Exchange and RAID Levels

I have recently been trying to put together the logic and reasoning behind the clear Microsoft recommendation that Exchange 2003 services are hosted on RAID 1+0 and not RAID 5. RAID 5 being for those people who had money constraints. All of the reasoning stated seems to be based on the need for IO performance.

Due to the way that a write occurs each RAID level can be given a factor. If RAID 0 is a factor of 1, then RAID 1 (or 1+0) is 0.8 in terms of performance and RAID 5 is 0.57.

So the logic goes a bit like this:

Each online user does something like 0.5 IOPS and a 10K drive does about 100 IOPS. So a single drive can support 200 concurrent users. With RAID factors built in the users per drive are 160 for RAID 1 and only 114 for RAID 5. So you get better performance from RAID 1 disks. But here is the rub; no-one would implement RAID 1 they would all implement RAID 1+0 so you actually only get 80 users per disk for RAID 1+0, or in other words you need two disks to get the 160. In a RAID 5 set the resilience is already factored into the performance numbers.

So in order to support 1000 concurrent users you need:

• RAID 1 = 7 disks
• RAID 1+0 = 14 disks
• RAID 5 = 9 disks

What’s more, as every Exchange administrator knows, you need loads of spare space on a server to do all of those maintenance tasks, and also it adds a little to the performance. With RAID 1+0 you get 7 disks worth, with RAID 5 you get 8 disks worth of storage.

Having blown the IOPS argument out of the water, there must be something more to it, and this is where I think that the real reasoning comes in. In the case of a failure a RAID 5 set will slow down significantly because all of its information is having to be calculated from parity information on the other disks. A RAID 1+0 set will not suffer from such an overhead because it isn’t having to calculate from parity information, it’s just reading the other disk that is a mirror of the one that has failed. The really worrying part is that I suspect Microsoft have focussed on the IO issue because they don’t want money strapped managers to implement RAID 5 because the disaster tolerance sell isn’t high on their priority list.