--On Tuesday, September 28, 2010 11:10 PM +0000 quanah@zimbra.com wrote:

--On Tuesday, September 28, 2010 8:04 PM +0200 masarati@aero.polimi.it wrote:

...

Each paged results control request mallocs a small amount of memory on the slab which is not explicitly freed. It should not be an issue, since it occurs so early in the life of an operation that the slab should never be exhausted. In any case, I've committed a fix. Can you test and report, please? Thanks, p.

Hi Pierangelo,

I'm testing a build with your patch now. In the meantime, you can see a callgraph generated from tcmalloc's heap profiling tool at http://www.stanford.edu/~quanah/pprof.pdf. I baselined the graph to start after the first initial searches have been to load the cache, etc. So you can see that slapd grew by 14.3GB once I started doing the paged results searches.

It appears to be a problem with the entry cache, which is set to 25,000:

olcDatabase={2}hdb.ldif:olcDbCacheSize: 25000

yet:

olmBDBEntryCache: 209759

from cn=monitor.

--Quanah

--

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