Ceph

Since sync thread will cause unstable iops and latency performance curve, we may want to WBThread do more(or moderate?) writeback and avoid sync thread flush too much which will cause journal io long tail.

It looks difficult to leverage writeback and sync thread, maybe we can firstly add some perf counter to WBThrottle. For example, we may want to know during a period, how much object or bytes are write backed by WBThread and how much are flushed by SyncThread. Via viewing these counter, we can have a great tuning for "*bytes_start_flusher", "*ios_start_flusher" and "*inodes_start_flusher". What we want to see is that in-memory data can writeback into disk with a moderate rate.

So maybe we can have a more stable latency curve?