[Date Prev][Date Next] [Chronological] [Thread] [Top]

Advise on tuning

I have a box (Proliant DL560) with this features:

- 3 x Xeon 2.8 GHz
- 2 x HD SCSI 10k RPM (No RAID)

I have the BDB in one of those HD and the logs in the other one. When I
installed the server I wasn't enough sure about a correct value for
cachesize so I decided to play safe.

set_cachesize 0 1073741824 0
set_lk_max_objects 2500
set_lk_max_locks 2500
set_lk_max_lockers 2500
set_lg_dir /storage/bdb-log/sldcu

accesslog DB_CONFIG
set_cachesize 0 715827883 0
set_lk_max_objects 2500
set_lk_max_locks 2500
set_lk_max_lockers 2500
set_lg_dir /storage/bdb-log/accesslog

I was reading this[1] yesterday and found out that a I have a very large
value. For my dc=sld,dc=cu DB I have a 6.2 MB dn2id.bdb and a 56 MB
id2entry.bdb, the DIT contains about 60k entries.

Normally I larger value is a Good Thing(TM) but I'm specially worried
about slapd using around 40%-50% time of (one) CPU and eventually a
message of 'deferring operation' gets logged.

The server is attending 30 connections per second sustained during
working hours. Read Waiters are about 100 sustained during this same
period. Yesterday I sent a message to the list asking for opinions about
that value for Read Waiters because it doesn't sound good to me, but
well, I know that 'sounds good' is very subjective.

I was running 2.3.38 yesterday morning and upgraded to 2.4.11. With the
same setup (only the slapd changed) the mail service got deeply
affected. Dovecot began to take several (+10) seconds to authenticate,
in other cases, the POP3/IMAP connection was dropped due to timeout.

While Dovecot was failing my slapd.log were solid filled by 'deferring
operation' messages. I stopped Dovecot, the messages dissapears, I
started Dovecot, the message began again to fill logs.

So, I thank it could be some problem on Dovecot side (although it was
playing nice with slapd 2.3.38).

Thanks God I hit man slapd-hdb and learn about cache tuning values
there. I then raised olcDbCacheSize / olcDbDNCacheSize /
olcDbIDLCacheSize and things began to calm down.

I wonder how slapd 2.3.38 were nicely serving Dovecot without the tuning
I made.

On the other hand I have the suspiction, based on benchmark data
published on the list, that there is something wrong with that large
cache plus 50% CPU for less than 30 connections per second.

I very much appreciate any help/advise/hint someone could give me about
this issue.


[1] http://www.openldap.org/faq/data/cache/1075.html