It's a rather annoying problem in general. When your view into your software alters the behavior of the software, how can you be sure you're measuring actual program behavior?

A healthy distrust in profilers is a good starting point. Having decent external measurement and using it to test with/out profiler attached is also recommended. Some profilers are less intrusive than others. In this particular case the issue should be visible in GC logs, which should always be monitored.