Guido: "So what does that benchmark measure? For me, python 3.6 startup is 44 ms and python 2.7 startup is 78 ms (real time; user time is proportionally less)."

It measures something like "time python -c pass".

The performance module creates a virtual environment to run benchmarks to get reproductible timings. If you benchmark the Python startup time using your system Python, the timing depends on the number of installed .pth files.

The code of python_startup benchmark:
https://github.com/python/performance/blob/master/performance/benchmarks/bm_python_startup.py

Note: I didn't write the benchmark, it comes from the old benchmark suite.