There are special integration tests for testing the workings of parallelized components of Rosetta with the MPI JobDistributors.
These can be run by passing the
--mpi-tests flag to integration.py.
These tests require Rosetta to be compiled first with the
-extras=mpi flag passed to scons.
Where standard integration tests are set up based on the
command files in the
tests/integration/tests/<testname>/ directories, MPI integration tests are defined in
Where a typical
command file runs the Rosetta executable directly,
command.mpi files call
mpirun, specify the number of processes to launch, and pass the Rosetta executable to
Note that MPI-level parallelism is inherently stochastic: any slave process may finish its work before any other, meaning that one needs to set up integration tests carefully to test only for unexpected variation and not for this normal variation.
A few things can help you to do this:
The -jd2:sequential_mpi_job_distribution flag tells the MPI JobDistributors that slave processes are to request jobs in sequential order (i.e. slave 1 first, then slave 2, then slave 3). This ensures that the same job is always assigned to the same slave, yielding consistent trajectories. Note that this should *NOT** be used for production runs, since this can result in idle processes, particularly on massively-parallel systems.*
The output log can be split by grepping '(0)', '(1)', '(2)', etc. While the master process (process 0) is expected to produce a log with some variability in its output, slave processes (process 1, 2, 3, etc.) should have consistent output.
Scorefiles may need to be re-sorted, since they will be written in the order in which jobs complete, which could vary from run to run.
For examples of how to set up MPI integration tests, see tests/integration/tests/bundlegridsampler_design_nstruct_mode/.
As a final note, most integration tests do not run when the
--mpi-tests flag is passed to integration.py.
Only those tests that have a
command.mpi file will run in this mode.