hi,
so after upgrade to the current cvs all of our problem disappear:-)
but we don't have time to test the ssl part...
although after I compile and install the new dovecot I stop it I do a
--------------------
find /home/*/Maildir -name '.imap*' -exec rm {} \;
find /home/*/Maildir -name 'dovecot-uidlist' -exec rm {} \;
find /home/*/Maildir -name 'msgid.cache' -exec rm {} \;
find /home/*/Maildir -name '.customflags' -exec rm {}\;
--------------------
just to be sure that we use a fresh version and no other problem can be
caused by the old index or any other cached files.
this was a BIG mistake!!!
we've got about 300 users and about 100 online at the same time. after a
two minutes break (while I reinstall and delete) every user check his
mail which casue a load about 90!!! the cpu usage 1-2% but the avarage
load was 90 (curently about 0.22)! the reason was IO. we have to stop
imap and wait for a while...
I don't know the exact reason bu I assume dovecot is regenerate
everybody's index a other cache at the same time. is there any way to
reduce the possible number of IO load? this would be useful if never
happend again!:-)
this happend around 4:40.
the output of sar -b:
--------------------
03:00:00 PM tps rtps wtps bread/s bwrtn/s
03:10:00 PM 48.18 29.19 18.99 1456.53 771.29
03:20:00 PM 41.66 25.80 15.86 1054.88 830.05
03:30:00 PM 42.64 25.67 16.97 1223.03 759.77
03:40:00 PM 44.98 31.95 13.03 1959.15 467.40
03:50:00 PM 26.68 14.63 12.05 741.48 424.12
04:00:00 PM 25.80 11.70 14.10 516.21 450.04
04:10:00 PM 59.55 34.91 24.64 1755.07 939.11
04:20:00 PM 95.98 61.15 34.83 4453.24 1563.25
04:30:00 PM 95.08 55.45 39.63 2182.22 1591.56
04:40:00 PM 168.85 145.99 22.86 11155.73 1037.25
04:50:00 PM 135.98 106.50 29.48 7908.21 1604.76
05:00:00 PM 125.87 97.65 28.23 5876.03 1060.50
05:10:00 PM 80.83 62.92 17.90 4321.82 590.19
05:20:00 PM 102.04 77.65 24.40 4907.27 1076.92
05:30:00 PM 129.04 106.58 22.46 6387.18 883.13
05:40:00 PM 89.44 73.20 16.24 4905.64 782.94
05:50:00 PM 88.23 73.16 15.06 5539.93 965.34
06:00:00 PM 92.49 72.67 19.81 5486.25 692.18
06:10:00 PM 45.24 32.82 12.42 1711.35 618.31
06:20:00 PM 42.17 28.16 14.01 960.11 622.33
06:30:00 PM 29.90 20.24 9.66 1538.27 376.17
06:40:00 PM 25.89 12.14 13.74 796.73 833.25
06:50:00 PM 20.50 4.98 15.52 264.91 993.62
07:00:00 PM 15.08 7.57 7.52 453.41 282.63
--------------------
the output of sar -d
--------------------
04:10:00 PM dev22-3 0.00 0.00
04:10:00 PM dev33-0 1.76 109.55
04:10:00 PM dev34-0 8.51 919.94
04:20:00 PM dev3-0 8.56 183.93
04:20:00 PM dev22-2 67.62 3798.02
04:20:00 PM dev22-3 0.00 0.00
04:20:00 PM dev33-0 4.80 463.81
04:20:00 PM dev34-0 15.01 1570.73
04:30:00 PM dev3-0 4.40 126.63
04:30:00 PM dev22-2 77.02 2440.08
04:30:00 PM dev22-3 0.00 0.00
04:30:00 PM dev33-0 1.94 124.80
04:30:00 PM dev34-0 11.73 1082.28
04:40:00 PM dev3-0 4.78 124.56
04:40:00 PM dev22-2 137.27 10306.18
04:40:00 PM dev22-3 0.00 0.00
04:40:00 PM dev33-0 13.50 410.44
04:40:00 PM dev34-0 13.29 1351.80
04:50:00 PM dev3-0 6.33 194.27
04:50:00 PM dev22-2 111.50 7961.44
04:50:00 PM dev22-3 0.00 0.00
04:50:00 PM dev33-0 5.64 156.51
--------------------
--
Levente "Si vis pacem para bellum!"
On Tue, 2003-04-15 at 20:21, Farkas Levente wrote:> hi, > so after upgrade to the current cvs all of our problem disappear:-) > but we don't have time to test the ssl part... > although after I compile and install the new dovecot I stop it I do a > -------------------- > find /home/*/Maildir -name '.imap*' -exec rm {} \;This is fine, but as you noticed regenerating indexes isn't very fast :) Also currently Dovecot doesn't behave too nicely if the index gets deleted while it was open.> find /home/*/Maildir -name 'dovecot-uidlist' -exec rm {} \;There shouldn't be any need to do this. Point of it was that the UIDs will be remembered if indexes are deleted so client doesn't have to invalidate local cache.> find /home/*/Maildir -name 'msgid.cache' -exec rm {} \;Not created by Dovecot.> find /home/*/Maildir -name '.customflags' -exec rm {}\;Should never be deleted with maildir. This contains user-defined flag names so it could cause some data loss.> I don't know the exact reason bu I assume dovecot is regenerate > everybody's index a other cache at the same time. is there any way to > reduce the possible number of IO load? this would be useful if never > happend again!:-)Currently it re-reads all the mails in the directory and indexes them. I'll make this configurable some day.