KGB paral processor assignment
Posted: Fri Feb 25, 2011 10:21 pm
This may be in the wrong forum, but I wasn't sure where to put it.
When running in parallel beyond just k-points I assume there is some degree of communication between different processors dealing with different bands or fft grids of the same k-point. As it stands it appears that ABINIT uses a round-robin type of assignment. For example, with npkpt 8, npband = 8, processors 1, 9, 17, etc. will be dealing with different bands for k-point 1.
Is there a way to control how ABINIT assigns band/fft parallelization or should I just pass mpirun flags that try to work around the default to maximize locality on a chip or in a node?
When running in parallel beyond just k-points I assume there is some degree of communication between different processors dealing with different bands or fft grids of the same k-point. As it stands it appears that ABINIT uses a round-robin type of assignment. For example, with npkpt 8, npband = 8, processors 1, 9, 17, etc. will be dealing with different bands for k-point 1.
Is there a way to control how ABINIT assigns band/fft parallelization or should I just pass mpirun flags that try to work around the default to maximize locality on a chip or in a node?