Wednesday, November 18, 2009

Octave is getting faster!


GNU Octave "is a high-level language, primarily intended for numerical computations. It provides a convenient command line interface for solving linear and nonlinear problems numerically, and for performing other numerical experiments using a language that is mostly compatible with Matlab. It may also be used as a batch-oriented language. "

In numerical calculations speed is very important. I remember that Octave 2.x, although very useful, used to be slow. What about now when there is Octave 3.2? To answer to this question I decided to compare the time performance of this new Octave release with the performances of versions 3.0 and 2.9 using custom made benchmark script. The tests were run on Intel Mac OS X 10.4.11 (Intel Core Duo 2GHz, 2GB RAM).

 The results of 21 small tests for matrix and I/O operations are below or can be downloaded from here (gnumeric file):

The bar plot of normalized geometric mean of the 21 tests is as follows (lower is better):

The graph shows that Octave 3.2 is a bit faster than the two previous versions! Additionally, since the time values used were normalized against times obtained using MATLAB 7.4.0.287 (R2007a) on Intel Mac OS X 10.4.11 (Intel Core Duo 2GHz, 2GB RAM), it can be seen that Octave 3.2 is faster than the MATLAB R2007a. However, based on the data in the table it can be seen that the lowest performance of the Octave, as compared to Matlab is, for loop (row 13 in the table). The Octave 3.2 was nearly 700 times slower than the Matlab in execution of nested for loops! At the same time Matlab was about 1000 times slower than Octave 3.2 (but not Octave 2.9 and 3.0) in creation of sparse matrices (row 14 in the table)!!!

Should we be very excited about this! Probably not very much, since my analysis has a number of limitations. The main one is that the question whether Octave 3.2 is faster than the older versions should be answered using some real statistical approach e.g. Mann-Whiten U tests or Ordinary least products (OLP) regression analysis, influence of outliners (row 13 and 14) should be investigated etc. There is also a question whether the differences in the performance obtained are statistically significant, or whether Octave 3.2 is statistically significantly faster/slower that Matlab?

Nonetheless, I think that the results obtained clearly show a trend of increased performance for Octave.

I'm also planning to repeat the above tests for Octave 3.2, but this time with inclusion of image processing functions such as: image creation, rotation, resize, image saving etc. I could not do this for is post, because there were problems compiling image package for Octave 2.9 and 3.0 on my Mac.

5 comments:

  1. Anonymous11:23 PM

    Why did you use geometric mean?

    ReplyDelete
  2. Geometric mean is a recommended way of summarizing benchmark results: PJ Fleming, JJ Wallace, How not to lie with statistics: the correct way to summarize benchmark results, Communications of the ACM, 1986.

    ReplyDelete
  3. Anonymous11:09 AM

    Octave 3.2 was released in Jun-2009, but 2007a was released in Mar-2007. Comparing with the version available in Jun-2009 (ie 2009a) will be more valid and interesting. Also as far as I know Matlab in 2007a did not have much multithreading.

    ReplyDelete
  4. Anonymous11:25 AM

    i prefer englab
    http://englab.bugfest.net

    ReplyDelete
  5. Anonymous2:48 PM

    Needs more exclamation marks. And cowbells.

    ReplyDelete