Sometimes an OSD may start performing very poorly for no apparent reason. If there is nothing obvious being revealed by your monitoring tools, consult ceph.log and the Ceph health detail output. You can also run Ceph osd perf, which will list all the commit and apply latencies of all your OSDs and may also help you identify a problematic OSD.
If there is a common pattern of OSDs referenced in the slow requests, then there is a good chance that the mentioned OSD is the cause of the problems. It is probably worth restarting the OSD in case that resolves the issue; if the OSD is still problematic, it would be advisable to mark it out and then replace the OSD.