Dear all,
I haver encountered a problem when using the different number of cores in my simulation. The simulation I did was getting the rotation speed of carbon nanotubes. I ran the same input script on 2,16 cores. The simulation both worked fine for 2 and 16 cores. But the results that got from the output files are different. The rotation speed got from 16 cores are almost twice as much as the speed got from 2 cores.
Does anyone know the difference of the output come from? And how to choose the right cores?
Dear all,
I haver encountered a problem when using the different number of cores in
my simulation. The simulation I did was getting the rotation speed of carbon
nanotubes. I ran the same input script on 2,16 cores. The simulation both
worked fine for 2 and 16 cores. But the results that got from the output
files are different. The rotation speed got from 16 cores are almost twice
as much as the speed got from 2 cores.
Does anyone know the difference of the output come from? And how to
choose the right cores?
i think you are asking the wrong question here. there is not enough
information here, but my guess is that the overall setup of your
simulation is flawed, i.e. that both results are bogus.
I think I made a bad description. I have the exactly same data file and input file, which is a 2000 atoms nanotube structure. I ran the script both on my desktop and on a cluster. My desktop use Intel dual-core cpu and have 4GB memory. On the cluster I use 1 node with two hex-core cpu with 16GB memory. but the simulation result is totally different.I know there will be some differences if we ran simulation on different computers, but those results are totally different. Also, I don’t think my desktop have any problem, because I ran a example which lammps gives use the same number of cores on it and it gives the same(or say similar) result as I run on that cluster. I am confused about when I use the differernt number of cores, the results are different.
Could you please give some explanations? How to solve such problem or something I should be beware of?
Wan
I think I made a bad description. I have the exactly same data file and
input file, which is a 2000 atoms nanotube structure. I ran the script both
on my desktop and on a cluster. My desktop use Intel dual-core cpu and have
4GB memory. On the cluster I use 1 node with two hex-core cpu with 16GB
memory. but the simulation result is totally different.I know there will be
some differences if we ran simulation on different computers, but those
results are totally different. Also, I don't think my desktop have any
problem, because I ran a example which lammps gives use the same number of
cores on it and it gives the same(or say similar) result as I run on that
cluster. I am confused about when I use the differernt number of cores, the
results are different.
so you have an isolated nanotube in vacuum?
then please explain, where does the rotation come from?
Could you please give some explanations? How to solve such problem or
something I should be beware of?
you confirmed my suspicion. the problem is not the number of parallel
tasks, the problem is that your "experiment" is flawed at the core
(pun intended) and that it can only produce mostly random results.
How did you initialize the velocities? Are you thermostatting,
e.g. with the Langevin thermostat. Both those procedures use
random numbers, which will typically be different if you
run on different #s of processors.
If you start two simulations on different # of procs, but
with identical velocities, and do not change them randomly
(e.g. fix langevin), then the 2 simulations should remain
nearly identical for some time, then diverge (statistically)
slowly over longer times. As I said before the Section errors
doc page discusses this.