Re: HP-UX performance problem

[Gordon Guthrie <gdg@soco.agilent.com>]

> 
> 
> --On Monday, March 14, 2005 4:25 PM -0800 Gordon Guthrie 
> <gdg@soco.agilent.com> wrote:
> 
> > I've discovered a performance problem with openldap 2.2.20 running on
> > HP-UX 11i.
> >
> > If logging level is turned up high enough,
> > bdb(o=xyzzy.com): unable to lock mutex: Invalid argument
> > shows occasionally in the log during a search for 2900 entries.  This
> > search takes ~ 1.5 sec.
> >
> > However if I make the same search simultaneously from 2 different
> > machines, I get > 100000 "unable to lock mutex: Invalid argument"
> > messages and the elapsed time goes to > 55 sec.
> >
> > I discovered that removing PHKL_25367 restores the performance.  The
> > mutex entries disappear, and the search takes 1.2 sec for a the
> > single query and 1.9 sec for the double.
> >
> > PHKL_25367 has been supeseeded by PHKL_30837 which has worse performance
> > The single search takes 17 sec with > 5800 mutex entries while the
> > double search takes >> 23 min with > 27000 mutex entries.
> > (I abandonded the double search after 23 min).
> >
> >
> > These experiments were performed on openldap 2.2.20 which used
> > Berkley DB 4.2.52, but I've seen the same symptoms with
> > openldap 2.2.23 using Berkley DB 4.3.27.
> >
> > Any idea what is going on?
> 
> What does your DB_CONFIG file look like?

set_cachesize 0 134217728 0 

#   For Loading indexing only
#set_flags       DB_TXN_NOSYNC

set_lg_dir     /hfs/d1/agux/logs/openldap2220

> 
> What does your OpenLDAP slapd.conf file look like?

It's 225 lines long.  Before or while I sanitize it, what lines are of 
interest?

Guessing about the most significant ones:

#################################
#   database definitions
################################
database        bdb
...


cachesize    50000
idlcachesize 150000


> 
> Is your version of BDB 4.2.52 correctly patched?

If you mean the sleepycat patches, yes.
If you mean the patch you support, no.

> 
> And, avoid BDB 4.3. ;)  OL 2.2.23 works just fine with BDB 4.2.52 +Patches.

Right, I did get caught by the README note, but that experiment was just
to show that OL 2.2.23 hadn't fixed the problem.  Virtually all of 
my tests are on OL 2.2.20 and BDB 4.2.52.


Thanks for the attention,

Gordon