Parallelize in 'predict' calculation
Issue #186
duplicate
I want to run MD in nanosecond scale with time step of 1(fs) (i.e. 1,000,000 steps)(about 60 days)
To do that, it is desirable to do parallel calculation. But as I know, Amp is not yet parallelized in 'Energy & Force predict' step
Is there any plan to parallelize predict step?
Comments (2)
-
-
repo owner - changed status to duplicate
Duplicate of #79.
- Log in to comment
Parallelization is implemented in image level (and not atom level). That is, a single image goes to a single core for calculations (and not multiple cores each taking care of a number of atoms in the system). That is why calls for energy & force are not parallelized.
To have parallelization in energy and force calls, we need to revisit our parallelization scheme and make it atom-based. As long as only energies are trained or called, this should be easy to change. But an atom-level parallelization is a bit challenging when training or calling forces, because in the case of forces, pair-atoms interactions are important.
I am not aware of any clear plan from our side for changing parallelization scheme to the atom-level, but this is an open project and user contributions are very welcome. So feel free to give it a shot if you like!