We do this to maximize the time spent to actual operations in comparison to other expenses. The most important one among of such expenses is transaction commit. For example, if we would measure each operation alone, transaction commit time there should be much bigger than the operation time itself.
On the other hand, we can't exclude transaction commit time from tests, because this is the only way to ensure all the changes were really sent to RDBMS: many ORM tools can delay sending the changes until certainly moments. Moreover, similar "delaying" could happen on any other layer, such as client-server connection protocol and even database server itself. So transaction commit is the easiest way to ensure everything is done.
Note: we do not disable write caches on machine we run tests on. So transaction commit there is relatively fast (~ up to 5000 commits per second). You can see the same result on the production system only if there is RAID controller with independently powered write cache or something similar. On the other hand, this is even better for purposes of our tests.
