Closed Ig-dolci closed 2 months ago
Could anyone with macOS help verify if this firedrake-install
script fixes the parallel issue with MUMPS?
Please try running test_poisson_analytic_linear_parallel()
from tests/regression/test_poisson_strong_bcs.py
. My machine works correctly with mpiexec -n 2 tests/regression/test_poisson_strong_bcs.py
, but fails when run with pytest
.
If the pytest also fails for others, it might indicate a parallel issue with pytest rather than with MUMPS.
You can try pytest and python executions also for test_dg_advection_cubed_sphere_parallel()
from tests/regression/test_dg_advection.py
.
Obs: Without -Wl,-dead_strip_dylibs
flag, neither python nor pytest executions work.
We also have this Colin's test:
from pyop2.mpi import MPI
size = MPI.COMM_WORLD.size
ensemble = Ensemble(COMM_WORLD, size//2)
rank = ensemble.ensemble_comm.rank
mesh = UnitSquareMesh(20, 20, comm=ensemble.comm)
V = FunctionSpace(mesh, "CG", 1)
n_Js = [3, 3]
Js_offset = [0, 3]
u = TrialFunction(V)
v = TestFunction(V)
a = inner(grad(u), grad(v))*dx
for i in range(n_Js[rank]):
val = Js_offset[rank]+i+1
x = Function(V)
L = x*v*dx
u0 = Function(V)
bc = DirichletBC(V, Constant(val), "on_boundary")
solve(a == L, u0, bcs=[bc])
print(rank, norm(u0), "NOOORMM")
with correct output:
~ mpiexec -n 2 python3 mfe.py
1 3.9999999999999716 NOOORMM
1 5.0 NOOORMM
1 6.000000000000031 NOOORMM
0 0.9999999999999929 NOOORMM
0 1.9999999999999858 NOOORMM
0 3.0000000000000155 NOOORMM
~ mpiexec -n 4 python3 mfe.py
0 0.9999999999999949 NOOORMM
0 1.9999999999999898 NOOORMM
0 2.999999999999998 NOOORMM
0 0.9999999999999949 NOOORMM
0 1.9999999999999898 NOOORMM
0 2.999999999999998 NOOORMM
1 3.9999999999999796 NOOORMM
1 5.0 NOOORMM
1 5.999999999999996 NOOORMM
1 3.9999999999999796 NOOORMM
1 5.0 NOOORMM
1 5.999999999999996 NOOORMM
After a fresh install on my M3 Mac, parallel pytest
also fails for me, but running test_poisson_analytic_linear_parallel()
directly with mpiexec -n 2 python ...
works. Colin's test also passes.
After a fresh install on my M3 Mac, parallel
pytest
also fails for me, but runningtest_poisson_analytic_linear_parallel()
directly withmpiexec -n 2 python ...
works. Colin's test also passes.
So, I believe we have two options here:
Any thoughts?
1
, I think, if this PR fixes a major issue.
Description
Additional flag to fix macOS issue involving MUMPS.