In my opinion when you want such precision, you need to stablish strict constraints to the measurements, for example memory fences: https://www.kernel.org/doc/Documentation/memory-barriers.txt
If you do not do this, the times will never be consistent.
The author produced a faulty benchmark.