Here is an update to a previous post that now includes a graph along with some additional quick tests. Even though the MD3000 line is getting a bit old, I got a killer deal on a unit had some free time for testing.
Quick info: Dell PowerEdge 2950 with PERC 5/e SAS card dual connected to a Dell MD3000 disk array with qty 15 73GB 15K drives running CentOS 5.5 all with kernel 2.6.18-194.26.1.el5. The MD3000 has 1 global host spare setup and is performing the RAID 1, 5 or 6 portion of the test and linux md is the RAID 0 part. The MD3000 really only performs well when both controller are active in all of the tests that I have run. In addition, mirrorCacheEnabled=FALSE is set on the MD3000 controllers.
A picture does speak a thousand words. Higher is better.
Very nice for ext4! I did test with the "nobarrier" option just to make sure there was not the performance regression in this kernel. The issue does not appear in this kernel since the iozone numbers were right on.
As a note, this is mostly "stock" setup changing the RAID types. You *can* have even better performance numbers by playing with simply items like " blockdev --setra xxxx". For example, a change of "blockdev --setra 8192" produced about a 20% increase under ext4 for both read and re-read in iozone ("--setra 32768 got 30%)! Also, a scheduler change to "deadline" may help depending on your workload.