Re: [mpls] Relation between the number of special purpose labels and the depth of the label stack - Was:Re: comments on draft-kompella-mpls-special-purpose-labels-02

[Jeff Wheeler <[e-mail]>]

On Wed, Apr 3, 2013 at 6:45 AM, Loa Andersson <[e-mail]> wrote:
> I want to see separate discussions on separate problems.

There are only two proposals which create different, related problems.
 Discussing them separately is only useful to understand what those
problems are.  I think that's been thoroughly dealt with.  Thus, you
are left with a choice: which problems are the lesser set of evils?
There is no point in handling them separately at this stage.

> The way I read some of the mails in this discussion is that if we
> increase the number of with one, the label stack will be one label
> deeper.

> I seriously doubts this, it is potentially not true even in the
> were the new special purpose label is present in the label stack.

You don't know what future special-purpose labels will be used for.
None of us do.  If one is needed for something like a future iteration
of Entropy Labels, then the cost of forwarding all traffic using that
feature will be increased.  That means the data-plane needs more
power, area, and produces more heat; all of which decrease the useful
forwarding density.

You should only have a doubt about this if you foresee a system that
reserves all the remaining "legacy" special-purpose labels (not
needing 15 prefix) for high traffic-bearing features, and allocates
the extended ones (requiring 15 prefix) to future features that are
unlikely to be included in packets with enough frequency to impact
data-plane performance or density.

You have not proposed such a system yet.  Do you think one is needed?
If so, then every "legacy" value handed out now decreases the number
available for future high-traffic special purpose labels before the
costs get high.  This dramatically changes the urgency of this draft.

> Do we have any hard figures on the relationship between the number
> of available number of special purpose labels and the depth of the
> label stack?

I do not understand your question.

> Same for the number of labels to be processed, do we know which
> effect special purpose labels have on how many labels that need to
> be processed.

Yes, we do know this.  If your proposal is adopted, then every packet
labeled with an extended special-purpose label must process one
additional label.  This is obvious.  Did I misunderstand your question
again?

This is a very simple issue.  Imagine if the [7] Entropy Label
Indicator were, in fact, [15][7].  Then, every packet using Entropy
Labels for hashing would have a larger size (uses bandwidth) and would
require more NPU time to process, unless the LSR is already beyond its
maximum label processing depth, in which case the forwarding behavior
might be unspecified because it will not be able to find the Bottom of
Stack.

I hope you can understand why I am so concerned about this.  It is at
least *possible* to upgrade control-plane software and implement my
suggestion.  It is *impossible* to make deployed hardware able to
process a deeper label stack than its original design requirement.

It is further *impossible* to process a deeper label stack in FUTURE
LSRs without additional costs (head/power/area.)  That's why I think
your proposed solution is, frankly, very foolish.

--
Jeff S Wheeler <[e-mail]>
Sr Network Operator  /  Innovative Network Concepts