[Pdns-users] rec_control with forking

[bert hubert]

On Wed, Oct 11, 2006 at 08:43:26PM +0200, Nico van Royen wrote:

> Using rec_control we're trying to get some accurate data on how much the
> pdns-recursor can take on varrious hardware setups.  We are however
> running into some hickups when running the recursor with the "fork=yes"
> options.

The 'fork' option is experimental. We have concrete plans to improve it, but
as it gives large financial benefits for users (you can basically run with
40% less servers), we are hoping someone is willing to sponsor 'enterprise
grade multi processor' support. 

> When the recursor is forked, there seems so way that the rec_control can
> utilize the socket since it basicly has 2 different processes (thus also 2
> pid's) and 2 different socket files with its pid attached.  Any stats made
> now (when running with no forking) are picked up by rrdtool to give us
> some insight in what pdns can manage. (we're attempting to stress out the
> server using the queryperf tool from bind9).

rec_control can be used to query a specific instance, using the --socket-pid
option. You need to get the pid yourself. As said, the 'fork' option is
experimental.

> Running it from several sources to a single instance if the recursor gives
> us an average of 8500 qps at 99% (single-)cpu usage on low-specs hardware. 

That sounds more or less ok.

> We might asume this will increase by (guessed) 75% when running it forked
> (since its a dual cpu setup) but accurate stats would be nice.  Question

75% would be a safe bet. At those packet rates ("16000qps", which is around
40000pps), network driver and kernel version start to matter. We've seen a
90% increase on modern hardware (Opteron) and modern kernel (Linux,
2.6.recent).

> pdns recursor (when stressed) likes: big cpu + big mem (for caching) ?
> pdns authorative (using gmysql from external sql server): likes: big cpu + low/medium memory ?

Depends a bit. Both recursor and auth consume lots of memory *bandwidth*. It
appears there might even be a performance drop beyond 2G of memory use by
the recursor.

The authoritative server needs a lot less memory, and loves memory
bandwidth. 

In short -> get dual opterons.

Good luck!

-- 
http://www.PowerDNS.com      Open source, database driven DNS Software 
http://netherlabs.nl              Open and Closed source services