Skip to content

Commit

Permalink
Issue python#13528: rework the performance question in the programmin…
Browse files Browse the repository at this point in the history 
…g FAQ
  • Loading branch information
@pitrou
pitrou committed Dec 9, 2011
1 parent 720f34a commit 432259f
Showing 1 changed file with 62 additions and 153 deletions.
215 changes: 62 additions & 153 deletions Doc/faq/programming.rst
View file
Original file line number Diff line number Diff line change
Expand Up @@ -115,159 +115,6 @@ Yes. The coding style required for standard library modules is documented as
:pep:`8`.


My program is too slow. How do I speed it up?
---------------------------------------------

That's a tough one, in general. There are many tricks to speed up Python code;
consider rewriting parts in C as a last resort.

`Cython <http://cython.org>`_ and `Pyrex <http://www.cosc.canterbury.ac.nz/~greg/python/Pyrex/>`_
can compile a slightly modified version of Python code into a C extension, and
can be used on many different platforms. Depending on your code, Cython
may be able to make it significantly faster than when run by the Python
interpreter.

The rest of this answer will discuss various tricks for squeezing a bit more
speed out of Python code. *Never* apply any optimization tricks unless you know
you need them, after profiling has indicated that a particular function is the
heavily executed hot spot in the code. Optimizations almost always make the
code less clear, and you shouldn't pay the costs of reduced clarity (increased
development time, greater likelihood of bugs) unless the resulting performance
benefit is worth it.

There is a page on the wiki devoted to `performance tips
<http://wiki.python.org/moin/PythonSpeed/PerformanceTips>`_.

Guido van Rossum has written up an anecdote related to optimization at
http://www.python.org/doc/essays/list2str.html.

One thing to notice is that function and (especially) method calls are rather
expensive; if you have designed a purely OO interface with lots of tiny
functions that don't do much more than get or set an instance variable or call
another method, you might consider using a more direct way such as directly
accessing instance variables. Also see the standard module :mod:`profile` which
makes it possible to find out where your program is spending most of its time
(if you have some patience -- the profiling itself can slow your program down by
an order of magnitude).

Remember that many standard optimization heuristics you may know from other
programming experience may well apply to Python. For example it may be faster
to send output to output devices using larger writes rather than smaller ones in
order to reduce the overhead of kernel system calls. Thus CGI scripts that
write all output in "one shot" may be faster than those that write lots of small
pieces of output.

Also, be sure to use Python's core features where appropriate. For example,
slicing allows programs to chop up lists and other sequence objects in a single
tick of the interpreter's mainloop using highly optimized C implementations.
Thus to get the same effect as::

L2 = []
for i in range(3):
L2.append(L1[i])

it is much shorter and far faster to use ::

L2 = list(L1[:3]) # "list" is redundant if L1 is a list.

Note that the functionally-oriented built-in functions such as :func:`map`,
:func:`zip`, and friends can be a convenient accelerator for loops that
perform a single task. For example to pair the elements of two lists
together::

>>> list(zip([1, 2, 3], [4, 5, 6]))
[(1, 4), (2, 5), (3, 6)]

or to compute a number of sines::

>>> list(map(math.sin, (1, 2, 3, 4)))
[0.841470984808, 0.909297426826, 0.14112000806, -0.756802495308]

The operation completes very quickly in such cases.

Other examples include the ``join()`` and ``split()`` :ref:`methods
of string objects <string-methods>`.

For example if s1..s7 are large (10K+) strings then
``"".join([s1,s2,s3,s4,s5,s6,s7])`` may be far faster than the more obvious
``s1+s2+s3+s4+s5+s6+s7``, since the "summation" will compute many
subexpressions, whereas ``join()`` does all the copying in one pass. For
manipulating strings, use the ``replace()`` and the ``format()`` :ref:`methods
on string objects <string-methods>`. Use regular expressions only when you're
not dealing with constant string patterns.

Be sure to use the :meth:`list.sort` built-in method to do sorting, and see the
`sorting mini-HOWTO <http://wiki.python.org/moin/HowTo/Sorting>`_ for examples
of moderately advanced usage. :meth:`list.sort` beats other techniques for
sorting in all but the most extreme circumstances.

Another common trick is to "push loops into functions or methods." For example
suppose you have a program that runs slowly and you use the profiler to
determine that a Python function ``ff()`` is being called lots of times. If you
notice that ``ff()``::

def ff(x):
... # do something with x computing result...
return result

tends to be called in loops like::

list = map(ff, oldlist)

or::

for x in sequence:
value = ff(x)
... # do something with value...

then you can often eliminate function call overhead by rewriting ``ff()`` to::

def ffseq(seq):
resultseq = []
for x in seq:
... # do something with x computing result...
resultseq.append(result)
return resultseq

and rewrite the two examples to ``list = ffseq(oldlist)`` and to::

for value in ffseq(sequence):
... # do something with value...

Single calls to ``ff(x)`` translate to ``ffseq([x])[0]`` with little penalty.
Of course this technique is not always appropriate and there are other variants
which you can figure out.

You can gain some performance by explicitly storing the results of a function or
method lookup into a local variable. A loop like::

for key in token:
dict[key] = dict.get(key, 0) + 1

resolves ``dict.get`` every iteration. If the method isn't going to change, a
slightly faster implementation is::

dict_get = dict.get # look up the method once
for key in token:
dict[key] = dict_get(key, 0) + 1

Default arguments can be used to determine values once, at compile time instead
of at run time. This can only be done for functions or objects which will not
be changed during program execution, such as replacing ::

def degree_sin(deg):
return math.sin(deg * math.pi / 180.0)

with ::

def degree_sin(deg, factor=math.pi/180.0, sin=math.sin):
return sin(deg * factor)

Because this trick uses default arguments for terms which should not be changed,
it should only be used when you are not concerned with presenting a possibly
confusing API to your users.


Core Language
=============

Expand Down Expand Up @@ -938,6 +785,68 @@ What does 'UnicodeDecodeError' or 'UnicodeEncodeError' error mean?
See the :ref:`unicode-howto`.


Performance
===========

My program is too slow. How do I speed it up?
---------------------------------------------

That's a tough one, in general. First, here are a list of things to
remember before diving further:

* Performance characteristics vary accross Python implementations. This FAQ
focusses on :term:`CPython`.
* Behaviour can vary accross operating systems, especially when talking about
I/O or multi-threading.
* You should always find the hot spots in your program *before* attempting to
optimize any code (see the :mod:`profile` module).
* Writing benchmark scripts will allow you to iterate quickly when searching
for improvements (see the :mod:`timeit` module).
* It is highly recommended to have good code coverage (through unit testing
or any other technique) before potentially introducing regressions hidden
in sophisticated optimizations.

That being said, there are many tricks to speed up Python code. Here are
some general principles which go a long way towards reaching acceptable
performance levels:

* Making your algorithms faster (or changing to faster ones) can yield
much larger benefits than trying to sprinkle micro-optimization tricks
all over your code.

* Use the right data structures. Study documentation for the :ref:`bltin-types`
and the :mod:`collections` module.

* When the standard library provides a primitive for doing something, it is
likely (although not guaranteed) to be faster than any alternative you
may come up with. This is doubly true for primitives written in C, such
as builtins and some extension types. For example, be sure to use
either the :meth:`list.sort` built-in method or the related :func:`sorted`
function to do sorting (and see the
`sorting mini-HOWTO <http://wiki.python.org/moin/HowTo/Sorting>`_ for examples
of moderately advanced usage).

* Abstractions tend to create indirections and force the interpreter to work
more. If the levels of indirection outweigh the amount of useful work
done, your program will be slower. You should avoid excessive abstraction,
especially under the form of tiny functions or methods (which are also often
detrimental to readability).

If you have reached the limit of what pure Python can allow, there are tools
to take you further away. For example, `Cython <http://cython.org>`_ can
compile a slightly modified version of Python code into a C extension, and
can be used on many different platforms. Cython can take advantage of
compilation (and optional type annotations) to make your code significantly
faster than when interpreted. If you are confident in your C programming
skills, you can also :ref:`write a C extension module <extending-index>`
yourself.

.. seealso::
The wiki page devoted to `performance tips
<http://wiki.python.org/moin/PythonSpeed/PerformanceTips>`_.

.. _efficient_string_concatenation:

What is the most efficient way to concatenate many strings together?
--------------------------------------------------------------------

Expand Down

0 comments on commit 432259f

Please sign in to comment.