Hi Massimo,

On Thu, Oct 30, 2025 at 2:10 PM Massimo Sgaravatto
<[email protected]> wrote:
>
> Hello Venky
>
> Thanks for your help in debugging this issue
>
> I am using the default value for mds_cache_memory_limit (4 GiB).
> Should I increase this value since the server hosting the MDS has much more 
> memory ?

Yes, that's generally recommended. The defaults are pretty low for any
real production use. Note that the MDS can still use more than this
configured cache limit (upto 150%), but it should start trimming its
cache when it is nearing the cache memory limit.

>  Are there some guidelines on how to set this parameter wrt the physical 
> memory available in the server hosting the mds daemon ? I can't find this in:
>
> https://docs.ceph.com/en/reef/cephfs/cache-configuration/

Unfortunately, no. There have been requests to add some kind of a
performance tuning guide which I think should be available in the near
future.

>
>
> Thanks, Massimo
>
>
>
> On Thu, Oct 30, 2025 at 9:07 AM Venky Shankar <[email protected]> wrote:
>>
>> Hi Massimo,
>>
>> On Wed, Oct 29, 2025 at 8:19 PM Massimo Sgaravatto
>> <[email protected]> wrote:
>> >
>> > Hi Venky
>> > Thanks for your answer
>> >
>> > No: we are not using snapshots
>>
>> Was the MDS cache memory running close to mds_cache_memory_limit? This
>> is available in perf dump via
>>
>>         $ ceph tell mds.<id> perf dump
>>
>> and look for mds_co_bytes. Or if you can recreate the issue and
>> capture these details.
>>
>> And what's mds_cache_memory_limit set to BTW?
>>
>> >
>> > Regards, Massimo
>> >
>> > On Wed, Oct 29, 2025 at 3:05 PM Venky Shankar <[email protected]> wrote:
>> >>
>> >> Hi Massimo,
>> >>
>> >> On Tue, Oct 28, 2025 at 5:30 PM Massimo Sgaravatto
>> >> <[email protected]> wrote:
>> >> >
>> >> > Dear all
>> >> >
>> >> > We have a portion of a Cephfs file system that maps to a ceph pool 
>> >> > called
>> >> > cephfs_data_ssd.
>> >> >
>> >> >
>> >> > If I perform a "du -sh" on this portion of the file system, I see that 
>> >> > the
>> >> > value matches the "STORED" field of the "ceph df" output for the
>> >> > cephfs_data_ssd pool.
>> >> >
>> >> > So far so good.
>> >> >
>> >> > I set a quota of 4.5 TB for this file system area.
>> >> >
>> >> >
>> >> > During the weekend, this pool (and other pools of the same device class)
>> >> > became nearfull.
>> >> >
>> >> > A "ceph df" showed that the problem was indeed in the the 
>> >> > cephfs_data_ssd
>> >> > pool, with a reported usage of 7 TiB of data (21 TiB in replica 3):
>> >> >
>> >> > cephfs_data_ssd 62 32 7.1 TiB 2.02M 21 TiB 89.06 898 GiB
>> >> >
>> >> >
>> >> > This sounds strange to me because I set a quota of 4.5 TB in that area, 
>> >> > and
>> >> > because a "du -sh" of the relevant directory showed a usage of 600 GB.
>> >> >
>> >> >
>> >> > When I lowered the disk quota from 4.5 TB to 600 GB, the jobs writing in
>> >> > that
>> >> > area failed (because of disk quota exceeded) and after a while the 
>> >> > space was
>> >> > released.
>> >> >
>> >> >
>> >> > The only explanation I can think of is that, as far as I understand,
>> >> > cephfs can take a while to release the space for deleted files
>> >> > (https://docs.ceph.com/en/reef/dev/delayed-delete/).
>> >> >
>> >> >
>> >> > This would also be consistent with the fact that it looks like some jobs
>> >> > were performing a lot of writes and deletions (they kept writing a ~ 5GB
>> >> > checkpoint file, and deleting the previous one after each iteration).
>> >>
>> >> That's likely what is causing the high pool usage -- the files are
>> >> logically gone (du doesn't see them), but the objects are still lying
>> >> in the data pool consuming space which aren't getting deleted by the
>> >> purge queue in the MDS for some reason. Do you use snapshots?
>> >>
>> >> >
>> >> >
>> >> >
>> >> > How can I understand from the log files if this was indeed the problem ?
>> >> >
>> >> > Or do you have some other possible explanations for this problem ?
>> >> >
>> >> > And, most important, how can I prevent scenarios such as this one ?
>> >> >
>> >> > Thanks, Massimo
>> >> > _______________________________________________
>> >> > ceph-users mailing list -- [email protected]
>> >> > To unsubscribe send an email to [email protected]
>> >> >
>> >>
>> >>
>> >> --
>> >> Cheers,
>> >> Venky
>> >>
>>
>>
>> --
>> Cheers,
>> Venky
>>


-- 
Cheers,
Venky
_______________________________________________
ceph-users mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to