Architecting for Performance: Storage - Virtualizing SQL Server with VMware - page 187

Temp DB, and another 100GB for the OS, page file, and so on. In totally, the capacity
required is approximately 2.5TB.
From Table 6.14 , you can see the usable capacity after taking into consideration the
redundant or parity disks of the various RAID types needed to achieve 5,000 IOPS
based on the previous examples. The 2.5TB usable capacity requirement for our sample
SQL Server can easily be met by any of the selected RAID levels based on the number
of disks required to achieve 5,000 IOPS. In fact, all of the RAID levels provide a lot
more capacity than is actually required—some in the extreme.
Table 6.14 Usable Capacity Deployed to Achieve 5,000 SQL IOPS
Table 6.14 shows that a large amount of the deployed usable capacity is actually
unusable from a performance perspective. Or to put it another way, you have way too
much capacity at the end of your performance . This clearly demonstrates Principle 3
of sizing for performance, and in doing so, capacity will usually take care of itself.
Now that we have calculated the usable capacity that needs to be provisioned to
achieve the 5,000 SQL IOPS, we can calculate the IOPS per TB. As mentioned,
previously using IOPS per TB is a good way to communicate with applications teams
how much performance they should expect for each TB of data based on the different
available storage policies available. For this example, we will take a conservative
approach so that the application teams are planning on a worst-case scenario and their
performance surprises will be positive. You'll remember the quote from Principle 3:
“The bitterness of poor performance lasts long after the sweetness of a cheap price is
forgotten.”
To illustrate this, we will define three tiers of storage or storage policies:
Gold —RAID10 300GB 15K RPM disks
Silver —RAID5 7+1 300GB 15K RPM disks
Wood —RAID6 8+2 1TB 7.2K RPM disks
We will base the IOPS per TB calculation on a 30% read-biased IO pattern. This will
mean our DBAs and applications teams will likely get better performance than the