Am 07.06.2013 16:31, schrieb Sage Weil:
On Fri, 7 Jun 2013, Oliver Schulz wrote:
Btrfs is the longer-term plan, but we haven't done as much testing there yet, and in particular, there is a bug in 3.9 that is triggered by a power-cycle and the fixes aren't yet backported to 3.9 stable. Until we have done more validation, we still recommend XFS. The last time we did aging tests on btrfs performance was very good (better than xfs) initially but then trailed off as things fragmented. This was ~3.2 era. We haven't repeated that yet for newer kernels. I suspect it is better now, but I don't know how much better...
I had a ceph cluster running with SSD journals and btrfs (3.8 kernel) and it started to heavily slow down after just a week. After 3 weeks it was unusable (going from 900 iops/cephnode up to 3500 iops/cephnode).
With XFS we have constant 900 iops over month. Also the avg. performance is more constant. With btrfs i had strange up and downs to 0 iop/s for short periods. While under XFS the performance is absolutely constant.
Greets, Stefan _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
