Latency jump after upgrade from 1.5.8 to 1.9.0
George Thessalonikefs
george at nlnetlabs.nl
Mon Dec 2 10:55:10 UTC 2019
Hi Matthew,
1.5.8 to 1.9.0 is quite the jump with a lot of changes and added
functionality. The best place to start would be the download page
(https://www.nlnetlabs.nl/projects/unbound/download/) where you can see
an overview of features and bug fixes per release.
I did a quick diff for the runtime configuration options and the
differences in default values that change behavior between 1.5.8 and
1.9.0 are:
- harden-below-nxdomain: no -> yes
- qname-minimisation: no -> yes
- minimal-responses: no -> yes
If you weren't using the above you could try configuring them to "no"
and see if that changes anything.
If you could also share your configuration it would shed some more light
and maybe we could spot something.
-- George
On 26/11/2019 20:03, Matthew Ghali via Unbound-users wrote:
> Hi folks- we're working on an internal upgrade of our Unbound instances
> from the Ubuntu xenial default of 1.5.8 to 1.9.0 (backported from eoan)
> so we can start using dnstap and prefetching.
>
> Our initial test upgrades (without either dnstap or prefetching enabled
> yet) consistently show a large jump in response latency though. Adjacent
> un-updated Unbounds left as controls do not show the same changes.
>
>
> While the query types and response codes show no corresponding change:
>
>
>
> And overall cache efficiency looks unchanged:
>
>
>
> We do see a dramatic change in latencies:
>
>
> These results do not change over time as the upgraded instances become
> "warmed".
>
> We've looked at the Ubuntu maintainer scripts for the packages and don't
> see any red flags in build option differences between versions. Were
> there perhaps build defaults that changed between versions? We're
> puzzled since a performance change like this surely would produce some
> mailing list traffic at the least, which we also couldn't find- so that
> must mean there is some difference local to us somehow. Are there
> suggestions on where we can look?
>
> Thanks!
More information about the Unbound-users
mailing list