[lammps-users] Question about large pencentage of communication time

Dear users,

When I was doing my calculation on a cluster containing 18 computers each with 1 processor, I observed that communication time take about 70% of all the time used to finish a job. I simulated a system containing about 50000 atoms, 40 in x and y direction and 100 in z direction (periodic in x and y direction and nonperiodic in z direction), and write the command ‘processors 3 3 2’ in the input file.

Since when performing DFT tasks, I can specify that each processor write their temporary files into its own hard disk and hand in the final results to the ‘leader’ of the cluster, my question is :

Is there any similar commands in Lammps or MPI settings to specify it ,or this kind of settings does not help to accelerate MD calculation in Lammps?

Any reply is highly appreciated!

Dear users,

dear chen chen,

When I was doing my calculation on a cluster containing 18 computers each

the most important information is missing here:
what kind of interconnect do your nodes have?

with 1 processor, I observed that communication time take about 70% of all
the time used to finish a job. I simulated a system containing about 50000
atoms, 40 in x and y direction and 100 in z direction (periodic in x and y
direction and nonperiodic in z direction), and write the command 'processors
3 3 2' in the input file.

Since when performing DFT tasks, I can specify that each processor write
their temporary files into its own hard disk and hand in the final results
to the 'leader' of the cluster, my question is :

LAMMPS does not create temporary files. please explain what you mean
with "DFT tasks" running a different program? if yes, which one?

Is there any similar commands in Lammps or MPI settings to specify it ,or
this kind of settings does not help to accelerate MD calculation in Lammps?

you need to have a fast interconnect with low latencies.
gigabit ethernet is not, it has very high latencies.

Any reply is highly appreciated!

hope this helps,
   axel.

Unless you have a specific reason for it, I wouldn't use
the "processors" command. Let LAMMPS figure out
how it wants to split the domain across processors.

Steve

Axel and Steve,

Thank you for your reply, really.

To answer Axel’s questions, I would say:

  1. The computers are connected through gigabit ethernet.
  2. The ‘DFT tasks’ I mentioned was performed by a programme named ‘Dacapo’. I know little about DFT calculations, and when discussing with other people I came up with the question. Since this kind of optimization as specifying that processors write temporory file to their own hard disks would greatly enhence the performance.
    Given that LAMMPS does not create any temporary files as you said, I think this kind of optimization would not help anyway.

Steve, I have tried what you suggested but the ratios are nearly the same.
Actually, I didn’t know LAMMPS would split the domain itself before you teachered me.