2.13. Parallelization with MPI

Particularly with larger simulation times and system sizes, complex calculations are computationally intensive. In order to keep the actual simulation time manageable, tkwant is parallelized with the Message Passing Interface (MPI). With MPI, tkwant can run it in parallel, e.g. on the local computer or on a cluster.

Parallel programming and in particular MPI is a vast subjet and the following tutorial cannot explain these techniques here, but we refer to dedicated material which can be found in the web. This tutorial will explain however the very basic concept which is sufficient to run tkwant simulations in parallel without having a deeper knowledge in MPI. This is possible since compute-intensive routines are natively MPI parallelized in tkwant, such that only minor changes to a simulation script are required. We will explain them in the following.

2.13.1. Running code in parallel with MPI

As an example, let us focus on the tkwant example script fabry_perot.py. To execute this script on 8 cores, the scripts must be called with the following command:

mpirun -n 8 python3 fabry_perot.py

Calling the script with:

python3 fabry_perot.py

will run it in standard serial mode.

2.13.2. Enabling output only from the MPI root rank

Running a script with the prefix mpirun -n x will execute the entire script x times. While tkwant is designed to benefit from this parallelization, any output in the script as printing, plotting or writing to a file will be also repeated x times. This is unpractical, as e.g. a print() call in the simulation script will print some information x times instead of only once. Moreover, not all of the x parallel runs are equivalent, but the result of a calculations from the tkwant solvers is returned by default only on one MPI rank, the so-called master or root rank, which has the rank index 0. It is sufficient however to add a few additional lines of code, in order to redirect all plotting and printing to the MPI root rank, such that both serial and parallel execution will lead to the same result.

As an example, we look again at the script fabry_perot.py. In this script, a few additional lines of code redirect all plotting and printing to the MPI root rank. For plotting and saving the result, the following block of code can be used:

import tkwant

comm = tkwant.mpi.get_communicator()
def am_master():
    return comm.rank == 0

# do the actual tkwant calculation

if am_master():
    # plot or save result

Quite similar, printing a message only by the master rank is possible by the following lines of code:

import sys

def print_master(*args, **kwargs):
    if am_master():
        print(*args, **kwargs)

print_master('this message is printed only by the master rank')
this message is printed only by the master rank

Note the flush command to prevent buffering of the messages.

2.13.3. MPI communicator

The following information is not relevant for tkwant users, but inteded for tkwant developers working with MPI. Tkwant initializes automatically the MPI communicator, if needed. To uses MPI, the function mpi.get_communicator() returns tkwant’s global MPI communicator which is used by all routines by default:

import tkwant

comm = tkwant.mpi.get_communicator()
print('rank={}, size={}'.format(comm.rank, comm.size))
rank=0, size=1

comm is basically a copy of the MPI COMM_WORLD communicator.

If tkwant should be used as an external library with a different MPI communicator, the routine mpi.communicator_init() allows to change the default communicator:

import tkwant
from mpi4py import MPI

my_comm = MPI.COMM_WORLD
WARNING:tkwant.mpi:56:rank=0: tkwant MPI communicator cannot be reinitialized.

Note that the MPI communicator must be set after importing the tkwant module and before executing any tkwant code.

2.13.4. Multi-threading

Tkwant does not support multi-threading as OpenMP. The environment variable OMP_NUM_THREADS must be set to one:


2.13.5. Examples

The example scripts fabry_perot.py and voltage_raise.py can both be executed in parallel using MPI.