Ceph

- generate some write/overwrite workload with many concurrent writes
- extend ceph_manager to pause (kill STOP) an osd process
pause one replica, wait long enough for the other to commit several writes
- kill the unpaused one
- unpause the paused one
- let it recover, and continue writing
- restart killed osd

This should at least exercise the divergent log handling. Not sure how we can verify the result...