When running the full compiler regression suite it can take a lot of memory.  Randomly sort the regression test order to make it more likely that the times will iron out (since similar named tests tend to have similar overheads, so you want to avoid executing all the very intensive queries at the same time).