Running fast code in Python

As I have suspected, the numbers I got for execution time of the Haskell fibonacci program (look at my previous posts) are wrong, in the sense that I compiled the program without optimisation, as most normal users will do in real applications. That is why I feel that I should redeem myself and give the numbers for the program with optimisation turned on. So, when compiling the program using

ghc -O2 fib.hs -o fib -no-recomp (thanks to Don Stewart)

execution time is 0.64s (mean time of five runs). This time is much closer to the one reported on the Haskell blog (0.48s). Ok, so my hat goes off to Haskell---it really is faster than Python some 50 times.
But, being a Python fan, I couldn't help but feel that something is wrong here. Is Haskell really that good? It's fast, easy to parallelize, pure functional and it probably cooks for you after a hard day at work. So, why should I use Python?
Well, the first thing that comes to my mind is the fact that ghc generated a 925kB exe file from a 1kB file that computes fibonacci numbers. In comparison, gcc (3.4.2) generated 16 kB for a fibonacci program that ran 0.75s (compiled using -O2). While it is still impresive that Haskell was faster than C, I cannot help but think that it is just too much to have an exe file almost 1000 times larger than the source. While this is certainly a drawback of Haskell, it still wasn't enough to return my faith in Python. Then, another thing came to my mind---if Haskell can use optimisations, why can't Python? I mean, if I used -O2 in ghc, why can't I do somehting similar in Python?
Of course, running Python with -OO gave no real speedup---it shook off about one second in execution time. Instead of 26s, I got 25s for the non-parallel case, and instead of 18s, I got 17s for the non-parallel case. But, the great thing about Python are its packages. I remember a friend of mine telling me about a thing called Psyco. It was some kind of a package that optimised execution of Python code, so I decided to give it a shot. After downloading and installing Psyco (http://psyco.sourceforge.net/), I just added two lines to my source (I give the full source here):

from time import time
import psyco
psyco.full()

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

if __name__ == "__main__":
    start_time = time()
    for i in range(36):
        print "n=%d => %d" % (i, fib(i))
        
    print "Time elapsed: %f" % (time() - start_time)

So, the lines I added were import psyco, and psyco.full(). I ran this thing and voila---the time of execution was 1.93s (mean time of five runs). Wow! That was all I could think. Remember, this was the non-parallel case, so it went from 25 seconds to 1.9s. Or, in terms of Haskell speed, Haskell now beat Python by on 3x. And with no extra effort (ok, adding two lines to the source) and no parallelization.
Unfortunately, psyco didn't help the parallel version---it still took 18s. But still, I think this is pretty amazing and it definitely brought back faith in Python. So, it's three times slower, but the synatx is much cleaner and it doesn't generate a thousand times bigger executable. I'm sticking to Python :)
(Note: some parts of this post are biased and down right offensive to Haskell fans (just kidding about the last one:) ). My intention is not to start a flame war or anything like that, I just felt that someone should stand up and take Python's side:). Seriously though, if you have any objections on my methodology, I would like to hear from you.)

Holy shmoly Haskell doesn't smoke Python away (that much)

Recently, there has been some talk about how Haskell smokes Python and Ruby away in the naive fibonacci algorithm. While it is obvious that Haskell will beat both due to the code being compiled and not interpreted, I decided to see if I got the same results as reported on the Haskell hacking blog, and also see how Python handles parallel processing. Note that I will refer to the Haskell hacking blog simply as Haskell blog from now on.

Since I have no experience in Ruby, I will only compare Python and Haskell. So, here is the code I used to test the two.
Python:

from time import time

def fib(n):
   if n == 0 or n == 1:
      return n
   else:
      return fib(n-1) + fib(n-2)
      
if __name__ == "__main__":
    
    start_time = time()
    for i in range(36):
        print "n=%d => %d" % (i, fib(i))
        
    print "Time elapsed: %f" % (time() - start_time)

So, as we can see, this code is pretty much the same as that reported in the blog about Haskell.

Haskell:

import Control.Monad
import Text.Printf

fib :: Int -> Int
fib 0 = 0
fib 1 = 1
fib n = fib (n-1) + fib (n-2)
main = forM_ [0..35] $ \i ->
    printf "n=%d => %d\n" i (fib i)

This is also the same as in the blog, the only difference being that I import what is needed to make it run (for some reason, the import statements aren't included in the Haskell blog). Since I have no idea how to measure time in Haskell (any hint on how to do that is appreciated), I used the ptime command for Windows (available at http://www.pc-tools.net/win32/ptime/).

So, here are the results of the test:

Python 2.5: 26.07 s
Haskell (GHC 6.8.1): 3.873 s (mean of 5 runs)

So, the results for Python are pretty much the same as reported in the Haskell blog (25.16 s is reported there), but the ones for Haskell are not, in fact the Haskell blog reports 0.48 s, which is some four times faster than I got on my computer. What is the reason for this? I have no idea. I compiled the program using "ghc fib.hs", maybe they used some code optimizations? Anyway, the Haskell advantage seems to be melting, although Haskell is still some 6-7 times faster. However, the Haskell blog proceeds to use parallel processing in Haskell, but uses no such thing in Python. So, I decided to see if Python can take advantage of my two cores to shake some time off this 26 seconds. I used the Parallel Python package (available at http://www.parallelpython.com/) to make use of my two cores. Here is the modified Python program:

from time import time
import pp

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


if __name__ == "__main__":
    start_time = time()
    ppservers = ()
    job_server = pp.Server(ppservers=ppservers)
    print "Starting pp with", job_server.get_ncpus(), "workers"
    jobs = [(input, job_server.submit(fib, (input,), (), ())) for input in range(36)]
    for input, job in jobs:
        print "n=%d => %d" % (input, job())
        
    print "Time elapsed: %f" % (time() - start_time)

The result: 17.86 s
So, Python went from 26.07s to 17.86s, which means it ran 31.5% faster than on only one core. Unfortunately, I had no success in running the parallel Haskell code reported on the Haskell blog (for some reason, I always got "`pseq` not in scope" error while compiling). However, I can use the numbers given on the Haskell blog, where Haskell went from 0.48s to 0.42s on two cores. This means a 12.5% reduction in execution time. Although I admit that is takes slightly more intervention from the programmer to make the parallelism in Python work, the gain in speed is also more significant. As a final note, let us compare 3.873s (Haskell) and 17.86s (parallel Python) - the ratio is around 4.6, let's say 5. So, Haskell doesn't smoke Python away THAT much.