Feature #5437
closed
ceph-mon performance on ARM
Added by Mark Nelson almost 11 years ago.
Updated about 10 years ago.
Description
Scaling tests with 168 OSDs show bottlenecks when ceph-mon is running on ARM processors. Requests are not being processed quickly enough to keep OSDs from dropping at default heartbeat and timeout intervals during PG creation. This can lead to further overhead causing the cluster to never become healthy. Overhead in the monitor and improved scaling may be necessary to support large deployments with monitors on ARM processors.
Latest wip-mon-pgmap tests got through PG creation without issue. The mons have been quite stable on ARM with 168 OSDs through PG creation and read/write tests. CPU usage varied from about 20-140% during activity and idled lower. Memory usage per mon was about 600-900MB. mon usage did go up during small reads and writes, but not enough to severely limit performance. With more clients and more OSDs it is possible we could see some performance limitations with the mons, but at least at this level we are now seeing good stability and behavior.
- Priority changed from Urgent to Normal
Downgraded to Normal, due to landing of patch. Keeping open as a placeholder for any future issues identified for improving ARM performance.
- Status changed from New to Resolved
Also available in: Atom
PDF