OOM-killer going crazy. (was: Re: memory not released after using cdrecord/cdrdao)

[Jan-Frode Myklebust <janfrode@parallab.uib.no>]

On Sun, Jul 25, 2004 at 09:30:38PM -0400, Ed Sweetman wrote:
> 
> Indeed, i burned a smaller cd and got very similar results.  

Same here.. After upgrading to 2.6.8-rc2 the OOM-killer is going crazy.
It's particularly angry at the backup client 'dsmc' (from Tivoli Storage
Manager).  I'm monitoring its usage with 'top', and 'dsmc' is not using
more than ~150MB in either size or RSS when the OOM-killer takes it down.

The 'dsmc'-process is reporting that it's processed 2,719,000 files, and
transfered 164.34 MB when it gets killed. i.e. it's traversed a lot of
files, but only read about 164 MB data, so it shouldn't have filled up any
buffer cache... 

The system still has lots of free memory (~900 MB), and also 2 GB of
unused swap. Actually there's 0K used swap..??  

I've tried turning on vm.overcommit_memory, but it had no effect. Also
tried changing the swappiness both up to 90% and down to 10%, but it
never uses any swap.. ???

BTW: I had no OOM-killer problems on 2.6.7.


  -jf