how to choose the best number of nodes for parallel processing

Dear LAMMPS users,

Is there a way to determine how to choose the best number of nodes for parallel processing? or choosing number of nodes is based on experience and try and error?

Thank you,

Sahand

Dear LAMMPS users,

Is there a way to determine how to choose the best number of nodes for
parallel processing? or choosing number of nodes is based on experience and
try and error?

there is no single "best" choice. there is always a range. using more
nodes will make your job ran faster, but you rarely get ideal (strong)
parallel scaling (i.e. using twice as many processors will not make
your job run twice as fast), until you reach the scaling limit. thus
using less nodes will allow you to run more jobs at the same time and
have overall more throughput. using more nodes, will make the
individual job run and complete faster. which of those is the better
choice, depends on individual preferences and overall needs.

on top of that, this strongly depends on the specifics of your system.
certain systems have load imbalances (some CPUs/nodes need to wait for
others), that can be addressed by changing the processor distribution
or using the balance command. certain features of LAMMPS have better
parallel scaling than others.

finally, it also depends strongly on the kind of hardware you are
using, how this is set up and managed and what policies about node
usage exist.

experience helps a lot, but doing tests and benchmarks is better. even
the most experienced LAMMPS users, do so. LAMMPS helps with the
information at the output of a job, which shows how the workload is
distributed across different facilities in LAMMPS, but also it
provides a measure of load (im)balance.

axel.

Thank you so much Axel.