Re: Getting to consensus on packet number encryption

[Spencer Dawkins at IETF <[e-mail]>]

FWIW,

On Tue, May 1, 2018 at 7:19 AM, Benjamin Kaduk <
bkaduk=[e-mail]> wrote:

> On Mon, Apr 30, 2018 at 09:55:31PM +0000, Praveen Balasubramanian wrote:
> > I disagree that we need any more data for not doing PNE in the
> datacenter. Why would we add an extra encrypt-decrypt step for no obvious
> benefit? The cost of this operation will keep magnifying as UDP performance
> improves. We have several optimizations planned for UDP, but short of
> hardware offload, none for crypto – this is an important trend to account
> for performance wise. For short packets like ACKs an extra encrypt decrypt
> step can become a large fraction of CPU cost over time.
>
> There is benefit in uniformity and simplicity of implementation; I concede
> that
> the claimed potential benefits about the middlebox ecosystem are
> speculative,
> but the simplicity seems clear to me.
>
> With respect to the cost of the crypto, yes we have some numbers, but
> AFAICT we
> don't have a great view for how they will fit into the broader picture as
> implementations mature, so to some extent the estimate of the magnitude of
> growing cost of crypto is also a speculative position.  (Also, who is the
> "we"
> that has several optimizations planned?)
>
> In other words, how do we know this is not premature optimization?
> Shouldn't
> we wait until we have actual performance numbers before doing lots of
> optimization?


I can't speak to whether this optimization is premature, but I do suspect
that if more people were looking at more actual performance numbers, this
conversation might converge more quickly ...


> I was under the impression that QUIC versions were going to be
> lightweight, so putting out v1 with mandatory PNE would not preclude
> quickly
> following with a V3 that adapts to any discovered performance concerns,
> whether
> for datacenter usage or otherwise.
>

Yeah, about that ...

I've been asking in a few places - most recently, at
https://datatracker.ietf.org/meeting/101/materials/minutes-101-quic-00.html
- about version appearance rates, and I haven't heard answers that sounded
like the working group had strong agreement that versions were going to
roll out pretty quickly. For example, from that conversation, I got

Spencer (as AD): *[deleted first three points, to get to] *4. (as Spencer)
don't have a good feeling about version negotiating and how much we might
roll them.

mnot: not going to wait 5 years for v2.


So, the chairs can correct me (and should, and they have before on other
topics), but I think I've heard a lot of "we don't have to get everything
perfect in QUICv1, because there will be other versions", and not a lot of
"and the other versions will appear pretty quickly".

I don't know what all the participants are thinking, but that's what the
people who are talking have been saying to me.

IMO, of course.

Spencer, as individual


> -Ben
>
>