Hi, On 07/21/13 09:05, Dan van der Ster wrote:
This is with a 10Gb network -- and we can readily get 2-3GBytes/s in "normal" rados bench tests across many hosts in the cluster. I wasn't too concerned with the overall MBps throughput in my question, but rather the objects/s recovery rate --
they are necessarily linked, and that was too close to 1G for not mentioning it. But then it means there's another cap somewhere...
I assumed that since we have so many small objects in this test system that the per-object overhead is dominating here... or maybe I am way off??
I'm not sure, I haven't read ceph's source code, but I think the overhead is (mostly) to the PG count in a cluster. Do you monitor your cluster with something like munin ? What remains after the network is the drive utilisation/latency (look at "iostat -mx 5" for instance) and the CPU usage (using top for instance).
_______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
