dovecot - Mar 2009 - dbox benchmarks

http://hg.dovecot.org/dovecot-dbox-redesign/

Looks like multi-dbox scales pretty nicely. Even after 100k messages the
peak saved msgs/sec is the same as the initial saved msgs/sec, even if
the average slows down somewhat.

I tested this by first deleting mailbox, then running "imaptest" for a
second to get saving to start writing several fields to
dovecot.index.cache file. Then ran imaptest for 60 seconds that did
nothing but save messages using IMAP APPEND commands. The summary is
that multi-dbox is almost 10x as fast as maildir when fsync is disabled
and 2,5 faster with fsyncing:

fsync_disable=yes:

multi-dbox : ~1700 -> ~1500 msgs/sec, 90842 msgs/60 sec
single-dbox: ~1700 -> ~1400 msgs/sec, 86149 msgs/60 sec
maildir    :  ~330 ->  ~110 msgs/sec,  9789 msgs/60 sec

fsync_disable=no:

multi-dbox : ~220 -> ~170 msgs/sec, 10113 msgs/60 sec
maildir    : ~100 ->  ~70 msgs/sec,  4123 msgs/60 sec

Also I just noticed that saving messages with two clients gets a peak of
2200 msgs/sec, most likely because two CPU cores can be used then. Hmm.
I guess if there were no locking it should have reached 3400 msgs/sec,
but it's better than no performance increase.

And the raw data:

multi-dbox, fsync_disable=yes

imaptest clients=1 secs=1;imaptest clients=1 msgs=1000000 - logout=0
append=100,0 secs=60
Logi Sele Appe 
100% 100% 100% 
   1    1 1759   1/  1
   0    0 1424   1/  1
   0    0 1734   1/  1
...
   0    0 1701   1/  1
   0    0 1640   1/  1
   0    0  778   1/  1
   0    0 1305   1/  1
   0    0 1598   1/  1

Totals:
Logi Sele Appe 
100% 100% 100% 
   1    1 90842 

multi-dbox, fsync_disable=no

imaptest clients=1 secs=1;imaptest clients=1 msgs=1000000 - logout=0
append=100,0 secs=60
Logi Sele Appe 
100% 100% 100% 
   1    1  209   1/  1
   0    0  227   1/  1
   0    0  208   1/  1
...
   0    0  203   1/  1
   0    0  145   1/  1
   0    0  156   1/  1
   0    0  180   1/  1
   0    0  162   1/  1

Totals:
Logi Sele Appe 
100% 100% 100% 
   1    1 10113 

maildir, fsync_disable=yes

Logi Sele Appe 
100% 100% 100% 
   1    1  331   1/  1
   0    0  311   1/  1
   0    0  288   1/  1
...
   0    0   96   1/  1
   0    0  119   1/  1
   0    0  111   1/  1

Totals:
Logi Sele Appe 
100% 100% 100% 
   1    1 9789 


maildir, fsync_disable=no

Logi Sele Appe 
100% 100% 100% 
   1    1   77   1/  1
   0    0  101   1/  1
   0    0   45   1/  1
...
   0    0   52   1/  1
   0    0   67   1/  1
   0    0   89   1/  1

Totals:
Logi Sele Appe 
100% 100% 100% 
   1    1 4123 

single-dbox, fsync_disable=yes

imaptest clients=1 secs=1;imaptest clients=1 msgs=1000000 - logout=0
append=100,0 secs=60
Logi Sele Appe 
100% 100% 100% 
   1    1 1765   1/  1
   0    0 1715   1/  1
   0    0 1764   1/  1
...
   0    0 1722   1/  1
   0    0 1679   1/  1
   0    0 1678   1/  1
   0    0  828   1/  1
   0    0 1467   1/  1
   0    0 1469   1/  1
   0    0 1149   1/  1
   0    0 1135   1/  1

Totals:
Logi Sele Appe 
100% 100% 100% 
   1    1 86149 

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 197 bytes
Desc: This is a digitally signed message part
URL:
<http://dovecot.org/pipermail/dovecot/attachments/20090330/67d8423c/attachment-0002.bin>

On Mon, 2009-03-30 at 18:26 -0400, Timo Sirainen wrote:
> maildir    :  ~330 ->  ~110 msgs/sec,  9789 msgs/60 sec
With maildir_very_dirty_syncs=yes (just committed to v1.2 hg):

~1100 -> ~110 msgs/sec, 11904 msgs/60 sec.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 197 bytes
Desc: This is a digitally signed message part
URL:
<http://dovecot.org/pipermail/dovecot/attachments/20090330/fd6221e6/attachment-0002.bin>

I can see where large mailservers would benefit from significant write 
performance increases - but unless the server is being actively limited 
by the local delivery agent, what other performance benefits does this 
offer?

In particular, is there any increase in READ performance by using dbox?  
Does it reduce memory consumption at all?

--
Daniel