On Fri, 23 Dec 2016 11:24:13 -0800
Doug Dumitru [off-list ref] wrote:

With hard drives, I suspect your single disk tests are taking
advantage of the disks' on-controller cache and is doing read-ahead
and thus streaming.  With the array in place, you are probably doing
512K reads (check the array chunk size) so the disks will see bursts
of 512K reads with big gaps.  The gaps are large enough that the
rotation has gone too far and the caching makes you wait a rotation.
This is just a guess.

You can compensate for that via setting a higher read-ahead for the array
itself, such as 

  blockdev --setra 8192 /dev/mdX 

(or more)

-- 
With respect,
Roman