Re: ITS#5860 slapd in openldap-2.4.16 still doesn't respect memory bounds

[Quanah Gibson-Mount <quanah@zimbra.com>]

--On Friday, April 17, 2009 10:25 PM +0300 sscdvp@gmail.com wrote:
> Why do you have so many threads configured?  How many cpu's do you have??
>
>
>  
> > I have tried openldap-2.4.16 (latest) BDB 4.7.25 with all patches from
> > Oracle on Dell AMD64 RAM 4Gb, 8 CPUs, OS Solaris 10.
>  
> I have an application which generates heavy load in slapd (many
> ldap_search'es and very very many ldap_add, ldap_modify, ldap_delete).
> BDB databases are situated on separate disk and BDB logs are on the other
> disk. But slapd performance (I mean time to answer) still cuts
> down during large disk IO periods 'cause these disks are used by other
> applications at the same time. And so I have increased number of threads
> from the defaults threads=32 up to 64, just to save some time for my
> application which requires not more then 30 ms for one search to
> complete.
> Just to be sure I have tried the default settings for threads and it's
> behaviour didn't changed - slapd heap memory just continuously grows (I
> have seen through pmap)...
> Also I have tried libtcmalloc which is usually recommended and solaris
> built in' libmtmalloc, but it didn't helped.

Please copy replies to the bug tracker if you expect help.

I will note that every thread consumes a large chunk of memory.  I will 
also note that the default is 16 threads, not 32.  I will also note, that 
as far as read operations are concerned, usually *fewer* threads is faster, 
not more.  You appear to have quite an excessive amount of threads 
configured.

Others will have to continue the conversation on what you're seeing growth 
wise. ;)

--Quanah

--

Quanah Gibson-Mount
Principal Software Engineer
Zimbra, Inc
--------------------
Zimbra ::  the leader in open source messaging and collaboration