dovecot - May 2003 - Speed and memory probs writing large Maildir

I have Dovecot 0.99.9.1 compiled from the source, without SSL, running
on a Red Hat 9.0 system - EXT3 file system, Pentium Pro 233 MHz, Intel
motherboard, 128M RAM, IBM 40Gig 7200 RPM HD.   I find reading and
searching to be nice and fast, and comparing it with Courier IMAP on a
Celeron 824 MHz, I think Dovecot is generally not much slower, which
means it may well be faster than Courier in some or many ways if it was
running on the same machine.

My only problem is when writing mailboxes with large numbers of
messages.  I use Mailir - the same arrangement as Courier IMAP.  (I also
find that Netscape 4.77 can't use mailboxes which are subfolders of
others, but that is probably a Netscape config problem.)

I am using Netscape 4.77 (Windows 2000 1.3 GHz Celeron, via 100Mbps
switched Ethernet to the Linux machines) to do these tests since I find
Netscape 7.02 to be much slower.  Netscape 7.02 seems to want to resynch
its notion of the mailbox repeatedly as it is writing to it.  Without
looking at the IMAP traffic (I don't know how) I would say that Netscape
4.77 is not doing this at all.  Indeed, after the write, looking at the
mailbox causes Netscape 4.77 to read the headers from the IMAP server -
so I feel sure that its way of writing the messages is purely "write".


Writing a mailbox with a small number of large messages, such as 20 x
100k messages is fine. There is no obvious speed problem or excessive
memory use.

When writing 2000 x 1k messages, it is totally different.  Here are some
times for writing from a Netscape 4.77 local mailbox to Dovecot and to
Courier IMAP.


Message        Total    Time       CPU max     Memory RSS
size   number  size     Seconds    Approx      Megs - max

Dovecot

100k  100      10M        23       15%         ~2
100k   20       2M        ~5       ?            1.2
 10k  200       2M        20       15%          4    !
  1k 2000       2M       230       60          32    !!!!

Courier IMAP

100k  100      10M        14        5%         ~1
100k   20       2M        ~2        ?           ?
 10k  200       2M         9        5%         ~0.6
  1k 2000       2M        88        2%         ~0.6


This 2000 message test is small compared to some of the mailboxes I have.

I used a real-live mailbox, 6878 messages - 44 Megabytes in a single
Mbox for a more demanding test.  Using Netscape 4.77 to copy this to a
Dovecot Maildir took a very long time - over 30 minutes.  The process
was fast at first and got slower and slower.  As the number of messages
in the destination Maildir increased, the memory usage went up and the
speed reduced.  It seems that RSS memory usage (as reported by "top -d
0.3" hit some kind of limit at about 49 Megs after 4000 messages and
then went down a little to about 48 or less.  CPU usage was 30 to 82%,
fluctuating according to the moment top looked.  At the end, I think it
was only writing 1.5 messages a second.

RSS is the "total amount of physical memory used by the task".  The
"SIZE" was even bigger - "code plus data plus stack".  This
got up to
114 Megs by the end of the writing process!

After the write had finished, I made Netscape 4.77 view the destination
mailbox.  This caused Dovecot to send it all the headers, and again
Dovecot had these huge memory usage figures.   It seemed to read the
mailbox at a decent speed - 20 seconds or so.

But if I made Dovecot look at another mailbox, and then again at this
big target mailbox (say with another client - Netscape 7), it would read
the big mailbox at Dovecot's usual high speed - without the high memory
usage and I think without high CPU usage as well.


Without knowing anything about how Dovecot works, I imagine that there
is some kind of caching algorithm for writes, and that this is retained
until Dovecot (or at least the instance serving this client) is asked to
look at another Maildir.  It seems that reading is not slowed by this
process much or at all - but writing gets progressively slower as the
cache gets bigger.

The memory usage - even 49 Megs for RSS - seems excessive.  It reached
that about halfway through writing the mailbox, so the entire message
contents at that state would have been about 22 Megs.  But the "SIZE"
memory usage is even larger - perhaps limited by the available RAM.


This slow speed for writing Maildirs with large numbers of messages is
currently a barrier to me using Dovecot.  I am about to go back and try
again to get Courier IMAP to do what I want . . .

The Courier IMAP documentation and what I regard as the difficult
"configure -> build -> install -> wonder why it doesn't
run" process has
cost me so much time that I would be very happy to use something simpler
and easier (for me, at least) to install.

Thanks for developing this new IMAP server!


  - Robin

P.S. I have not altered the dovecot.conf file at all.  The only
deviation from standard is no SSL.  I find it remarkable that Dovecot
works out of the box and automagically finds the user's Maildirs!

  - Robin

On Wed, 28 May 2003 15:18:22 +1000,
Robin Whittle <rw at firstpr.com.au> wrote:
> My only problem is when writing mailboxes with large numbers of
> messages.  I use Mailir - the same arrangement as Courier IMAP.  (I
> also find that Netscape 4.77 can't use mailboxes which are subfolders
> of others, but that is probably a Netscape config problem.)
> When writing 2000 x 1k messages, it is totally different.  Here are
> some times for writing from a Netscape 4.77 local mailbox to Dovecot
> and to Courier IMAP.
I had problems with big (2000+ messages) folders on OpenBSD. I updated
to 0.9.99.10-test6 and it is fine.

http://dovecot.procontrol.fi/test/

-- 
http://cemetery.homeunix.org
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
URL:
<http://dovecot.org/pipermail/dovecot/attachments/20030528/6cb49d18/attachment-0003.bin>

On Wed, 2003-05-28 at 08:18, Robin Whittle wrote:
> When writing 2000 x 1k messages, it is totally different.  Here are some
> times for writing from a Netscape 4.77 local mailbox to Dovecot and to
> Courier IMAP.
Thanks, this test helped to find several problems.

First problem was a memory leak with APPEND command. This caused the
huge memory usage. Here's a fix:

--- cmd-append.c        14 Feb 2003 08:00:52 -0000      1.24
+++ cmd-append.c        28 May 2003 11:01:03 -0000
@@ -197,6 +197,7 @@
 
                count++;
        }
+        imap_parser_destroy(save_parser);
 
        if (!box->save_deinit(ctx, failed)) {
                failed = TRUE;

Then I noticed that each APPEND command has to reopen and resync the
mailbox every time which gets just slower and slower. I think I'll make
it leave the mailbox open for a few seconds just in case another APPEND
comes. That should speed it up quite a lot.

Actually APPEND wouldn't even need the mailbox to be synced. I should
try fixing that too..

And finally there's a nasty bug in 0.99.10-test's new maildir syncing
code. I think I'll have to rewrite it a bit more to avoid wasting
memory..

On Wed, 2003-05-28 at 08:18, Robin Whittle wrote:
> When writing 2000 x 1k messages, it is totally different.  Here are some
> times for writing from a Netscape 4.77 local mailbox to Dovecot and to
> Courier IMAP.
Want to try this with 0.99.10-test11? :) It now keeps the index file
opened for 10 seconds before really closing it, that should help with
the speed quite a lot. I just tested copying 4000 messages myself, I
didn't notice any actual increase in speed, but CPU usage dropped from a
few percent to 0.4%.

It still does a few unnecessary things every time mailbox is opened (ie.
at every APPEND command). I think I shouldn't try to mkdir() the cur,
new and tmp dirs immediately. Those could be more easily created if/when
stat() fails while syncing.