Re: speed comparison of IDL, numPy, Matlab

[Christopher Lee <clee(at)gnwy100.wuh.wustl.edu>]

>>>>> "Gavrie" == Gavrie  <gavrie@my-deja.com> writes:

    Gavrie> In article <3A7EEE5C.6419B707@pacific.jpl.nasa.gov>, Benyang
    Gavrie> Tang <btang@pacific.jpl.nasa.gov> wrote:


    Gavrie> So, this difference seems a bit strange, doesn't it?  I'm using
    Gavrie> MATLAB 6 (whichis supposed to be *slower* than 5.3), and Python
    Gavrie> 1.5.2.

I believe there's an explaination for this:

While I've heard that Matlab 6's footprint is much larger, especially the
interface, I have heard that Matlab 6 does (finally) use LAPACK and BLAS
which are optimized for modern machines w/ hierarchical memory.  I would
therefore expect that it would do substantially better than the
out-of-the-box numpy build for matrix multiplication.  By default, numpy
uses a naive algorithm for matrix multiplication which is quite cache
unfriendly.  When numpy is linked to ATLAS's BLAS routines and LAPACK, it's
more cache-friendly---and much faster.

I did some benchmarks myself: For matrix inversion of a 1000x1000 matrix,
numpy-atlas is 7 times faster than matlab 5.3 (no lapack).  The difference
is greater if you have a dual processor machine because ATLAS now has
options for multi-threaded operation.  And for matrix multiplication you
can compare the naive numpy multiplication with the optmized-dgemm based
version on the same matrix:

#  1000 x 1000 matrix 
time elapsed for matrixmultiply (sec) 24.142714
time elapsed for dgemm-matrixmultiply (sec) 5.997188


-chris

[BTW: I'm working on putting together an optimized-numpy distribution that
uses ATLAS by default.]