Date: Thu, 28 Feb 2008 12:21:02 +0100 From: Kris Kennaway <kris@FreeBSD.org> To: Adrian Chadd <adrian@freebsd.org> Cc: Sam Leffler <sam@errno.com>, freebsd-performance@freebsd.org, Oliver Herold <oliver@akephalos.de>, freebsd-questions@freebsd.org, Ted Mittelstaedt <tedm@toybox.placo.com> Subject: Re: FreeBSD bind performance in FreeBSD 7 Message-ID: <47C6991E.1050502@FreeBSD.org> In-Reply-To: <d763ac660802271746l229c1e1hca44db5438e142e9@mail.gmail.com> References: <BMEDLGAENEKCJFGODFOCIEGKCFAA.tedm@toybox.placo.com> <47C59591.6040600@errno.com> <d763ac660802271746l229c1e1hca44db5438e142e9@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
Adrian Chadd wrote:
> (Sorry for top posting.)
>
> Its not actually -that- bad an idea to compare different applications.
> It sets the "bar" for how far the entire system {hardware, OS,
> application, network} can be pushed.
>
> If nsd beats bind9 by say 5 or 10% over all, then its nothing to write
> home about. If nsd beats bind9 by 50% and shows similar
> kernel/interrupt space time use then thats something to stare at. Even
> if its just because nsd 'does less' and gives more CPU time to
> system/interrupt processing you've identified that the system -can- be
> pushed harder, and perhaps working with the bind9 guys a little more
> can identify what they're doing wrong.
>
> Thats how I noticed the performance differences between various
> platforms running Squid a few years ago - for example, gettimeofday()
> being called way, way too frequently - and I compare Squid's
> kernel/interrupt time; syscall footprint; hwpmc/oprofile traces; etc
> against other proxy-capable applications (varnish, lighttpd, apache)
> to see exactly what they're doing differently.
Yep, and in this case NSD is currently 90% faster with prospects to push
it even higher with some further kernel changes (so far we have improved
it by 45%). BIND is limited by its own architecture, so improvements
cannot be made by modifying the kernel.
Anyway, the motivation here is not a DNS deathmatch, but part of our
ongoing effort to look for aspects of FreeBSD performance that can be
improved. Currently we are looking at UDP performance, and DNS serving
was thought to be a good model for that. It turns out that BIND does
not stress the kernel, but NSD does.
Kris
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?47C6991E.1050502>