Jupyter notebook %timeit 1000 times slower than command line, why?

Question

I'm trying to time a python 3.6. program from jupyter notebook, but it seems like the magic command %timeit adds a lot of extra overhead, giving me wrong stats.

From Jupyter notebook:

%timeit a=1
10000000 loops, best of 3: 84.1 ns per loop

From cmdline

python -m timeit 'a=1'
100000000 loops, best of 3: 0.0163 usec per loop

So in this case the command line timeit runs millions of times faster than the jupyter notebook timeit. What is the reason for this, and is there a way to fix it so timeitfrom jupyter notebook can give correct measurements?

Answer 1

You are not reading those numbers correctly. IPython is reporting timings in nanoseconds (note the ns abbreviation). Python is reporting the timings in microseconds (usec).

1 microsecond is 1000 nanoseconds; normalising to nanoseconds Python reported 16.3 nanoseconds, so it was only 5 times as fast.

However, I can't reproduce your findings. Using the same Python binary in a virtualenv to both run IPython and directly:

venv-3.6 $ bin/ipython
Python 3.6.2 (default, Jul 18 2017, 14:26:50)
Type "copyright", "credits" or "license" for more information.

IPython 5.2.2 -- An enhanced Interactive Python.
?         -> Introduction and overview of IPython's features.
%quickref -> Quick reference.
help      -> Python's own help system.
object?   -> Details about 'object', use 'object??' for extra details.

In [1]: %timeit a=1
100000000 loops, best of 3: 11.9 ns per loop

In [2]:
Do you really want to exit ([y]/n)? y

venv-3.6 $ bin/python -m timeit 'a=1'
100000000 loops, best of 3: 0.0121 usec per loop

and in a Jupyter notebook, again with the same virtualenv; this essentially drives ipython, so as expected there is no real difference:

%timeit a=1
100000000 loops, best of 3: 11.8 ns per loop

So thats 11.9 vs 12.1 vs 11.8 nanoseconds; too close to call a difference.