But it's not just interpreting linear code. It is also classloading which involves a non-trivial amount of work before any code in a class can even run. Also I never said the testing code doesn't have any loops. There often are loops but they need to run for a really long time until hotspot kicks in. So if you lose a few tens of millisecond here and there in a single test, you will get minutes or hours in total in a codebase that has hundreds of thousands tests. And it gets terribly worse if many of your tests are end-to-end tests that e.g. launch new JVMs.