On 04/23/2012 12:20 PM, Russell Beall wrote:

On Apr 23, 2012, at 10:28 AM, Rich Megginson wrote:

That's very interesting.  Does Sun DS have some sort of tuning parameter for number of values?  That is, they may have some threshold for number of values in an attribute - once the number hits that threshold, they may switch to using some sort of ADT to store the values, like a AVL tree or hash table, rather than the simple linked list used by default.

I've compared the dse.ldif of both servers looking specifically for attributes that I should transfer from our production environment to 389.  The configurations for major components are virtually identical and I have seen no attribute that relates to the number of values in a multi-valued attribute.  I expect that the optimization is a behind-the-scenes code improvement.

Also during this testing I have noticed a memory leak when running large quantities of ldapmodify operations.  When I set up a loop to delete and then re-add the eduPersonEntitlement attribute across 100K entries, I found that memory consumption continuously increased and the server crashed after the fifth iteration of this loop.  (And this one really is with ldapmodify and is not related to my earlier issues with excessively creating tombstones by deleting and adding entire entries).  Before digging into this too deeply and making another ticket, I wanted to ask if this had been noticed and fixed in the 1.2.10 release?  I am using the default release. I'm guessing it hasn't since I didn't see it in the release notes.

Try increasing your nsslapd-cachememsize and monitoring it closely.  Using the size of id2entry.db4 is a good place to start, but that will not be enough.

Early on in the process of setting up 389 I optimized the cachememsize.  I configured a 12G cache, and the cache usage after loading all 600K entries is just under 10G.  While the ldapmodify operations are in progress, I am pretty sure I did not have an increase in the cacheentryusage monitor attribute under cn=config, but I'd have to re-check to be sure.

You will see an increase due to replication metadata and possibly other factors.

Unfortunately, with valgrind attached, the server uses much extra memory on startup and does not complete the startup operation before running out of memory on my 32GB machine.  I have had to reduce the cachememsize so that it will start.  It's been starting up for two hours and finally stopped allocating more memory at 24G (with only a 3G cachememsize configured).  I'll probably have to delete out a large quantity of entries to run the test within the bounds of the cachememsize.

Ok - so valgrind is probably not an option.


See also https://fedorahosted.org/389/ticket/51 and https://bugzilla.redhat.com/show_bug.cgi?id=697701

This bug appears very different from what I am looking at.  The ldapmodify I run makes a single connection and transmits a large file of operations to perform value deletions on 100K entries, followed by a new connection to transmit value additions to 100K entries contained within a single large file, and then loop around to do the same thing again.  This emulates the behavior of our directory synchronization script which calculates large quantities of necessary modifications and then submits them all in an ldif file.

The thing in common is this - when the cache usage hits the cache max size, you see unbounded memory growth.

I am starting up the server with the valgrind command you recommended a few messages back to see if I can spot the leak, though of course with valgrind in the mix, the overhead and runtimes are, as might be expected, much increased.

Yes, and valgrind will report many false positives that are hard to weed through.

The issue you are seeing may not be a memory leak per se - see the ticket/bug above.

Ok.  I'll see if there is anything I can pull from the rough.


389 users mailing list