Hello all, I have created a —with-slurm option when running (See updated README). In order to set new defaults for collective algorithms, we will need data from those who wish to provide it. We have created the following package that allows for collecting data: https://github.com/open-mpi/ompi-collectives-tuning
Please run the package as soon as possible. Details on how to run are in the README.md<http://README.md>. If data collection fails, the output of the analyze script (either analyze.sh.o* for SGE or the ouput of ./run_and_analyze if using slurm) will report "Error parsing <filename>. Data format doesn't match. Exiting..”. Please make sure data collection succeeds and a decision file is written entirely. Please provide me with either the output directory or if it’s inconvenient to share this data, provide me a list of optimal switchover points at different message sizes for each algorithm (This can be in the form of the output/decision.file which only contains switchover points and no specific performance numbers) Thanks, William Zhang