On Thu, 2006-03-16 at 20:24 -0800, Cary, Kim wrote:

In summary, tuning the solaris udp hiwat and max_buf had no effect on the continuous udpInOverflows loss.

At present, changing some parameters in syslog-ng.conf has helped, it seems.

We have now gone from

log_fifo_size(100000); gc_busy_threshold(3000000); gc_idle_threshold(300);

and having

A1) udpInOverFlows increment by 3000 every 4 sec A2) Garbage collecting while busy for 3 sec every half hour at peak

to

log_fifo_size(1000000); # up an order of magnitude gc_busy_threshold(300000); # down an order of magnitude gc_idle_threshold(500); # log was reporting objects 310-322 range, so some headroom

and having

B1) udpInOverflows increment by 50 every half hour B2) Garbage collecting while busy for a fraction of a second every 4-8 minutes

This seems better. Hope that helps someone out there.

My first question - we are losing messages during GC while busy - do we only lose the messages that arrive between these two log entries:

Mar 16 20:00:30 syslog-ng[9688]: Garbage collecting while busy... Mar 16 20:00:31 syslog-ng[9688]: Objects alive: 310, garbage collected: 299695

My second question is, what can we do better to tune this further?

Thanks to all (& Balasz for the wonderful software).

IIRC someone pointed out that the receive buffer cannot forcedly be increased on Solaris outside the application, so syslog-ng should call SO_RCVBUF setsockopt which it does not at the moment. You could hand-patch a hard-coded number though. -- Bazsi