Run in parallel mode

KLIFF supports parallelization over data. It can be run on shared-memory multicore desktop machines as well as HPCs composed of multiple standalone machines connected by a network.

Physics-based models

We implement two parallelization schemes for physics-based models. The first is suitable to be used on a desktop machine and the second is targed for HPCs.

multiprocessing

This scheme uses the multiprocessing module of Python and it can only be used on shared-memory desktop (laptop). It’s straightforward to use: simply set nprocs to the number of processes you want to use when instantiate Loss. For example,

calc = ...  # create calculator
loss = Loss(calc, ..., nprocs=2)
loss.minimize(method="L-BFGS-B")

See also

See Train a Stillinger-Weber potential for a full example.

MPI

The MPI scheme is targeted for HPCs (of course, it can be used on desktops) and we use the mpi4py Python wrapper of MPI. mpi4py supports OpenMPI and MPICH. Once you have one of the two working, mpi4py can be installed by:

$ pip install mpi4py

See the mpi4py package documentation for more information on how to install it. Once it is successfully installed, we can run KLIFF in parallel. For example, for the tutorial example Train a Stillinger-Weber potential, we can do:

$ mpiexec  -np 2  python example_kim_SW_Si.py

Note

When using this MPI scheme, the nprocs argument passed to Loss is ignored.

Note

We only parallelize the evaluation of the loss during the minimization. As a result, the other parts will be executed multiple times. For example, if kliff.models.Model.echo_model_params() is used, the information of model parameters will be repeated multiple times. If this annoys you, you can let only of process ( say the rank 0 process) to do it by doing something like:

from mpi4py import MPI

rank = MPI.COMM_WORLD.Get_rank()
if rank == 0:
    model.echo_model_params()

Machine learning models