The other day I posted an article on Exchange disk performance and something that was puzzling me. Stu assisted me in finding the correct answer (because the last one was a little flawed).
The flaw was in some information that I didn’t communicate, which lead me to a wrong conclusion. I concluded that the number of disks required was double what the calculation for RAID 0+1 had produced.
I think what I had done was to assume, in my head, that in a mirrored pair that it was only the ‘front’ disk in the mirror that responded to reads, which is of course ridiculous (for most modern hardware). On top of this I didn’t communicate where I got the RAID overhead ratios from.
Clarifying the RAID overhead ratios. For Exchange database storage the ratio of reads to write is something like 3 to 1 according the Optimising Storage for Exchange 2003.
In a RAID 0+1 infrastructure a write requires 2 IO’s. So for 4 IO operations you actually do 5 disk IO’s (3 reads and 2 writes) giving you the RAID impact of .8 (you get 80% of the performance from the volume of disks that you have).
In a RAID 5 configuration a write require 4 IO’s. So for 4 IO operations you actually undertake 7 disk IO’s (3 reads and 4 writes) giving you a RAID impact of .57 (you get 57% of the performance from the volume of disks that you have).
So my calculations were actually correct, but I made the wrong conclusion because of a false assumption – not for the first time, unfortunately not for the last either.
So in order to support 1000 concurrent users you need:
- RAID 1 = 7 disks (well 8 actually because you can’t have an odd number)
- RAID 0+1 = 7 disks NOT 14 as a first concluded (well 8 actually because you can’t have an odd number)
- RAID 5 = 9 disks
“One pound of learning requires ten pounds of common sense to apply it”. Persian Proverb