Originally Posted by hdawg
IOPs = Input Output / Sec ... hence the same as disk transfers / sec.
Good to know, that helps!
The multiplier of 3.64 is what RIM has in the doc you previously mentioned ... I always add buffer though.
So we agree that it is a multiplier, and not a static additional IOPS per user number then. My architects believe it is a static number of IOPS per user, and not a multiplier. This is where there is confusion for us.
OK, now I'm going to give you a real live example of one of our regional datacenter Exchange 2007 IOPS requirements analysis. We are architecting for 800 users in one of our regional DC's. 60% BB use within the user population. 800 x .60 = 480 BB users.
So, 480 mailboxes * 3.64 = 1747.20 IOPS
We calculated baseline IOPS at 435.8 for a mix of medium (20%), heavy (60%), and very heavy (20%) user profiles. Total came out to be roughly 2183 IOPS for 800 users in a CCR Exchange cluster using 600MB mailboxes, 100% user activity, 20% maintenance space for the LUN's, etc., and so forth (basically we took all of the standard settings in the Microsoft IOPS calculator). We're looking to use an EVA 8000 for back-end disk. According to the latest HP whitepaper on Exchange 2007 and EVA 8000, each 15k RPM disk will provide approximately 110 IOPS.
2183/110 = 20 disks for the active CCR node, 35 disks for the passive CCR node. Plus, a minimum of 16 disks for the tlogs.
This adds up to 71 disks for Exchange 2007 alone. 800 users/71 disks ~ 11 users per disk. That seems high to me. Given our space requirements, we need to order 300GB 15k RPM fiber channel disks. This is about 100k of cost in disk alone.
If I take the IOPS requirements for BB out of the equation completely, my disk IOPS requirements go down by a factor of ten, down to two disks for the active CCR node and three disks for the passive CCR node. Something just doesn't seem right about BB users multiplying my physical disk requirements by a factor of ten, hence my posts here.
If I use the multiplier as an actual multiplier and NOT as a static number like I did for the numbers above, then here's how it breaks out:
435.8 baseline IOPS
0.48 IOPS per user for Exchange 2007 cached mode very heavy user profile
0.48 x 3.64 = 1.74 additional IOPS per Blackberry user
480 mailboxes x 1.74 = 835.2 IOPS additional for BB user
Total IOPS required: 435.8 + 835.2 = 1271 IOPS
1.37 IOPS per user for Exchange 2007 online very heavy user profile
1.37 x 3.64 = 4.99 additional IOPS per BB user
480 mailboxes x 4.99 = 2395.20 IOPS
Total IOPS required: 435.8 + 2395.20 = 2831 IOPS
The difference between 1271 IOPS, 2183 IOPS (what we ended up using), and 2831 IOPS from a disk purchasing perspective is rather large. It's the difference between 11 disks for the active CCR node and 26 disks for the very same active CCR node, plus the cost of the passive CCR node. The difference is financially significant, hence my questions and conundrums here on this forum.
right ... Check out the Performance Benchmarking guide to see values for 2003 requirements. I would honestly take this data and use it to come up with what you'd need for 2003 ... then take what MSFT says 2007 needs and multiple by the 2003 factor.
Yep, I am VERY familiar with this guide. I've pestered RIM for performance data for Exchange 2007, they don't have anything to share just yet. Once again, thanks very much for your assistance on this matter!