On Sun, Dec 28, 2014 at 1:25 AM, Lindsay Mathieson <lindsay.mathieson@xxxxxxxxx> wrote: > On Sat, 27 Dec 2014 06:02:32 PM you wrote: >> Are you able to separate log with data in your setup and check the >> difference? > > Do you mean putting the OSD journal on a separate disk? I have the journals on > SSD partitions, which has helped a lot, previously I was getting 13 MB/s > No, I meant XFS journal, as we are speaking about filestore fs performance. > Its not a good SSD - Samsung 840 EVO :( one of my plans for the new year is to > get SSD's with better seq write speed and IOPS > > I've been trying to figure out if adding more OSD's will improve my > performance, I only have 2 OSD's (one per node) Erm, yes. Two OSDs cannot be considered even for a performance measurement testbed setup, neither should three or any other small number. This explains numbers you are getting and impact from nobarrier option. > >> So, depending on type of your benchmark >> (sync/async/IOPS-/bandwidth-hungry) you may win something just for >> crossing journal and data between disks (and increase failure domain >> for a single disk as well ). > > One does tend to foxus on raw seq read/writes for becnhmarking, but my actual > usage is solely for hosting KVM images, so really random R/W is probably more > important. Ok, then my suggestion may not help as much as it can. > > -- > Lindsay _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |