Thanks for the info, that's why we love Dovecot ;)
asko wrote:
Hi,
I did some performance testing with dovecot and qpopper, if anyone is interested..
Hardware: Dual Xeon 2.8 with HT enabled (dmesg sees 4 cpu-s), 1 GB RAM, SCSI RAID 1 with 2x Maxtor Atlas 10K V disks.
Software: CentOS 4.3, Qpopper 4.08 in standalone mode vs Dovecot 1.0-0_19.rc6, ext3 filesystem
Mailbox format Mbox, tested Pop3 with PAM authentication
My setup had 15 test users with 15 identical 40 MB mbox files. Each mbox had 370 messages, 35 % of them had attachments. So average message size is about 110 KB, average attachment size 220 KB.
"rabid" utility is part of postal package (http://www.coker.com.au/postal/). Command line, which fetches and deletes all the messages from users in the textfile:
rabid -p 1 192.168.x.x ./users -
Resulting fields are: time,messages,data(K),errors,connections,SSL connections,IMAP connections
Results (only the best one shown):
Qpopper:
1 parallel processes: 12:36,1020,111470,0,6,0,0
5 parallel clients: 12:44,3246,366207,0,13,0,0
15 parallel clients: 13:06,4099,481347,0,48,0,0
Dovecot: p1: 12:19,3316,367859,0,13,0,0
p5: 12:50,4653,524456,0,34,0,0
p15: 12:56,5063,578889,0,84,0,0
As you can see, Dovecot is faster.. With 15 parallel processes it seems to have saturated my 100Mbit/s LAN (578889 kB data in one minute). Dovecot had noticeably smaller disk IO and system load was smaller too, probably thanks to smaller IO.
Qpopper performed better than I hoped.. It is probably due fast disks of the test system, I wouldn't expect such numbers using single IDE disk.
Conclusion: The hardware was too fast for testing, LAN became a bottleneck atleast once. The results could be different with different mbox structure.. Anyway, I was convinced, Dovecot wins.
-- asko