Compiling Parallel Version of LAMMPS

Dear All,

I’ve done some MD simulations of SPC water molecules in CNT by running in serial. I want to increase the dimensions and number of particles in my simulation. Therefore, I compiled the Parallel version of LAMMPS to make simulation faster. To do this, I loaded module of mpich/3.0.4 in our cluster and then type “make mpi”. Then, “lmp_mpi” was produced in src directory successfully. I used the executable file of “lmp_mpi” to repeat my previous simulation (in which I used executable lmp_serial). The results are very similar to serial version, however the computational times in parallel version take longer in comparison to serial one, even I increase the number of cores. I am wondering how can I run parallel version properly? I am using the version of 17 Nov, 2016.

Regards,

Hamed.

Dear All,

I've done some MD simulations of SPC water molecules in CNT by running in
serial. I want to increase the dimensions and number of particles in my
simulation. Therefore, I compiled the Parallel version of LAMMPS to make
simulation faster. To do this, I loaded module of mpich/3.0.4 in our cluster
and then type "make mpi". Then, "lmp_mpi" was produced in src directory
successfully. I used the executable file of "lmp_mpi" to repeat my previous
simulation (in which I used executable lmp_serial). The results are very
similar to serial version, however the computational times in parallel
version take longer in comparison to serial one, even I increase the number
of cores. I am wondering how can I run parallel version properly? I am using
the version of 17 Nov, 2016.

how to run LAMMPS in parallel is described in the manual. you didn't
say what kind of time you are measuring and how many more CPU cores
you are using and on what kind of machine. LAMMPS provides some
standard benchmark inputs and you can compare their performance and
speedup from those to your machine.

if you are not getting the expected performance, it is most likely an
issue that is specific to the cluster you are running on and thus you
should work with the people providing support for that cluster to
improve your parallel performance. outside of that, you may be running
into load balancing issues. for that, please have a look at the
processors, and balance commands for explanations possible problems
and how to address them.

axel.

Dear Dr. Kohlmeyer,

Thanks for your comments.

The time I am comparing for serial and parallel, are the run time (The time that the job is actually running). I started from 4 cores and increased the number of cores to 8, 24, 30 and 80 cores. Overall, the run time also increases with increasing the number of cores. Our GHPCC cluster is using Redhat 6.4 .

About compiling LAMMPS in parallel, my question is that do I need to include additional package/library before compiling LAMMPS in parallel for my simulation? The manual says “If you get no errors and an executable like lmp_mpi or lmp_g++_serial or lmp_mac is produced, then you’re done”(Page 14). I also got lmp_mpi file without error after compiling; however, I can’t get the expected performance. Please advise.

Regards,
Hamed.

Dear Dr. Kohlmeyer,

Thanks for your comments.

The time I am comparing for serial and parallel, are the run time (The time
that the job is actually running). I started from 4 cores and increased the
number of cores to 8, 24, 30 and 80 cores. Overall, the run time also
increases with increasing the number of cores. Our GHPCC cluster is using
Redhat 6.4 .

About compiling LAMMPS in parallel, my question is that do I need to include
additional package/library before compiling LAMMPS in parallel for my
simulation? The manual says "If you get no errors and an executable like
lmp_mpi or lmp_g++_serial or lmp_mac is produced, then you're done"(Page
14). I also got lmp_mpi file without error after compiling; however, I can't
get the expected performance. Please advise.

i already advised you. talk to your local cluster experts and let them
help you with the tests.

there is plenty of evidence that LAMMPS can work well in parallel, at
the very least with the bundled benchmark examples and according to
the instructions given in chapter 8 of the manual and on the LAMMPS
homepage.

axel.