Re: [tsvwg] A review of draft-ietf-tsvwg-udp-options-12

[Paul Vixie <[e-mail]>]

On Tue, Jun 15, 2021 at 09:58:55AM -0700, Tom Herbert wrote:
> On Tue, Jun 15, 2021 at 9:32 AM Joseph Touch <[e-mail]> wrote:
> > ...
> > And, FWIW, trailers are just as easy if not *EASIER* for zero-copy to
> > handle; they just DMA the user data first then the UDP options.

> I am dubious of that statement. You'll have to show running code and
> performance results. Software stacks and hardware have long been
> designed around the fact that networking protocols are in headers; UDP
> options is an outlier in that regard. While it is certainly possible
> to make a functional stack to handle trailers, it is going to be hard
> if not impossible to leverage many implementation optimizations and
> techniques we've done over the years.

just as ip trailers were a temporary crutch to help the VAX BSD networking
stack use page-aligned mbufs and led to a great improvement in performance,
so it is that there will always be some implementation detail which will
attract our protocol design attention.

however, we must recognize that all such attractions are temporary. just as
the current linux networking stack has had several decades in which to
optimize for its operating conditions, so it is that other networking stacks
will eventually supplant it and take a commanding market share.

the zero-copy feature joe is working on should last 50 to 100 years. linux
will almost certainly be revised if not replaced on that time scale. let's
make the best protocol we can, so that both current _and future_
implementations can get maximum benefit from our efforts.

640KB was a bad architectural assumption, as was 16 bits, as is 1500 octets.
let us, please, treat any current operating condition as likely to be wrong
as a guidepost. we must think and act more fundamentally and more eternally.

-- 
Paul Vixie