Matlab speed test

Hi,

I repeated Dan Spielman's experiment for comparing speed of Matlab under 4 different modes in Matlab 2008a. In my experiment, contrary to his results, I got nearly the *same* performance no matter which of the following I chose:
* non vectorized .m code
* vectorized code,
* mex file
* java class.

Here is the website that the sample files can be downloaded:
http://www-math.mit.edu/~spielman/ECC/speedTests.html

This suggests that, going into the trouble to write mex file does not help to improve speed in such a long for loop of adding 10^8 random numbers. I doubt that it is true in general. What might be special here?

I am especially curious about how using Java classes improve performance, when vectorizing the code is not possible. It gave the worst performance in my experiment. Is it true in general?

I append the sample output. I would appreciate your opinion.

Gorkem

n = 1e8

n =

100000000

tic, speedTest(n),toc

ans =

4.9999e+07

Elapsed time is 7.306400 seconds.
tic, speedTest2(n),toc

ans =

4.9999e+07

Elapsed time is 7.327862 seconds.
tic, mex_speedTest(n),toc

ans =

5.0001e+07

Elapsed time is 6.271364 seconds.
import SpeedTest
s = SpeedTest

s =

SpeedTest@15e3abb

tic, s.sumRand(n),toc

ans =

4.9993e+07

Elapsed time is 8.452087 seconds.
.

Relevant Pages
Loading