Thank you for your reply.
I have add the log files simulated with different nodes. It means other time
has occupied the biggest part.
I use the nonperiodic boundary and the size of the box is
-500.000000000000 500.0000000000000 xlo xhi
-500.000000000000 500.0000000000000 ylo yhi
-500.000000000000 500.0000000000000 zlo zhi ,
From papers I read, they suggest 'the cutoff is chosen such that all the
ions in the system are included in force and energy calculations', so I
chose quit a large value. Here I knw that too long a cutoff may lead to
higher computational cost.
it also causes a lot of overhead. however, you have a system with only
a few thousand atoms, so they are likely within reach, even at a
shorter cutoff.
there are a number of issues, you need to look into.
- with a few thousand atoms and a "cheap" potential model, there is
not much parallel scaling to be expected.
- with such a large box and m/m/m boundaries, yet a fairly small
number of atoms, you are very likely to have load balancing issues.
LAMMPS will be default divide the total volume of your simulation cell
into equal sub-boxes and then parallelize over them. if the number of
atoms in these sub-boxes varies largely, then your parallel efficiency
is limited.
- for a small-ish system and non-periodic boundaries, your choice of
cutoff is quite legit. it doesn't add much overhead, since there are
not a lot of atoms around. and since you have non-periodic boundaries,
there are no atoms from periodic boxes included. it is the latter that
makes a long cutoff particularly expensive when you run a periodic
system.
- you need to find out what time contributes to "Other". that is best
done by removing all computes and other components that are not part
of the plain simulation. again, this is possibly a load balancing
issue, but may also be due to something that doesn't parallelize as
well as the plain MD. e.g. sorting atoms in a dump file.
For the neighbor skin distance is may related to the timestep. I found if
the distance is quite short, the atoms may lost. Prabaply just as you
suggest I need a smaller value of timestep.But for the buck/coul/cut I have
no idea to calculate the value of timestep. I referred to some papers, the
cutoff is about 1-2fs for most cases.
the length of the time step is chose so that the error from the
discretization of the differential equations describing the motion of
the atoms is sufficiently small. this is a very fundamental MD issue
and topic of every MD text book worth reading.
Actually, I have little idea of rules to choose the parameters. Could you
give me some suggestion?
as you seem to be somewhat lost with running MD in general and
efficient parallel computing, you need to get training in these
topics. this is not something that you can learn from a few e-mails,
you need proper (local) guidance and work through text books.. we
cannot teach you from remote since it is very inefficient and time
consuming. without suitable guidance, you will have a hard time.
axel.