Re: [BSD] Levelezés IMAP-al

[Bognár Zsolt]

Ez azért elég megdöbbento...:

Hamm, i don't think this test completely utilize performance of every MTA.
It did not tuning, both linux system itself and MTA. And the result could
not reflect the truth.

In my test, a PIII 933*2 SMP/512RAM/18G U160 SCSI disk linux box. with some
system tuning( enable the /var/spool/postfix and /var/qmail/queue with async
write, enlarge fs.file-max to very high, enlarge system socket slots to very
high ..) and MTA tuning:

qmail: disable loging for multilog using syncing write??(with multilog only
50% performance as disable multilog),using cdb to store user information for
fast lookup, enlarge qmail-smtpd concurrency to 400, and etc..

postfix: raise process_limit to 250 or higher, use hash to store users
information for fast lookup, decrease delivery concurrency to 100 or 50 or
lower for causing less i/o..etc

result:(using postal and smtp-source, from a linux client)
1.postal
2200 mails/minute for qmail without logging
1120 mails/minute for qmail with multilog
qmail with cdb, only 5% performance improvement.

5320 mails/minute for postfix with async syslog setting
1320 mails/minute for postfix with syncing syslog !!so slow
enlarge delivery destination limit will slow the speed.

2.smtp-source
1000 mails to 1 system account, 100 concurrency connection.

without multilog and using cdb
qmail using about 39 seconds to finish the test, 26 msgs/s

with async syslog using hash
postfix using only 5.3 seconds to finish the test,188 msgs/s

my explain:
from the result , i think sync logging is the killer to qmail and postfix;
and disk i/o will also slowing down smtp injection. Postfix is much faster
than qmail when doing concurrency injection for less disk i/o than qmail.

every message will cause qmail to do at least 3 writes to the system,
default is sync write! Postfix only do 1 sync write, as i know...

Please correct me if i'm wrong. and my result could only be an reference
when you are trying to choose which mta is the best for you. But the best
way is to *test* it yourself!