Unresponsive child and accounting

[Alan DeKok]

Christophe Saillard wrote:
> Our first experience with Freeradius on a FreeBSD server was a nightmare
> (it seemed to be a thread related problem, the server stopped working
> with a lot of "unresponsive child" error logs).

  Hmm... the code *did* work well on FreeBSD at one point.

> So, we tried on a Linux server (kernel 2.6.22-14-server ubuntu feisty
> fawn)  and it worked fine since last week :
...
> Wed Nov 21 15:33:22 2007 : Error: WARNING: Unresponsive child (id
> 2841623456) for request 2419798 (in component accounting module
> rlm_radutmp)

  That's a lot more helpful than earlier versions.

  My guess is that the radumtp file is *huge*.  If you don't need it, it
should not be used.  If you do need it, it should be rotated frequently.
 Or, use a database to store that information.

> The CPU went up to 100%.

  Consistent with spending lots of time reading the radutmp file

> There was about 300 802.1X clients connected (with a 2 minutes reauth
> period).

  Any commodity machine should be able to handle that.

> At this time we had no other choice than upgrading the hardware, it runs
> now on a 8 processor server but even with more CPU power we noticed a
> 20% system load.

  300 clients shouldn't need that much CPU power.

> Here's the threading part of the radiusd.conf :
...
>         max_requests_per_server = 300

  Please set that to zero.  That portion of the code isn't very well tested.

  Alan DeKok.