Performance Expectations?

kollross · Post by **kollross** » Thu Dec 01, 2005 9:02 pm

This is an open question to anybody, i'm just trying to a get a estimation if our install and setup is completely jacked or the Zimbra product is underperforming. Currently we have a two server setup. A dedicated spam/virus filter and a dedicated mailstore which also does imap,pop, and outgoing smtp. Currently we are using lmtp to deliver messages between the two machines. The transfer is running about three message per second. After the queues reach over 9,000 in roughly 6 hours we were forced to shut the system down. The mailstore machine specs are Dual Xeon with 2 gig of ram and the spam filter is a Dual Opteron with 4 gig or ram. Our old system which used essentially all the same software, postfix, amavis, clamav, and transported over smtp between the 2 machines with slower cpu machines (1 filter machine, 1 mailstore). With these slower machines we are processing around 30 emails a second. We are royally perplexed at this point and are trying to narrow down if this is a configuration issue or a product issue.
Let me know please.

14319KevinH · Post by **14319KevinH** » Thu Dec 01, 2005 9:27 pm

We see much higher numbers in the lab, and the deployments we've seen. Something is wrong. What kind/number of disks do those machines have? Are you using normal traffic to drive the load or a load test tool?

14319KevinH · Post by **14319KevinH** » Thu Dec 01, 2005 9:27 pm

BTW: also sent you a PM.

kollross · Post by **kollross** » Thu Dec 01, 2005 9:30 pm

on the Mailstore we have a Dell Hardware Raid5 perc controller. Capacity wise we have 500 gig. All benchmarks for disk reponse are fine.

marcmac · Post by **marcmac** » Fri Dec 02, 2005 7:58 pm

I saw these results after our conversation today.
One workaround I saw was changing log_level = 0 in amavisd.conf.in, then restarting amavis (zmamavisdctl stop/start).
I need to test more, with larger message sizes, but this increased my performance quite a bit.

marcmac · Post by **marcmac** » Sun Dec 04, 2005 11:10 am

I've run some more tests, and I suggest the following changes (which will become the new defaults):
set log_level to 1 in amavisd.conf.in

set max_servers to 10 in amavisd.conf.in

set smtp-amavis processes to 10 in postfix/conf/master.cf:

smtp-amavis unix - - n - 10 smtp
Restart postfix and amavis.
This will cut down on the IOWAIT on the mta, which is what seems to be the biggest throttle.
If you want the absolute maximum performance from amavis, you can set the log_level to 0, but you'll lose the AS/AV info from the logs (which we use in message tracing).

kollross · Post by **kollross** » Sun Dec 04, 2005 4:25 pm

I passed this info onto Tron (he was unfortuatly gone during our conversation friday) he made the changes and ran the script that we supplied you with to test the queues, according to tron the queues never went over 3 and all the messages were delivered instantly. He set logging to 1 so we could use the mailtrace and reran the test with amavis running and the queues never went over 4 so i think this fixed the issue.
Thanks alot for the help.

marcmac · Post by **marcmac** » Sun Dec 04, 2005 5:42 pm

I'm glad you guys got the same results. As I mentioned, this will be the default for future releases, so you should see this performance gain persist.