTs) happens to be used in most non-SSM deployments, which is why this scenario certainly is relevant beyond SSM alone.

If you think that you can make a useful contribution to the document by looking at a non-ASM case, then please go ahead (e.g. something that would result in exposing orders of magnitude significantly different that the one we obtained).


Further,  if  you want  to  draw any  conclusions  whatsoever  based on  the
analysis of a small subset of  the scalability issues, you have to show that
that subset contains the bottleneck issues.  If the MVPN bottleneck is PE-CE
PIM, then you can't say "I'm only considering PE-PE PIM, PE-CE PIM is out of
scope".
I am not the one saying that PE-CE PIM is out of scope. The working group charter and the drafts on mVPN being worked on in the WG, do not cover the PE-CE part. The cases where PE-CE scaling is an issue are *not* specific to mVPN, and logically happens to be addressed in the PIM Working group.

Following your claim, we would end up saying that "since CE-PE is the bottleneck, then we don't need anything special on the PE-PE side, and then all options for PE-PE mVPN routing can be thus said to scale equally well, and well... let's say that this is true even for deployments will no particular CE-PE scaling issue". And we would totally miss out the fact that (a) for the PE-PE scaling part we have different options with different scaling characteristics impacting *real* deployment scenarios and (b) that the CE-PE scaling part can find solutions on its own. *That* would be irrelevant.

It also helps if the analysis  uses a measure of "resource utilization" that
has some practical meaning.  The sum of the number of cycles expended by all
routers has no practical  meaning at all; if it did, no  one would ever have
heard of such notions as "distributed computing" or "datagram networks".
I totally agree with you that a total across multiple equipments has not an interesting meaning in itself. For instance, the total amount of time a message is processed, across all equipements, is in O(#PEx#PEs-joined-to-(S,G)) has no meaning *in itself*. It is not a big or small number intrinsically. It is only when you compare "the total with option X" and "the total with option Y" that you get something carrying a meaning.

And... this is what we do: *comparisons* between comparable numbers.
I stand that the comparison done is meaningful.


(feel free to suggest a better measure, but it seems to me that the *comparison* will lead to very similar results)

If  you do  a scalability  analysis without  attending to  any of  this, the
analysis is simply irrelevant
"If,..., if ..."
That makes many "if"s, in your email, for such a strong judgment.


Of course, if you are doing  a purely theoretical analysis, with no intended
practical application, then of course you can focus in on any isolated issue
that you choose.  However, I don't  think the IETF is the proper publication
venue for that.
Oh, big words again...

The analysis in appendix A is not purely theoretical, and your claim on the relevance of the analysis does not appear to me as founded. So, while I don't think I am in a position to discuss of the appropriateness of publication by the IETF of a document that I myself co-edit, let me just remind that there has been consensus in this working group to adopt this document, including Appendix A, and that the only comments on the list on this section where made by you, one week before the end of WG last call (nearly seven month after WG adoption).


A  further  issue  is that  the  analysis  provided  in  Appendix A  of  the
"considerations" draft is not correct even within its limited scope, for all
the reasons  I have detailed  in a prior  message.  These mostly have  to do
with the  failure to  properly understand the  effects of  join suppression.
But I'm not sure it's worthwhile to repeat all those points again as part of
the same last call.
Except that I extensively replied to you to point some mistakes you made, and explain that "join suppression" is actually taken into account in the analysis. It would be worthwhile if you were able to point at something precise that you think is incorrect (ie. quote and show why/how incorrect).