ORCA can take advantage of MATLAB’s parallel toolbox. The parallelism is done at dataset partition level. The method runExperiments of the class Utilities includes the following optional arguments, apart from the name of the configuration file:
'parallel': false or true to activate CPU parallel processing of databases’s folds. Default value is ‘false’.'numcores': default maximum number of cores or desired number. If true and numcores is lower than 2, this paramter sets the maximum number of cores.'closepool’: whether to close or not the pool after the experiments. Default true. Disabling it can speed up consecutive calls to runExperiments saving the time of opening and closing pools.The improvement is done in models fitting and prediction. However, the reports have to be generated sequentially. Given that lots of metrics are obtained in these reports, this non-parallelizable operation is very costly.
In Octave, the parfor tool is not yet implemented. However, we have adapted the code to use the parallel package which provides similar functionality. If you want to parallelize experiments in Octave, you will have to install the corresponding package:
pkg install -forge parallel
These are some examples measuring the performance improvement:
cd src
addpath('Utils')
% Launch experiments sequentially
tic;Utilities.runExperiments('tests/cvtests-30-holdout/kdlor.ini');toc
...
Elapsed time is 318.869864 seconds.
% Launch parallel experiments with maximum number of cores
tic;Utilities.runExperiments('tests/cvtests-30-holdout/kdlor.ini', 'parallel', true);toc
...
Elapsed time is 190.453860 seconds.
% Runs parallel folds with max workers and do not close the pool
Utilities.runExperiments('tests/cvtests-30-holdout/kdlor.ini', 'parallel', 1, 'closepool', false)
Utilities.runExperiments('tests/cvtests-30-holdout/svorim.ini', 'parallel', 1, 'closepool', false)