From the dualstack-is-fun department...

[Guillaume.Leclanche at swisscom.com]

> -----Original Message-----
> From: ipv6-ops-
> bounces+guillaume.leclanche=swisscom.com at lists.cluenet.de [mailto:ipv6-
> ops-bounces+guillaume.leclanche=swisscom.com at lists.cluenet.de] On
> Behalf Of Daniel Roesen
> Sent: Monday, February 28, 2011 11:34 PM
> To: ipv6-ops at lists.cluenet.de
> Subject: From the dualstack-is-fun department...


> DS-Website having one AAAA RR:  ~3 minutes time to IPv4 fallback(!)

I analyzed this behavior, the reason is that the browser is calling the connect() API from the OS. This function will then try the complete TCP opening process, including TCP retries. Linux (at least debian flavour) will do 5 retries, with exponentially growing time between them.

You can play with the number of retries; I put it down to 2, and the fallback-to-ipv4 time was roughly 10s.

I opened a feature request on the kernel bug tracker to request the possibility to tune this parameter for dualstack websites, and put it by default to 2, but nothing happened afaik.

The request :
https://bugzilla.kernel.org/show_bug.cgi?id=23242

The sysctl control is tcp_syn_retries (sorry forgot the hierarchy, you'll have to grep :)

Guillaume