Re: matlab/octave performance hit

[Quentin Spencer]

Victor Zhorin wrote:
> We tried to see whether we could make existing matlab code to run 
> using octave on parallel cluster.
> Unfortunately we found that performance hit is too strong.
>
> It seems that slowness is most likely due to some instrinsic 
> system/octave  issues.
>
> Here is a quick example:
>
> tic; test = 1:200; nc = cumsum(test);toc
>
> octave (linux based box, latest release compiled)
> Elapsed time is 0.023521 seconds.
> matlab (mac, powerpc)
> Elapsed time is 0.000064 seconds.
> octave (same mac, latest binary package for Mac X installed as-is)
> Elapsed time is 0.050179 seconds.
>
> That's a very significant difference in performing the most basic 
> operations.
> Is there a way we can speed it up or could we possibly miss something 
> in installation?

I should point out that your test is much too small to get an accurate 
characterization. The tic and toc functions are m files in octave and 
builtin functions in matlab (BTW, I think this might be a worthwhile 
enhancement for anyone who wants a small project). Running nothing more 
than tic; toc takes 0.02 sec in octave and 0.000011 sec in matlab on my 
computer. So, If I make your test large enough to make that difference 
negligible:
tic; test = 1:1e7; nc = cumsum(test); toc
I get 0.73 seconds in octave and 0.51 seconds in matlab. For any heavily 
vectorized operation like this, the times should be not too different. 
Matlab will be considerably faster on code that is not heavily 
vectorized because its interpreter is faster. Any contributions to the 
project to make octave's interpreter faster are welcomed.

Quentin