[lustre-discuss] lester on mds with low memory

[Alexander Oltu]

> On 27 Mar 2015, at 16:52, Mohr Jr, Richard Frank (Rick Mohr) <rmohr at utk.edu> wrote:
> 
> 
>> On Mar 27, 2015, at 4:38 AM, Alexander Oltu <Alexander.Oltu at uib.no> wrote:
>> 
>> yes, right before starting Lester. I tried sending 1 and sending 3 to /proc/sys/vm/drop_caches, maximum I can get is 1.5 GB.
>> 
>> I don’t know where the buffer memory is being used. Running “sync” didn’t change anything. I was thinking maybe this is something Lustre related.
> 
> I wonder if it is possible that you are hitting LU-5726.  You might want to check the “Inactive(file)” line in /proc/meminfo to see if that accounts for most of the buffer usage.  When I was hitting this bug, dropping caches would have no effect and I could not reclaim memory on my MDS server.  After I applied the patch, I could consistently run “echo 3 > /proc/sys/vm/drop_caches” and reclaim almost all the memory on the MDS.

Thanks Rick. The LU-5726 description looks very familiar to what is happening on the MDS. Unfortunately it crashed this weekend due to OOM, so I can’t check Inactive pages now, will monitor it closely, especially in the light that we are going to run massive purge on this filesystem.

Thanks,
Alex.