Does KIM allow the simulator (eg, LAMMPS) to run in parallel? How is this reflected within the model?
This question comes from reading the current LAMMPS documentation for pair_kim, which hints to running in serial (i.e. lmp_serial).
Along the same lines, can tests be written to take advantage of parallelization (either MPI, OpenMP or both)?
KIM Models can be used with LAMMPS in parallel. This is handled by LAMMPS through its domain decomposition with “ghost” atoms at the boundary. We’ll take a look at the documentation to make sure its not misleading.
Just to be clear, there are two different cases discussed here: (1) using KIM Models outside openkim.org with your atomistic simulation package of choice; (2) Running KIM Tests within openkim.org.
It is possible to use KIM Models externally to the OpenKIM system to work in parallel mode with LAMMPS, IMD, DL_POLY and other MD codes and lattice dynamics codes like GULP that support the KIM API.
Within the OpenKIM system (i.e. in openkim.org), OpenKIM Tests are currently limited to serial calculations. However, we are working with collaborators to add support for parallel computations in order to make it possible to have more computationally intensive properties within OpenKIM.
-Ellad
This is one of my interests in the KIM project and a primary focus of my involvement in the upcoming content carnival.
I have fully automated property calculations that require the distribution of parallel tasks. The question is now how (best) to implement this within the KIM framework.
I will report to the list regarding progress following the carnival.
Cheers,
-Zach
Yes, this clears up my prior confusion. I look forward to using KIM models (in parallel) within LAMMPS for my research on local machines.
In the future, I hope I can contribute a KIM test to characterize complex structures by computing 3D diffraction intensities; however the computation required for this would be too much to be run in serial on openkim.org.
Thanks for your clarification,
Shawn