Skip to content

Latest commit

 

History

History
32 lines (27 loc) · 4.16 KB

orca_condor.md

File metadata and controls

32 lines (27 loc) · 4.16 KB
Raw

Experiments parallelization with HTCondor

ORCA can be easily integrated with a High Throughput Computing (HTC) environment, such as HTCondor, so extensive experiments can be speed up. The parallelization is done at dataset partition level, i.e. each partitions of each dataset is run in a different slot of the cluster.

The src/condor folder contains a set of scripts that automate the use of ORCA with HTCondor. The script condor-matlabExperiment.sh allows to run an ORCA set of experiments by using any of the configuration files. To use the script, you will need to have the configuration file in the src/condor folder. We will combine two different algorithms by concatenating two ini files:

~/orca/orca$ cd src/condor/
~/orca/orca/src/condor$ cat ../tests/cvtests-30-holdout/svorim.ini > test.ini
~/orca/orca/src/condor$ cat ../tests/cvtests-30-holdout/kdlor.ini >> test.ini

Then, you have to edit the test file, so that the path of the experiments is correct with respect to the current path (replace ../example-data by ../../example-data). This can be done with a text editor or using the following sed command:

~/orca/orca/src/condor$ sed -i 's/\.\.\//\.\.\/\.\.\//g' test.ini

Now the script is ready to be used. The following command:

~/orca/orca/src/condor$ ./condor-matlabExperiment.sh test.ini

will create a HTCondor work and will add this work to the HTCondor queue. Each work consists of a task for dividing the work into different independent configuration files, a train-test task for each dataset partition and an extra task to collect all the data and create the reports. Most of the experimental results will be compressed, with the exception of the CSV files. To adapt the set of scripts to your HTCondor system please set up environment variables corresponding to MATLAB's path, universe, requirements and edit the .sh file.

Additionally, the src/condor folder includes the following files:

  • condor-matlabFramework.dag: this HTCondor dag file will run the submit files into the appropriate order, that is, condor-createExperiments.submit, condor-runExperiments.submit and condor-joinResults.submit.
  • condor-createExperiments.sh: this a bash script invoking ORCA for separating the configuration file into as many configuration files as the number of datasets by the number of partitions. The script receives two command line arguments, the name of the configuration file and the name of the working directory.
  • condor-createExperiments.submit: this HTCondor submit file will run condor-createExperiments.sh into the HTCondor cluster.
  • condor-runExperiments.sh: this HTCondor submit file will run condor-joinResults.sh into the HTCondor cluster. The script receives two command line arguments, the name of the working directory and the number of experiment to be run.
  • condor-runExperiments.submit: this HTCondor submit file will run condor-runExperiments.sh into the HTCondor cluster.
  • condor-joinResults.sh: this a bash script invoking ORCA for joining the results of all the experiments run. The script receives a command line argument, the name of the working directory.
  • condor-joinResults.submit: this HTCondor submit file will run condor-joinResults.sh into the HTCondor cluster.

Experiments parallelization with other cluster environments

The design of ORCA allows easy parallelization with other cluster environments, given that all the intermediate results of the different partitions are saved to disk. In this way, you can use the scripts condor-createExperiments.sh, condor-runExperiments.sh and condor-joinResults.sh to run the code in your own cluster environment.