happy CGN -- beating happy eyeballs and trending toward E2E success

[Erik Nygren]

I'd strongly discourage the use of DNS hackery like this.
It could have bad side-effects for sites using short DNS TTLs
for load balancing and fault response (and having a much longer AAAA
TTL could result in much worse performance/behavior and thus possible
blacklisting of AAAA responses).

It would seem like getting improved Happy Eyeballs implementations
would be the best approach forwards.  A "good" Happy Eyeballs implementation
should compare IPv4 vs IPv6 TCP (HTTP) performance and availability
with some preference towards IPv6 (whether it is 30ms or 300ms).
This should do the right thing for CGN since a direct path over IPv6
should generally be faster than an indirect IPv4 path through the CGN,
especially when a bias for IPv6 is added in.

Anything using A vs AAAA DNS response times here seems like a bug
(unless the AAAA response is *much* slower to come back) as that bears
no correlation to what the actual TCP/HTTP availability and performance
will be like for IPv4 vs IPv6.

       Erik


On Sat, Jun 9, 2012 at 1:24 AM, Cameron Byrne <cb.list6 at gmail.com> wrote:
> Hi,
>
> During v6launch, i heard a fair amount of chatter that "happy
> eyeballs" has had a bit of a negative impact on traffic levels.
> Meaning, people who move a lot of content and keep stats on who does
> what and when, know that Apple's implementation of happy eyeballs is
> moving traffic over IPv4 when it can be over IPv6.
>
> This is not news
> https://labs.ripe.net/Members/emileaben/hampered-eyeballs, but now
> that it is past v6launch, it is real and traffic impact is real, and
> ... for people with CGNs, it's even more real since the offload from
> CGN to IPv6 is not happening (business case for IPv6 looking a little
> weak, promises about CGN off load  not kept, traffic projections and
> peering builds-- back to the slow track , blah...)
>
> So, there are explorations on how to beat happy eyeballs, especially
> Apple's happy eyeballs.  This is becoming an arms race, but so be it.
> Maximizing on 5ms delta to a web browser is not an objective of any
> network providers that has a CGN that i know of.  Minimizing state /
> scale in a CGN is an objective... in fact, there are many who say IPv6
> only gets deployed because they fear the CGN.  So, if we deploy IPv6
> to not have to deal with the CGN, but happy eyeballs does not get the
> traffic off of the CGN .... something must be done.
>
> Being a hack, i always look to solve problems with DNS trickery first:
>
> Would aggressive caching of AAAA and aggressive non-caching of A
> result in beating happy eyeballs on OSX?
>
> I don't own any Apple products, so i cannot test this myself.
>
> i setup a very simple test web page
>
> http://test1.cgn.strangled.net/b.shtml
>
> DNS TTL = 1 day for A and AAAA (the control)
>
> and
>
> http://test2.cgn.strangled.net/b.shtml
>
> DNS TTL = 1 day for AAAA and 1 second for A (the experiment)
>
> To get the effect, you would need to refresh a few times to prime the DNS cache.
>
> Let me know your thoughts and results.
>
> CB