Dear all,
I am using MPI communication in pair calculation to share coordinate of each atom between all cores by MPI_Allreduce, however, the speed doesn’t enhance significantly.
By what and how does LAMMPS generally communicate? Is there any key point?
Why would you expect that to help? For pair styles, each processor already knows about all of the atoms it needs to.
Which pair style are you using?
Dear all,
I am using MPI communication in pair calculation to share coordinate of each
atom between all cores by MPI_Allreduce, however, the speed doesn’t enhance
significantly.
why should it?.
By what and how does LAMMPS generally communicate? Is there any key point?
the key point is that LAMMPS uses a distributed data communication
based on a domain decomposition. the details are laid out in the
LAMMPS paper and many presentations held by steve and other LAMMPS
developers held at the LAMMPS workshops and elsewhere. allreduce will
lead to a replicated data distribution which doesn't scale beyond a
small number of processors.
this is not LAMMPS specific, but a general property of parallel MD
that is well known for decades.
axel.