On Tue, 12 Apr 2016 10:53:50 +0200 Alwin Antreich wrote: > > On 04/12/2016 01:48 AM, Christian Balzer wrote: > > On Mon, 11 Apr 2016 09:25:35 -0400 (EDT) Jason Dillaman wrote: > > > > > In general, RBD "fancy" striping can help under certain workloads > > > where small IO would normally be hitting the same object (e.g. small > > > sequential IO). > > > > > > > While the above is very true (especially for single/few clients), I > > never bothered to deploy fancy striping because you have to plan it > > very carefully, as you can't change it later on. > > > > For example if you start with 8 OSDs and set your striping accordingly > > (as Alwin's example suggested) but later add more OSDs you won't be > > taking full advantage of the IOPS availalbe. > > I didn't think about that, thanks for pointing that out. As we have a > mixed workload on our new cluster, VMs and cephfs for login directories > and sources, I am definitely going to test these settings. > Note that once you get to a point where all your OSDs are somewhat busy all the time and/or you have many clients, stripping starts to make somewhat less sense. That said, one way around the issue above is of course to come up with a planned, maximum size of a cluster (like 64 OSDs) and set the stripe count accordingly. This SHOULD result in the same thing with 8 initial OSDs, but I haven't tested it myself. Also there is the question about what happens if you wind up with OSD numbers that are uneven, will the odd one wind up with more data? Test this if you have the chance. Christian -- Christian Balzer Network/Systems Engineer chibi@xxxxxxx Global OnLine Japan/Rakuten Communications http://www.gol.com/ _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |