BIND and UDP tuning
Browne, Stuart
Stuart.Browne at team.neustar
Thu Sep 27 08:01:13 UTC 2018
> -----Original Message-----
> From: Tony Finch [mailto:dot at dotat.at]
>
> > - { name: 'net.ipv4.tcp_sack', value: 0 }
>
> Why? SACK is super important for TCP performance over links that have any
> degree of lossiness, and I don't recall hearing of any caveats.
>
> Tony.
> --
> f.anthony.n.finch <dot at dotat.at>
If I recall correctly, it had to do with the fact that we were in a very-network-close test environment with very-small packets so it wasn't necessary to even consider resends. I don't recall whether it did anything at all to the results; it is just one of the various things I stuck into the blender in order to see if it made a difference and was still in at the end of testing. The number of test iterations I went through was in the hundreds and most of it was "Moar! MOAR!" rather than good arguments; more about proving a design could reach a theoretical limit than whether it would be 100% stable in production.
The environment design that these tests were preparing for haven't been implemented yet; that's what I'm working on over the next few weeks, so I'll be going over these settings with some kid-gloves and being a little gentler as we don't need a single location churning out 2M5 qps; we're quite happy with 2M.
Let's hear it for overkill!
Stuart
More information about the bind-users
mailing list