I think the really big thing is to actually create some infrastructure around your product to run performance tests whenever you're developing a feature. That's the only way you're ever going to good data.
As an example, the SQL tests will act very differently depending on if the table was in the buffer pool, or it had to be fetched from disk (I wrote my own tool to run tests on MySQL if anyone is interested, https://github.com/arianitu/sql-stress)