Tests for comparison of parallel efficiency of assembly routines in FEniCS/DOLFINX (https://www.fenicsproject.org/).

The list of options for the weak scaling test can be obtained by running

python3 test_assembly_routines.py -h

Remark: pandas (https://pandas.pydata.org/) has to be installed, e.g. with python3 -m pip install pandas, in order to store the results in a CSV file that can be further postprocessed with

./output_assembly_routines.py <filename>

The following results were obtained by running

test_assembly_routines.py -t <mono|block|nest> -r 10 --dofs 100000 --cpp-forms

in parallel. The last option --cpp-forms ensures that all forms are converted to dolfinx.cpp.fem.Form before passing them to the chosen assembly routine.

Figure 1: MPI-averaged wall time for matrix/vector assembly using the individual routines invoked with (bi)linear forms of the type dolfinx.cpp.fem.Form.

Table 1: Data for the case shown in Figure 1.

Things can go wrong if the option --cpp-forms is omitted. In that case, the forms remain of the type ufl.form.Form at the time of passing them to the assembly routine which can lead to an I/O overhead (due to dynamic linking of cached forms). This becomes perceptible with MPI internode communication, especially for block/nest assemblers.

Figure 2: MPI-averaged wall time for matrix/vector assembly using the individual routines invoked with (bi)linear forms of the type ufl.form.Form.

Table 2: Data for the case shown in Figure 2.