Re: Parallelization

[Jean-Noël Grad]

Dear Ahmad Reza,

Thank you for pointing that out. I can't find the information in the 
user guide. We need to write it down.

ESPResSo supports OpenMPI, MPICH and MPICH+UCX. To run a simulation in 
parallel, invoke the command `mpiexec -n 4 ./pypresso script.py`. It 
should also be possible to run a Jupyter notebook in parallel, but I 
can't remember the exact syntax, and restarting the Python session will 
disable parallelism if I recall correctly. To run GDB in parallel, 
invoke `mpiexec -np 2 xterm -fa 'Monospace' -fs 13 -e ./pypresso --gdb 
script.py`.

To control how the system volume is partitioned across MPI ranks, add 
`system.cell_system.node_grid = [i,j,k]` to the script, with the product 
i*j*k being equal to the number of MPI ranks. This is useful for example 
when particles move preferentially along a specific direction and the 
selected cell system only has one cell in that direction. When using 
FFT-based methods, an additional constraint i <= j <= k is enforced due 
to a limitation of our FFT algorithm. For example when running P3M 
electrostatics with 4 MPI ranks, [4,1,1] and [2,2,1] are the only valid 
partitions. When using GPU-based methods, there is the additional 
constraint that the GPU device is only visible on MPI rank 0, therefore 
one has to account for the overhead of gathering data from all ranks to 
rank 0 to communicate it to the GPU, followed by a broadcast of the GPU 
results from rank 0 to all ranks, at every time step.

Finally, the size of the Particle struct matters a lot in parallel 
simulations. Using a custom myconfig.hpp file containing only the bare 
minimum of features that are relevant to the simulation can 
significantly speed up parallel simulations.

Best regards,
JN

On 2021-12-29 18:08, Ahmad Reza Motezakker wrote:
> Dear EspressoMD users,
>
> Is there any tutorial or any document on EspressoMD parallelization? I
> really appreciate your help. Any suggestion or experience will be
> really helpful for me.
>
> P.S. I am not familiar with parallelization.
>
> Merry Christmas,
>
> Ahmad Reza