[SciPy-user] More on speed comparisons
Ivo Maljevic
ivo.maljevic@gmail....
Mon Jun 16 09:32:00 CDT 2008
I was planing to refrain from this, but I believe this might be an
interesting comparison for people who seriously think about switching to
SciPy/NumPy.
Few days ago I wrote about speed comparison between the scalar and
vectorized versions of function calls. Based on several comments, I
concluded that the same story that applies to Matlab and Octave applies
here: vectorize thy code, and speed gain will come.
Before I show the results of vectorized vs. non-vectorized results, just
want to go on the record and say that I am by no means sayhing that
SciPy/NumPy is not good. I still like what has been done here. There is a
particular scenario that I use at my work where SciPy, combined with
matplotlib, is extremely useful. That scenario is the following.
In my wireless lab, I have basestations, mobile stations, whole bunch of
instruments and PCs either connected via network or GPIB cables (for
instruments).I use Python here to automate test cases and data collection
and the ability to do SSH and GPIB communication is very useful. Once I
collect data, I use SciPy for some simple postprocessing and I generate PNG
plots, and finally, I generate HTML pages with results shown as tables and
plots. So, it is all done in a single language/script instead of having to
break the processing into several languages/scripts.
However, I wanted to see if SciPy would be good enough speedwise to
completely replace Matlab. An, at least for the type of processing I do, it
comes nowhere near it. I wrote a small toy program that does some simple
random variable manipulation in several languages. The python code consists
of two versions, one uses for loop and basic pathon libraries and the other
uses nympy's vectorized form (there was no difference between numpy and
scipy).
Here are the relative results after running the code on two machines:
64-bit Ubunty 8.04:
=================
Fortran C Octave SciPy Pure Python
=================================================
1 1.2 2.2 16 20
32-Bit openSUSE 10.3:
==================
Fortran C Octave SciPy Pure Python
=================================================
1 1.2 2.4 15 19.4
The numbers are rounded a little bit, but they are in that range. I see two
problems here:
1. SciPy is very slow, even when compared to Octave 3.0
2. It is only sligtly faster than Python with a for loop.
Below is the source code for the two python versions. While this processing
is not from any real application,
it is not very different from the processing I normally do.
Now, it is very likely that for different type of processing people will
find SciPy fast enough (matrix inversions, eigenvalues, etc), but for the
type of
processing I need it is not fast enough.
Ivo
##########################################################################
# rand_test_1.py
from random import random
from math import sqrt, sin
N = 1000000
mean = 0
var = 0
for i in range(N):
x = random()
x = 3.14*sqrt(x)
x = sin(x)
mean += x
var += x**2
mean = mean/N
var = var/N - mean**2
print 'Mean=%g, var=%g' % (mean, var)
# rand_test_2.py
from numpy import random, sin, sqrt
N = 1000000
x = random.rand(N)
x = 3.14*sqrt(x)
x = sin(x)
mean = sum(x)/N
var = sum(x**2)/N - mean**2
print 'Mean=%g, var=%g' % (mean, var)
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://projects.scipy.org/pipermail/scipy-user/attachments/20080616/f2273631/attachment.html
More information about the SciPy-user
mailing list