Python is faster than C

[Armin Rigo]

Hi!

This is a rant against the optimization trend of the Python interpreter.

Sorting a list of 100000 integers in random order takes:

* 0.75 seconds in Python 2.1
* 0.51 seconds in Python 2.2
* 0.46 seconds in Python 2.3

Tim Peters did a great job optimizing list.sort(). If I try with a
simple, non-stable pure Python quicksort implementation, in Python 2.3:

* 4.83 seconds
* 0.21 seconds with Psyco

First step towards world domination of high-level languages :-)

The reason that Psyco manages to outperform the C implementation is not
that gcc is a bad compiler (it is about 10 times better than Psyco's).
The reason is that the C implementation must use a generic '<' operator
to compare elements, while the Psyco version quickly figures out that it
can expect to find ints in the list; it still has to check this
assumption, but this is cheap and then the comparison is done with a
single machine instruction.

Similarily, here are some results about the heapq module, which is
rewritten in C in the CVS tree for Python 2.4:

l = [random.random() for x in range(200000)]
heapq.heapify(l )

This code executes on my laptop in:

* 1.96 seconds on Python 2.3 (pure Python)
* 0.18 seconds on Python 2.4cvs (rewritten in C)
* 0.16 seconds on Python 2.3 with Psyco

So this is not so much a plug for Psyco as a rant against the current
trend of rewriting standard modules in C. Premature optimization and
all that.

Worse, and more importantly, the optimization starts to become visible
to the programmer. Iterators, for example, are great in limited cases
but I consider their introduction a significant complication in the
language; before, you could expect that some function from which you
would expect a sequence returned a list. Python was all lists and
dicts, with dicts used as namespaces here and there. Nowadays you have
to be careful. Moreover, it is harder to explain:

zip([1,2,3], [4,5,6]) # easy to understand and explain [(1, 4), (2, 5), (3, 6)]
enumerate([6,7,8,9]) # uh ?

<enumerate object at 0x401a102c>

I know you can always do list(_). My point is that this is a
user-visible optimization. enumerate() should return a normal list, and
it should be someone else's job to ensure that it is correctly optimized
away if possible (and I'm not even talking about Psyco, it could be done
in the current Python implementation with a reasonable amount of
effort).
Protesting-ly yours,

Armin

[Michel Claveau/Hamster]

Hi !

For the 1st April, it's finish.

[Armin Rigo]

Michel Claveau/Hamster wrote:

For the 1st April, it's finish.

Check it for yourself. Find yourself an Intel machine and grab Psyco
from http://psyco.sf.net. Here is the source of my test:

# Python Quicksort Written by Magnus Lie Hetland
# http://www.hetland.org/python/quicksort.html

def _partition(list , start, end):
pivot = list[end]
bottom = start-1
top = end

done = 0
while not done:

while not done:
bottom = bottom+1

if bottom == top:
done = 1
break

if pivot < list[bottom]:
list[top] = list[bottom]
break

while not done:
top = top-1

if top == bottom:
done = 1
break

if list[top] < pivot:
list[bottom] = list[top]
break

list[top] = pivot
return top
def _quicksort(list , start, end):
if start < end:
split = _partition(list , start, end)
_quicksort(list , start, split-1)
_quicksort(list , split+1, end)

def quicksort(list) :
if len(list) > 1:
_quicksort(list , 0, len(list)-1)

# _______________ _______________ _______________ _______________

import random, time, psyco
l = range(100000)
random.shuffle( l)

#TIMEIT = "l.sort()"
#TIMEIT = "quicksort( l)"
TIMEIT = "psyco.proxy(qu icksort)(l)"
print TIMEIT, ':',
t = time.time()
exec TIMEIT
print time.time() - t

assert l == range(100000)

[Paul Rubin]

Armin Rigo <ar***@tunes.or g> writes:

enumerate([6,7,8,9]) # uh ?

<enumerate object at 0x401a102c>

I know you can always do list(_). My point is that this is a
user-visible optimization. enumerate() should return a normal list, and
it should be someone else's job to ensure that it is correctly optimized
away if possible (and I'm not even talking about Psyco, it could be done
in the current Python implementation with a reasonable amount of
effort).

I think enumerate(xrang e(1000000000)) returning a normal list would
exhaust memory before some later optimizer had a chance to do anything
with it.

[Josiah Carlson]

> For the 1st April, it's finish.

That wasn't a joke, psyco does in fact work /really/ well for some things.

- Josiah

[Joe Mason]

In article <40************ ***@tunes.org>, Armin Rigo wrote:

The reason that Psyco manages to outperform the C implementation is not
that gcc is a bad compiler (it is about 10 times better than Psyco's).
The reason is that the C implementation must use a generic '<' operator
to compare elements, while the Psyco version quickly figures out that it
can expect to find ints in the list; it still has to check this
assumption, but this is cheap and then the comparison is done with a
single machine instruction.

Why can't the C implementation do the same thing?

Joe

[Armin Rigo]

Hello Paul,

Paul Rubin wrote:

I think enumerate(xrang e(1000000000)) returning a normal list would
exhaust memory before some later optimizer had a chance to do anything
with it.

There are two levels: the language specification and its implementation.
My point is that there is no reason why everything that is (at the
language level) a list, should always be implemented as a plain array of
objects. The list returned by range(1000000) doesn't have to use 4MB of
memory when the same information can be encoded in a couple of ints.
The presence of xrange() is the oldest of these user-visible
optimization hack that should have been optimized in the implementation
instead. Similarily, if you do 'a'*999999999 I can think of a better
way to encode the resulting string than with 100MB of RAM.

On your specific example, we could also argue that a slightly involved
but reasonable amount of effort would allow C functions to internally
receive a context hint, which would tell enumerate() that its result
will only ever be used for iteration when this is the case, so that it
can internally return an iterable instead of a list -- but this would
only be a hack, a workaround for the limits of CPython's internal object
representations .
Armin

[Paul Rubin]

Armin Rigo <ar***@tunes.or g> writes:

There are two levels: the language specification and its implementation.
My point is that there is no reason why everything that is (at the
language level) a list, should always be implemented as a plain array of
objects. The list returned by range(1000000) doesn't have to use 4MB of
memory when the same information can be encoded in a couple of ints.
The presence of xrange() is the oldest of these user-visible
optimization hack that should have been optimized in the implementation
instead.

I think you're saying that instead of having xrange return a special
object, range should return that special object instead. I'm not too
clear on the distinction. Also, how can the compiler or interpreter
know whether the program will only access the object sequentially?
It has to predict the behavior of the program, instead of being told
explicitly.

[Armin Rigo]

Paul Rubin wrote:

I think you're saying that instead of having xrange return a special
object, range should return that special object instead. I'm not too
clear on the distinction.
No, range should return an object that is a list, as far as you can tell
from Python, but which is represented more efficiently than an array of
objects internally. The distinction is between the language level (it
would be a list, with all operations, etc.) and the implementation
(there is no reason why all lists should be arrays of PyObjects
internally).

Another example would be 'a'*999999999: the result is a string, but
there is no reason that it takes 100MB of memory. Instead, store it
into a C structure that contains a pointer to the original string object
'a' and the repetition counter, but still give this C structure the
Python type str, so that the difference doesn't show up and the Python
language remains simple. (This is a bit difficult to implement
currently in CPython, but not impossible.)
Also, how can the compiler or interpreter
know whether the program will only access the object sequentially?
It has to predict the behavior of the program, instead of being told
explicitly.

Ideally: If you do x=range(100); x[50]='hi' then the interpreter first
builds this optimized range representation and assigns it to x; and when
in the next statement you modify this list x it says 'oops! i cannot do
that with this representation' , so it reverts to an array-like
representation (i.e. it creates all 100 elements) and then changes the
50th. No gain here. If on the other hand you only ever do 'easy'
things with your list, like iterate over it or read elements, then it
can all be done with the range representation, without falling back to
the array representation.

Again I'm not saying it is trivial to implement it, but that not having
to expose for optimization purposes 'xrange' and the whole 'iterator'
part of the language would be worth it, in my opinion.
Armin

[Armin Rigo]

Hi,

Joe Mason wrote:

The reason is that the C implementation must use a generic '<' operator
to compare elements, while the Psyco version quickly figures out that it
can expect to find ints in the list; it still has to check this
assumption, but this is cheap and then the comparison is done with a
single machine instruction.

Why can't the C implementation do the same thing?

You could, if you wanted to optimize specifically lists of integers. If
you did the result would probably be really fast. The problem is that
you can only really special-case so many types: the C code has to deal
with all cases without knowing which cases are likely. The Psyco
version quickly figures out that for this list it pays off to make a
special case for integers; with another list, the machine code would be
different, special-cased differently.

However, in most real examples, you are not sorting a list of integers
but of something more complex anyway, where the built-in sort wins
easily. My message was intended as a long-term hint that maybe, at some
point, the built-in sort will actually be more often faster than the C
one if rewritten in Python. The advantage would then be that (as Psyco
does in a limited fashion) you can specialize the code for the
particular kind of list you are dealing with.
Armin