Friday, December 18, 2009

syslog-ng OSE 3.1beta2 release

I've mentioned shortly in my previous post, but here's a more official announcement: I've released syslog-ng OSE 3.1beta2, containing some important bugfixes.

The list of changes:

Thanks to Martin Holste for the feedback he provided, hopefully we can forget about the "beta" part soon.


/-djs-/ said...

Bazsi- I've had some problems with the OSE and udp syslog receive speed. Since so many features have been added in 3.x, could that be a regression? I'm getting like 30KB/sec receive on a fast box that can receive 100x that amount of data with a Splunk listener. I've looked at all the tuning info out there and have tried all kinds of different options including disabling dns, etc., but I could only get an improvement from 20KB/sec to 30KB/sec. This is on RHEL 4.6 x86_64. I look forward to testing 3.1, but I'm really hoping you have some real "AHA!" info that will explain and fix this for me.



Bazsi said...

is this 3.1.x or 3.0.x ? I don't know about any kind of UDP related performance regressions, but it may happen that noone told me.

But that rate seems very low indeed.

What are the symptoms you are experiencing? Message drops, or syslog-ng simply maxes out your processor?

/-djs-/ said...

3.0.5 It's just not getting all the events. CPU util is 0.5% - 0.7%. netstat -su doesn't show receive errors. The only thing I've thought of since is instead of specifying the ip to listen on, just use, but that should really only increase cpu util in the ip stack to id the dest ip on the packet, right, so I don't think that will do much to fix it.

Bazsi said...

hmm... does the syslog-ng dropped counter for the specific destination count any messages?

since you said that the kernel is not reporting receive errors, that pretty much means that syslog-ng itself is dropping those messages for some reason.

if you enable stats_freq(), you should see the message drop counters.

If syslog-ng itself is dropping messages, that should mean that the destination FIFO is undersized, though the default should pretty much suffice.

I've tried to reproduce the problem here and syslog-ng happily processed a couple of thousand msg/sec.