Re: [ojAlgo-user] Performance tests

[Niko B. <nik...@gm...>]

Hi Anders

I have a general comment about  performance when working with large matrices.

************************************************************************
*  It matters a lot in what order you access the data in memory. *
************************************************************************

Consider the following example, using MATLAB version 5.2.0 on a 3Ghz Pentium 4.

>> M = ones(10000); % This matrix has a size of about 800 MB.
>> v = ones(10000,1);

% multiply
>> tic;M*v;toc
elapsed_time =

    6.6870

% If M is symmetrical, this gives the same answer, but accesses memory
differently
>> tic;(v'*M)';toc
elapsed_time =

    0.2350

I have done similar matrix multiplication experiments in Java, with
similar results. Although in Java you don't have very good control
over how and where matrix elements are stored in memory, it is still
often possible by re-arranging the order in which elements are
accessed, to speed up things considerably.


Niko Brummer