Mehul Choksi <[EMAIL PROTECTED]> wrote:
> The purpose is to introduce qmail as an SMTP server in outbound email
> distribution system – just to send a million mails daily to subscribed
> users.

Do you mean send identical copies of one message to a million users, or send
one million unique emails, each to one user?  The difference is enormous.

> We tested the performance of the qmail by sending a few thousands mails to
> non-existent email-ids and to bounce back to another mail server. Strangely
> enough, qmail just built up a huge queue, without even preprocessing.

qmail doesn't start new deliveries if there are messages waiting to be
preprocessed (in todo).  If todo grows large, linear directory scan times can
slow the system down significantly; Russell Nelson's big-todo patch might help
here.  Others have used various schemes, such as injecting X at a time,
pausing a minute or two in between injections to allow qmail to catch up with
the todo contents, or trying first delivery with qmail-remote and only queuing
the mail if that delivery fails, saving queue disk bandwidth.

> Once the application stopped pumping to the Qmail server, it started
> processing and clearing queue, which took very long.

You may be running into a queue disk bandwidth limitation.  What sort of
hardware are you using?  Is the queue on a disk by itself?  Is that disk a
15kRPM SCSI disk, sitting on its own U160 controller?  Is that filesystem
mounted noatime?  What filesystem are you using?  What OS?

How are you logging?  What does the system load reach while running your
injection?  Have you read the section on large servers at www.qmail.org?  Is
/var/log on a separate disk?

> Is it a normal behavior (since all the email addresses were non-existent) or
> are we missing something somewhere? Concurrency is set to 120 for local and
> remote.

It's not trivial, but a million unique mails a day can be handled by qmail if
you set it up properly.  We just need _way_ more information than you've
provided to start guessing at what your limiting factor(s) is.

Charles
-- 
-----------------------------------------------------------------------
Charles Cazabon                            <[EMAIL PROTECTED]>
GPL'ed software available at:  http://www.qcc.sk.ca/~charlesc/software/
Any opinions expressed are just that -- my opinions.
-----------------------------------------------------------------------

Reply via email to