[Numpy-discussion] speed of numpy vs matlab on dot product (original) (raw)

JJ josh8912 at yahoo.com
Sat Jun 10 18:15:07 EDT 2006

Hello. I am a new user to scipy, thinking about crossing over from Matlab. I have a new AMD 64 machine and just installed fedora 5 and scipy. It is a dual boot machine with windows XP. I did a small test to compare the speed of matlab (in 32 bit windows, Matlab student v14) to the speed of scipy (in fedora, 64 bit). I generated two random matrices of 10,000 by 2,000 elements and then took their dot product. The scipy code was:

python import numpy import scipy a = scipy.random.normal(0,1,[10000,2000]) b = scipy.random.normal(0,1,[10000,2000]) c = scipy.dot(a,scipy.transpose(b))

I timed the last line of the code and compared it to the equivalent code in Matlab. The results were that Matlab took 3.3 minutes and scipy took 11.5 minutes. Thats a factor of three. I am surprised with the difference and am wondering if there is anything I can do to speed up scipy.

I installed scipy, blas, atlas, numpy and lapack from source, just as the instructions on the scipy web site suggested (or as close to the instructions as I could). The only thing odd was that when installing numpy, I received messages that the atlas libraries could not be found. However, it did locate the lapack libraries. I dont know why it could not find the atlas libraries, as I told it exactly where to find them. It did not give the message that it was using the slower default libraries. I also tried compiling after an export ATLAS = statement, but that did not make a difference. Wherever I could, I complied it specifically for the 64 bit machine. I used the current gcc compiler. The ATLAS notes suggested that the speed problems with the 2.9+ compilers had been fixed.

Any ideas on where to look for a speedup? If the problem is that it could not locate the atlas ibraries, how might I assure that numpy finds the atlas libraries. I can recompile and send along the results if it would help.
Thanks. John

PS. I first sent this to the scipy mailing list, but it didnt seem to make it there.

Do You Yahoo!? Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com

More information about the NumPy-Discussion mailing list