[dns-operations] FlagDay 2020 UDP Size

[Viktor Dukhovni]

On Mon, Aug 03, 2020 at 09:44:17PM +0100, Tony Finch wrote:

> jack tavares <tavares at gmail.com> wrote:
> >
> > I have gone through the archives, is there consensus on this at this time?
> > For both the date of Flag Day (Which appears to be 1st October 2020,
> > pending confirmation from google)
> > and for the suggested default?
> 
> There are some interesting measurements in
> https://rp.delaat.net/2019-2020/p78/report.pdf

What I haven't seen reported is measurements of problems that occur when
the EDNS(0) UDP buffer size is *too small*.

There are lots of measurements with lost UDP datagrams when the buffer
size is too large, but given a "too small" buffer size servers truncate
responses, and some don't also support TCP.  This causes lookup failures
when the buffer size is sufficiently low.

I've posted a few examples to this list of failure cases with a buffer
size of 1232 that are resolved with buffer sizes of ~1400.

So it is not entirely obvious where the most practical tradeoff lies.
And indeed that best value likely depends on where the resolver (or auth
server) sits on the network.  So while we may be able to converge on a
recommended default, unfortunately some users may want/need to move the
needle up or down to meet their needs.

My sense at the moment is that 1232 is too conservative.  I see
that (thanks Tony):

    https://rp.delaat.net/2019-2020/p78/report.pdf

recomments 1372 for IPv4 and 1332 for IPv6 on internal networks and 1232
for stub resolvers using external IPv6 recursive servers out on the
public Internet.  The suggested bounds look more promising than a flat
1232 across the board, but I wonder whether the downside of being too
conservative has also been taken into account.  So the presently optimal
values may be somewhat higher than recommended in that paper.

-- 
    Viktor.