On Thu, 7 Jul 2016 12:53:33 +0100 Nick Fisk wrote: > Hi All, > > > > Does anybody else see a massive (ie 10x) performance impact when either > deleting a RBD or running something like mkfs.xfs against an existing > RBD, which would zero/discard all blocks? > > > > In the case of deleting a 4TB RBD, I'm seeing latency in some cases rise > up to 10s. > > > > It looks like it the XFS deletions on the OSD which are potentially > responsible for the massive drop in performance as I see random OSD's in > turn peak to 100% utilisation. > > > > I'm not aware of any throttling than can be done to reduce this impact, > but would be interested to here from anyone else that may experience > this. > I haven't tested this since firefly and found RBD deletions, discards and snapshots all to be very expensive operations.
See also: http://ceph.com/planet/use-discard-with-krbd-client-since-kernel-3-18/ I would think that the unified queue in Jewel would help with this. But how much this is also an XFS amplification and thus not helped by proper queuing above I can't tell, all my production OSDs are Ext4. Christian -- Christian Balzer Network/Systems Engineer [email protected] Global OnLine Japan/Rakuten Communications http://www.gol.com/ _______________________________________________ ceph-users mailing list [email protected] http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
