I had a lot of "slow MDS request" (MDS, nos OSD) when writing a lot of
small files to cephfs.
The problem was an I/O stuck when flushing XFS buffers. I had the same
problem in other (non ceph) xfs systems when asking for a lot of inode
operations (rm -rf for example).
The solution was to destroy the filestore (xfs) OSDs and recreate them
in bluestore. Is a slow process. First you have to stop the OSDs, wait
for rebalancing, destroy, recreate, and go on with the next OSD. It took
a 10 days to recreate 8 OSDs with 3Gb each (In production, with the RBD
and CEPHFS services in use).
Hope this helps.
--
Alfredo Daniel Rezinovsky
Director de Tecnologías de Información y Comunicaciones
Facultad de Ingeniería - Universidad Nacional de Cuyo
null
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
