Closed kevinstratford closed 4 months ago
The mpirun -np 2
github action seems unreasonably slow compared with the serial version (4-5 minutes cf 40-50 seconds).
Modest investigation suggests the increase in time is across-the-board, althouhg halo exchanges are particularly bad.
I've removed a number of unused tests from
./tests/regression/d3q19
which are mainly longer versions of active tests. These tests all still pass at the point of removal.
This is updated at #312
I have left mpi tests for d2q9
and d3q19
only to keep the time under control.
Note on regression tests.
This is the first attempt to unify the serial and parallel test cases. The aim is to run the same input files in either serial and parallel and not care about details.
I've removed decomposition, cell list dimensions and so on which do not match the serial case in the
test-diff.sh
script. However, this may not catch all decomposition-dependent output.Unit tests
1 x mpi x 2 x threads 2 x mpi x 2 x threads 4 x mpi x 2 x threads
All ok.
Regression tests
d2q9
Good at 2 x mpi x 2 x threads
d3q15
Good at 2 x mpi x 2 x threads
d3q19-short
Gross failure:
Marginal failures (will pass at e.g.
TOLERANCE=1.0e-08
)Otherwise ok.
d3q27
Good at 2 x mpi x 2 x threads