Is a "real" C-Python possible?

[Raymond Hettinger]

Re: Is a "real&quot ; C-Python possible?

On Dec 9, 1:14 pm, "Jack" <nos...@invalid .comwrote:

I wonder if it's possible to have a Python that's completely (or at
least for the most part) implemented in C, just like PHP - I think
this is where PHP gets its performance advantage. Or maybe I'm wrong
because the core modules that matter are already in C and those Python
files are really a think wrapper. Anyhow, if would be ideal if Python
has performance similar to Java, with both being interpreted languages.

-1 This would seriously muck-up the evolution of the language.
Having a few building blocks written in C provides a basis
for writing very fast pure python (for example, sets, heapq,
itertools).
Beyond those building blocks, it is a step backwards to write in C.

Also, if you really need performance, the traditional solutions are to
use tools like Psyco or Pyrex.


Raymond

[hdante]

Re: Is a &quot;real&quot ; C-Python possible?

On Dec 9, 10:07 pm, "Jack" <nos...@invalid .comwrote:

I think most Java-Python benchmarks you can find online will indicate
that Java is a 3-10 times faster. A few here:
>http://mail.python.org/pipermail/pyt...ry/125789.html
>http://blog.snaplogic.org/?p=55

>

There are lies, damn lies and benchmarks. :)

>

Pure Python code is not going to beat Java code until the Python core
gets a JIT compiler. If you want fair results you have to either
disable the JIT in Java or use Psyco for Python. Otherwise you are
comparing the quality of one language implementation to the quality of a
JIT compiler.

>
The second articple does have a column for Psyco. It helps in some areas
but still not good enough to stand up against Java. Plus, Psyco is not the
main stream and has stopped development.
>
I'm also wondering, if Psyco is the right way to do, any reason it's not
being integrated into standard Python?

Instead of recurring to benchmarks, I recommend that you read the
following:

http://highscalability.com/youtube-architecture

There are no comparisons there, just a sample of what python and
psyco can achieve. For a language that isn't designed with speed in
mind, I think that's quite impressive.

[Bruno Desthuilliers]

Re: Is a &quot;real&quot ; C-Python possible?

Jack a écrit :

>>>I'm not sure
>>>how much of the C-Python is implemented in C but I think the more
>>>modules implemented in C, the better performance and lower memory
>>>footprint it will get.

>>
>>Prove it. ;-)

>
>
I guess this is subjective :)

If yes, benchmarks are not an argument. Else, you'll have hard time
making your point !-)

(hint: doing objective benchmarking is really a difficult art)

- that's what I felt in my experience
with web applications developed in Python and PHP. I wasn't able to
find a direct comparison online.

Could it be the case that you are comparing Python CGI scripts with
mod_php ? Anyway, since php is also usable (hem... maybe not the
appropriate qualificative) outside Apache, it should quite easy to make
a more serious test ?

Seriously: I never saw any benchmark where php was faster than Python
for any kind of stuff - unless of course you're trying to compare Zope
running as a CGI script with an hello world PHP script runned by mod_php.

[Bruno Desthuilliers]

Re: Is a &quot;real&quot ; C-Python possible?

hunter.grubbs@g mail.com a écrit :
(snip)

I'd like to provide some evidence that Python is *faster* than Java.

Then benchmark the time taken for the interpreter (oops, sorry: "VM") to
start !-)

[sturlamolden]

Re: Is a &quot;real&quot ; C-Python possible?

On 9 Des, 22:14, "Jack" <nos...@invalid .comwrote:

I understand that the standard Python distribution is considered
the C-Python. Howerver, the current C-Python is really a combination
of C and Python implementation. There are about 2000 Python files
included in the Windows version of Python distribution. I'm not sure
how much of the C-Python is implemented in C but I think the more
modules implemented in C, the better performance and lower memory
footprint it will get.

Donald Knuth, one of the fathers of modern computer science, is famous
for stating that "premature optimization is the root of all evil in
computer science." A typical computer program tends to have
bottlenecks that accounts for more than 90% of the elapsed run time.
Directing your optimizations anywhere else is futile.

Writing a program in C will not improve the speed of your hardware. If
the bottleneck is a harddisk or a network connection, using C will not
change that. Disk i/o is a typical example of that. It is not the
language that determines the speed by which Python or C can read from
a disk. It is the disk itself.

I had a data vizualization program that was slowed down by the need to
move hundreds of megabytes of vertex data to video RAM. It would
obviously not help to make the handful of OpenGL calls from C instead
of Python. The problem was the amount of data and the speed of the
hardware (ram or bus). The fact that I used Python instead of C
actually helped to make the problem easier to solve.

We have seen several examples that 'dynamic' and 'interpreted'
languages can be quite efficient: There is an implementation of Common
Lisp - CMUCL - that can compete with Fortran in efficiency for
numerical computing. There are also versions of Lisp than can compete
with the latest versions of JIT-compiled Java, e.g. SBCL and Allegro.
As it happens, SBCL and CMUCL is mostly implemented in Lisp. The issue
of speed for a language like Python has a lot to do with the quality
of the implementation. What really makes CMUCL shine is the compiler
that emits efficient native code on the fly. If it is possible to make
a very fast Lisp, it should be possible to make a very fast Python as
well. I remember people complaining 10 years ago that 'Lisp is so
slow'. A huge effort has been put into making Lisp efficient enough
for AI. I hope Python some day will gain a little from that effort as
well.

We have a Python library that allows us to perform a wide range of
numerical tasks at 'native speed': NumPy (http://www.scipy.org). How
such array libraries can be used to get excellent speedups is
explained here: http://home.online.no/~pjacklam/matl...mtt/index.html

We obviously need more effort to make Python more efficient for CPU
bound tasks. Particularly JIT compilation like Java, compilation like
Lisp or data specialization like Psyco.

But writing larger parts of the standard library in C is not a
solution.

[sturlamolden]

Re: Is a &quot;real&quot ; C-Python possible?

On 9 Des, 23:34, Christian Heimes <li...@cheimes. dewrote:

Nevertheless it is just one algorithm that beats Python in an area that
is well known to be slow. Python's numbers are several factors slower
than C code because the overhead of the dynamic language throws lots of
data out of the cache line. If you need fast and highly optimized int
and floating point operations you can rewrite the algorithm in C and
create a Python interface for it.

Lisp is a dynamically typed language. CMUCL can compete with Fortran
for numerical work. SBCL can compete with the Java server VM. If the
current CPython VM throws data out of the cache line, then it must be
a design flaw in the VM.

[Aahz]

Re: Is a &quot;real&quot ; C-Python possible?

In article <a324fa50-fde9-4173-81f3-f7e756bd46d7@l1 6g2000hsf.googl egroups.com>,
sturlamolden <sturlamolden@y ahoo.nowrote:

>
>Donald Knuth, one of the fathers of modern computer science, is famous
>for stating that "premature optimization is the root of all evil in
>computer science."

From my .sig database:

"Premature optimization is the root of all evil in programming."
--C.A.R. Hoare (often misattributed to Knuth, who was himself quoting
Hoare)
--
Aahz (aahz@pythoncra ft.com) <* http://www.pythoncraft.com/

"Typing is cheap. Thinking is expensive." --Roy Smith

[Bruno Desthuilliers]

Re: Is a &quot;real&quot ; C-Python possible?

sturlamolden a écrit :

On 9 Des, 23:34, Christian Heimes <li...@cheimes. dewrote:
>
>

>>Nevertheles s it is just one algorithm that beats Python in an area that
>>is well known to be slow. Python's numbers are several factors slower
>>than C code because the overhead of the dynamic language throws lots of
>>data out of the cache line. If you need fast and highly optimized int
>>and floating point operations you can rewrite the algorithm in C and
>>create a Python interface for it.

>
>
Lisp is a dynamically typed language. CMUCL can compete with Fortran
for numerical work. SBCL can compete with the Java server VM. If the
current CPython VM throws data out of the cache line, then it must be
a design flaw in the VM.

Or a lack of time and money. Lisp is one of the older programming
languages around, and at a time had BigBucks(tm) invested on it to try
and make it practically usable.

[sturlamolden]

Re: Is a &quot;real&quot ; C-Python possible?

On 10 Des, 23:54, Bruno Desthuilliers
<bdesth.quelque ch...@free.quel quepart.frwrote :

Or a lack of time and money. Lisp is one of the older programming
languages around, and at a time had BigBucks(tm) invested on it to try
and make it practically usable.

Yes. But strangely enough, the two Lisp implementations that really
kick ass are both free and not particularly old. CMUCL and SBCL proves
that you can make a dynamic language implementation extremely
efficient if you try hard enough. There are also implementations of
Scheme (e.g. Bigloo) that shows the same.

[sturlamolden]

Re: Is a &quot;real&quot ; C-Python possible?

On 10 Des, 23:49, a...@pythoncraf t.com (Aahz) wrote:

"Premature optimization is the root of all evil in programming."
--C.A.R. Hoare (often misattributed to Knuth, who was himself quoting
Hoare)

Oh, I was Hoare? Thanks. Anyway, it doesn't change the argument that
optimizing in wrong places is a waste of effort.

[sturlamolden]

Re: Is a &quot;real&quot ; C-Python possible?

On 9 Des, 23:34, Christian Heimes <li...@cheimes. dewrote:

http://antoniocangiano.com/2007/11/2...19-smokes-pyth...

>
The Ruby developers are allowed to be proud. They were able to optimize
some aspects of the implementation to get one algorithm about 14 times
faster. That's good work. But why was it so slow in the first place?

The thing to notice here is that Congiano spent 31.5 seconds computing
36 Fibonacci numbers in Python and 11.9 seconds doing the same in
Ruby. Those numbers are ridiculous! The only thing they prove is that
Congiano should not be programming computers. Anyone getting such
results should take a serious look at their algoritm instead of
blaming the language. I don't care if it takes 31.5 seconds to compute
36 Fibonacci numbers in Python 2.5.1 with the dumbest possible
algorithm.

[Kay Schluehr]

Re: Is a &quot;real&quot ; C-Python possible?

We obviously need more effort to make Python more efficient for CPU

bound tasks. Particularly JIT compilation like Java, compilation like
Lisp or data specialization like Psyco.

Given that the Python core team has been mostly silent about JIT
compilation and Armin Rigos work in particular which started 5 years
ago ( Psyco will not be promoted towards Python 3.0 and there is no
indication that anyone but Armin would maintain Psyco ) I wonder about
this sudden insight. But maybe you can tell us more when you work on
such stuff for CPython yourself? Given your status in the core Python
team this would have a chance of not being just a loosely coupled
independent project as almost all the interesting stuff that happens
around Python these days.

[Duncan Booth]

Re: Is a &quot;real&quot ; C-Python possible?

sturlamolden <sturlamolden@y ahoo.nowrote:

On 9 Des, 23:34, Christian Heimes <li...@cheimes. dewrote:
>

>http://antoniocangiano.com/2007/11/2...19-smokes-pyth
>...

>>
>The Ruby developers are allowed to be proud. They were able to
>optimize some aspects of the implementation to get one algorithm
>about 14 times faster. That's good work. But why was it so slow in
>the first place?

>
The thing to notice here is that Congiano spent 31.5 seconds computing
36 Fibonacci numbers in Python and 11.9 seconds doing the same in
Ruby. Those numbers are ridiculous! The only thing they prove is that
Congiano should not be programming computers. Anyone getting such
results should take a serious look at their algoritm instead of
blaming the language. I don't care if it takes 31.5 seconds to compute
36 Fibonacci numbers in Python 2.5.1 with the dumbest possible
algorithm.
>

Quite so.

Take something like

http://aspn.activestate.com/ASPN/Cookbook/Python/Recipe/498110

and then modify the Python code from the "Ruby smokes Python" article by
the addition of @memoize(3) to decorate the otherwise unchanged fib
function: the Python runtime drops down to 0.002 seconds.

That is just slightly better than Ruby's 11.9 seconds although I'm sure the
Ruby code would also gain as much from a memoize decorator.


from memoize import memoize

@memoize(3)
def fib(n):
if n == 0 or n == 1:
return n
else:
return fib(n-1) + fib(n-2)

from time import clock
start = clock()
for i in range(36):
print "n=%d =%d" % (i, fib(i))
print clock()-start


When I run this (with output directed to a file: I'm not trying to time
windows console speed), the output is:

n=0 =0
n=1 =1
n=2 =1
n=3 =2
n=4 =3
n=5 =5
n=6 =8
n=7 =13
n=8 =21
n=9 =34
n=10 =55
n=11 =89
n=12 =144
n=13 =233
n=14 =377
n=15 =610
n=16 =987
n=17 =1597
n=18 =2584
n=19 =4181
n=20 =6765
n=21 =10946
n=22 =17711
n=23 =28657
n=24 =46368
n=25 =75025
n=26 =121393
n=27 =196418
n=28 =317811
n=29 =514229
n=30 =832040
n=31 =1346269
n=32 =2178309
n=33 =3524578
n=34 =5702887
n=35 =9227465
0.0022642542557 8

[MonkeeSage]

Re: Is a &quot;real&quot ; C-Python possible?

On Dec 11, 3:10 am, Duncan Booth <duncan.bo...@i nvalid.invalidw rote:

sturlamolden <sturlamol...@y ahoo.nowrote:

On 9 Des, 23:34, Christian Heimes <li...@cheimes. dewrote:

>

http://antoniocangiano.com/2007/11/28/holy-shmoly-ruby-19-smokes-pyth

...

>

The Ruby developers are allowed to be proud. They were able to
optimize some aspects of the implementation to get one algorithm
about 14 times faster. That's good work. But why was it so slow in
the first place?

>

The thing to notice here is that Congiano spent 31.5 seconds computing
36 Fibonacci numbers in Python and 11.9 seconds doing the same in
Ruby. Those numbers are ridiculous! The only thing they prove is that
Congiano should not be programming computers. Anyone getting such
results should take a serious look at their algoritm instead of
blaming the language. I don't care if it takes 31.5 seconds to compute
36 Fibonacci numbers in Python 2.5.1 with the dumbest possible
algorithm.

>
Quite so.
>
Take something likehttp://aspn.activestat e.com/ASPN/Cookbook/Python/Recipe/498110
and then modify the Python code from the "Ruby smokes Python" article by
the addition of @memoize(3) to decorate the otherwise unchanged fib
function: the Python runtime drops down to 0.002 seconds.
>
That is just slightly better than Ruby's 11.9 seconds although I'm sure the
Ruby code would also gain as much from a memoize decorator.
>
from memoize import memoize
>
@memoize(3)
def fib(n):
if n == 0 or n == 1:
return n
else:
return fib(n-1) + fib(n-2)
>
from time import clock
start = clock()
for i in range(36):
print "n=%d =%d" % (i, fib(i))
print clock()-start
>
When I run this (with output directed to a file: I'm not trying to time
windows console speed), the output is:
>
n=0 =0
n=1 =1
n=2 =1
n=3 =2
n=4 =3
n=5 =5
n=6 =8
n=7 =13
n=8 =21
n=9 =34
n=10 =55
n=11 =89
n=12 =144
n=13 =233
n=14 =377
n=15 =610
n=16 =987
n=17 =1597
n=18 =2584
n=19 =4181
n=20 =6765
n=21 =10946
n=22 =17711
n=23 =28657
n=24 =46368
n=25 =75025
n=26 =121393
n=27 =196418
n=28 =317811
n=29 =514229
n=30 =832040
n=31 =1346269
n=32 =2178309
n=33 =3524578
n=34 =5702887
n=35 =9227465
0.0022642542557 8

Another point is, the reason the ruby code shows such a performance
increase is because of the way it wraps native (C) types for integers
in the the new byte compiler; i.e., it's a directed optimization,
which the example code exploits to its full extent. But with
dictionary access, for example, python still creams ruby (by a 2/1
factor in my tests). Speaking as someone who uses both python and
ruby, I can say that ruby 1.9 is approaching python's speed, which is
very cool, but is still not quite as fast as python in general (the
whole "smokes python" bit is just propaganda that utilizes a specific
feature vector, and is generally unhelpful).

Regards,
Jordan

[John Nagle]

Re: Is a &quot;real&quot ; C-Python possible?

sturlamolden wrote:

On 10 Des, 23:49, a...@pythoncraf t.com (Aahz) wrote:
>

>"Premature optimization is the root of all evil in programming."
>--C.A.R. Hoare (often misattributed to Knuth, who was himself quoting
>Hoare)

We're ten years into Python, and it's still a naive interpreter.
It's time for a serious optimizing compiler. Shed Skin is going
in the right direction. But for some reason, people seem to dislike the
Shed Skin effort. Its author writes "Am I the only one seeing the potential
of an implicitly statically typed Python-like-language that runs at
practically the same speed as C++?"

"For a set of 27 non-trivial test programs (at about 7,000 lines in total;
.... measurements show a typical speedup of 2-40 times over Psyco, about 10 on
average, and 2-220 times over CPython, about 35 on average." So that's
what's possible.

I'm surprised that Google management isn't pushing Guido towards
doing something about the performance problem.

John Nagle