Maintained by: NLnet Labs

[Unbound-users] No answers from unbound occasionally

W.C.A. Wijngaards
Fri Jun 5 16:15:16 CEST 2009


-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Joe Abley wrote:
> This suggests to me that for large, anycast-distributed servers there is
> an opportunity for unbound to give up to early whenever services are
> moved between far-distant anycast nodes.
> 
> What is the actual algorithm?

Exponential backoff.
(I think internet software should do exponential backoff because that is
good behaviour.)

> For example, if the RTT to an anycast server suddenly changes from
> reliably <20ms to something substantially higher, what is the threshold
> RTT at which we can expect unbound to stop being able to get answers?

If it becomes more than 16x or 32x slower, then you can expect a first
servfail as lots of exponential backoff probes are failing.  And then
things work again.  If there is some other server in the NS set that
does not move like this, then the (1 second of) servfail does not
happen.  In that case, unbound finds out the new rtt for the moved
server gradually, using the other server if the slow server does not
respond quickly enough.

Best regards,
   Wouter

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (GNU/Linux)
Comment: Using GnuPG with Fedora - http://enigmail.mozdev.org

iEYEARECAAYFAkopKHQACgkQkDLqNwOhpPinyACdGvfUZcOyZDMF5zUBtWRHtZNd
yooAmQGTvjkpex2oQsefdOtmjBq/mA9g
=RRyl
-----END PGP SIGNATURE-----