That's the problem we had - some results were better some were worse but from server to server the better or worse ones were not the same.

At this stage I tested only on non production systems so that I could test the results outside of working time to ensure my load testing was the only thing running on both the before and the after tests so that the comparison was valid.

For instance this test;

Test 5 - Sequential writing for 360 seconds using 8k blocksize and 8 outstanding I/O requests per thread

This one varied from +10% to -84% in the 4 separate arrays we tried.

We ran 16 separate runs - Random Writing, Sequential Writing, Random Reading & Sequential Reading all using each of 8k, 64k, 128k & 256k blocksizes.

I'm no expert on SQLIO but I do believe the tests were done in a representative manner so was a bit disappointed when the magic wand didn't work 😉