High CPU usage after SQL failure when using buffered-sql

[Stephen Fulton]

Hi all,

First, I've used FreeRADIUS for a number of years in a number of installations, 
and I'm fairly comfortable with it.  I have looked through the archives, as well 
as read the documentation, FAQ, wiki and the notes within each of the 
configuration files that make up FR, such as the virtual servers and modules 
that would be used.  Now onto the issue:

During testing of a new RADIUS cluster, I implemented the buffered-sql virtual 
server, which works fine, with the detail file poll interval set to 10 seconds. 
However, when I simulate a failure of the accounting MySQL server to which it 
normally writes, CPU usage spikes significantly, from an average of ~ 1-2% to 
30+ over time.  If the accounting MySQL server is unavailable for longer than 
2-3 minutes, the spikes occur much more frequently.

Once the SQL server returns to service, accounting records stored in the detail 
file are properly read and inserted into the DB, then the files are cleared, as 
I expected.  However, the RADIUS process still maintains a higher than normal 
CPU load (according to top) of around 10%, which spikes to 30-40% when it is 
handling a new request (note the singular).  The server is a dual Xeon 
3.something, with 2GB RAM and SCSI disks.  The server is running Debian Linux 
2.6.26-2-686.  FreeRADIUS is running version 2.1.4, compiled manually.

During previous debugging, while running FR in debugging mode, I noted that it 
was attempting to read the detail file every second, sometimes less.  This 
matches a comment made by Alan in a similar sounding issues discussed on the 
list in April 2008.

(http://www.mail-archive.com/freeradius-users@lists.freeradius.org/msg44137.html)

Is there any way to mitigate these CPU issues in version 2.1.4?

-- Stephen