Re: [tsvwg] Status of ECN encapsulation drafts (i.e., stuck)

["Holland, Jake" <[e-mail]>]

Hi Bob,

> To be unambiguous, I've provided a schematic here:
> http://bobbriscoe.net/projects/netsvc_i-f/consig/encap/ecn-reassembly.pdf

Thanks for clarifying, I think your message confirms the speculation Jonathan
and I landed on as the source of the misunderstanding.

As I said there, it turns out this was just about semantics, not a disagreement
about what's happening on the wire, and so I don't think the point needs more
discussion, but I just wanted to confirm I believe I now understand what you
meant, and your message reinforced that belief.

And thanks for the schema diagram, I was able to see it and it was helpful.

As I tried to say, I think the reason I got confused about your meaning was
that I thought separating the number and proportion of CE-marked packets into
2 values would only make sense for 2 values that were able to vary differently
from each other, and so thought you were saying by "number" (as a meaningfully
different value from "proportion") that the reassembly process was increasing
the report to the receiver of the number of CE marks actually emitted by the
network.

I agree that when fragmented and getting a fixed proportion of marked packets
with an even random distribution of markings, it obviously increases the number
of marked packets relative to an equal unfragmented stream.  Although I wouldn't
have thought to phrase it as "both number and proportion" since it's so nearly
the same thing as just "proportion", I acknowledge that your phrasing is not an
invalid construction, and that for some reasonable AQMs, you're accurately
describing a sense in which the number of marks can be increased due to
fragmentation.

I hope you likewise agree that the logical-OR reassembly doesn't ever increase
the count of CE marks seen by the receiver to a value above the count of CE
marks actually emitted by the network (and for threshold-based systems often
reduces that number, since fragments will often be marked together), and that
the concern you were addressing comes because fragmentation provides more
opportunities to the network to emit CE marks, not because logical-OR reassembly
can increase the raw count of CE marks actually delivered.  And also that you
can see this as a reasonable interpretation of "number of CE-marked packets"
that can reasonably result in confusion.

Best,
Jake

PS: I also think I understand the proposal outlined in section 5 of update-shim-10,
but I think allowing generalized behaviors like that text tried to do, in pursuit
of preserving the proportion of marking instead of the fact of marks on a packet,
requires a much more precise and careful treatment than the one outlined there, and
would be better done as a separate proposal if it's to be done at all (I agree with
Jonathan that it's not clear it's worthwhile).

The sample algorithm permitted by that text clearly would require a 3168 update,
but also makes some very specific assumptions about the AQM's behavior that may
not be justified without specific knowledge of the network.

It's not hard to imagine a situation where the later fragment(s) of a packet are
more likely to receive a mark (for instance if the later fragments are back to back
more often than the head, so that later fragments arrive with a systematically
fuller queue).

If that's the case, examining only the head of the fragment would systematically
suppress delivery of CE marks in fragmented packets, because of the violated
assumption about an even probability distribution for marks.

This further seems like it could give perverse throughput incentives to degenerate
behaviors like fragmenting packets at the sender, to get the benefits of reduced loss
from ECN but receive fewer marks, even though the network might emit more, and thus
pushing out unfragmented competing flows.  So the specific example algorithm given
seems like an unsafe or at least counterproductive thing to encourage reassemblers
to do.

But that's a discussion for another thread, really.