performance testing new server, results

[Ken A]

Chris Yuzik wrote:
> Hi everyone,
> 
> As I mentioned in a previous thread or two, we're almost ready to roll 
> out our new mail server. To stress test it, I have essentially captured 
> several hours worth of live queue files from the current production 
> server, both spam and ham. To be even more mean to the new box, I have 
> disabled spamassassin caching. Finally, (and I feel like a bit of a mad 
> scientist), I simply copied all of those queue files (about 4500) into 
> /var/spool/mqueue.in. Mwahahaha.
> 
> The server is chewing through them in batches of 30, with the 
> recommended of 5 children per CPU (we have dual 3 GHz Xeons in this box) 
> there are 10 children. With all the features turned on, including 
> fuzzyocr (cranked up), bayes and clamav, but spamassassin caching off, 
> it seems to be averaging about 140 seconds per batch of 30. Server load 
> is averaging 2.56 over last 15 minutes. So in past 20 minutes, it's 
> handled about 2500 messages.

Seems close. If you don't have > 2gb of ram, consider it because if 
these are dual core processors, then you can up the children to 15 or so 
if you have enough ram. Have you done other performance tweaks - 
mailscanner incoming in tmpfs?, spools and logs on separate spindles?, 
Fuzzyocr.cf tweaks? We put about 200k messages through a similar box 
yesterday. Load avg stayed ~2 and batches were consistently very small, 
so it wasn't really breaking a sweat. We don't use bayes, but do split 
recipients, and use SA caching.
Try 'vmstat 2', 'iostat -x 2' or similar to see what, if anything is 
bottlenecking.

Ken A.
Pacific.Net


> 
> Does this performance seem about right? Or should it be significantly 
> better?
> 
> Thanks,
> Chris